THE AI TOOLKIT / AGENTS
Agents.
A clear role.
A reviewable brief.
Find reusable agent definitions for real project work. Inspect the role, expected output and review notes, then reuse an attributed definition where its licence allows it.
01 / START WITH THE JOB
What are you working on?
Explore a category, then refine your shortlist.
All categories are available in the filters.
BUILT BY UNDOMINATED
Evidence tools you can inspect.
Original source, MIT licence and complete downloads.
Each review names the checks performed and their limits.
- 3D parallelism architect ↗
Architects balanced 3D parallelism configurations (TP, PP, DP, SP) for massive LLM serving clusters.
- Adversarial Red Teamer ↗
Classifies supplied reports of prompt-injection and jailbreak failures. It does not generate bypasses or run attacks.
- Agent cascade failure auditor ↗
Audits multi-agent autonomous chains for failure cascades, hallucinated handoffs, and infinite retry loops.
- Agent State Budget Architect ↗
Architects conversation state pruning, sliding context windows, and tool output summarization to control multi-turn token costs.
- Agent trajectory pruning specialist ↗
Multi-turn log auditor specializing in identifying conversational bloat, redundant tool executions, memory leaks, and obsolete context across agent trajectories.
- Agent workflow economist ↗
Autonomous agent economist specializing in multi-step agent orchestration efficiency, tool schema minimization, function call consolidation, and token-preserving memory architectures.
- AI resource curator ↗
Review skills, agents and MCP servers against a concrete use case, with pinned identity, rights and runtime evidence.
- Attention Architecture Economist ↗
Models serving economics and hardware footprint trade-offs across Multi-Head Attention (MHA), Grouped-Query Attention (GQA), and Multi-Head Latent Attention (MLA).
- Attention Sink Custodian ↗
Manages initial token attention sink preservation and sliding window buffers in infinite-horizon streaming generation.
- Batch pipeline architect ↗
Systems architect specializing in asynchronous batch inference pipelines, SLA deadline management, partial batch ingestion, and automated synchronous fallback cascades.
- Benchmark rigor auditor ↗
Evaluation integrity specialist auditing public benchmark leaderboards for test set contamination, cherry-picked task splits, and uncalibrated significance claims.
- Cache retention evaluator ↗
Audits needle-in-a-haystack retrieval accuracy under aggressive KV cache compression strategies.
- Canary Rollout Commander ↗
Orchestrates traffic splitting, statistical drift detection, and automated rollbacks for model migrations.
- Chunked Prefill Tuner ↗
Calibrates chunked prefill batch parameters to achieve target ITL p99 while maintaining prefill throughput.
- Comparison editor ↗
Edit model-comparison copy so ties, same-seller tiers and dropped requirements are not published as dominance.
- Constrained Decoding Optimizer ↗
Optimizes JSON schema compilation, regex FSM caching, and token masking to maximize structured output throughput.
- Context parallelism planner ↗
Plans context-parallel serving topologies for 1M+ token context windows using Ring and Ulysses algorithms.
- Continuous Eval Guardian ↗
Maintains continuous regression evaluation harnesses across production endpoints to detect silent capability degradation.
- Crawl auditor ↗
Review a sitemap or llms.txt before publication, so an English-only path is not locale-prefixed and a markdown link names a file that was written.
- Delivery-class reviewer ↗
Review a dominance sentence before it treats batch delivery, or a model's own other delivery, as a like-for-like beater.
- Disaggregated Serving Architect ↗
Designs separate prefill and decode GPU clusters connected via RDMA to eliminate TTFT and ITL scheduling interference.
- Distributed Context Engineer ↗
Architects sequence-parallel, RingAttention, and context-parallel clusters for million-token inference pipelines.
- Draft Model Calibrator ↗
Aligns draft model tokenizer, temperature, and logit distributions with the verifier model to maximize acceptance rates.
- Draft verification auditor ↗
Profiles draft verification kernels to prevent compute overhead from outpacing verification gains.
- Dynamic Routing Economist ↗
Architects real-time cost-and-latency aware multi-provider routing proxies that arbitrage spot price shifts.
- Edge AI deployment architect ↗
Sizes models for edge hardware (Apple Silicon, mobile NPUs, local workstations) to eliminate cloud API spend.
- Enterprise Compliance Guardian ↗
Audits Zero Data Retention (ZDR), HIPAA, SOC2, and sovereign data boundary commitments across commercial AI providers.
- Enterprise compliance officer ↗
Enterprise compliance and regulatory risk auditor evaluating AI inference pipelines against GDPR data sovereignty, EU AI Act risk classifications, HIPAA BAA requirements, and SOC 2 Type II controls.
- Eval harness auditor ↗
Audit LLM evaluation harnesses, judge prompt rubrics, scoring determinism, and benchmark data hygiene to eliminate evaluation leaks and spurious rankings.
- Evidence reviewer ↗
Independently challenge numerical claims using source identity, denominator reconstruction and repeatable calculations.
- Expert Load Balancer Auditor ↗
Monitors router entropy, expert skew, and load balancing auxiliary loss regularization in production MoE models.
- Failover router architect ↗
Systems resilience architect designing active-active multi-provider inference topologies with zero-loss retries, automated circuit breaking, and cross-region failover.
- GPU Footprint & Sharding Architect ↗
Calculates required GPU VRAM, tensor parallelism sharding (TP=2, 4, 8), and minimum cluster configurations for hosting open-weight models.
- GPU Utilization Optimizer ↗
Implements autoscaling, sleep modes, and batch backfilling to maximize hardware utilization and eliminate idle power waste.
- Graph RAG Cost Optimizer ↗
Tunes entity extraction prompts and graph traversal depths to balance retrieval precision and token spend.
- Guardrail Overhead & Latency Analyst ↗
Analyzes the throughput, latency, and financial token overhead of enterprise AI safety and moderation layers.
- Hardware Infrastructure Planner ↗
Sizes on-premise GPU clusters based on enterprise peak and off-peak token demand profiles and CapEx payback models.
- Hierarchical Memory Architect ↗
Designs multi-tiered memory architectures combining working memory, episodic summaries, and graph knowledge.
- Hybrid SSM serving architect ↗
Architects hybrid Transformer-Mamba/SSM inference pipelines for extreme token throughput.
- In-Context Example Pruner ↗
Audits the marginal accuracy contribution of few-shot exemplars and prunes low-ROI examples to trim token costs.
- Indirect Injection Defender ↗
Designs untrusted data isolation boundaries and document sanitization filters to prevent indirect prompt injection.
- Inference CI/CD Architect ↗
Integrates model quality, latency, and cost regression gates into continuous integration and automated deployment pipelines.
- Inference Firewall Engineer ↗
Architects low-latency security firewalls, prompt injection classifiers, and rate limiters at the API gateway layer.
- Inference Scaling Strategist ↗
Analyzes test-time compute scaling laws, verifier passes, and search budgets to optimize accuracy per dollar spend.
- Inference SLO Gate Architect ↗
Defines and enforces strict latency, availability, and error budget SLO gates for production AI gateways.
- Inference SRE architect ↗
Designs production observability, SLO alert thresholds, and automated traffic shedding for inference platforms.
- Judge Model Calibrator ↗
Calibrates LLM judge prompts with few-shot calibration anchors to eliminate position bias and rating drift.
- KV eviction policy architect ↗
Architects dynamic KV cache pruning and eviction policies under strict hardware VRAM limits.
- Library intake editor ↗
Decide whether a skill, agent profile, or MCP server belongs in the public install set, using pinned identity, a licence, and a real checker or an explicit reason it has none.
- Licence derivation auditor ↗
Open-source legal and IP compliance specialist auditing fine-tuning data lineage, synthetic data derivation rights, model weight inheritance, and commercial redistribution boundaries.
- LLM security red teamer ↗
Designs adversarial jailbreak probes and evaluates model alignment under automated attacks.
- Long context bandwidth auditor ↗
Audits inter-node network requirements to guarantee communication is fully masked by attention compute.
- Low-bit fidelity auditor ↗
Audits mathematical reasoning and coding benchmarks under aggressive sub-4-bit quantization.
- Merged capability auditor ↗
Audits merged model checkpoints for catastrophic forgetting and multi-task capability collisions.
- Mixed Precision Architect ↗
Designs layer-wise quantization profiles balancing perplexity retention against memory savings across FP4, FP8, and INT4.
- Model Cascade Architect ↗
Architects multi-tier model cascading pipelines with confidence-gated fallback routing to optimize cost-per-task.
- Model merging architect ↗
Architects zero-training model merges using TIES, DARE, and SLERP to combine specialized domain models.
- Model Merging Evaluator ↗
Evaluates merged model weights across benchmark suites to prevent catastrophic interference and capability loss.
- Model migration planner ↗
Plan workload-specific model migrations while preserving required capabilities and defining observable rollback criteria.
- Model Provenance Investigator ↗
Investigates training data lineage, open-weights licensing boundaries, and attribution requirements across model releases.
- MoE Serving Architect ↗
Specializes in Expert Parallelism (EP), Tensor Parallelism (TP), and expert capacity factor calibration for large MoE clusters.
- Multi-agent systems economist ↗
Designs multi-agent workflows to minimize redundant inter-agent chatter and prevent runaway token loops.
- Multi-LoRA Serving Specialist ↗
Configures multi-adapter serving clusters serving hundreds of fine-tuned LoRA adapters concurrently on a single base model.
- Multi-Needle Retrieval Validator ↗
Generates and evaluates multi-needle retrieval benchmarks to expose lost-in-the-middle context blindness in long-context models.
- Multimodal streaming architect ↗
Architects real-time streaming audio and video inference pipelines with bounded lookahead latency.
- NCCL topology auditor ↗
Audits NVLink, NVSwitch, and InfiniBand rail-optimized communication topologies for distributed serving.
- On-device performance auditor ↗
Audits thermal throttling, unified memory pressure, and battery consumption on consumer edge devices.
- Population reviewer ↗
Review a count, rank or correlation before publication, so the computed rows are the population the sentence names.
- Preference alignment auditor ↗
Audits reward models and preference datasets for length bias, sycophancy, and mode collapse.
- Pricing source reviewer ↗
Review a provider price feed before any rate is admitted, including units, currency, tiers and prose quotes that are not USD.
- Prompt Compression Economist ↗
Audits client-side prompt compression pipelines to minimize billed input token volume without losing semantic fidelity.
- Prompt firewall architect ↗
Constructs low-latency pre-inference guard firewalls to intercept injection and extraction attacks.
- Provider Resilience Architect ↗
Designs automatic failover, circuit breaking, and idempotency guarantees across redundant inference providers.
- Quantization Evaluation Reviewer ↗
Audits LLM evaluation protocols to verify whether published benchmark scores match the model's serving quantization or upstream unquantized weights.
- Quantized KV Cache Calibrator ↗
Calibrates per-channel and per-group scale factors for sub-8-bit KV caches to prevent long-context reasoning drift.
- Quote identity reviewer ↗
Review a proposed join between a vendor quote and a catalogue model, and refuse any match that is not exact.
- RAG Context Economist ↗
Optimizes context window packing, chunk overlap boundaries, and token budgets for enterprise RAG systems.
- Reasoning budget planner ↗
Architect agent that models reasoning token expenditure, dynamically calibrates thinking budgets across prompt difficulty classes, and prevents runaway cost expansion.
- Reasoning Trace Compressor ↗
Audits chain-of-thought reasoning traces to detect circular deliberation loops, repetitive tokens, and distillation opportunities.
- Reasoning verifier auditor ↗
Audits process reward model verifiers for false-positive validation passes and search path collapse.
- Recurrent state auditor ↗
Audits recurrent hidden state numerical drift and precision degradation during continuous streaming.
- Regulatory Compliance Officer ↗
Audits AI deployments for EU AI Act, US Executive Order, and ISO 42001 governance and transparency frameworks.
- Resource release verifier ↗
Independently verify packaged resources, clean installation and meaningful public availability after a release.
- Router Threshold Calibrator ↗
Calibrates draft model confidence thresholds to balance API cost savings against quality degradation.
- Safety Refusal Calibrator ↗
Calibrates safety classifier thresholds to minimize benign query refusals while maintaining strict compliance boundaries.
- Schema Contract Verifier ↗
Verifies tool schemas, argument enums, and parameter constraints to prevent runtime validation failures.
- Serving Engine Tuner ↗
Specializes in tuning vLLM, SGLang, and TensorRT-LLM parameters including chunked prefill, radix trees, and batching scheduling.
- Sovereign AI TCO auditor ↗
Models national sovereign AI cluster CapEx/OpEx, power grid contracts, and 5-year hardware lifecycle TCO.
- Sparse Attention Evaluator ↗
Audits accuracy retention and memory bandwidth reduction across Native Sparse Attention (NSA) and block-sparse mechanisms.
- Speculative Architecture Engineer ↗
Architects speculative decoding topologies including Multi-Token Prediction (MTP), Medusa heads, and Eagle tree verifiers.
- Speculative tree architect ↗
Designs multi-candidate speculative draft trees to maximize generation speed on constrained GPUs.
- Stack cost optimizer ↗
Full-stack AI system cost optimization specializing in multi-tier cascading, retrieval pruning, speculative drafting, and prompt caching to slash API spend while preserving benchmark quality.
- Streaming Reliability Engineer ↗
Specializes in Server-Sent Events (SSE) connection health, keep-alive heartbeat spacing, and client reconnection handling.
- Sustainable inference architect ↗
Optimizes AI workload scheduling for renewable power availability and lowest carbon footprint.
- Synthetic curation architect ↗
Architects synthetic training dataset filtering pipelines to eliminate contamination and redundancy.
- Synthetic judge calibrator ↗
Evaluation scientist specializing in LLM-as-a-judge alignment, position bias mitigation, length normalization, and inter-rater agreement calibration against human experts.
- Tail latency auditor ↗
Profiles p99.9 tail latencies to eliminate micro-stalls caused by garbage collection and multi-tenant contention.
- Tail latency SLA auditor ↗
Production observability auditor comparing provider Service Level Agreements against measured p95/p99 tail latency, serverless cold starts, and outage windows.
- Ternary inference architect ↗
Architects native ternary and low-bit compute pipelines for low-energy edge and cloud inference.
- Test-time compute economist ↗
Optimizes test-time compute allocation across search width and reasoning depth to balance accuracy and cost.
- Tool Registry Architect ↗
Designs dynamic tool indexing and semantic retrieval to keep system prompt tokens minimal across large API catalogs.
- Vision token economist ↗
Optimizes vision patch resolution and temporal subsampling to minimize multi-modal token billing.
- VRAM Fragmentation Analyst ↗
Analyzes PagedAttention block size allocation, CUDA graph memory pools, and virtual memory fragmentation to prevent GPU OOMs.
02 / THE SELECTION
Find a fit. Read the reasoning.
Alphabetical by default. Every match is shown.
Compare up to three before opening a full review.
Runtime, access and execution
Labels summarise reviewed source wording. A local process can still send data to external services. Read each review's prerequisites and permissions; these filters do not certify compatibility or enforce access.
401 of 401 selected definitions
Migration · GitHub community
.NET solution upgrade planner
A solution-upgrade reference organizing framework discovery, dependency order and validation checkpoints.
Why we selected itIt sequences independent libraries before shared components and application projects to expose compatibility issues incrementally.
Weigh upThe embedded shell examples and netstandard migration rule are not universally appropriate; current support windows, consumer compatibility and multi-targeted projects require a real assessment.
Accessibility · github
Accessibility Expert
Guides web accessibility work through keyboard, focus, screen-reader, zoom and contrast checks, with developer examples and a structured review checklist.
Why we selected itOperating rules require an a11y pre-check before code, verification steps with edits, and rejecting requests that remove focus outlines or otherwise reduce access.
Weigh upThe definition can edit files and run commands; it is broader than an advisory accessibility audit.
Accessibility · github
Accessibility Runtime Tester
Exercises keyboard journeys, focus changes and form errors, then records reproducible accessibility findings and retest steps.
Why we selected itRequires observed runtime behaviour and explicitly rejects a Lighthouse pass as proof of accessibility.
Weigh upThe body expects browser automation but the original allowlist does not enable the preferred Chrome DevTools or Playwright MCP tools. Configure and allow those tools in a working copy first.
Design and data · Agency Agents community
Accessible data visualization engineer
A chart-design profile matching the analytical question to honest encodings and accessible interaction.
Why we selected itIt distinguishes zero-based bar lengths from deliberately bounded line-chart axes and asks whether aggregation hides multimodal data.
Weigh upSVG/canvas/WebGL count thresholds and frame-rate targets are heuristics, not measured limits; population prevalence claims and palette names do not prove accessibility for a specific chart.
Security operations · VoltAgent
Active Directory security posture reviewer
An AD review profile tracing privilege, delegation, protocol and service-account risks into a remediation plan.
Why we selected itThe review distinguishes privileged-group membership from orphaned ACLs, delegation and service-account exposure.
Weigh upThis is not a read-only tool profile: it declares Bash/Edit/Write and requests implementation scripts. A checklist does not prove exploitability or compliance.
Media engineering · Agency Agents community
Adaptive video delivery engineer
A streaming-design profile connecting encoding, packaging, device playback and measured quality of experience.
Why we selected itIt treats segment length as a latency/cache/request-overhead tradeoff and requires testing the protected playback path on actual devices.
Weigh upThe ffmpeg example writes to the null muxer and does not create the v360.mp4/v720.mp4 files consumed by the next command; it is an illustration, not an executable end-to-end pipeline. Storage and latency figures are not measured guarantees.
Architecture and decisions · github
Adr Generator
Turns a concrete architecture decision into a numbered ADR with context, alternatives, consequences and references, asking for missing decision inputs first.
Why we selected itIt will not start until title, context, decision, alternatives, and stakeholders are present, and it tells the agent to ask rather than invent those fields.
Weigh upThe /docs/adr/ path is ambiguous outside its expected project layout; choose an explicit repository-relative output directory before use.
Security review · Trail of Bits
Adversarial modeler
Pins Trail of Bits' differential-review agent that models high-risk changes and says not to run on low or medium risk diffs.
Why we selected itThe catalogue has the differential-review skill and not this adversarial-modeler agent.
Weigh upThis review did not execute the resource.
Agent development · Affaan Mustafa
Agent evaluator
Scores an agent answer on five axes and requires evidence for the score.
Why we selected itThe role scores another agent's answer and says not to re-perform the original task.
Weigh upThe report must match scripts/evaluate.py, which is not contained in this file.
Agent development · VoltAgent
Agent Memory Curator
Maintains a confirmed agent memory store with dated facts, visible corrections and user-reviewed pruning proposals.
Why we selected itRequires confirming the existing store rather than creating a competing memory system, and forbids storing secrets.
Weigh upThe file is a curation procedure, not a memory database, recall service or cross-session hook installer.
Agent development · Anthropic
Agent SDK verifier (Python)
Checks a Python Claude Agent SDK app for SDK usage, environment setup and secret handling, and reports PASS, PASS WITH WARNINGS or FAIL.
Why we selected itThe environment checks are concrete: .env.example with ANTHROPIC_API_KEY, .env listed in .gitignore, and no API key hardcoded in source files.
Weigh upNo tools field is declared, so in Claude Code it inherits every tool available to subagents, including write and shell tools.
Agent development · anthropics
Agent Sdk Verifier Ts
Checks a TypeScript Claude Agent SDK project against its configuration, current SDK guidance, type-checking results and environment-variable hygiene before a structured readiness report.
Why we selected itChecklist is SDK-specific: package name, "type":"module", engines, MCP/subagent/permissions patterns, ANTHROPIC_API_KEY in .env.example and .env gitignore.
Weigh upNo tool allowlist is declared, so documentation retrieval and shell execution use inherited host permissions.
Agent development · VoltAgent
Agent-Log Knowledge Synthesizer
Finds recurring patterns in an explicitly scoped set of agent logs and writes cited findings with counts derived from the actual files.
Why we selected itStops on an empty or unspecified source scope and distinguishes recurring evidence from a single unconfirmed anecdote.
Weigh upIt cannot run analytics jobs or query remote services with its declared file tools; inputs must be accessible files.
Repository assessment · GitHub community
AgentRC readiness report renderer
Turns actual AgentRC assessment JSON and repository context into a static HTML report with policy-aware findings and remediation priorities.
Why we selected itIt requires real JSON input and retains that input in the report for inspection.
Weigh upThe source invokes an unpinned GitHub package through npx and depends on a sibling HTML template; no standalone bundle is offered. Pin and review the executable separately.
Security and governance · awslabs
AI-DLC Compliance Agent
Organizes data classification, requirement-to-control mapping and compliance evidence into an advisory role within an AI-DLC project.
Why we selected itConnects each proposed control to its technical implementation and required evidence such as logs, configuration and test results.
Weigh upRequires the AI-DLC framework, active-space policies and resolved knowledge references; the raw persona download is not a complete installation.
Documentation · wshobson
API Documenter
Builds API documentation as a contract-first artefact — OpenAPI 3.1 specifications, SDK generation and developer portals — with documentation-driven testing that validates examples and responses against the schema.
Why we selected itTreats the OpenAPI specification as the source of truth: contract-driven development, schema-validated examples, and documentation-driven testing with mock servers generated from the spec.
Weigh upA capability-list persona without an output schema or evidence gate; it does not itself verify that generated docs match a running API.
Architecture and decisions · Agency Agents
API Platform Engineer
Designs contract-first public APIs with a backward-compatibility check on every change, forcing breaking changes through versioning and a humane deprecation runway.
Why we selected itDefault requirement that every API change is checked against the contract for backward compatibility, with breaking changes routed through versioning and deprecation, never silent breaks.
Weigh upConvention choices (naming case, pagination style) are consistency mandates; adopting projects must pick and record their own standards.
Documentation and knowledge · wshobson
API Reference Builder
Builds a versioned API or configuration reference from an inventory of the implementation’s public interfaces.
Why we selected itQuality standards say every public interface is documented and that accuracy is verified against the implementation.
Weigh upThe Haiku model alias and absent tool allowlist are host-specific configuration choices.
System administration · GitHub community
Arch Linux maintenance guide
An Arch-specific administration profile for rolling updates and package-state diagnosis.
Why we selected itIt distinguishes full rolling upgrades from unsupported partial-upgrade patterns and inspects recent package changes first.
Weigh upThe declared shell/edit tools can change the real machine; no system snapshot, privilege boundary or package trust check is enforced by the Markdown.
Architecture and decisions · anthropics
Architecture Critic
Challenges modernization designs and transformed code against actual requirements, failure paths and simpler alternatives, ending with prioritized actionable findings.
Why we selected itSeparate review lenses distinguish architecture proposals from already-transformed code.
Weigh upRead-only shell use is an instruction; Bash remains available and must be constrained by the host.
Geospatial automation · Agency Agents community
ArcPy geoprocessing workflow designer
Designs ArcPy toolbox and batch-processing workflows with parameter validation, explicit environments and lock cleanup.
Why we selected itThe definition calls for dependent-parameter validation and user-facing error/progress reporting before long-running processing.
Weigh upThe source mixes legacy Data Driven Pages terminology into Pro guidance; current Esri documentation calls this Map Series and says it is not a geoprocessing toolset.
Migration · GitHub community
Arm workload migration assistant
An Arm migration workflow that checks source, dependency and container architecture assumptions through the Arm MCP service.
Why we selected itIt distinguishes a language client package version from the corresponding server product version.
Weigh upIts body names learning_path_server although the embedded allowlist does not, uses a mutable latest image and mounts the workspace writable. CPU feature advice must be checked against the exact target; migration does not guarantee speedup.
Service integration · VoltAgent
Authentication integration engineer
Designs identity flows around actual providers, client types and tenant boundaries, including negative tests and staged rollout.
Why we selected itThe checklist follows assertion validation, token lifecycle, server-side data access and recovery flows rather than stopping at login success.
Weigh upThe source is a broad design checklist, not a provider implementation or security guarantee; verify each flow against current provider/standard documentation.
Diagnostics · GitHub Next
AWF diagnosis
Fingerprints an Agentic Workflow Firewall failure, matches one finding, and reports the symptom, the finding id, and one next probe before naming a fix.
Why we selected itThe procedure records unknown rather than guessing, classifies one failing boundary, and matches the narrowest finding before it names a remedy.
Weigh upThe frontmatter has no tool allowlist. A host that grants write tools is not stopped by this file. Probes named in the catalog include docker exec, awf --version, findmnt, grep of a lock file, and gh aw logs. This review did not run them.
Operations and reliability · GitHub
AWS Incident Triage
Time-boxes an AWS alarm investigation and defaults to read-only operations.
Why we selected itThe persona defaults to read-only work and forbids infrastructure mutation without explicit approval.
Weigh upNo tool allowlist is declared. Enforce read-only AWS permissions; the source itself requires explicit approval before infrastructure mutation.
Testing and review · HashiCorp
AWS provider maintainer review
A short review persona for AWS provider changes. It says not to open pull requests.
Why we selected itThe file includes secrets in code, logs and version control in its review checklist.
Weigh upThe sibling @contributor persona was not opened. This file is a checklist, not a complete provider workflow.
Cloud architecture · GitHub community
AWS serverless architecture planner
A serverless design profile for event flows, managed state and observable failure handling.
Why we selected itIt asks for current AWS primary guidance before recommending service patterns.
Weigh upWorking examples are an output aspiration; IAM, quotas, retries, costs and service integrations must be validated in the target account.
Workflow integration · GitHub community
Azure Logic Apps workflow designer
A Logic Apps profile for JSON workflow definitions, connector integration and failure paths.
Why we selected itIts scope makes workflow definition structure and expressions explicit rather than treating integration as an opaque diagram.
Weigh upThe gpt-4 alias and named Microsoft/Azure documentation tools may need host adaptation; connector authentication and workflow execution are not bundled.
Security and governance · GitHub
Azure Policy analyzer
Requires evidence in the report and has a guardrails heading against invented policy data.
Why we selected itGuardrails forbid fabricated policy data and forbid Azure writes unless the user asks.
Weigh upTools include edit, execute, and a wildcard azure-mcp/*, which is broader than the no-write guardrail.
Infrastructure and delivery · github
Bicep Planning
Writes Azure Bicep implementation plans as deterministic, machine-readable Markdown under a write-scope guardrail — only .bicep-planning-files/ — grounding every resource in Microsoft docs and pinning Azure Verified Module versions from their changelogs.
Why we selected itA hard write-scope guardrail: only files under .bicep-planning-files/ may be created or modified, and the plan itself is out of deployment scope — no pipelines, no processes, no next steps.
Weigh upA planner, not a reviewer: it produces the plan file and stops; validating or deploying the resulting Bicep is out of its declared scope.
Creative tooling · Agency Agents community
Blender add-on engineer
A Blender tooling profile for asset validators, exporters and repeatable artist workflows.
Why we selected itIt prefers the direct data API over context-sensitive operators and requires actionable cancellation instead of silent success.
Weigh upThe snippets are illustrative and were not run in Blender; confirmation rules are prompt instructions, and omitted tools inherit host capabilities.
Evaluation workflows · anthropics
Blind skill-output comparator
Compares two anonymized task outputs against an explicit content and structure rubric, then writes a reasoned comparison JSON.
Why we selected itThe role keeps the skill identities hidden while comparing concrete output artifacts.
Weigh upThis is an orchestrated role prompt without custom-agent YAML frontmatter; pass it to a bounded worker or adapt a separate working copy.
Modernization and specifications · anthropics
Business Rules Extractor
Extracts domain calculations, validation and policy rules from legacy code into concrete test scenarios with source references and uncertainty questions.
Why we selected itSeparates business rules from infrastructure behavior and requires exact code locations.
Weigh upObserved legacy behavior is not automatically the business behavior that should be retained.
Security review · Trail of Bits
C review worker
Pins the c-review producing worker that reads source, writes one part file, and states that it has no shell.
Why we selected itThe catalogue has the c-review skill and not this producing worker.
Weigh upThis review did not execute the resource.
Language development · Seth Hobson
C systems developer
A compact C profile centered on ownership, system-call errors and measured performance.
Why we selected itThe short checklist prioritizes allocation ownership and checking system-call return values.
Weigh upEvery malloc needs free is shorthand rather than a lifetime proof; no compiler, test framework, debugger or Valgrind execution is supplied.
Testing and review · affaan-m
C# Reviewer
Reviews C# changes by running dotnet build and dotnet format verification before reading the diff, then ranks findings from SQL injection and async misuse to nullable reference types and EF Core traps.
Why we selected itGates the review on executed commands — dotnet build and dotnet format --verify-no-changes run before the diff is read, so findings rest on toolchain output.
Weigh upClosing pointers to sibling dotnet-patterns and csharp-testing skills reference companion files that are not part of this standalone definition.
Build and runtime diagnostics · Affaan Mustafa
C++ build error resolver
Diagnoses compilation, CMake, linking and template errors through a small-edit, rebuild and test loop.
Why we selected itIt distinguishes missing symbols, duplicate definitions, include problems and template deduction failures.
Weigh upIts example clang-tidy command hardcodes C++17 and truncates/suppresses diagnostics; adapt it to the actual compile database and preserve complete logs.
Testing and review · affaan-m
C++ Reviewer
Reviews C++ changes by running clang-tidy and cppcheck before reading the diff, then ranks findings from raw new/delete, buffer overflows and use-after-free to format-string attacks and integer overflow on untrusted input.
Why we selected itLeads with memory safety as the CRITICAL tier — raw new/delete, C-style arrays and strcpy/sprintf without bounds, use-after-free and dangling iterators, uninitialized reads, missing RAII — which is the correct priority order for C++ review.
Weigh upThe closing pointer to a sibling cpp-coding-standards skill references a companion file that is not part of this standalone definition.
Architecture documentation · Seth Hobson
C4 component documentation synthesizer
Groups code-level documentation into components with explicit responsibilities, interfaces and source-document links.
Why we selected itIt explicitly scopes each component diagram inside a single container and records external dependencies separately.
Weigh upThe definition supplies templates but requires your code-level documents; it does not include or run the sibling C4-Code agent.
Architecture and decisions · wshobson
C4 System Context Documenter
Synthesizes system and component documentation into a stakeholder-facing context diagram, actor list and external-dependency map.
Why we selected itDefines a full c4-context.md output covering system purpose, personas, features, journeys and external systems.
Weigh upAssumes container, component and system documentation are available; the file does not include those inputs or the other C4 agents.
Geospatial design · Agency Agents community
Cartography design reviewer
Plans thematic map styling, label hierarchy and layouts for a defined audience and delivery medium.
Why we selected itIt separates sequential, diverging and categorical encodings and connects basemap detail to the audience and subject.
Weigh upThe sample style JSON is descriptive, not a validated MapLibre/ArcGIS import format.
Security workflow components · Anthropic
Claude Security code explorer
A Claude Security search component that maps one code path for a caller and returns file/line evidence.
Why we selected itRepository instructions, comments, filenames and commit messages are treated as evidence rather than commands to change the task.
Weigh upRead-only is a prompt policy: Bash remains declared and can execute writes unless the host enforces the stated command restrictions.
3 sources · 2026-10-14
Read the full review Upstream setup linkedProprietary — Anthropic internal-use termsSecurity workflow components · Anthropic
Claude Security patch generator
A security-fix component that stages one narrow change in an orchestrator-created scratch checkout.
Why we selected itIt locates drifted findings by snippet and symbol rather than trusting old line numbers.
Weigh upThe source stages all scratch changes with git add -A and has Bash/Edit/Write; its workspace-name check is not an OS-level containment boundary.
5 sources · 2026-10-14
Read the full review Upstream setup linkedProprietary — Anthropic internal-use termsSecurity workflow components · Anthropic
Claude Security patch verifier
A security-patch verifier that derives the staged diff and separates tested behavior from source-only confidence.
Why we selected itIt rejects unstaged/untracked changes, derives the diff with external drivers disabled and checks the pre-change exploit before judging a fix.
Weigh upPASS is a model judgment, not a security guarantee; project tests can execute code and dependencies, and the no-edit instruction does not sandbox Bash.
5 sources · 2026-10-14
Read the full review Upstream setup linkedProprietary — Anthropic internal-use termsInteractive programming · GitHub community
Clojure REPL pair programmer
Reads a Clojure source file, develops and checks expressions in a live REPL, then applies the change with structural editing tools.
Why we selected itIt requires reproducing current behaviour and checking multiple inputs before modifying files.
Weigh upThe definition assumes a live REPL and structural editing tools but configures neither. It is not a REPL server or editor extension.
Operations and reliability · GitHub community
Cloud and SaaS Outage Triage
Separates upstream cloud/SaaS incidents from application failures with a timestamped evidence snapshot before any code is changed.
Why we selected itRequires a timestamped dependency-health snapshot and separates confirmed facts, plausible hypotheses and unknowns before proposing code changes.
Weigh upRequires the OutageDeck MCP server (remote http) plus repository/log access; without it only the local-investigation half of the workflow runs.
Network diagnostics · Seth Hobson
Cloud network engineer
Structures cloud and hybrid network diagnosis around layered connectivity, DNS resolution, TLS trust and observed traffic.
Why we selected itIt requires tracing resolution from client to authoritative DNS and validating certificate chains.
Weigh upPacket captures and flow logs may contain sensitive data; give the host only the intended interfaces, accounts and diagnostic targets.
Architecture and decisions · anthropics
Code Architect
Turns observed repository conventions into a concrete feature blueprint with file changes, interfaces, trade-offs and an implementation sequence.
Why we selected itRequires existing patterns and conventions to be supported by file-and-line references before proposing a design.
Weigh upThe prompt asks for one decisive design; use another review process when competing designs or unresolved requirements need comparison.
Codebase understanding · anthropics
Code Explorer
Traces an existing feature from entry points through state changes and storage, producing a source-linked map for the next developer.
Why we selected itRequires call-chain and data-transformation tracing, so the output goes beyond a directory summary.
Weigh upIt produces an explanatory map rather than implementing changes or proving behavior with executed tests.
Code maintenance · anthropics
Code Simplifier
Refines recently changed code for clearer structure and naming while explicitly keeping its behavior and scope intact.
Why we selected itLimits the default pass to code changed in the current session and explicitly prioritizes behavior preservation.
Weigh upIts embedded coding preferences target JavaScript, TypeScript and React; they need review against other languages and the project’s own conventions.
Code review · Trail of Bits
Code slice worker
Pins the Trailmark worker that answers from one source packet, returns a single JSON object, and says it cannot read the repository.
Why we selected itThe agents index has no entry for plugins/trailmark/agents/code-slice-worker.md.
Weigh upThis review did not execute the resource.
Developer education · Seth Hobson
Code tutorial designer
Turns code-backed learning goals into progressive tutorials with runnable examples, practice and troubleshooting checkpoints.
Why we selected itIts sequence introduces a minimal example before guided practice, variations and independent challenges.
Weigh upThe source provides educational heuristics, not evidence validating learning-style categories or completion-time estimates.
Codebase understanding · Agency Agents
Codebase Archaeologist
Audits codebases touched by many AI sessions for silent logic mismatches, dead code and doc-vs-code divergence, producing findings only.
Why we selected itExplicitly does not rewrite or refactor; it produces precise, evidenced, prioritized findings plus a maintained drift registry answering which files are safe to build on.
Weigh upThe file declares no tools allowlist; run it with host-default tools restricted to read-only, since the role is findings-only by instruction, not by enforcement.
Codebase understanding · agency
Codebase Onboarding Engineer
Orients new developers in an unfamiliar repository with a fixed three-level briefing — one line, five minutes, deep dive — where every statement names the files it was read from and unvisited files are listed as not inspected.
Why we selected itEvidence discipline is structural, not aspirational: it may never state that a module owns behaviour without pointing at the implementing file, and a partial answer must list which files were inspected and which were not.
Weigh upStrictly read-only and descriptive: it will not evaluate code quality, suggest changes, or test that documented setup steps actually run.
Documentation and knowledge · GitHub community
CodeTour Guide Builder
Plans and writes guided code walkthroughs with file references and a maintenance checklist.
Why we selected itThe schema marks the tour title and each step description as required.
Weigh upNo tool allowlist is declared; host permissions determine file access and command execution.
Pull-request review · OpenAI
Codex Rust review
A Codex label prompt for one codex-rs pull request: a short change summary, then a friendly review that pushes mixed changes apart and flags unsafe and crate placement.
Why we selected itThe review has to ask for a one-to-two sentence change summary and a one-to-two sentence review, and it pushes back when one pull request mixes a refactor with a new feature.
Weigh upThis is a repository-specific Codex review prompt for codex-rs; adapt crate assumptions deliberately for other Rust projects.
Model evaluation · GitHub partners
Comet Opik
Instruments an LLM app with Opik, versions its prompts, and investigates traces, token cost and latency through the Opik MCP server.
Why we selected itIt does not continue with MCP commands until a configuration path is confirmed, validates setup with masked keys, and never initialises git or rewrites repository history.
Weigh upThe Bronze, Silver and Gold instrumentation levels are the file's own scheme, not an external standard.
Documentation · anthropics
Comment Analyzer
Checks comments and docstrings against implementation behavior, then separates factual errors, worthwhile improvements and obsolete explanations.
Why we selected itExplicitly cross-checks signatures, side effects, error conditions and complexity claims against source code.
Weigh upThis is advisory review: the body explicitly forbids modifying code or comments.
Change communication · Affaan Mustafa
Commit message reviewer
Derives a commit message from staged changes and repository history, flagging breaking contracts and mixed concerns without creating a commit.
Why we selected itIt infers conventions from recent non-merge subjects instead of forcing a fixed subject length.
Weigh upRecent history can contain inconsistent conventions; inspect the inferred style before accepting the message.
Payments · Stripe
Company researcher
Stripe Claude agent that reads a company URL or description and maps the business to a Connect integration shape with a confidence cap for description-only input.
Why we selected itThe catalogue's Stripe skill is stripe-pay, not this Connect research agent.
Weigh upThis review did not execute the agent or fetch a company site.
Security and governance · 0xSteph
Compliance mapper
Maps a finding to a control only as far as the finding supports, and names the frameworks it will not pretend to certify.
Why we selected itThe honesty rule says to map only what the finding supports and not to claim a control passed or failed beyond the evidence.
Weigh upThe file says to confirm the current control revision before citing numbers, and it does not bundle those standards.
Planning workflow components · Seth Hobson
Conductor artifact validator
A Conductor review profile checking project context, track metadata and status consistency.
Why we selected itIt cross-checks track registry entries against directories, plans and metadata rather than trusting a completion checkbox alone.
Weigh upIt claims read-only operation but declares Bash; directory/grep examples are not an enforced shell allowlist, and prompt heuristics may drift from the installed Conductor version.
Change planning · GitHub community
Context Architect
Maps the files, dependency paths, existing patterns and tests for a proposed change before requesting permission to implement it.
Why we selected itThe required context map separates directly changed files from related files and describes why each belongs in scope.
Weigh upThe definition can edit files, run terminal commands and fetch the web. Its plan-first wording is a prompt rule, not a read-only sandbox.
Pull-request review · GitHub
Contribution checker
Reads one pull request against that repository's contributing guide and returns one JSON verdict from checklist answers.
Why we selected itThe agent must answer from the pull request metadata, diff, and CONTRIBUTING.md, and must not browse the rest of the repo.
Weigh upuser-invokable is false: an orchestrator must call the role.
Resilience testing · VoltAgent
Controlled failure-experiment designer
Frames failure experiments around a steady-state hypothesis, bounded impact, observable recovery and explicit stop controls.
Why we selected itThe workflow begins in non-production and expands scope gradually rather than beginning with combined failures.
Weigh upThe profile provides no executable safety controller or permission boundary; the host must enforce the approved failure targets and abort mechanism.
Security operations · Agency Agents community
Credential lifecycle and leak-response planner
A credential-lifecycle profile connecting detection, ownership, provider revocation and controlled rotation.
Why we selected itIt distinguishes removing a key from source from revoking it at the provider and asks for an exposure-window usage audit.
Weigh upShort lifetimes do not make blast radius near zero; sample scanner versions and grants require environment-specific verification. Revocation, deployment and history rewriting are consequential actions.
Security review · 0xSteph
Crypto Analyzer
Reviews cryptographic design and code for weak primitives and mishandled keys, and points exploitation elsewhere.
Why we selected itThe hard refusal is defeating cryptography outside the authorized scope or weakening production cryptography.
Weigh upRead/search/web tools support advisory review; the prompt is not a cryptographic implementation or assurance proof.
Build and runtime diagnostics · Affaan Mustafa
Dart/Flutter build error resolver
Diagnoses analyzer, platform build, package-resolution and generated-code failures with bounded changes.
Why we selected itThe workflow separates null/type errors from pub solver conflicts, build_runner output and Android/iOS configuration.
Weigh upIts examples include force unwraps, dependency overrides and deletion of generated outputs; do not apply these without establishing the intended data and generator contract.
Data and analytics · wshobson/agents
Data Engineer
A broad design prompt for batch and streaming pipelines, data quality, schema evolution and operational handoffs.
Why we selected itCovers failure handling, incremental processing and schema evolution alongside the happy-path transformation.
Weigh upThe file supplies instructions, not an executable pipeline or verified expertise across every listed cloud and engine.
Security and privacy · Trail of Bits
Data Flow Analyzer
Traces a suspected vulnerability from sink to sources with trust levels and validation points, producing structured read-only evidence.
Why we selected itDeclares only Read, Grep and Glob and states read-only explicitly; the four-phase trace (sink, sources with trust levels, per-validation pass/fail/bypass, full path) is checkable line by line.
Weigh upDesigned as a Phase-1 worker for the fp-check skill; used standalone the caller must supply the bug description (claim, class, file:line, trigger, impact) as dispatch input.
3 sources · 2026-10-07
Read the full review Attributed definition availableCC-BY-SA-4.0Database architecture · Seth Hobson
Database architecture designer
Designs schemas, storage choices and migration plans from application access patterns and operational constraints.
Why we selected itThe role defaults to advisory work and prohibits file changes or migration execution unless explicitly requested.
Weigh upThe broad product inventory is not a current compatibility matrix; verify exact engine/provider support before selecting a design.
Data and retrieval · wshobson
Database Cloud Optimization Database Optimizer
Investigates database bottlenecks from query plans and workload evidence, then proposes query, index, caching or partitioning changes with validation.
Why we selected itBehavioral trait “Measures performance first using appropriate profiling tools before making optimizations” is the right order for this role.
Weigh upThis is an implementation-capable role and it declares no tool allowlist; database credentials and migration authority must be scoped in the host.
Database operations · Seth Hobson
Database reliability administrator
Plans database availability and maintenance around tested recovery, observed resource limits and documented operational procedures.
Why we selected itIt calls for restoring backups regularly rather than treating backup-job success as recovery evidence.
Weigh upThe role spans many engines without supplying engine-specific runbooks or validated commands; select the exact deployment before acting.
Operations and reliability · Agency Agents
Database Reliability Engineer
Keeps stateful datastores available and recoverable with tested restores, drilled failovers and lock-budgeted online migrations, each carrying a written back-out plan.
Why we selected itPairs backfill and rollback rules explicitly: batched backfills kept out of exclusive-lock transactions, plus a written back-out plan and blast-radius estimate for every migration, failover and large delete.
Weigh upThe RPO and RTO figures in the template are illustrative targets, not guarantees; each project must set and then prove its own numbers with drills.
Data and retrieval · affaan-m
Database Reviewer
Reviews PostgreSQL queries, schemas and migrations for performance, integrity and Row Level Security, reporting findings only.
Why we selected itDeclares only Read, Grep, Glob and Bash (no Write/Edit), so the review-only posture is enforced by the tool list, not just by instruction.
Weigh upDiagnostic queries need a readable DATABASE_URL; the definition does not gate who may supply production credentials.
System administration · GitHub community
Debian maintenance guide
A Debian administration profile emphasizing package state, reversible changes and vendor-file preservation.
Why we selected itIt distinguishes interactive apt usage from apt-get scripting and records package pinning when suites are mixed.
Weigh upThe profile’s default firewall and support preferences are not a compatibility guarantee; terminal/edit access remains capable of changing services and packages.
Documentation · Microsoft
Deep Wiki catalogue architect
A documentation component that builds a code-cited wiki outline and audience-specific onboarding structure.
Why we selected itEvery catalogue prompt must name source file/line references and diagram/table explanations retain source provenance.
Weigh upIts suggested HEAD branch command does not discover the remote default branch; pin the actual revision before constructing citations. Diagram quotas do not prove clarity.
Security query drafting · GitHub community
Defender Scout KQL
Drafts and explains Defender Advanced Hunting queries from a security question or existing query, with time-window, projection and join checks.
Why we selected itThe role separates query generation, inspection, optimisation suggestions and explanation, and supplies a repeatable response format.
Weigh upNo query execution tool is configured. “Validates” in the upstream description means model inspection here; syntax, table availability and detection quality were not measured.
Documentation and knowledge · GitHub community
Demonstrate Understanding
Checks a developer’s understanding of a codebase through focused questions and evidence-backed feedback.
Why we selected itThe role is questioning, and the close says the goal is understanding rather than testing.
Weigh upThis is an interactive tutoring role; it requires the learner to answer and cannot independently demonstrate their knowledge.
Infrastructure and delivery · wshobson
Deploy with Verification
Refuses to say shipped until the live system confirms the new revision.
Why we selected itA test failure stops the flow, and the words deployed and shipped are forbidden until live verification and the state-doc update both succeed.
Weigh upTools are Bash, Read, and Edit, so the file can change the state doc and run the deploy command.
Design systems · Seth Hobson
Design system architect
A profile for organizing design tokens, component APIs and multi-brand theming.
Why we selected itIt distinguishes primitive, semantic and component tokens instead of treating every token as a raw colour.
Weigh upFigma, Style Dictionary and visual-regression services are suggested ecosystem tools; the profile does not configure or authenticate them.
Security review · 0xSteph
Detection engineer
Writes detection content for a named technique. It is not a licence to attack a system.
Why we selected itEvery rule must include the Sigma-shaped fields in the file, including log source, false positives, and level.
Weigh upTools include Write, Edit, WebFetch, and WebSearch, so the file can change the workspace and use the network.
Developer tools · Agency Agents community
Developer CLI tooling engineer
Designs command interfaces with stable machine output, actionable errors, configuration precedence and measured startup cost.
Why we selected itThe role treats flags, JSON output and exit codes as versioned interfaces and requires explicit deprecation paths for changes.
Weigh upStartup budgets and time-saved examples are illustrative targets, not measured results for your tool.
Architecture and decisions · github
Devils Advocate
Stress-tests a proposal through one objection at a time, then switches to a balanced recap when the user ends the debate.
Why we selected itA clear dialogue protocol keeps each challenge focused and gives the user an explicit way to stop adversarial mode.
Weigh upIt deliberately challenges rather than neutrally weighs a proposal during debate; use the final recap to recover balance.
Browser and performance · github
Devtools Regression Investigator
Reproduces browser regressions and links screenshots, console or network evidence to likely code paths, returning a structured bug report before any requested fix.
Why we selected itThe file forbids declaring root cause without browser evidence or code correlation, and forbids treating a flaky repro as solved.
Weigh upChrome DevTools MCP and optional Playwright are described but not installed or explicitly named in the tool allowlist; configure the required browser tools separately.
Operations and reliability · Seth Hobson
DGX Spark ops engineer
Diagnoses one named hardware stack and does not generalise its checks to every GPU.
Why we selected itHardware identity is checked before any other diagnosis, and a Spark finding must cite a G-number.
Weigh upThe checks depend on spark-environment-setup, spark-training-gotchas, and spark-memory-thermal-ops, which this file does not contain.
Testing and review · GitHub partners
Diffblue Cover Java Unit Test Agent
Collects the Java packages, classes or methods to cover, runs Diffblue Cover through its MCP server, reports the result and commits the tests.
Why we selected itTools are limited to the DiffblueCover MCP server. Test generation and validation are delegated to Diffblue Cover rather than written by the model.
Weigh upStep 4 commits the generated tests with no review step in between. Review the diff first.
Code maintenance · Trail of Bits
Dimension propagator
Propagates dimensional comments through assigned files, refuses executable edits, and reports mismatches instead of fixing them.
Why we selected itIt requires a status for every assigned file and says not to skip a file silently.
Weigh upA Markdown profile does not register a native subagent by itself.
Testing and review · affaan-m
Django Reviewer
Reviews Django changes for ORM, migration-safety, DRF, Celery and security pitfalls with severity-graded findings and merge-blocking approval criteria.
Why we selected itMigration-safety rules are rollback-aware: RunPython requires reverse code, column drops go two-phase across deployments, and every model change must carry a migration.
Weigh upReferenced companion skills such as django-patterns, django-security and django-tdd are not bundled with the standalone file; treat those pointers as optional.
Infrastructure and delivery · voltagent
Docker Expert
Reviews and rebuilds Dockerfiles for production — multi-stage builds, layer caching, non-root execution, distroless bases — and pairs image optimization with supply-chain hardening: SBOMs, cosign signing, SLSA provenance and CIS benchmark compliance.
Why we selected itThe review loop is concrete: read existing Dockerfiles and compose files, analyze security posture and build performance, then implement — with .dockerignore, HEALTHCHECK, BuildKit and non-root execution named as first-class checks.
Weigh upIts excellence checklist prescribes fixed numeric targets (images under 100MB, builds under 5 minutes, cache hit rate over 80%, CIS over 90%) that many legitimate workloads cannot meet; treat them as prompts, not gates.
Documentation · wshobson
Documentation Generation Docs Architect
Builds a structured architecture manual from an existing codebase, covering design rationale, component interactions, operational behavior and paths into the source.
Why we selected itOutput contract is specific: Markdown, heading hierarchy, code blocks, tables, links as file_path:line_number, 10-section skeleton.
Weigh upThe source favors extensive manuals; choose a bounded audience and scope to avoid documentation that is costly to maintain.
Evidence review · GitHub community
Doublecheck
Extracts the claims in AI-generated text, searches for sources, and reports links and risk flags for a person to check, without issuing verdicts.
Why we selected itTools are web_search and web_fetch only. The agent reads and searches; it cannot edit files or run commands.
Weigh upThe pipeline itself is in the doublecheck skill, skills/doublecheck/SKILL.md, which the agent follows; that skill was not opened.
Observability · GitHub community
Dynatrace evidence-guided incident analyst
A Dynatrace investigation reference connecting traces, exceptions, deployments and current security findings.
Why we selected itIt asks to inspect exception span events and correlate logs, metrics and traces rather than infer cause from one aggregate.
Weigh upThe definition embeds a development-tenant URL; never send a real token there by copying it unchanged. DQL examples, latest-scan assumptions and wildcard MCP capabilities require validation against the actual tenant.
Finance operations · Anthropic
Earnings reviewer
Updates a coverage model from an earnings print and drafts the post-earnings note, marking any figure without a FactSet, Daloopa or filing source as [UNSOURCED].
Why we selected itEvery number must cite a source. A figure that cannot be sourced from FactSet, Daloopa or a filing is marked [UNSOURCED] rather than filled in.
Weigh upDeclared tools include Write and Edit, and it updates the live coverage workbook through the model-update skill. Work on a copy.
Service integration · VoltAgent
Email delivery infrastructure engineer
Organizes email authentication, provider events, suppression and monitoring around an observed sending profile.
Why we selected itThe workflow connects SPF/DKIM/DMARC changes to provider events, global suppression and transactional-versus-bulk domain separation.
Weigh upThe fixed performance thresholds and success-message percentages are illustrative aspirations, not measured outcomes or universal provider requirements.
Embedded software · GitHub community
Embedded C code reviewer and implementer
Reviews or implements embedded C against the existing compiler, MCU constraints and project analysis rules.
Why we selected itThe role preserves compiler flags and target settings unless requested, avoids generated RTE/MCAL code and checks that new files enter the build.
Weigh upNo licensed coding standard, compiler, static analyzer or hardware test fixture is bundled.
Identity systems · Agency Agents community
Enterprise identity and authorization designer
Plans enterprise identity flows, session revocation and tenant authorization with explicit failure-path tests.
Why we selected itThe source consumes the login state before async exchange and explicitly requires atomic consumption in shared session stores.
Weigh upThe code contains application-owned helpers and incomplete WebAuthn ceremony wiring; use maintained protocol libraries and verify current APIs.
Operations and reliability · VoltAgent
Error Coordinator
Mines logs and agent output for recurring failure and cascade patterns, citing evidence paths and refusing to invent quantities.
Why we selected itRequires every reported count to be Grep/Glob-counted with path:line evidence, a two-independent-source minimum for patterns, and explicitly forbids reporting recovery rates, MTTR or cascades prevented.
Weigh upNo Bash tool is declared, so mining is limited to what Read/Glob/Grep can see; live detection, retries and circuit-breaking are explicitly out of scope.
Architecture and decisions · wshobson
Event Sourcing Architect
Plans event-sourced systems through aggregate boundaries, immutable event records, projections, sagas and event-versioning decisions.
Why we selected itCovers the practical linkage between command handling, read projections and cross-aggregate workflows.
Weigh upThe role assumes event sourcing is appropriate; it does not compare simpler persistence designs first.
Code review · Anthropic
Expert Rust Code Reviewer
Reviews a Rust diff for ownership, unsafe invariants, concurrency, error handling and public APIs, with compiler or Clippy checks for testable findings.
Why we selected itThe source covers ownership, unsafe contracts, async cancellation, lock ordering, secret-bearing Debug output and dependency/MSRV concerns in one review checklist.
Weigh upThe source contains 18 review categories. Its Excellent/Good/Needs Improvement/Poor labels are subjective report labels, not measured quality scores.
Security review · Trail of Bits
Exploitability verifier
Pins the fp-check phase-2 agent that stays read-only while deciding whether a suspected bug is exploitable, and that skips a proof the bug class does not need.
Why we selected itThe agents catalogue has no sourcePath for plugins/fp-check/agents/exploitability-verifier.md.
Weigh upThis review did not execute the resource.
Language-specific review · Affaan Mustafa
F# code reviewer
Reviews F# changes for exhaustive domain modelling, computation-expression flow, disposal and safe boundaries.
Why we selected itIt checks incomplete union matches, smart constructors, Option/Result handling and async binding.
Weigh upFunction-length and nesting thresholds are style heuristics; immutable/functional rewrites and sequence materialization must preserve semantics and measured resource costs.
Testing and review · affaan-m
FastAPI Reviewer
Reviews FastAPI changes for async blocking, dependency wiring, schema exposure and test setup, with file references and explicit unverified risks.
Why we selected itFocuses on framework-specific failure modes such as blocking clients in async routes, incorrect test overrides and sensitive response fields.
Weigh upIts severity checklist includes architectural preferences that require context; an inline database session is not automatically a demonstrated vulnerability.
Code review · Anthropic
Feature-dev code reviewer
Reviews the unstaged git diff for bugs and against CLAUDE.md rules, and reports only issues it rates 80 or higher.
Why we selected itThe frontmatter tools are Glob, Grep, LS, Read, NotebookRead, WebFetch, TodoWrite, WebSearch, KillShell and BashOutput. No Write, Edit or Bash tool is declared, so the review role does not include changing files.
Weigh upThe 0 to 100 confidence is the model's own rating of each issue, not a measured precision. The threshold of 80 hides lower-rated findings; it does not prove they are false.
Research and product · msitarzewski
Feedback Synthesizer
Organizes a supplied customer-feedback corpus into themes, product priorities and audience-specific reports for product and support teams.
Why we selected itSpecifies separate executive, product-team and customer-success deliverables rather than one undifferentiated summary.
Weigh upThe definition's accuracy, growth and service-level targets are aspirations; none were demonstrated by this source review.
Security review · 0xSteph
Fix Verifier
Retests a declared finding after a patch, and only during authorized testing.
Why we selected itCommands against a target are refused until the user declares scope.
Weigh upBash/Write/Edit are available, so declared target scope and non-destructive rules require host enforcement.
Code maintenance · Trail of Bits
Fixer
Applies dispatched blocking fixes only inside the workflow's scope globs, and records a verdict instead of skipping a finding or discarding the worktree.
Why we selected itThe description limits dispatch to the code-improver workflow and says the agent is not for ad-hoc editing.
Weigh upA Markdown profile does not register a native subagent by itself.
Cross-platform applications · Seth Hobson
Flutter application engineer
Plans Flutter widget architecture, native integration and testing around the actual target platforms.
Why we selected itThe profile pairs widget identity and rebuild concerns with real-device profiling and platform-channel boundaries.
Weigh upThe definition references a 2025 ecosystem and Riverpod 2.x; verify installed package APIs rather than treating its inventory as current guidance.
Testing and review · affaan-m
Flutter Reviewer
Reviews Flutter and Dart changes state-management-agnostically — detecting the project's BLoC, Riverpod, Provider or other setup first — then ranks widget rebuild, architecture and security findings with a per-severity verdict table.
Why we selected itAdapts to the project's chosen state management (BLoC, Riverpod, Provider, GetX, MobX, Signals or built-in) and its routing and DI approach before reviewing, explicitly to avoid flagging idiomatic usage of a different solution as a violation.
Weigh upThe closing pointer to a sibling flutter-dart-code-review skill references a companion file that is not part of this standalone definition.
Security review · 0xSteph
Forensics analyst
Starts from evidence acquisition and chain of custody, not from a guessed timeline.
Why we selected itBehavioral rules require hashes, chain of custody, and work on copies rather than original evidence.
Weigh upUse only with evidence the caller is allowed to handle; this file is not that authorization.
Browser and performance · github
Frontend Performance Investigator
Diagnoses a specific slow web flow from traces, network activity and code paths, then produces prioritized fixes with a validation method for each.
Why we selected itRequires a concrete reproduction path and recorded environment assumptions before recommending optimizations.
Weigh upChrome DevTools MCP and the optional Playwright fallback require separate configuration; neither is explicitly included in this file’s tool allowlist.
Codebase understanding · Trail of Bits (Omar Inuwa)
Function Analyzer
Builds a line-cited account of one function’s invariants, assumptions and callee dependencies before a later security review.
Why we selected itRequires reading callees and their branches before treating a precondition as enforced; function names are explicitly insufficient evidence.
Weigh upThe file requires sibling analysis formats and domain references plus a resolved {baseDir}; the download alone does not install the plugin.
6 sources · 2026-10-13
Read the full review Attributed definition availableCC-BY-SA-4.0Game development · Agency Agents community
Game economy designer
A game-economy profile that connects currency sources and sinks to simulated player cohorts.
Why we selected itEach currency needs a purpose, source, sink and rationale linked to a curve or simulation.
Weigh upSample balances, ratios and simulation durations are illustrative design targets, not validated outcomes or universal game-economy constants.
Game design · Agency Agents community
Game level layout designer
Creates level and encounter specifications that can be tested as grey boxes before art production.
Why we selected itThe workflow asks for before/after layout evidence tied to playtest observations and marks gameplay-critical geometry for the art team.
Weigh upFixed visibility/timing thresholds and success percentages in the source are aspirational examples, not validated outcomes.
Game development · Agency Agents community
Game narrative systems designer
A game-narrative profile for branch maps, character voice and coherent optional lore.
Why we selected itIt maps branch structure before writing lines and requires a dramatic function for each dialogue node.
Weigh upThe sample is labeled Ink/Yarn/generic and is not validated language-specific syntax; persona experience claims and success targets are not evidence.
Geospatial operations · Agency Agents community
GIS analyst
Handles scoped map production, layer hygiene and basic spatial queries with provenance and export checks.
Why we selected itIt distinguishes basic buffer/clip/query work from statistical modelling and automated ETL.
Weigh upA matching CRS label is insufficient evidence of correct coordinates; verify spatial plausibility and suitable distance/area units.
Geospatial quality · Agency Agents community
GIS data quality reviewer
Structures geospatial QA around reproducible feature-level findings, metadata, topology and delivery checks.
Why we selected itIt distinguishes declared CRS from actual coordinate plausibility and compares completeness to expected rows/extent.
Weigh upThe Python examples contain pass-only stubs; there is no executable validation suite in the definition.
Code maintenance · agency
Git Workflow Master
Establishes branch hygiene with copy-safe commands — trunk-based versus Git Flow diagrams, --no-track branch creation, worktrees for parallel work, force-with-lease over force-push — and warns before every destructive operation.
Why we selected itThe commands are copy-safe and explain their own sharp edges: --no-track is justified (a new branch must not inherit origin/main as upstream) and worktree creation is shown instead of a second checkout of the same branch.
Weigh upA workflow guide, not an auditor: it does not inspect an existing repository's history or branch state and report hygiene findings.
Infrastructure and delivery · github
Github Actions Expert
Designs and reviews GitHub Actions workflows around minimal token permissions, immutable action references, OIDC, deployment concurrency and validation.
Why we selected itAction pinning section forbids `@main`/`@latest`/moving tags, requires full commit SHA plus a version comment, and explains tag-move supply-chain risk.
Weigh upThe GitHub tool wildcard and terminal access can perform repository actions; the definition itself does not enforce a read-only review.
CI maintenance · GitHub
GitHub Actions Node runtime upgrade
Nine-step upgrade of a JavaScript or TypeScript GitHub Action to a newer Node runtime, treated as a breaking major version.
Why we selected itThe upgrade treats a Node runtime change as a major version bump in package.json.
Weigh upEdit and terminal tools can modify the action and run project npm scripts; inspect script side effects.
Documentation editing · GitHub
GitHub Docs builder writer
Writes GitHub Docs for the Builder persona and keeps enterprise-admin material out of that article.
Why we selected itBuilder content is for people doing the work themselves, and organization-wide administration is sent to the Driver persona.
Weigh upFrontmatter sets include-custom-instructions and does not name tools.
Documentation editing · GitHub
GitHub Docs content pipeline update
Updates official reference articles from ephemeral source docs and refuses instructions in those docs that ask for secrets or unrelated actions.
Why we selected itMissing prompt fields stop the run, and edits are limited to TARGET_ARTICLES plus reusables.
Weigh upThe security section says there is no shell, while step 4 still says to run npm run lint-content and a content render test.
Documentation editing · GitHub
GitHub Docs driver writer
Writes Driver documentation for people who administer GitHub at enterprise scale, and flags specific high-risk claims for a writer to verify.
Why we selected itThe audience is people who administer GitHub, and the file says to flag one risky claim rather than a whole article.
Weigh upThe file sets include-custom-instructions and does not show those instructions.
Documentation editing · GitHub
GitHub Docs GHES release notes
Drafts GitHub Enterprise Server release notes from a releases issue or a changelog pull request.
Why we selected itA private-preview title is skipped and returns an empty list, and the word deprecated is forbidden.
Weigh upThe file tells the agent not to ask follow-up questions when invoked programmatically.
Documentation editing · GitHub
GitHub Docs readability editor
Edits only the markdown files it is given, and keeps a sentence when a shorter wording would change the meaning.
Why we selected itEdits are limited to the specific markdown files provided. The file says not to move or delete files, and to keep the original sentence when an edit would change its meaning.
Weigh upThe definition requests a pull request and grants edit, execute and github/* capabilities. Select a restricted host policy for a local editing pass; prose scope is not a tool sandbox.
Finance operations · Anthropic
GL reconciler
Compares general ledger and subledger balances for a trade date, traces each break and drafts an exception report for controller sign-off.
Why we selected itDeclared tools are Read, Grep, Glob and the internal-gl and subledger MCP servers. The file says it produces a report and that ledger adjustments need human approval outside the agent.
Weigh upThe reader, critic and resolver workers named in the body are defined in managed-agent-cookbooks/gl-reconciler/subagents/, not in the plugin's agents directory. Installed as a plugin, the split between untrusted readers and the Write-holding resolver depends on that setup.
Build and runtime diagnostics · Affaan Mustafa
Go build error resolver
Resolves Go compiler, module, vet and linter failures while keeping changes bounded to the reported build problem.
Why we selected itThe workflow checks compilation before narrowing into interfaces, imports, cycles and module resolution.
Weigh upgo mod tidy, go get and cache cleaning can change dependencies or local state; review their scope before execution.
Testing and review · affaan-m
Go Reviewer
Reviews Go changes by running go vet, staticcheck and the diff first, then ranks findings from SQL injection and ignored errors to goroutine leaks, channel deadlocks and non-idiomatic patterns.
Why we selected itRuns go vet and staticcheck before reviewing and lists golangci-lint, -race builds and govulncheck as diagnostic commands, so findings are grounded in executed tooling.
Weigh upThe closing pointer to a sibling golang-patterns skill references a companion file that is not part of this standalone definition.
API architecture · Seth Hobson
GraphQL schema and resolver architect
Plans GraphQL schemas and resolver behaviour with batching, authorization, evolution and testing considerations.
Why we selected itThe definition covers DataLoader/N+1 concerns alongside field authorization, subscriptions and schema evolution.
Weigh upNo GraphQL server, schema registry or framework dependency is included.
Agent development · affaan-m
Harness Optimizer
Tunes agent configuration with an explicit evaluation definition, repeated trials, baseline checks and rollback instructions for failed experiments.
Why we selected itExplicit non-goals: no product-code rewrites; no slash-command invocation; run node scripts/harness-audit.js instead.
Weigh upRequires the upstream-style scripts/harness-audit.js, tests/run-all.js and skills/eval-harness/SKILL.md; a standalone download cannot run that workflow by itself.
Language development · Seth Hobson
Haskell typed-design developer
A Haskell profile that keeps pure logic separate from IO and favors total functions.
Why we selected itIt asks for safe total alternatives to partial functions and language extensions only when their purpose is explained.
Weigh upHigh-assurance wording describes intent; it is not a formal verification result or a compiled package.
Agent development · Anthropic
Hookify conversation analyzer
Reads a Claude Code conversation for user corrections and proposes regex hook rules, graded by severity, for the user to choose from.
Why we selected itTools are Read and Grep only. The agent returns findings, and the /hookify command then asks which rules to create.
Weigh upThe regex patterns are the model's proposals. They were not compiled or tested against a transcript here, and a broad pattern can block legitimate commands.
Cloud architecture · Seth Hobson
Hybrid-cloud architecture planner
Plans workload placement, connectivity and recovery across public, private and edge environments using explicit constraints.
Why we selected itThe response weighs data gravity, latency, operational needs and cost before assigning workloads to environments.
Weigh upThe broad provider/product inventory may contain dated names and does not prove integration support; verify current availability and contracts for each selected service.
Testing and review · wshobson
Hypothesis Debugging Investigator
Investigates one assigned bug hypothesis and sends the team lead a cited account of supporting evidence, contradictions and remaining gaps.
Why we selected itDefines confirmation, falsification and ambiguous evidence before investigation, reducing one-sided hypothesis confirmation.
Weigh upDepends on a lead and the host’s messaging/task tools. The single file does not create a team or an orchestration workflow.
Operations and reliability · Seth Hobson
Incident Responder
Runs the first five minutes of an incident as assessment and communication, not as an immediate change.
Why we selected itThe first section is a five-minute sequence: severity, a single incident commander, then stabilization.
Weigh upThe Sonnet model alias and absent tool allowlist are host-specific.
Game audio · Agency Agents community
Interactive game audio planner
Plans event-driven game audio, state parameters, voice budgets and target-hardware profiling.
Why we selected itIt connects gameplay state to middleware parameters and tests transitions such as combat, death and scene changes.
Weigh upThe budgets, codecs and middleware-only rules are example policies, not measured platform requirements or universal audio guidance.
Code maintenance · Agency Agents
Internationalization Engineer
Makes codebases translation-ready: complete ICU messages instead of concatenated fragments, CLDR-backed formatting, logical-property layouts that survive RTL, and pseudo-localization wired into CI so hardcoded strings fail the build.
Why we selected itStates eight concrete engineering rules — no concatenated fragments, CLDR plurals through ICU, Intl-only formatting, logical CSS properties, expansion budgeting, translator context, Unicode normalization, locale negotiation — each tied to the failure it prevents.
Weigh upThe success metrics (zero hardcoded strings, clarification requests below 2 per 1,000 strings) are targets for the codebase, not measured outcomes of running this agent.
Infrastructure · Agency Agents community
IoT fleet and staged-update architect
A fleet-design profile for intermittent devices, scoped identity and health-gated firmware rollouts.
Why we selected itIt combines bounded edge buffering with original timestamps and sequence-based backend deduplication.
Weigh upA/B partition diagrams do not prove a device is unbrickable; bootloader, flash, power-loss and recovery behavior require hardware testing. Rollout sizes are examples.
Testing and review · affaan-m
Java Reviewer
Detects Spring Boot versus Quarkus from the build file first and runs the framework build check, then reviews Java changes against framework-specific rules spanning JPA, Panache, MongoDB, security and concurrency.
Why we selected itFramework detection is an explicit first step — the build file is read and [SPRING] or [QUARKUS] rulesets applied, with both-present treated as a finding and neither-present noted as ambiguity — so framework-specific rules are never applied blind.
Weigh upClosing pointers to sibling springboot-patterns and quarkus-patterns skills reference companion files that are not part of this standalone definition.
Security and governance · GitHub partners
JFrog Security Agent
Checks a dependency upgrade against the organisation's JFrog Curation policy before recommending it, then applies CVE-specific remediation guidance.
Why we selected itBefore any change it must check the upgrade version against the organisation's Curation Policy with a JFrog MCP tool, and recommend only the policy-compliant version.
Weigh upIt is told to use JFrog MCP tools exclusively and not npm audit, CodeQL or the GitHub Advisory Database, so a finding is not cross-checked against a second source.
Project planning · Agency Agents community
Jira delivery traceability steward
A delivery-planning profile linking a confirmed Jira ticket to branch, commit, PR and release evidence.
Why we selected itThe profile refuses to invent ticket identifiers and separates Git-facing work from unrelated tasks that need no Jira process.
Weigh upThe supplied hook only checks name patterns; it does not verify a Jira issue exists or that branch and commit IDs match, and release subjects bypass its commit regex. Its develop/main convention is not universal.
Research and product · github
Jobs-to-be-Done UX Planner
Turns user context and jobs to be done into a journey map, flow specification and design handoff documents.
Why we selected itStarts by asking about users, context and pain points instead of treating a requested interface feature as the underlying goal.
Weigh upThe agent drafts research artifacts; it does not conduct interviews, validate personas or create Figma designs.
Language development · Seth Hobson
Julia scientific-code developer
A Julia profile emphasizing type-stable numerical code and reproducible package environments.
Why we selected itIts constraints require Pkg APIs rather than hand-editing Project.toml and explicitly request type-stability inspection.
Weigh upThe version and ecosystem descriptions are the prompt author’s scope, not a tested compatibility matrix; GPU and scientific packages are not included.
Testing and review · affaan-m
Kotlin Reviewer
Reviews Kotlin and Android/KMP changes for coroutine cancellation mistakes, Flow and Compose recomposition traps and clean-architecture violations, reporting only findings it holds with more than 80% confidence.
Why we selected itEncodes noise control as a rule — only findings held above 80% confidence are reported — and closes with a per-severity verdict table whose Block verdict is bound to any CRITICAL or HIGH issue.
Weigh upThe clean-architecture module rules assume a domain/data/presentation layering; projects with other layouts will get spurious boundary findings.
Finance operations · Anthropic
KYC screener
Extracts an onboarding packet, runs the firm's KYC rules and sanctions, PEP and adverse-media screening, and packages gaps for compliance sign-off.
Why we selected itEach KYC rule result carries pass or fail with an evidence reference. The agent recommends a risk rating, and the file says the compliance officer decides.
Weigh upScreening results and match confidence come from the firm's screening MCP and its lists. Their coverage was not assessed.
Code maintenance · GitHub community
LaunchDarkly Flag Cleanup
Removes obsolete feature flags by reading each environment’s live LaunchDarkly configuration first, deriving the forward value from what is actually served, and opening a PR that preserves production behaviour.
Why we selected itGates every removal on live evidence: per-environment flag configuration, lifecycle status and code references come from the LaunchDarkly MCP tools, and inconsistent state across critical environments stops the removal.
Weigh upRequires the LaunchDarkly MCP server and an LD_ACCESS_TOKEN; the frontmatter mcp-servers block is Copilot-specific configuration and needs porting to other hosts.
Codebase understanding · anthropics
Legacy Analyst
Builds an evidence-linked account of an unfamiliar legacy system by tracing entry points, stable data structures and actual control flow.
Why we selected itRequires source citations and explicitly distinguishes observed behavior from inferred intent.
Weigh upThe role is an analyst, not a migration implementer or an executable behavior test.
Model development · Seth Hobson
LLM fine-tuning architect
A fine-tuning planning component that requires a baseline and can recommend a retrieval or prompting alternative.
Why we selected itIt owns planning phases and must hand off a written brief rather than silently train or invent dataset-size guidance.
Weigh upNo model, dataset rights, hardware or evaluation harness is bundled; a training plan does not establish that fine-tuning helps.
Model evaluation · Seth Hobson
LLM Fine-Tuning Eval Engineer
Owns the pre-training evaluation baseline and post-training checkpoint verdict, with explicit per-trace records and no authority to train or automatically retrain.
Why we selected itDefines a stable eval/ measuring set outside disposable runs/ outputs and requires full assistant completions in the trace records consumed downstream.
Weigh upThis Markdown profile does not install an evaluation harness, companion skills or infrastructure, and declares model: sonnet with no tool allowlist. Host defaults and permissions remain decisive.
Model development · Seth Hobson
LLM fine-tuning training engineer
A training component that follows an approved brief and hands checkpoint promotion to a separate evaluator.
Why we selected itIt decodes packed examples to inspect formatting and requires a hardware/environment preflight before launching a job.
Weigh upThe source demands committing configuration before training; that is a separate repository mutation requiring task authorization. GPU jobs, downloads and checkpoints incur cost and consume storage.
Finance operations · Anthropic
LP statement auditor
Ties out a batch of LP capital-account statements against the fund NAV pack and recommends pass or hold for each statement.
Why we selected itDeclared tools are Read, Grep, Glob and the nav MCP server. It recommends pass or hold, and the file says investor relations distributes only after human sign-off.
Weigh upThe statement-reader and flagger workers are defined in managed-agent-cookbooks/statement-auditor/subagents/, not in the plugin agent file.
Accessibility · github
Markdown Accessibility Assistant
Reviews existing Markdown for accessible headings, links, lists and image descriptions, while reserving meaning-changing rewrites for human approval.
Why we selected itMission forbids generating docs from scratch and limits work to improving existing markdown, which is a real role boundary.
Weigh upThe scope is selected Markdown accessibility practices, not full web accessibility conformance.
Mathematics · Anthropic
Math proof worker, deep
Pins the higher-effort math-proof worker that answers one question per launch and writes the answer to a file the orchestrator reads.
Why we selected itThe catalogue has math-proof-worker and math-proof-judge and not this deep worker.
Weigh upThis review did not execute the resource.
Mathematics · Anthropic
Math-proof judge
Runs one named step of a math-proof siege: writes only the files that step names, keeps notes under 4000 words, and requires proof.md to stand alone.
Why we selected itEach launch is one step with no memory except the named files, and every query file must stand alone because the worker sees only that query plus the problem statement.
Weigh upThis is one seat in the siege protocol. A standalone copy cannot orchestrate the workflow, and ledger bookkeeping is not a mathematical proof checker.
Mathematics · Anthropic
Math-proof worker
Answers one self-contained mathematics question into a file, separates proved claims from sketches and open questions, and appends the task's end marker only when finished.
Why we selected itThe answer file has to separate what is proved, what is only sketched or conjectured, and what remains open, and a literature result must be labeled as cited from memory.
Weigh upRead, Write and Edit are the only declared tools; no code execution or web access is available to verify calculations or literature.
Meeting records · Agency Agents community
Meeting notes extractor
Extracts explicit decisions, assigned actions and unresolved questions from meeting material without filling missing facts.
Why we selected itIt distinguishes agreement from discussion and refuses to infer an owner from who usually handles a topic.
Weigh upThe original frontmatter name contains spaces; adapt it to a valid Claude Code name in a separate working copy.
Documentation · Seth Hobson
Mermaid diagram author
A diagram-authoring profile that asks for readable Mermaid source and a rendered check.
Why we selected itIt chooses the diagram type from the data and explicitly asks for a rendering test before delivery.
Weigh upNo renderer is bundled; a model can emit invalid syntax, and a requested preview is not evidence that rendering happened.
Infrastructure · VoltAgent
Microsoft 365 workload automation planner
A Microsoft 365 profile for scoped mailbox, collaboration and license-lifecycle automation.
Why we selected itIt distinguishes Graph from Exchange Online connection requirements before changing workload objects.
Weigh upNo tenant connector, admin consent or module is bundled; Bash and edit permissions can make real account, access or license changes when credentials exist.
Documentation contribution · GitHub community
Microsoft Learn contribution guide
Guides a documentation contribution through article structure, Microsoft style, metadata and review feedback.
Why we selected itIt distinguishes conceptual articles, tutorials, quickstarts and reference material before reviewing structure.
Weigh upThe legacy MCP/tool aliases need mapping to the installed client; the definition does not configure the Learn connection.
Game server development · Seth Hobson
Minecraft server plugin engineer
Develops version-aware server plugin plans covering events, commands, configuration and measured tick performance.
Why we selected itThe source requires server/build inspection and profiling before optimization, and distinguishes Paper APIs from Bukkit/Spigot fallbacks.
Weigh upOfficial Paper guidance warns that world-state APIs are unsafe from asynchronous tasks; Folia requires its own schedulers.
Model development · Seth Hobson
ML serving and lifecycle engineer
A production-ML planning and implementation profile covering serving, feature validation and model lifecycle checks.
Why we selected itIt includes holdout and temporal validation alongside online comparisons, rather than treating offline accuracy as the only deployment criterion.
Weigh upListed serving frameworks are not bundled or verified current; production-ready wording is a target, and example throughput is not a measured result.
Model evaluation · Affaan Mustafa
MLE Reviewer
Reviews an ML change for leakage, reproducibility and serving safety, and starts by checking the diff is reviewable.
Why we selected itThe decision line is APPROVE, APPROVE WITH WARNINGS, or BLOCK, and BLOCK includes leakage or a missing rollback.
Weigh upBash can execute repository code; use existing approved tools and inspect test side effects. The prompt forbids installing new packages without approval.
Machine learning operations · Seth Hobson
MLOps pipeline engineer
Plans reproducible training and serving infrastructure with artifact lineage, approval gates and operational monitoring.
Why we selected itIt connects dataset/model versioning and lineage to registry promotion and rollback rather than treating deployment as a one-step upload.
Weigh upThis broad instruction profile includes no runnable pipeline, cloud account or vetted tool integration.
Operations and reliability · Agency Agents
Mobile Release Engineer
Engineers the signing-to-store path for iOS and Android with phased rollouts and a forward-fix-only recovery model, since shipped binaries cannot be rolled back.
Why we selected itStates the no-rollback constraint explicitly and requires phased rollouts with pre-defined halt-on-crash-spike thresholds plus a forward-fix path before shipping.
Weigh upFastlane, match or secrets-manager storage, store accounts and CI runners must be configured in the host project; the file supplies the workflow, not credentials.
Data and retrieval · github
MongoDB Performance Advisor
Combines repository query patterns with read-only MongoDB diagnostics to explain query plans and index trade-offs.
Why we selected itStops when a read-only MongoDB MCP connection is unavailable instead of fabricating database observations.
Weigh upRequires the MongoDB MCP server configured read-only; Atlas Performance Advisor also needs the appropriate account access.
Architecture and decisions · Seth Hobson
Monorepo architect
Chooses a monorepo tool from the repository's constraints instead of naming one by fashion.
Why we selected itThe workflow says to assess codebase size and team structure before selecting Nx, Turborepo, Bazel, or Lerna.
Weigh upThe file does not install Nx, Turborepo, Bazel, or Lerna.
Mobile applications · Seth Hobson
Native iOS application engineer
Plans native SwiftUI/UIKit work with lifecycle, persistence, accessibility and device validation concerns.
Why we selected itIt includes SwiftUI/UIKit interoperability, background work and storage choices rather than treating the UI in isolation.
Weigh upThe source targets an iOS 18/Xcode 16-era baseline; confirm SDK availability and platform policies for the actual project.
Modernization and specifications · github
Neon Migration Specialist
Tests schema changes on a temporary Neon database branch and prepares migration files for a separate production rollout.
Why we selected itExplicitly forbids running migrations on the main Neon database branch and separates database branches from Git branches.
Weigh upWorkflow adjustment is required: the description promises automated production application and zero downtime, while the body prohibits main-branch migrations. Use only the test-branch workflow, with no downtime guarantee.
Data and retrieval · GitHub partners
Neon Performance Analyzer
Finds slow Postgres queries on a temporary Neon database branch, tests fixes on a second branch, and reports before-and-after metrics in a pull request.
Why we selected itAnalysis and tests run on Neon database branches with a four-hour TTL, never the main database branch, and both branches are deleted at the end.
Weigh upIf pg_stat_statements is missing, it enables the extension on the analysis branch and tells the user.
Infrastructure and delivery · affaan-m
Network Config Reviewer
Audits router and switch configuration — Cisco IOS-style interfaces, ACLs, VTY, AAA, SNMP and routing blocks — returning severity-ranked findings with quoted evidence and a PASS/WARNING/BLOCK verdict.
Why we selected itAnchors every finding to a quoted config snippet or block, and separates hard blockers from best-practice improvements before issuing a PASS/WARNING/BLOCK verdict.
Weigh upScoped to Cisco IOS and IOS-XE style configuration; other vendors’ syntaxes are out of scope.
Operations and reliability · GitHub community
New Relic Incident Response Agent
Assesses a New Relic alert, investigates traces, errors and deployments, and proposes a code fix only after the engineer agrees to each phase.
Why we selected itIf the New Relic MCP server is unavailable or misconfigured, it stops and names the missing server instead of guessing.
Weigh upThe example incident report contains illustrative figures, such as a 15% error rate and an 8.5s trace. They show the format and are not results.
Operations and reliability · wshobson/agents
Observability Engineer
A planning prompt for telemetry coverage, service objectives, alert design and incident runbooks.
Why we selected itConnects metrics, logs and traces with service-level indicators and actionable incident response.
Weigh upThis is a broad instruction profile, not an installed collector, dashboard, integration or compliance control.
Changelog editing · GitHub partners
Octopus Deploy release notes
Reads the last release deployed to an Octopus project and environment and writes Markdown release notes from its commits.
Why we selected itThe Octopus MCP tool list is read-only: every allowed tool is a get_ or list_ call, such as list_releases_for_project and get_release_by_id.
Weigh upIt fetches commit diffs from GitHub, but no GitHub tool is declared. Whether that step works depends on the host's tools.
Planning and delivery · GitHub
One-shot feature issue planner
Writes one issue-ready plan and states operating rules before the workflow.
Why we selected itThe operating rules forbid follow-up questions and forbid code changes.
Weigh upThe no-questions rule can hide a missing fact; the file requires that fact to be labeled as an assumption.
Security and privacy · affaan-m
Open-Source Sanitizer
Independently audits a project before open-source release — 20+ secret-pattern scans, PII and internal-reference detection, dangerous-file checks, .env.example completeness, a git-history audit and a list of files that do not read as text — issuing PASS, PASS WITH WARNINGS or FAIL where any critical finding fails the release.
Why we selected itIt is an adversarial second stage by design: it never trusts the forker’s work, re-scans every text file itself, and a single critical finding in any category is an overall FAIL.
Weigh upRegex scanning produces false positives by design — the definition accepts false positives over false negatives, and high-entropy heuristics are flagged for manual review rather than auto-failing.
Modernization and specifications · GitHub
Oracle-to-PostgreSQL migration expert
Phases an Oracle-to-PostgreSQL migration and keeps the source as the behaviour to match.
Why we selected itThe file forbids applying database changes for the user and says to hand over scripts instead.
Weigh upTools include edit and execute, so the no-direct-database-change rule is an instruction.
Operations and reliability · GitHub partners
PagerDuty Incident Responder
Pulls a PagerDuty incident, correlates it with commits from the 24 hours before it started, and proposes a fix or rollback pull request.
Why we selected itIt forms a triage hypothesis by root-cause category (code change, configuration, dependency, infrastructure) and compares the incident start with deployment times before searching code.
Weigh upDeclared tools include github/create_pull_request, create_issue, create_branch and create_or_update_file, which write to GitHub. Remove them unless a pull request is intended.
Marketing analysis · Agency Agents community
Paid search query analyst
A query-audit profile that requires actual account reports before suggesting negative keywords or expansion opportunities.
Why we selected itIt considers account, campaign and ad-group negative scopes and cross-campaign overlap rather than producing one undifferentiated exclusion list.
Weigh upThe source also instructs pushing negatives through an available integration; that can change paid campaigns. Zero conversions alone does not prove wasted spend, and percentage reduction targets are unmeasured.
Application integration · Agency Agents community
Payments and billing engineer
Models payment mutations, subscriptions and fulfillment as retryable state machines with durable webhook intake and reconciliation.
Why we selected itThe example separates verified event acceptance from leased processing and only marks completion after side effects succeed.
Weigh upThe inbox adapter, scheduler, domain handlers and ledger schema are application dependencies, not implemented by the prompt. Example API versions, PCI scope claims and business thresholds need current provider/qualified review; no payments were run.
Browser and performance · wshobson
Performance Engineer
Runs performance work measurement-first — establish a baseline, profile, then optimize — across browser Core Web Vitals, backend hot paths and load tests, with budgets and regression gates to hold the gains.
Why we selected itThe response protocol is measurement-first by construction: establish a performance baseline with comprehensive measurement and profiling before any optimization, then validate improvements and set budgets to prevent regression.
Weigh upA capability-list persona: it prescribes the measurement discipline in prose but carries no enforced evidence-citation contract, unlike a log-grounded monitor.
Operations and reliability · voltagent
Performance Monitor
Reads existing metric, log and output files to surface performance anomalies, then writes an evidence-cited observability plan — every reported number must have been read from a file or counted, never invented.
Why we selected itBinds every reported figure to file evidence — a number may only be stated if it was read from a file or counted with Grep/Glob, and each finding carries path:line citations plus a high/medium/low confidence grade.
Weigh upAnalyses only files that already exist; it cannot collect live metrics, query time-series databases or build dashboards.
Testing and review · affaan-m
PHP Reviewer
Reviews PHP changes by running PHPStan, Psalm and Pint where available, then ranks findings from SQL injection and mass assignment to Eloquent N+1, with Laravel, Livewire and Filament framework checks.
Why we selected itIs the one reviewer in the family whose approval criteria bind to the automated checks — Block on any failed PHPStan, Psalm, PHPUnit or Pint run OR any CRITICAL/HIGH finding — so code the tools reject can never be approved.
Weigh upClosing pointers to sibling laravel-patterns, laravel-security and laravel-tdd skills reference companion files that are not part of this standalone definition.
Modernization and specifications · anthropics
Pilot-Driven Uplift Migrator
Applies an approved pilot playbook to one migration unit and returns build evidence, playbook gaps and shared-file needs to the coordinating owner.
Why we selected itA missing pilot playbook is an explicit no-edit stop condition, while every gap must be reported even when the worker resolved it.
Weigh upRequires analysis/<system>/PLAYBOOK.md and DELTA_CATALOG.md plus an assigned unit. The agent and workflow prompt instruct checking the playbook; the JavaScript itself does not independently read or attest its existence or approval.
Operations and reliability · GitHub
Platform SRE for Kubernetes
Asks for the cluster facts before a deploy and treats security defaults as non-negotiable.
Why we selected itEvery change must include a plan, validation, rollout, and a rollback, and images must not use the latest tag.
Weigh upTools include edit/editFiles and terminalCommand, so the file can change manifests and run kubectl.
Testing and review · github
Playwright Test Explorer
Explores a working interface through Playwright before generating and running TypeScript tests for its observed user flows.
Why we selected itProhibits writing test code until the agent has navigated the interface and identified its actual flows.
Weigh upRequires a configured Playwright MCP connection and a reachable application; the file does not install a browser server or the test framework.
Agent development · Anthropic
Plugin agent creator
Turns a description of a needed agent into an agents/<identifier>.md file with frontmatter, triggering examples and a system prompt.
Why we selected itIts own tools are Write and Read only, and its quality standards ask for least-privilege tool selection in the agent it writes.
Weigh upThe file also offers to omit tools for full access. Unless you set the list, an agent it writes can inherit every tool.
Agent development · Anthropic
Plugin skill reviewer
Reviews a skill file for structure, description triggers and progressive disclosure.
Why we selected itFrontmatter tools are only Read, Grep and Glob, so the declared role does not include Write, Edit or Bash.
Weigh upThe file is a Claude Code agent definition. It does not install a skill or certify one. A Markdown profile does not register a native subagent by itself.
Security and governance · wshobson
Policy Enforcer
Authors Cedar rules for what an agent may do, and audits an existing policy instead of assuming it is enforced.
Why we selected itThe help steps start by asking for the risk profile and then prefer allow-listing over deny-listing.
Weigh upFrontmatter sets model to opus and does not declare tools.
Evaluation · Anthropic
Post-hoc skill analyzer
Skill-creator agent that reads a finished blind comparison, then opens both skills and transcripts to say why the winner won.
Why we selected itThe catalogue has the skill-creator grader and not this post-hoc analyzer.
Weigh upThis review did not execute the agent.
Database administration · GitHub community
PostgreSQL extension operator
A small Copilot profile that routes database work through the PostgreSQL extension’s tools.
Why we selected itThe first gate checks that ms-ossdata.vscode-pgsql is installed and enabled before using its tools.
Weigh upThis is a thin routing profile without backup verification, destructive-action gates or an enforced read-only mode; claimed DBA tasks are not validated implementations.
Data modeling · GitHub community
Power BI data model planner
A Power BI profile for model grain, relationships and storage-mode decisions.
Why we selected itIt requires consistent fact-table grain and explicitly separates fact and dimension characteristics.
Weigh upThe profile requires Microsoft documentation MCP and legacy tool aliases; it does not configure that server or prove a PBIX model correct.
Data analysis · GitHub community
Power BI DAX reviewer
A DAX-focused profile for formula semantics, readability and calculation debugging.
Why we selected itIt distinguishes fully qualified column references from unqualified measure references.
Weigh upThe prompt’s broad optimization rules are suggestions, not measurements; documentation-tool availability and DAX execution remain external requirements.
Performance · GitHub community
Power BI performance investigator
A Power BI profile separating query, model, report and capacity performance investigations.
Why we selected itThe profile separates DAX/query work from visual rendering, model size and capacity pressure.
Weigh upExample queries and target thresholds were not run; performance improvement depends on representative data, capacity and repeated measurements.
Data visualization · GitHub community
Power BI report designer
A Power BI visualization profile for report hierarchy, interactions and accessible layouts.
Why we selected itIt distinguishes executive, analytical and operational report layouts rather than prescribing a single dashboard for every audience.
Weigh upEmbedded-report and Business Central snippets are separate integration examples, not a ready-to-deploy report; custom visuals and tenant permissions need review.
Programming · VoltAgent
PowerShell 7 cross-platform automation engineer
A scripting profile emphasizing portable paths, structured CI output and explicit tenant context.
Why we selected itIt checks subscription/tenant context and module compatibility before cloud automation.
Weigh upThe referenced .NET 6/7 knowledge is not a current runtime requirement; module portability and WhatIf/Confirm behavior must be verified per command.
Security operations · VoltAgent
PowerShell automation hardening reviewer
A PowerShell security profile for remoting boundaries, credential handling and sanitized operational logs.
Why we selected itIt considers scheduled tasks, WinRM endpoints and automation service accounts as part of the script’s security context.
Weigh upExecution Policy is not an authorization boundary; logging can itself capture sensitive values, and this profile grants shell and edit tools. Named compliance baselines are not bundled.
Code review · Anthropic
PR toolkit code reviewer
Pins the pr-review-toolkit code reviewer, which defaults to unstaged git diff and tells itself to drop issues below a confidence cutoff.
Why we selected itThe catalogue has feature-dev's code-reviewer at a different path and not this pr-review-toolkit agent.
Weigh upThis review did not execute the resource.
Security and privacy · agency
Privacy Engineer
Implements privacy as code — PII discovery across every store including logs and caches, consent enforced at the write path, fan-out deletion pipelines with verification scans, and re-identification testing before anything is called anonymized.
Why we selected itEvery rule is an engineering control with a named failure mode: consent must gate the write rather than just set a flag, deletion must fan out to replicas, indexes, caches, third parties and backups with an audit record, and retention must expire by automated job.
Weigh upEngineering guidance, not legal advice; the legal-basis and retention decisions it demands still need a qualified reviewer per jurisdiction.
Operations and reliability · wshobson
Prod Logs Health Check
Pulls recent production logs after a deploy or load test and reports errors grouped by root cause, counting distinct failures separately from retries — treating raw logs, not dashboards, as the only acceptable primary source.
Why we selected itMakes raw logs the mandatory primary source and refuses to present inference as fact — if logs were not checked, the finding must say so before anything is presented.
Weigh upThe definition is a template: the {{LOG_QUERY}} and {{PROJECT_SPECIFIC_MARKERS}} placeholders must be pointed at the project’s real log source before use.
Operations · Agency Agents community
Production incident coordination planner
A response-coordination profile with role, timeline, mitigation and follow-up templates.
Why we selected itIt separates commander, technical lead, communications and scribe responsibilities so troubleshooting and external updates have clear owners.
Weigh upThe sample severity thresholds, timing and error-budget policy are organization-specific templates; the profile does not grant emergency production or messaging authority.
Security workflow components · Seth Hobson
Protect MCP receipt verifier
A receipt-review component that separates valid signatures, invalid signatures and undecidable verification.
Why we selected itThe profile uses the external public key rather than trusting a key embedded in an untrusted receipt.
Weigh upA signature failure can mean a wrong key or malformed record, so the source’s tampering attribution must be softened. Signed tool names do not prove policy enforcement or actual tool effects.
Testing and review · anthropics
Pull Request Test Analyzer
Maps a pull request to its tests and identifies meaningful missing failure cases, brittle assertions and uncovered behavioral contracts.
Why we selected itPrioritizes tests by concrete regressions they would prevent rather than chasing line coverage.
Weigh upIts internal numerical criticality rubric is a prioritization instruction, not a measured quality score or a catalogue rating.
Data and analytics · GitHub community
PySpark Expert Agent
Reviews PySpark pipeline code for distributed-execution anti-patterns with confidence-graded hypotheses, Spark-native fixes, and a ban on inventing runtime evidence.
Why we selected itEvidence discipline is explicit: it must not invent Spark UI metrics or runtime evidence, and must ask for the minimum evidence such as explain output, stage symptoms, sizing and AQE state needed to confirm.
Weigh upStatic heuristics without cluster evidence stay hypotheses; confirm Medium and Low confidence findings against explain plans and the Spark UI before acting.
Data and analytics · GitHub community
Python Notebook Sample Builder
Builds a new Python demonstration notebook from snippets checked in the selected development environment.
Why we selected itThe file says not to put code in a notebook until it has been run in the terminal.
Weigh upThe prompt assumes a development container and broad package-install permissions; do not apply that assumption to an arbitrary host.
Testing and review · affaan-m
Python Reviewer
Reviews Python changes by running ruff, mypy, black and bandit where available before reading the diff, then ranks findings from f-string SQL injection and bare except to type-hint gaps and un-Pythonic patterns.
Why we selected itStructures the review around executed diagnostics — ruff, mypy, pylint and black --check run before the diff is read, and the documented command set adds bandit and pytest with coverage, so findings rest on tool output rather than opinion.
Weigh upThe closing pointer to a sibling python-patterns skill references a companion file that is not part of this standalone definition.
Build and runtime diagnostics · Affaan Mustafa
PyTorch build error resolver
Traces tensor shapes, devices and gradient paths to diagnose a bounded training or inference crash.
Why we selected itThe workflow starts at the failing line, traces shapes through layers, then reruns the failing script and checks autograd.
Weigh upSeveral fix-table suggestions can change semantics: clamping embedding indices, adding retain_graph or changing layer widths must be justified by the model/data contract.
Testing and review · github
QA Test Planner and Investigator
Builds a requirement-linked test plan, explores failure paths and reports reproducible bugs separately from potential improvements.
Why we selected itOrganizes tests by boundary, negative, failure, concurrency and security cases, then prioritizes by risk.
Weigh upIt grants broad edit, execute, delegation and web tool categories; the role’s focus is not a technical read-only restriction.
Data and retrieval · affaan-m
Rag Pipeline Reviewer
Reviews retrieval pipelines for grounding, pruning, fallback behavior and evaluation policy, returning a structured approval, conditional approval or block decision.
Why we selected itRole boundary is negative as well as positive: do not rewrite the LLM answer prompt; hand off to named specialists.
Weigh upThe read-only Bash rule is an instruction; the host must enforce the intended access boundary.
Testing and review · affaan-m
React Reviewer
Reviews React and JSX changes for hook correctness, dangerouslySetInnerHTML and URL-scheme security, server/client component boundaries, key stability and render performance — reporting findings without rewriting code.
Why we selected itOwns an explicit scope table that splits React-specific lanes (hooks, dangerouslySetInnerHTML, server/client boundaries, NEXT_PUBLIC leaks) from the generic TypeScript reviewer, so the two roles compose instead of duplicating.
Weigh upFull .tsx coverage needs the sibling typescript-reviewer for generic type-safety lanes; this file reviews React-specific concerns only.
Documentation · VoltAgent
Readme Generator
Builds a repository README from inspected manifests, scripts, tests and entry points, explicitly rejecting guessed commands, APIs and configuration.
Why we selected itRequires setup commands, configuration and usage examples to be extracted from actual repository evidence.
Weigh upThe opening context-manager request assumes a coordinator not supplied by the standalone file. Provide the repository purpose and audience directly when using it alone.
Realtime systems · Agency Agents community
Realtime collaboration protocol designer
Plans collaborative-state protocols around replay, operation identity, convergence tests and bounded fan-out.
Why we selected itIt separates ephemeral presence TTLs from durable document operations and explicitly rejects unbounded slow-consumer queues.
Weigh upThe snippets omit application-owned durable queue, operation application and authentication helpers; there is no complete sync engine here.
Model evaluation · GitHub
Research Harness Engineer
Treats a reported score as guilty until a null model and a published baseline have been checked.
Why we selected itThe file requires a null model and a reproduced baseline before any improvement claim.
Weigh upNo tool allowlist or executable harness is included; datasets, metrics and compute must be supplied.
Research and product · msitarzewski
Research Synthesist
Builds an auditable synthesis that traces repeated claims to their origin and separates independent agreement, disagreement and missing evidence.
Why we selected itExplicitly rejects counting repeated citations to one study as independent corroboration.
Weigh upThe frontmatter name is the display name “Research Synthesist” and its color is the hex value #9333EA, which is not one of the eight documented Claude Code colors; adapt both in a working copy.
Security workflow components · Seth Hobson
Review governance policy author
A policy-authoring component for identifying which review, merge and CI-edit commands need an approval gate.
Why we selected itIt distinguishes raw-command substring matching from reliable identity or semantic authorization: no human_approved attribute is available.
Weigh upThe approval flag bypasses policy until removed; a writable flag is not independently verified human approval. String matching can be bypassed, and the plugin deliberately permits some other public writes.
Code review · Anthropic
Rust API Ergonomics Reviewer
Reviews a Rust library from the downstream user’s side, and confirms a suspected lint with cargo before reporting it.
Why we selected itThe role is limited to what a downstream Rust user types, sees in compiler output, and can misuse. Correctness bugs, performance and unsafe soundness are left to other review.
Weigh upBash is enabled. The git/cargo allowlist and prohibition on repository edits live in prompt text; they are not an operating-system sandbox.
Build and runtime diagnostics · Affaan Mustafa
Rust build error resolver
Resolves Rust compilation, ownership, trait and Cargo-feature errors with a minimal-change verification loop.
Why we selected itIt inspects duplicate dependencies and enabled features, preferring a targeted package update to a full lockfile refresh.
Weigh upCloning, RefCell, added lifetime bounds and edition changes are contextual proposals, not interchangeable compiler fixes. Preserve API semantics and MSRV.
Security review · Trail of Bits
Rust review dedup judge
Pins the rust-review dedup judge that merges only obvious duplicates, aborts when the finding list cannot be loaded, and does not declare a shell.
Why we selected itThe agents index has the c-review worker and not plugins/rust-review/agents/rust-review-dedup-judge.md.
Weigh upThis review did not execute the resource.
Security review · Trail of Bits
Rust review FP judge
Pins the rust-review judge that runs after dedup, aborts when it cannot obtain a finding list, and must not invent the SARIF file.
Why we selected itThe agents catalogue has no sourcePath for plugins/rust-review/agents/rust-review-fp-judge.md.
Weigh upThis review did not execute the resource.
Security review · Trail of Bits
Rust review worker
Pins the rust-review worker that aborts a malformed or primer spawn, writes findings only inside its scope root, and does not filter them by severity.
Why we selected itThe agents catalogue has no sourcePath for plugins/rust-review/agents/rust-review-worker.md.
Weigh upThis review did not execute the resource.
Testing and review · affaan-m
Rust Reviewer
Reviews Rust changes by running cargo check, clippy, fmt and test before reading the diff, then ranks findings from unchecked unwrap and unjustified unsafe to ownership, concurrency and allocation issues.
Why we selected itGates the review on executed commands — cargo check, clippy with -D warnings, fmt --check and cargo test must run before review begins, so findings rest on tool output rather than opinion.
Weigh upThe closing pointer to a sibling rust-patterns skill references a companion file that is not part of this standalone definition.
Security review · Trail of Bits
Rust source analyzer
Pins the zeroize-audit Rust source analyzer that keeps a weak reference match, merges shared JSON without replacing C or C++ entries, and continues after rustdoc failure.
Why we selected itThe agents index has no entry for plugins/zeroize-audit/agents/2b-rust-source-analyzer.md.
Weigh upThis review did not execute the resource.
Rust maintenance · Agency Agents community
Rust workspace refactoring specialist
Plans and completes coherent Rust refactors while tracing public contracts, feature gates and observable behaviour.
Why we selected itIt treats drop timing, lock scope, cancellation, serialization and error behaviour as contracts, not just function signatures.
Weigh upThe profile requires explicit authorization for destructive Git operations, publishing/deployment and other named sensitive changes; these are prompt instructions, so reconcile them with task authorization and effective host permissions before execution.
Enterprise applications · GitHub community
Salesforce Apex and trigger engineer
Plans Apex changes from existing org patterns, with sharing/data-access review and positive, negative and bulk tests.
Why we selected itIt inspects sfdx-project.json, package.xml and existing handlers/factories before choosing a service, selector or asynchronous pattern.
Weigh upThe pinned model alias and legacy tool names need mapping to supported Copilot capabilities.
Modernization and specifications · anthropics
Scaffolder
Creates one modernization service from an approved architecture and specification, including its domain model, API stubs and rule-linked acceptance tests.
Why we selected itLimits the intended work to one caller-selected service directory so independently scaffolded services do not overwrite one another.
Weigh upRequires REIMAGINED_ARCHITECTURE.md, AI_NATIVE_SPEC.md and the caller’s modernized service path; the complete modernization workflow is not included.
Language development · Seth Hobson
Scala effects and distributed-systems developer
A Scala profile for effect boundaries, typed domain models and distributed streams.
Why we selected itThe profile separates Cats Effect or ZIO effects from pure domain models and calls out explicit error channels.
Weigh upIt lists many alternative stacks without selecting dependency versions; scalability language is aspirational and not benchmark evidence.
Security review · Anthropic
Scan inventory
Pins the Claude Security cartographer that partitions one tree and must place every listed top-level directory on a scanned or skipped ledger.
Why we selected itThe agents index has no entry for plugins/claude-security/agents/scan-inventory.md.
Weigh upThis review did not execute the resource.
4 sources · 2026-10-14
Read the full review Install command in reviewProprietary — Anthropic internal-use termsSecurity review · Anthropic
Scan loader
Pins the Claude Security loader whose only input is one directory plus one file name, and that returns that JSON unchanged.
Why we selected itThe agents index has no entry for plugins/claude-security/agents/scan-loader.md.
Weigh upThis review did not execute the resource.
4 sources · 2026-10-14
Read the full review Install command in reviewProprietary — Anthropic internal-use termsSecurity review · Anthropic
Scan redactor
Pins the Claude Security redactor that edits nothing unless the dispatch names three specific report files, and that tells itself not to repeat a secret in the reply.
Why we selected itThe agents index has no entry for plugins/claude-security/agents/scan-redactor.md.
Weigh upThis review did not execute the resource.
4 sources · 2026-10-14
Read the full review Install command in reviewProprietary — Anthropic internal-use termsSecurity review · Anthropic
Scan verifier
Pins the Claude Security voter that takes one candidate finding and tries to disprove it, with a false-positive default and a file-and-line citation rule.
Why we selected itThe agents index has no repository and sourcePath entry for plugins/claude-security/agents/scan-verifier.md.
Weigh upThis review did not execute the resource.
4 sources · 2026-10-14
Read the full review Install command in reviewProprietary — Anthropic internal-use termsSecurity and governance · github
Se: Responsible Ai
Screens user-facing changes for accessibility, privacy and exclusion concerns, then documents decisions and escalation needs in numbered responsible-AI records.
Why we selected itStep 2–4 give concrete tests: culturally varied names/ages/edge strings, keyboard vs clickable div, alt text, minimal vs excessive data collection, unbundled consent.
Weigh upThe illustrative name and demographic examples are screening prompts, not validated fairness tests or a compliance assessment.
Documentation · github
Se: Tech Writer
Turns code and design context into audience-specific documentation, tutorials, articles and decision records using explicit structures and a verification checklist.
Why we selected itContent-type templates are fully specified, including ADR status/deciders/consequences/alternatives and task-oriented user-guide workflows with verification checkpoints.
Weigh upThe frontmatter permits file editing and web retrieval but does not name an execution tool; testing or compiling examples needs a separate runner.
Accessibility · Agency Agents
Section 508 Accessibility Specialist
Audits and remediates web accessibility against the correct legal baseline (508 equals WCAG 2.0 AA) with assistive-technology testing, and reports honest VPAT conformance.
Why we selected itWCAG criterion references were screened against current W3C text with no misidentification found, including 1.3.1 Info and Relationships (A) and the 4.5:1 and 3:1 contrast thresholds, unlike the excluded a11y-architect's 2.4.11 error.
Weigh upThe 30 to 40 percent automated-tool coverage figures are the author's estimates, not independently measured here; treat them as cautionary guidance.
Security and governance · anthropics
Security Auditor
Reviews application code with a stack-specific security checklist, requiring source locations, exploit scenarios, concrete remediation and masked secrets in each finding.
Why we selected itDemands a plausible exploit scenario and a concrete code-level fix rather than unsupported vulnerability labels.
Weigh upBash and package auditors depend on host permissions and tool availability; the definition does not enforce its read-only boundary.
Browser and performance · ecc
SEO Specialist
Audits technical SEO from the repository's actual files — crawl blockers, canonical loops, redirect chains, malformed JSON-LD, heading hierarchy — and reports each finding as a severity-ranked issue with file location and exact fix.
Why we selected itThe audit is file-grounded and severity-ranked: critical crawl and index blockers, canonical loops and redirect chains outrank cosmetic issues, and every finding must carry a path:line or URL location plus an exact fix.
Weigh upReads source files; it does not crawl the deployed site or fetch live index state, so runtime-only issues (server headers, production redirect behaviour) need separate verification.
Infrastructure and delivery · Seth Hobson
Service mesh expert
Treats a mesh change as traffic, security and observability, not as a sidecar install.
Why we selected itThe workflow orders topology, mTLS and authorization, observability, traffic rules, failover tests, and runbooks.
Weigh upFrontmatter sets model to opus and does not declare tools or install Istio or Linkerd.
Security and governance · Trail of Bits
Sharp Edges Analyzer
Evaluates APIs, configurations and library interfaces for misuse resistance — probing zero, empty and negative edge cases and modelling three adversary types before reporting design-level security findings, each with a minimal misuse example.
Why we selected itTurns threat modelling into a repeatable four-phase procedure — surface identification, systematic edge-case probing (zero/empty/null, negative values, type confusion, defaults, error paths), threat modelling against three named adversary models, and validation that requires reproducing the misuse and confirming exploitability before reporting.
Weigh upThe language-specific footgun guides it reads on demand live in sibling reference files inside the same plugin (plugins/sharp-edges/skills/sharp-edges/references/), not in the agent definition itself.
3 sources · 2026-10-07
Read the full review Attributed definition availableCC-BY-SA-4.0Code review · Affaan Mustafa
Silent failure hunter
Looks for swallowed errors and bad fallbacks, and requires an output format.
Why we selected itThe hunt list names empty catches, log-and-forget handling, and fallbacks that hide a failure.
Weigh upTools include Bash, while the output is a report rather than a required code edit.
Evaluation · Anthropic
Skill eval grader
Grades each expectation pass or fail from cited evidence, treats uncertainty as a failure, and writes grading.json.
Why we selected itThe inputs are expectations, a transcript path, and an outputs directory, and each expectation is pass or fail.
Weigh upThis plain Markdown role has no native frontmatter or tool allowlist; provide it through an orchestrator with appropriate output inspection tools.
Testing and review · Trail of Bits
Skill reviewer fixture
Reviews one Claude Code skill from a loop dispatch, writes nothing, and leaves a rejected ledger finding closed unless the evidence is new.
Why we selected itThe description says it reviews one skill per dispatch and is not for general code review.
Weigh upA Markdown profile does not register a native subagent by itself.
Language and communication · Agency Agents community
Spanish–English context translator
A bilingual drafting profile that carries meaning, formality and regional context into Spanish–English translations.
Why we selected itIt preserves names and brands unless an established localized equivalent or user request applies.
Weigh upText-generated pronunciation is approximate, regional generalizations need review, and this profile does not provide certified translation or a qualified clinical interpreter.
Data and analytics · Agency Agents
Spatial Data Engineer
A geospatial ETL planning prompt centered on source preservation, explicit coordinate systems and transformation checks.
Why we selected itRequires original files to be preserved and transformations to write to a new location.
Weigh upThe frontmatter name “Spatial Data Engineer” contains spaces. Current Claude Code documentation describes name as a unique identifier, with examples such as code-reviewer, and forbids only “:”. Give the working copy an identifier such as spatial-data-engineer; this download remains byte-identical to upstream.
Spatial analysis · Agency Agents community
Spatial data scientist
A spatial-analysis profile that connects geographic patterns to diagnostics, uncertainty and reproducible research.
Why we selected itIt explicitly tests spatial dependence in residuals and sensitivity to aggregation boundaries rather than treating a visually persuasive map as evidence.
Weigh upThe blanket wording about non-spatial models is overbroad: appropriateness depends on assumptions and diagnostics. No model is fitted by the profile itself and named Python/R/GIS packages are not bundled.
Modernization and specifications · affaan-m
Spec Miner
Extracts a selected brownfield capability into source-linked OpenSpec requirements and invariants while limiting writes to the specification directory.
Why we selected itWrite may only create openspec/specs/<capability>/spec.md; Bash must stay read-only—rare path-scoped reviewer-writer.
Weigh upIts output is tailored to OpenSpec; a different specification workflow requires adaptation.
Documentation and knowledge · GitHub community
Specification Writer
Develops implementation specifications with explicit requirements, interfaces, constraints and acceptance examples.
Why we selected itThe specification must be self-contained Markdown and must follow the template sections in the file.
Weigh upFile-edit, test and web tools are declared; constrain them to the agreed documentation task.
Testing and review · Trail of Bits (Omar Inuwa)
Specification-to-Code Checker
Checks one quoted requirement against its implementation paths and records the enforcement evidence, gaps and unresolved questions.
Why we selected itSeparates implemented, partially enforced, contradicted, absent, stronger-than-specified and undecidable requirements instead of forcing a binary pass/fail.
Weigh upThe file requires the plugin’s sibling analysis format and domain references; {baseDir} must resolve before use.
6 sources · 2026-10-13
Read the full review Attributed definition availableCC-BY-SA-4.0Media engineering · Agency Agents community
Speech transcript pipeline designer
A speech-pipeline reference covering audio validation, chunked ASR, speaker assignment and structured output.
Why we selected itIt preserves timestamps and low-confidence review flags rather than silently deleting uncertain speech.
Weigh upThe sample requests word timestamps but discards them in its segment object, and whole-segment overlap trimming can lose boundary content. Local-only medical claims and improvement percentages are unsupported generalizations.
Data and analytics · wshobson/agents
SQL Pro
A SQL review prompt covering query plans, schema design, indexing and read/write performance tradeoffs.
Why we selected itAsks for usage patterns and realistic data before treating an index or rewrite as an optimization.
Weigh upSQL features and tuning advice are not portable across all named engines. Supply the exact engine, version, schema and workload.
Security and governance · GitHub partners
StackHawk Security Onboarding
Decides whether a repository is a deployed web app or API, and if so opens a pull request with a stackhawk.yml and a GitHub Actions scan workflow.
Why we selected itIts first step assesses attack surface: it declines for libraries and documentation or configuration repositories, offers a review when stackhawk.yml already exists, and asks when unsure.
Weigh upThe workflow template pins actions/checkout@v3 and stackhawk/hawkscan-action@v2 by tag, not by commit SHA.
Data analysis · Seth Hobson
Statistical analysis and experiment scientist
An analysis profile that connects data profiling, statistical assumptions and validation to a concrete business question.
Why we selected itIt explicitly separates statistical from practical significance and calls for assumption checks and robustness validation.
Weigh upThe long methods list is aspirational expertise, not evidence that a method was correctly applied; observational data alone cannot establish causal effects.
Security and governance · 0xSteph
STIG analyst
Writes a STIG-shaped finding and a keep-open justification, and does not certify the system.
Why we selected itThe finding format includes impact, a risk-to-remediate score, what could break, remediation, and a scriptable verification.
Weigh upA formatted finding is not a certification, and the file does not claim one.
Research and product · msitarzewski
Study Design and Evidence Statistician
Examines the chain from research question and measurement to comparison, analysis and decision, making assumptions and uncertainty explicit.
Why we selected itSeparates descriptive, predictive and causal questions before choosing a study or interpreting a result.
Weigh upThe prompt does not supply data, a statistical runtime or a validated analysis package. Calculations require suitable tools and reproducible inputs.
Framework development · Svelte
Svelte File Editor
A focused Svelte editing prompt that consults current documentation and iterates with the Svelte MCP static checker.
Why we selected itSeparates Svelte component work into a focused agent context with relevant documentation.
Weigh upThis listing recommends the configured Svelte MCP path. The prompt’s fallback CLI and normalized tool-name variants were not runtime-tested; map names to the tools your host actually exposes.
Build and runtime diagnostics · Affaan Mustafa
Swift build error resolver
Diagnoses Swift/Xcode compilation, package resolution and concurrency errors, with explicit stops for signing or ownership redesign.
Why we selected itIt distinguishes SPM resolution, protocol/type failures, actor isolation and Sendable issues.
Weigh upXcode diagnostics require the appropriate Apple toolchain and available destinations; signing identities are local account information.
Testing and review · affaan-m
Swift Reviewer
Reviews Swift changes by running swift build, swiftlint and swift test first — stopping to report if any fail — then ranks findings from force unwrapping and actor-isolation races to ARC cycles and release-build assertion traps.
Why we selected itRuns the hardest gate in the family — swift build, swiftlint and swift test must all pass or the review stops to report — so the reviewer never approves code the toolchain rejects.
Weigh upClosing pointers to sibling swift rules and skills (swift-concurrency-6-2, swiftui-patterns, swift-protocol-di-test) reference companion files that are not part of this standalone definition.
Testing and review · github
Systematic Debugging
Reproduces an application bug, tests root-cause hypotheses and verifies a minimal fix against the original failure.
Why we selected itRequires reproduction and an expected-versus-actual report before changing code.
Weigh upThe procedure is a general debugging framework, so the caller must provide a concrete symptom and a usable reproduction environment.
Decision support · Trail of Bits
Tarot draw
Pins the let-fate-decide agent that makes one script call for a twelve-house spread and returns a short verdict, or a fixed line if that call fails.
Why we selected itThe agents catalogue has no sourcePath for plugins/let-fate-decide/agents/draw.md.
Weigh upThis review did not execute the resource.
Testing and reliability · GitHub community
TDD Green Phase
Implements the smallest behavior change needed to satisfy an existing failing test.
Why we selected itGreen-phase rules say to implement only the current issue and not to modify the failing test.
Weigh upGreen tests demonstrate only the asserted examples; hidden requirements and untested behavior still need review.
Testing and reliability · GitHub community
TDD Red Phase
Writes a failing test tied to a requirement before implementation work begins.
Why we selected itThe red phase says to write one failing test at a time and never to start edits without user confirmation.
Weigh upAn issue number inferred from a branch name must be verified against the intended repository and issue.
Testing and reliability · GitHub community
TDD Refactor Phase
Improves a tested implementation through small changes while preserving the existing behavior checks.
Why we selected itRefactoring is allowed only after tests are green, and only after the user confirms the plan.
Weigh upSecurity suggestions and passing tests are not a security audit or proof of behavioral equivalence.
Creative tooling · Agency Agents community
Technical art pipeline planner
A cross-engine technical-art profile linking asset handoff to measured rendering budgets.
Why we selected itIt puts asset budgets before content production and requires in-engine review under target lighting.
Weigh upThe sample shader and numeric budget tables are examples, not portable working assets or universal requirements; the fixed LOD-count rule needs project-specific judgment.
Sales enablement · Agency Agents community
Technical evaluation and POC designer
A presales planning profile that turns buyer requirements into measurable proof-of-concept criteria and a focused demo.
Why we selected itThe POC template agrees success criteria, scope and measurement methods before setup and includes an interim alignment review.
Weigh upExample benchmarks, timelines and customer outcomes are illustrative; deal-closing pressure must not justify omitting a failed criterion or inventing a reference.
Research and product · github
Technical Spike Research Mode
Investigates a supplied technical-spike document through cited documentation, repository examples and separately approved experiments, recording findings and dead ends as work proceeds.
Why we selected itStops when no spike document is provided and ties the work to explicit research questions and success criteria.
Weigh upThe recursive research policy has no fixed time or cost budget; define a stopping condition for the chosen spike.
Documentation · Agency Agents
Technical Writer
Writes developer documentation with a tested-snippet rule and docs-as-code pipeline guidance, so examples and versioned docs stay runnable instead of drifting.
Why we selected itRequires every code example to be tested before it ships and docs to match the software version they describe, deprecating rather than deleting old docs.
Weigh upThe 5-second, 30-second and 15-minute usability bars are author heuristics, not measured outcomes; treat them as editorial guidance.
Maintenance planning · GitHub community
Technical-debt remediation planner
Turns a specified maintenance problem into a concise Markdown remediation plan with prerequisites, ordered steps and verification methods.
Why we selected itThe body limits the role to a remediation plan and says the work is analysis only, with no code modifications.
Weigh upThe frontmatter still lists edit, command, test, browser, and web/fetch tools, so the analysis-only sentence is not a tool allowlist.
Framework development · wshobson
Temporal Python Pro
Implements durable workflows on the Temporal Python SDK — deterministic workflow code, activity heartbeats and retry policies, the three worker execution patterns, saga compensation and testing strategies for long-running processes.
Why we selected itThe three worker execution patterns (async activities, sync multithreaded, sync multiprocess) are matched to workload types, with the stated anti-pattern that blocking the async event loop turns async programs into serial execution.
Weigh upA builder persona, not a reviewer; beyond its testing-strategy section it has no verification gate on what it produces.
Infrastructure and delivery · GitHub partners
Terraform Agent
Resolves provider and module versions from the Terraform registry before writing code, then manages HCP Terraform workspaces and runs through the Terraform MCP server.
Why we selected itVersions are resolved with get_latest_provider_version or get_latest_module_version before code is written, and the private registry is searched before the public one when a token is present.
Weigh upThe MCP server runs hashicorp/terraform-mcp-server:latest in Docker; the tag is not pinned.
Infrastructure and delivery · github
Terraform Iac Reviewer
Reviews and drafts Terraform changes with a plan summary, impact assessment, validation sequence, approval before apply, and an explicit recovery discussion.
Why we selected itEvery change must include plan summary (add/change/destroy counts), risk assessment, validation commands, and a rollback strategy.
Weigh upTerminal and editing access can change infrastructure. Approval before apply is an instruction, not an enforced host permission boundary.
Infrastructure orchestration · VoltAgent
Terragrunt orchestration engineer
Organizes Terragrunt units, stacks, includes and dependencies with attention to state and multi-environment execution.
Why we selected itIt distinguishes explicit stack blueprints from directory-based stacks and separates dependency outputs from ordering.
Weigh upThe source gives version-sensitive CLI/HCL examples and an undefined DRY percentage; verify syntax and report actual changes rather than repeating example metrics.
Infrastructure and delivery · github
Terratest Module Testing
Designs Terraform module tests around outputs, validation failures and isolated setup and cleanup, with cloud applies explicitly gated.
Why we selected itDistinguishes success paths, expected failures and staged integration tests rather than treating every module test as a deployment.
Weigh upThe original tool list declares only legacy codebase and terminalCommand aliases. Map them to current read/search/execute tools and explicitly add an editor for writing tests; verify its pinned model is available.
Testing and review · anthropics
Test Engineer
Writes characterization tests and dual-run harnesses that pin legacy behavior before modernization, preserving unresolved target behavior as explicitly pending tests.
Why we selected itThe body pins tests to literal inputs/outputs and dual-run comparison against legacy or a recorded trace, not to spec-as-intent.
Weigh upFrontmatter grants Write, Edit, and unrestricted Bash; the modernized/-only and never-edit-legacy/ rules are prompt text, not a sandbox.
Security review · Trail of Bits
Test generator
Pins the zeroize-audit test generator that stops when the final findings report is missing and still writes an empty manifest when nothing is confirmed.
Why we selected itThe agents index has no entry for plugins/zeroize-audit/agents/6-test-generator.md.
Weigh upThis review did not execute the resource.
Security and governance · wshobson
Threat Modeling Expert
Structures a security design review around trust boundaries, data flows, attack paths, mitigations and residual risks.
Why we selected itProvides a concrete sequence from scope and assets through STRIDE and attack trees to documented residual risk.
Weigh upThis is a design-review role, not a scanner, penetration test or compliance certificate.
Security and privacy · GitHub community
Trojan Skill Hunter
Reviews agent instructions, hooks and MCP configuration as untrusted text before installation.
Why we selected itThe file says reviewed content is untrusted data and must not be executed, fetched, or installed.
Weigh upPrompt instructions cannot make a model immune to prompt injection or replace isolated tooling and least-privilege access.
Security review · Trail of Bits
TU compiler analyzer
Pins the zeroize-audit per-file compiler analyzer that stops without compile flags, refuses a finding that lacks the evidence it names, and deletes that unit's temporary artifacts.
Why we selected itThe agents index has no entry for plugins/zeroize-audit/agents/3-tu-compiler-analyzer.md.
Weigh upThis review did not execute the resource.
Testing and review · anthropics
Type Design Analyzer
Reviews domain types by identifying their invariants, checking where those invariants are enforced, and weighing safer designs against complexity and compatibility.
Why we selected itFixed per-type report: invariants list, four rated axes with justification, Strengths, Concerns, Recommended Improvements.
Weigh upIts numerical design rubric is subjective guidance, not a measured quality score or compiler result.
Testing and review · affaan-m
TypeScript Reviewer
Reviews TypeScript/JavaScript diffs with merge-readiness, typecheck and lint gates first, reporting findings without rewriting code.
Why we selected itEstablishes review scope without hard-coding main, then stops and reports on red CI, merge conflicts, failed typecheck/lint or empty scope instead of reviewing blindly.
Weigh upRuns tsc/eslint via Bash; the project's own check commands execute, so sandbox or resource consequences belong to the host configuration.
Evidence review · Undominated.ai
Undominated · 3D parallelism architect
Architects balanced 3D parallelism configurations (TP, PP, DP, SP) for massive LLM serving clusters.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upProduces architectural blueprints and configuration matrices; does not provision physical clusters.
Evidence review · Undominated.ai
Undominated · Adversarial Red Teamer
Classifies supplied reports of prompt-injection and jailbreak failures. It does not generate bypasses or run attacks.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDoes not produce payloads, bypass steps, or a claim that a model is safe.
Evidence review · Undominated.ai
Undominated · Agent cascade failure auditor
Audits multi-agent autonomous chains for failure cascades, hallucinated handoffs, and infinite retry loops.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEvaluates task failure probabilities and handoff schema schemas; does not execute agent tool steps.
Evidence review · Undominated.ai
Undominated · Agent State Budget Architect
Architects conversation state pruning, sliding context windows, and tool output summarization to control multi-turn token costs.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns state compaction algorithms. Does not write domain-specific business application logic.
Evidence review · Undominated.ai
Undominated · Agent trajectory pruning specialist
Multi-turn log auditor specializing in identifying conversational bloat, redundant tool executions, memory leaks, and obsolete context across agent trajectories.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. Modifies no session storage directly.
Evidence review · Undominated.ai
Undominated · Agent workflow economist
Autonomous agent economist specializing in multi-step agent orchestration efficiency, tool schema minimization, function call consolidation, and token-preserving memory architectures.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. Executes no code directly.
Evidence review · Undominated.ai
Undominated · AI resource curator
Review skills, agents and MCP servers against a concrete use case, with pinned identity, rights and runtime evidence.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Attention Architecture Economist
Models serving economics and hardware footprint trade-offs across Multi-Head Attention (MHA), Grouped-Query Attention (GQA), and Multi-Head Latent Attention (MLA).
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upModels memory bandwidth and KV capacity math. Does not assess architectural training convergence.
Evidence review · Undominated.ai
Undominated · Attention Sink Custodian
Manages initial token attention sink preservation and sliding window buffers in infinite-horizon streaming generation.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upConfigures attention sink preservation. Does not modify model base architecture.
Evidence review · Undominated.ai
Undominated · Batch pipeline architect
Systems architect specializing in asynchronous batch inference pipelines, SLA deadline management, partial batch ingestion, and automated synchronous fallback cascades.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. Does not submit batch jobs or manage cloud infrastructure.
Evidence review · Undominated.ai
Undominated · Benchmark rigor auditor
Evaluation integrity specialist auditing public benchmark leaderboards for test set contamination, cherry-picked task splits, and uncalibrated significance claims.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. Treats published claims as hypotheses to verify.
Evidence review · Undominated.ai
Undominated · Cache retention evaluator
Audits needle-in-a-haystack retrieval accuracy under aggressive KV cache compression strategies.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEvaluates pass rates on synthetic needle-in-a-haystack and long-context benchmarks; does not generate answers.
Evidence review · Undominated.ai
Undominated · Canary Rollout Commander
Orchestrates traffic splitting, statistical drift detection, and automated rollbacks for model migrations.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upOrchestrates routing policies and statistical health monitoring.
Evidence review · Undominated.ai
Undominated · Chunked Prefill Tuner
Calibrates chunked prefill batch parameters to achieve target ITL p99 while maintaining prefill throughput.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upTunes engine flags and scheduling parameters.
Evidence review · Undominated.ai
Undominated · Comparison editor
Edit model-comparison copy so ties, same-seller tiers and dropped requirements are not published as dominance.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Constrained Decoding Optimizer
Optimizes JSON schema compilation, regex FSM caching, and token masking to maximize structured output throughput.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upOptimizes FSM compilation and vocabulary masking. Does not alter JSON Schema semantics.
Evidence review · Undominated.ai
Undominated · Context parallelism planner
Plans context-parallel serving topologies for 1M+ token context windows using Ring and Ulysses algorithms.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns cluster layout and network bandwidth contracts; does not allocate GPU memory dynamically.
Evidence review · Undominated.ai
Undominated · Continuous Eval Guardian
Maintains continuous regression evaluation harnesses across production endpoints to detect silent capability degradation.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upMaintains evaluation harnesses and generates alerts. Does not modify external provider endpoints.
Evidence review · Undominated.ai
Undominated · Crawl auditor
Review a sitemap or llms.txt before publication, so an English-only path is not locale-prefixed and a markdown link names a file that was written.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants. It does not fetch the live site to paper over a missing local file.
Evidence review · Undominated.ai
Undominated · Delivery-class reviewer
Review a dominance sentence before it treats batch delivery, or a model's own other delivery, as a like-for-like beater.
Why we selected itThe profile separates delivery class from price and capability, and it forbids a dominated sentence when the challenger is the same seller and the same weights on another delivery.
Weigh upThe profile does not compute a price, a score, or a rank, and it does not admit a rate.
Evidence review · Undominated.ai
Undominated · Disaggregated Serving Architect
Designs separate prefill and decode GPU clusters connected via RDMA to eliminate TTFT and ITL scheduling interference.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns cluster topology and network routing. Does not administer physical switches.
Evidence review · Undominated.ai
Undominated · Distributed Context Engineer
Architects sequence-parallel, RingAttention, and context-parallel clusters for million-token inference pipelines.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns distributed attention topologies. Does not manage Kubernetes cluster provisioning.
Evidence review · Undominated.ai
Undominated · Draft Model Calibrator
Aligns draft model tokenizer, temperature, and logit distributions with the verifier model to maximize acceptance rates.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upCalibrates sampling parameters and evaluates alignment. Does not retrain base models.
Evidence review · Undominated.ai
Undominated · Draft verification auditor
Profiles draft verification kernels to prevent compute overhead from outpacing verification gains.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upProfiles kernel execution times and memory bandwidth; does not alter kernel code directly.
Evidence review · Undominated.ai
Undominated · Dynamic Routing Economist
Architects real-time cost-and-latency aware multi-provider routing proxies that arbitrage spot price shifts.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns routing algorithms and fallback logic.
Evidence review · Undominated.ai
Undominated · Edge AI deployment architect
Sizes models for edge hardware (Apple Silicon, mobile NPUs, local workstations) to eliminate cloud API spend.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upFormulates hardware sizing blueprints; does not compile binary firmware directly.
Evidence review · Undominated.ai
Undominated · Enterprise Compliance Guardian
Audits Zero Data Retention (ZDR), HIPAA, SOC2, and sovereign data boundary commitments across commercial AI providers.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAudits technical and contractual provider commitments. Does not provide legal counsel.
Evidence review · Undominated.ai
Undominated · Enterprise compliance officer
Enterprise compliance and regulatory risk auditor evaluating AI inference pipelines against GDPR data sovereignty, EU AI Act risk classifications, HIPAA BAA requirements, and SOC 2 Type II controls.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. Provides technical guidance; does not replace legal counsel.
Evidence review · Undominated.ai
Undominated · Eval harness auditor
Audit LLM evaluation harnesses, judge prompt rubrics, scoring determinism, and benchmark data hygiene to eliminate evaluation leaks and spurious rankings.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Evidence reviewer
Independently challenge numerical claims using source identity, denominator reconstruction and repeatable calculations.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Expert Load Balancer Auditor
Monitors router entropy, expert skew, and load balancing auxiliary loss regularization in production MoE models.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAudits routing frequencies and logit entropy. Does not retrain router gating weights.
Evidence review · Undominated.ai
Undominated · Failover router architect
Systems resilience architect designing active-active multi-provider inference topologies with zero-loss retries, automated circuit breaking, and cross-region failover.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. Provisions no cloud infrastructure directly.
Evidence review · Undominated.ai
Undominated · GPU Footprint & Sharding Architect
Calculates required GPU VRAM, tensor parallelism sharding (TP=2, 4, 8), and minimum cluster configurations for hosting open-weight models.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEstimates exclude framework runtime overheads (CUDA context, PyTorch activation memory). Always add 15-20% headroom.
Evidence review · Undominated.ai
Undominated · GPU Utilization Optimizer
Implements autoscaling, sleep modes, and batch backfilling to maximize hardware utilization and eliminate idle power waste.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upImplements scheduling and autoscaling strategies.
Evidence review · Undominated.ai
Undominated · Graph RAG Cost Optimizer
Tunes entity extraction prompts and graph traversal depths to balance retrieval precision and token spend.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upOptimizes extraction templates and traversal parameters.
Evidence review · Undominated.ai
Undominated · Guardrail Overhead & Latency Analyst
Analyzes the throughput, latency, and financial token overhead of enterprise AI safety and moderation layers.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upReviews operational telemetry and token logs. Does not establish legal compliance guarantees.
Evidence review · Undominated.ai
Undominated · Hardware Infrastructure Planner
Sizes on-premise GPU clusters based on enterprise peak and off-peak token demand profiles and CapEx payback models.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upModels capacity, power, and amortization economics.
Evidence review · Undominated.ai
Undominated · Hierarchical Memory Architect
Designs multi-tiered memory architectures combining working memory, episodic summaries, and graph knowledge.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns memory schemas and compaction pipelines.
Evidence review · Undominated.ai
Undominated · Hybrid SSM serving architect
Architects hybrid Transformer-Mamba/SSM inference pipelines for extreme token throughput.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upConfigures hybrid serving parameters; does not modify model checkpoint weights.
Evidence review · Undominated.ai
Undominated · In-Context Example Pruner
Audits the marginal accuracy contribution of few-shot exemplars and prunes low-ROI examples to trim token costs.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEvaluates exemplar ablation datasets.
Evidence review · Undominated.ai
Undominated · Indirect Injection Defender
Designs untrusted data isolation boundaries and document sanitization filters to prevent indirect prompt injection.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns data isolation architectures and filtering rules.
Evidence review · Undominated.ai
Undominated · Inference CI/CD Architect
Integrates model quality, latency, and cost regression gates into continuous integration and automated deployment pipelines.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns CI/CD pipelines and quality gates.
Evidence review · Undominated.ai
Undominated · Inference Firewall Engineer
Architects low-latency security firewalls, prompt injection classifiers, and rate limiters at the API gateway layer.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upArchitects gateway security rules. Does not patch underlying model base weights.
Evidence review · Undominated.ai
Undominated · Inference Scaling Strategist
Analyzes test-time compute scaling laws, verifier passes, and search budgets to optimize accuracy per dollar spend.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upFocuses on cost-per-accuracy optimization. Does not train verifier reward models.
Evidence review · Undominated.ai
Undominated · Inference SLO Gate Architect
Defines and enforces strict latency, availability, and error budget SLO gates for production AI gateways.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upSLOs require observable telemetry data. Cannot guarantee performance on uninstrumented endpoints.
Evidence review · Undominated.ai
Undominated · Inference SRE architect
Designs production observability, SLO alert thresholds, and automated traffic shedding for inference platforms.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns observability architecture and alerting thresholds; does not configure live monitoring daemons.
Evidence review · Undominated.ai
Undominated · Judge Model Calibrator
Calibrates LLM judge prompts with few-shot calibration anchors to eliminate position bias and rating drift.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upCalibrates judge prompts and scoring guidelines.
Evidence review · Undominated.ai
Undominated · KV eviction policy architect
Architects dynamic KV cache pruning and eviction policies under strict hardware VRAM limits.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns cache eviction heuristics; does not alter hardware memory controllers.
Evidence review · Undominated.ai
Undominated · Library intake editor
Decide whether a skill, agent profile, or MCP server belongs in the public install set, using pinned identity, a licence, and a real checker or an explicit reason it has none.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upA portable role profile. It does not register a native subagent or certify third-party runtime behaviour.
Evidence review · Undominated.ai
Undominated · Licence derivation auditor
Open-source legal and IP compliance specialist auditing fine-tuning data lineage, synthetic data derivation rights, model weight inheritance, and commercial redistribution boundaries.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. Provides technical IP audit guidance.
Evidence review · Undominated.ai
Undominated · LLM security red teamer
Designs adversarial jailbreak probes and evaluates model alignment under automated attacks.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDevelops adversarial evaluation benchmarks; does not perform illegal cyber operations.
Evidence review · Undominated.ai
Undominated · Long context bandwidth auditor
Audits inter-node network requirements to guarantee communication is fully masked by attention compute.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upVerifies network saturation limits against compute execution times; does not capture live packets.
Evidence review · Undominated.ai
Undominated · Low-bit fidelity auditor
Audits mathematical reasoning and coding benchmarks under aggressive sub-4-bit quantization.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEvaluates benchmark regression matrices; does not perform weight quantization.
Evidence review · Undominated.ai
Undominated · Merged capability auditor
Audits merged model checkpoints for catastrophic forgetting and multi-task capability collisions.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEvaluates benchmark score matrices; does not perform model inference directly.
Evidence review · Undominated.ai
Undominated · Mixed Precision Architect
Designs layer-wise quantization profiles balancing perplexity retention against memory savings across FP4, FP8, and INT4.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns layer quantization strategies. Does not train quantization scales.
Evidence review · Undominated.ai
Undominated · Model Cascade Architect
Architects multi-tier model cascading pipelines with confidence-gated fallback routing to optimize cost-per-task.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upRequires access to historical query distributions and labeled validation datasets.
Evidence review · Undominated.ai
Undominated · Model merging architect
Architects zero-training model merges using TIES, DARE, and SLERP to combine specialized domain models.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upFormulates weight blending equations and pruning parameters; does not modify PyTorch tensors directly.
Evidence review · Undominated.ai
Undominated · Model Merging Evaluator
Evaluates merged model weights across benchmark suites to prevent catastrophic interference and capability loss.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAudits weight cosine similarity and downstream benchmark evaluations.
Evidence review · Undominated.ai
Undominated · Model migration planner
Plan workload-specific model migrations while preserving required capabilities and defining observable rollback criteria.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Model Provenance Investigator
Investigates training data lineage, open-weights licensing boundaries, and attribution requirements across model releases.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upInvestigates public training disclosures and license documentation.
Evidence review · Undominated.ai
Undominated · MoE Serving Architect
Specializes in Expert Parallelism (EP), Tensor Parallelism (TP), and expert capacity factor calibration for large MoE clusters.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns distributed topology and capacity factors. Does not write CUDA kernels.
Evidence review · Undominated.ai
Undominated · Multi-agent systems economist
Designs multi-agent workflows to minimize redundant inter-agent chatter and prevent runaway token loops.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns governance policies and token budgets; does not execute multi-agent swarms.
Evidence review · Undominated.ai
Undominated · Multi-LoRA Serving Specialist
Configures multi-adapter serving clusters serving hundreds of fine-tuned LoRA adapters concurrently on a single base model.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upConfigures S-LoRA/Punica serving frameworks. Does not train LoRA weights.
Evidence review · Undominated.ai
Undominated · Multi-Needle Retrieval Validator
Generates and evaluates multi-needle retrieval benchmarks to expose lost-in-the-middle context blindness in long-context models.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEvaluations use synthetic text corpora and require substantial inference tokens to complete comprehensive sweeps.
Evidence review · Undominated.ai
Undominated · Multimodal streaming architect
Architects real-time streaming audio and video inference pipelines with bounded lookahead latency.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns chunking topologies and latency budgets; does not stream live media.
Evidence review · Undominated.ai
Undominated · NCCL topology auditor
Audits NVLink, NVSwitch, and InfiniBand rail-optimized communication topologies for distributed serving.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAnalyzes NCCL topology graph dumps and bus profiles; does not reconfigure hardware switches.
Evidence review · Undominated.ai
Undominated · On-device performance auditor
Audits thermal throttling, unified memory pressure, and battery consumption on consumer edge devices.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEvaluates recorded thermal and memory telemetry; does not read hardware sensors in real time.
Evidence review · Undominated.ai
Undominated · Population reviewer
Review a count, rank or correlation before publication, so the computed rows are the population the sentence names.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Preference alignment auditor
Audits reward models and preference datasets for length bias, sycophancy, and mode collapse.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAnalyzes preference dataset distributions and response characteristics; does not execute training loops.
Evidence review · Undominated.ai
Undominated · Pricing source reviewer
Review a provider price feed before any rate is admitted, including units, currency, tiers and prose quotes that are not USD.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Prompt Compression Economist
Audits client-side prompt compression pipelines to minimize billed input token volume without losing semantic fidelity.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns compression and pruning policies. Does not modify upstream model weights.
Evidence review · Undominated.ai
Undominated · Prompt firewall architect
Constructs low-latency pre-inference guard firewalls to intercept injection and extraction attacks.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns firewall filter rules and routing architecture; does not run network proxies.
Evidence review · Undominated.ai
Undominated · Provider Resilience Architect
Designs automatic failover, circuit breaking, and idempotency guarantees across redundant inference providers.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns gateway resilience architectures.
Evidence review · Undominated.ai
Undominated · Quantization Evaluation Reviewer
Audits LLM evaluation protocols to verify whether published benchmark scores match the model's serving quantization or upstream unquantized weights.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upReview is strictly factual and evidence-bound. Missing quantization disclosures are flagged as unverified.
Evidence review · Undominated.ai
Undominated · Quantized KV Cache Calibrator
Calibrates per-channel and per-group scale factors for sub-8-bit KV caches to prevent long-context reasoning drift.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upCalibrates scaling parameters and clipping thresholds.
Evidence review · Undominated.ai
Undominated · Quote identity reviewer
Review a proposed join between a vendor quote and a catalogue model, and refuse any match that is not exact.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · RAG Context Economist
Optimizes context window packing, chunk overlap boundaries, and token budgets for enterprise RAG systems.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upRelies on documented retrieval metrics and rate cards. Does not alter upstream vector database schemas.
Evidence review · Undominated.ai
Undominated · Reasoning budget planner
Architect agent that models reasoning token expenditure, dynamically calibrates thinking budgets across prompt difficulty classes, and prevents runaway cost expansion.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Reasoning Trace Compressor
Audits chain-of-thought reasoning traces to detect circular deliberation loops, repetitive tokens, and distillation opportunities.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAudits reasoning trace structure and redundancy. Does not rewrite private model weight internal activations.
Evidence review · Undominated.ai
Undominated · Reasoning verifier auditor
Audits process reward model verifiers for false-positive validation passes and search path collapse.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEvaluates verification agreement and error distributions; does not execute reward models.
Evidence review · Undominated.ai
Undominated · Recurrent state auditor
Audits recurrent hidden state numerical drift and precision degradation during continuous streaming.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upMonitors state norm stability and accumulation variance; does not generate token predictions.
Evidence review · Undominated.ai
Undominated · Regulatory Compliance Officer
Audits AI deployments for EU AI Act, US Executive Order, and ISO 42001 governance and transparency frameworks.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAudits technical compliance requirements. Does not provide formal legal counsel.
Evidence review · Undominated.ai
Undominated · Resource release verifier
Independently verify packaged resources, clean installation and meaningful public availability after a release.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Router Threshold Calibrator
Calibrates draft model confidence thresholds to balance API cost savings against quality degradation.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upCalibration relies on empirical validation curves. Thresholds must be re-calibrated upon model updates.
Evidence review · Undominated.ai
Undominated · Safety Refusal Calibrator
Calibrates safety classifier thresholds to minimize benign query refusals while maintaining strict compliance boundaries.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDoes not loosen boundaries on core safety violations (CSAM, CBRN, self-harm). Calibrates benign edge cases.
Evidence review · Undominated.ai
Undominated · Schema Contract Verifier
Verifies tool schemas, argument enums, and parameter constraints to prevent runtime validation failures.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAudits JSON schema specifications and validation logic.
Evidence review · Undominated.ai
Undominated · Serving Engine Tuner
Specializes in tuning vLLM, SGLang, and TensorRT-LLM parameters including chunked prefill, radix trees, and batching scheduling.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upFocuses on runtime parameters and flags. Does not rewrite model CUDA kernels.
Evidence review · Undominated.ai
Undominated · Sovereign AI TCO auditor
Models national sovereign AI cluster CapEx/OpEx, power grid contracts, and 5-year hardware lifecycle TCO.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upModels multi-year financial statements and depreciation schedules; does not issue procurement tenders.
Evidence review · Undominated.ai
Undominated · Sparse Attention Evaluator
Audits accuracy retention and memory bandwidth reduction across Native Sparse Attention (NSA) and block-sparse mechanisms.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upEvaluates benchmark accuracy retention and hardware bandwidth savings.
Evidence review · Undominated.ai
Undominated · Speculative Architecture Engineer
Architects speculative decoding topologies including Multi-Token Prediction (MTP), Medusa heads, and Eagle tree verifiers.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns draft-verifier architectures and tuning parameters.
Evidence review · Undominated.ai
Undominated · Speculative tree architect
Designs multi-candidate speculative draft trees to maximize generation speed on constrained GPUs.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns tree topologies and candidate selection masks; does not compute forward passes.
Evidence review · Undominated.ai
Undominated · Stack cost optimizer
Full-stack AI system cost optimization specializing in multi-tier cascading, retrieval pruning, speculative drafting, and prompt caching to slash API spend while preserving benchmark quality.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. No automatic subagent registration or permission grants.
Evidence review · Undominated.ai
Undominated · Streaming Reliability Engineer
Specializes in Server-Sent Events (SSE) connection health, keep-alive heartbeat spacing, and client reconnection handling.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upOperates over observed network traces and server configuration. Does not execute intrusive network attacks.
Evidence review · Undominated.ai
Undominated · Sustainable inference architect
Optimizes AI workload scheduling for renewable power availability and lowest carbon footprint.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns carbon-aware scheduling models; does not interact with electrical power grids directly.
Evidence review · Undominated.ai
Undominated · Synthetic curation architect
Architects synthetic training dataset filtering pipelines to eliminate contamination and redundancy.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns filtering criteria and similarity thresholds; does not store external training corpora.
Evidence review · Undominated.ai
Undominated · Synthetic judge calibrator
Evaluation scientist specializing in LLM-as-a-judge alignment, position bias mitigation, length normalization, and inter-rater agreement calibration against human experts.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. Executes no judge calls directly.
Evidence review · Undominated.ai
Undominated · Tail latency auditor
Profiles p99.9 tail latencies to eliminate micro-stalls caused by garbage collection and multi-tenant contention.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAnalyzes latency distribution traces; does not modify kernel scheduling priorities.
Evidence review · Undominated.ai
Undominated · Tail latency SLA auditor
Production observability auditor comparing provider Service Level Agreements against measured p95/p99 tail latency, serverless cold starts, and outage windows.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upPortable profile; manually load or adapt to a native agent format. Executes no billing disputes directly.
Evidence review · Undominated.ai
Undominated · Ternary inference architect
Architects native ternary and low-bit compute pipelines for low-energy edge and cloud inference.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns compute pipelines and memory layouts; does not fabricate custom silicon.
Evidence review · Undominated.ai
Undominated · Test-time compute economist
Optimizes test-time compute allocation across search width and reasoning depth to balance accuracy and cost.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upModels cost-accuracy curves and search allocation policies; does not generate reasoning traces.
Evidence review · Undominated.ai
Undominated · Tool Registry Architect
Designs dynamic tool indexing and semantic retrieval to keep system prompt tokens minimal across large API catalogs.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upDesigns tool registry indexing and selection mechanisms.
Evidence review · Undominated.ai
Undominated · Vision token economist
Optimizes vision patch resolution and temporal subsampling to minimize multi-modal token billing.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upCalculates patch token billing models; does not run image computer vision models.
Evidence review · Undominated.ai
Undominated · VRAM Fragmentation Analyst
Analyzes PagedAttention block size allocation, CUDA graph memory pools, and virtual memory fragmentation to prevent GPU OOMs.
Why we selected itDefines a specialist role, required evidence, handoff contract and explicit limits on external actions.
Weigh upAudits memory allocation patterns. Does not debug hardware PCIe bus faults.
Creative tooling · Agency Agents community
Unity editor tooling developer
A Unity profile for editor extensions, asset-import checks and build-blocking validation.
Why we selected itEditor code is isolated through Editor folders, conditional guards and assembly definitions instead of leaking UnityEditor into player builds.
Weigh upExample code and adoption/time-saved claims were not executed or measured; the profile does not provide a Unity licence or CI runner.
Game development · Seth Hobson
Unity gameplay and performance engineer
Plans Unity gameplay and asset/rendering work using project-specific pipelines and target-hardware profiling.
Why we selected itThe profile connects Profiler/Frame Debugger evidence to CPU/GPU/memory work and includes Addressables asset-lifecycle concerns.
Weigh upThe broad capability inventory is not proof of version compatibility; some service names are legacy and require current official documentation.
Game development · Agency Agents community
Unreal world-streaming planner
An Unreal environment profile for world partition, HLOD and procedural population planning.
Why we selected itIt calls for configuring streaming structure before population and reviewing distant HLOD visually.
Weigh upFixed cell sizes, distances, layer limits and numeric performance targets are author heuristics, not measured limits for your project; examples need engine-version validation.
Changelog editing · Astral Software Inc.
uv changelog editorial
Rewrites only the newest uv changelog section, and keeps the version, date, pull-request numbers, and URLs as they are.
Why we selected itThe prompt edits nothing and uses no network; it returns only the newest release section.
Weigh upThis Markdown file has no frontmatter and is not a native subagent registration.
Pull-request labels · Astral Software Inc.
uv pull-request labels
Chooses at most the semantic labels for one uv pull request, and leaves the list empty when the diff does not support one.
Why we selected itLabels have to come from .pull-request-labels.json. An unclear case is an empty array. The prompt says not to recommend labels already on the pull request and not to suggest removals.
Weigh upThe prompt forbids file/GitHub changes and PR-code execution; the reviewed read-only profile permits only api.github.com and github.com.
Bug reproduction · Astral Software Inc.
uv reproduce-bug
Decides whether a reported uv behavior can be reproduced, and refuses to treat source inspection as a reproduction.
Why we selected itThe outcome is one of reproducible, not_reproducible, or needs_more_information. A reproduced behavior is not treated as proof the behavior is wrong.
Weigh upThe prompt is tied to astral-sh/uv issue context and test paths. Its workflow prepares temporary state from uv-dev; copying the prompt does not create that state.
CI diagnosis · Astral Software Inc.
uv workflow-failure
Classifies one failed uv workflow as flaky or deterministic, and as create, duplicate, or ignore.
Why we selected itflaky is allowed only when every independent root cause is transient. One deterministic failure makes the run deterministic. Dependent cancellations are not independent root causes.
Weigh upThe prompt assumes astral-sh/uv and uv-dev issue searches and repository labels. Another project must adapt those names, the input collector and reporting destinations.
Data and retrieval · wshobson
Vector Database Engineer
Plans and implements semantic retrieval systems through corpus analysis, embeddings, chunking, index selection, hybrid search, reranking and monitoring.
Why we selected itWorkflow is requirement→embedding→chunk→store→index→hybrid→rerank→monitor, which is a real RAG build path.
Weigh upThe broad implementation role declares no tool allowlist; access to databases, credentials and file changes comes from the host.
Infrastructure and delivery · Vercel
Vercel deployment expert
Diagnoses failed Vercel builds, function timeouts, environment variables and domains, and stages production builds before promoting them.
Why we selected itIt says to promote a staged production deployment, not a preview, because promoting a preview rebuilds it with production variables and the tested build is not the one released.
Weigh upOne build branch says to delete the lockfile, reinstall and commit a fresh one on a lockfile mismatch. That re-resolves every dependency; prefer fixing the mismatch.
Browser and performance · Vercel
Vercel performance optimizer
Walks LCP, INP and CLS decision trees for a Vercel or Next.js app, measuring first and reading the installed Next.js docs before recommending a change.
Why we selected itThe audit checklist starts by measuring real-user Core Web Vitals in Speed Insights and identifying the LCP element before any change.
Weigh upEach finding includes an Expected Improvement target metric. The file gives no method for that estimate; treat it as a hypothesis to re-measure.
Modernization and specifications · anthropics
Version Delta Analyst
Maps an exact same-stack version pair to the APIs and behaviors a particular tree uses, returning delta cards and explicitly lost tool coverage to the migration owner.
Why we selected itRequires file:line evidence for the intersection of version changes and actual code usage, excluding removed APIs the code never calls.
Weigh upRead-only wording conflicts with the instruction to create an rsync scratch copy and run tools that write build output/download plugins. Bash is declared; the profile does not enforce a filesystem or network sandbox.
Testing and review · affaan-m
Vue Reviewer
Reviews Vue single-file components for reactivity correctness, v-html and URL-binding security, composable cleanup, Router and Pinia patterns, and SSR safety — reporting findings without rewriting code.
Why we selected itOwns an explicit scope table that splits Vue-specific lanes (reactivity, v-html, composables, props/emits, Router, Pinia, SSR) from the generic TypeScript reviewer, so the two roles compose instead of duplicating.
Weigh upFull .vue coverage needs the sibling typescript-reviewer for generic type-safety lanes; this file reviews Vue-specific concerns only.
Systems performance · Agency Agents community
WebAssembly boundary and runtime planner
Evaluates Wasm adoption through measured workloads, explicit memory ownership and capability-scoped runtime design.
Why we selected itThe wasm-bindgen example explicitly performs bulk copies and disclaims a zero-copy ABI; ownership is part of the boundary design.
Weigh upThe Wasmtime fragment only builds configuration/context and does not supply complete linking, store, limits or module execution code.
Infrastructure and delivery · VoltAgent
Webhook Engineer
Builds webhook handlers and delivery paths that survive the real world: raw-body signature verification before parsing, durable enqueue before the 2xx, deduplication matched to the provider retry window, and dead-letter replay as a first-class operation.
Why we selected itTreats at-least-once delivery, duplicates and reordering as the default world and builds the checklist around them: constant-time signature checks with replay windows, event identifiers persisted before processing, a dedup store whose TTL matches the provider retry window, and database constraints as the final dedupe guarantee.
Weigh upThe integration section names sibling agents (backend-developer, api-designer and others) for coordination; those are separate files and not needed for the core method.
Documentation · Microsoft Corporation
Wiki writer
Microsoft deep-wiki agent that writes VitePress markdown with dark-mode Mermaid diagrams and a citation for each claim, and that refuses to write before the git remote is resolved.
Why we selected itThe catalogue has the deep-wiki researcher skill and not this writer agent.
Weigh upThis review did not execute the agent.
Build and packaging · GitHub community
Windows ARM64 wheel builder
Extends an existing Python wheel workflow with a native Windows ARM64 path, checking matrix filters, native dependencies and artifact names.
Why we selected itIt traces reusable workflows and composite actions before editing, and checks whether an ARM64 entry already exists to avoid duplicates.
Weigh upNo tools list is declared, so host defaults can include editing and execution. Scope workflow edits and CI triggering explicitly.
Infrastructure · VoltAgent
Windows infrastructure change planner
A Windows infrastructure profile organizing directory, DNS, DHCP and policy changes around inventories and rollback.
Why we selected itThe checklist requires enumerating affected objects before modification and reviewing a WhatIf preview.
Weigh upWhatIf support depends on the actual command and does not guarantee rollback; Bash/Edit/Write can alter infrastructure through available credentials.
Persistent work coordination · GitHub community
Workshop coordinator
Coordinates workstreams using shared artifacts, desk journals and structured escalation signals across independent sessions.
Why we selected itIt uses files and journals to carry state across sessions and treats peer findings as evidence to review.
Weigh upRequires workshop-create, desk-open, bench-read, desk-journal and signal-write; the optional signals dashboard is separately installed.
API engineering · VoltAgent
X data integration contract engineer
An integration profile organizing X-data reads and writes around explicit provider contracts and reproducible fixtures.
Why we selected itIt preserves partial results after late pagination failures and snapshots source version and request parameters for auditability.
Weigh upNo API access or quota is included; the alternative provider mention is not independent validation of its provenance or terms. The example completed-tests sentence is a template, not evidence.
Security review · Trail of Bits
Zeroize MCP resolver
Pins the zeroize-audit agent that resolves symbols through Serena and stops the pass when activation fails or the query budget is spent.
Why we selected itThe agents catalogue has no sourcePath for plugins/zeroize-audit/agents/1-mcp-resolver.md.
Weigh upThis review did not execute the resource.
Code review · Zoom
Zoom integration reviewer
Reviews a Zoom integration for architectural fit, token lifecycle, webhook verification and SDK or API misuse, findings first.
Why we selected itIt names what to flag first: wrong product-surface choices, broken token lifecycle assumptions, webhook verification bugs and unsupported SDK or API assumptions.
Weigh upThe file is a short prompt. It does not list Zoom API behaviour itself; correctness depends on the documentation and skills in the session.
Security review · Zoom
Zoom OAuth scope auditor
Reviews the Zoom scopes an app requests, configures and documents for least privilege, missing scopes and mixed product surfaces.
Why we selected itIt is told not to invent scopes, to say what evidence is missing, and to prefer the narrowest documented scope set that supports the workflow.
Weigh upThe file is a short prompt with no scope list of its own. Correctness depends on the Zoom documentation and skills available in the session.
03 / BEHIND THE SELECTION
The review is part of the product.
A useful recommendation tells you why it fits—and what still needs checking.
What we inspect
We read the complete selected definition, identify its role, workflow, output expectations and verification instructions, and check the applicable licence. Review notes distinguish instructions from enforcement: a prompt saying “read only” does not itself restrict the host’s tools.
What a selection means
Each resource has a concrete use case, original selection reasons, named limitations and a dated review record. We favour specific instructions, useful scope and inspectable setup. The list is curated; it is not a census of everything available or an independently benchmarked league table.
Before using it
Read the complete review, check the upstream revision and match the resource to your own host and permissions. Review notes apply to the material named at the recorded date. Installation commands may retrieve a newer upstream version.
Where we looked
- anthropics/claude-plugins-official ↗
- github/awesome-copilot ↗
- VoltAgent/awesome-claude-code-subagents ↗
- wshobson/agents ↗
- msitarzewski/agency-agents ↗
- affaan-m/everything-claude-code ↗
- AWS Labs AI-DLC Workflows ↗
Reviewed authored personas and the documented framework projection; retained the advisory compliance role.
- Trail of Bits Skills ↗
Reviewed full worker definitions, output contracts, plugin dependencies and the CC-BY-SA licence.
- Sentry Skills ↗
Screened agent originals; recorded frontmatter exclusions and selected the original Anthropic simplifier instead of a duplicate vendored copy.
- Microsoft Skills ↗
Reviewed infrastructure and wiki agent definitions; excluded project-specific or overlapping roles.
- VoltAgent/awesome-claude-code-subagents — expanded definition review ↗
Reviewed complete candidate definitions, licensing and required host/dependency boundaries; selections are source reviews, not runtime benchmarks.
- affaan-m/everything-claude-code — expanded definition review ↗
Reviewed complete candidate definitions, licensing and required host/dependency boundaries; selections are source reviews, not runtime benchmarks.
- anthropics/claude-plugins-official — expanded definition review ↗
Reviewed complete candidate definitions, licensing and required host/dependency boundaries; selections are source reviews, not runtime benchmarks.
- github/awesome-copilot — expanded definition review ↗
Reviewed complete candidate definitions, licensing and required host/dependency boundaries; selections are source reviews, not runtime benchmarks.
- msitarzewski/agency-agents — expanded definition review ↗
Reviewed complete candidate definitions, licensing and required host/dependency boundaries; selections are source reviews, not runtime benchmarks.
- wshobson/agents — expanded definition review ↗
Reviewed complete candidate definitions, licensing and required host/dependency boundaries; selections are source reviews, not runtime benchmarks.
- Svelte primary source ↗
Direct source screening; popularity was not treated as evidence of quality.
- Terraform AWS provider ↗
- Anthropic Buffa agent definitions ↗
Reviewed the Rust API ergonomics role and the Apache-2.0 licence. Repository-internal maintainer agents from other publishers were recorded as exclusions.
- GitHub Agentic Workflows ↗
- OpenAI Codex rust review label ↗
Read the rust review label and the Apache-2.0 licence at this commit. The workflow that substitutes the event-path token was not opened.
- Anthropic math-proof agents ↗
Read the math-proof judge and worker, and the plugin Apache-2.0 licence. The siege skill that launches them was not opened.
- GitHub AWF diagnosis agent ↗
Read the diagnosis playbook and the MIT licence. Sibling docs named by the playbook were not opened, and no probe was run.
- astral-sh/uv ↗
Four Codex prompts, config.toml, triage-issue.md, and LICENSE-MIT at the reviewed commit. Write-capable prompts were not selected.
- github/docs ↗
Read .github/agents/readability-editor.md, LICENSE-CODE, and README at b3ca4b7986534061037979e480c9d363fc95a909. The dependabot ecosystem agent at that commit was not selected.
- anthropics/claude-plugins-official at f359d43b ↗
Listed every plugins/*/agents/*.md. Selected four new agents. plugin-validator still ends with session text and silent-failure-hunter still fails YAML parsing, so their earlier exclusions stand.
- anthropics/financial-services at 574ed362 ↗
Formerly financial-services-plugins. Read five agent-plugin definitions, the README install section, two cookbook workers and the Apache-2.0 licence.
- zoom/zoom-plugin-codex at 10936829 ↗
Read both reviewer agents, agents/openai.yaml, the README, the plugin manifest and the MIT licence.
- vercel/vercel-plugin at b97d74d0 ↗
Read the three agents, the README (install, agents, telemetry), the plugin manifest and the Apache-2.0 notice. The ai-architect agent was not selected.
- openai/plugins at 0722921d ↗
Listed plugin-level agents/ directories. Vercel and Zoom agents were pinned at their publishers' own repositories. The Figma agents are under the Figma Developer Terms.
- openai/codex built-in agent roles at 21c16161 ↗
Read awaiter.toml and explorer.toml. Neither was selected.
- aws/agent-toolkit-for-aws at c34fbdfd ↗
Read the seven aws-startup-advisor agents and the Apache-2.0 licence. None was selected.
- stripe/ai at deae744b ↗
Read the company-researcher agent. Not selected.
- getsentry/sentry-for-ai at c2313d3a ↗
Read the descriptions of the skill-creator and skill-updater agents. Not selected.
- google-gemini/gemini-cli at 29315e53 ↗
Listed the tree. The only .gemini/agents file is tools/gemini-cli-bot/.gemini/agents/WORKER.md, inside the repository's own bot tooling. It was not opened, and no other agent definition was found.
- github/awesome-copilot at 156736fb ↗
Compared agents/*.agent.md with the library and its exclusions, read the partners plugin manifest and README, and read each candidate definition named in a record or an exclusion.
- GitHub community agent workflow review ↗
Reviewed complete definitions, MIT notice and current host boundaries. Selected six concrete workflows; four standalone originals include licensed downloads. No task execution or security certification.
Selections we withheld (118)
These are documented decisions from this review, not a blacklist of the wider ecosystem.
- anthropic-silent-failure-hunter: The full upstream file fails YAML parsing: the unquoted description contains embedded example text with colon-space syntax. Excluded from screened native-agent downloads; additional project-specific logging assumptions would also need adaptation. Source ↗
- anthropic-plugin-validator: The definition ends with unrelated generation-session chatter, refers to sibling validation utilities not included in a standalone download, and encodes host schema assumptions that need current validation. Kept as reviewed evidence, excluded from selected downloads. Source ↗
- github-api-architect: The role depends on a Code Interpreter environment that is not declared in its host tool configuration, and tightly couples its generation workflow to a particular service/manager architecture. Excluded in favor of more portable, source-aware architecture definitions. Source ↗
- github-scientific-paper-research: Primarily a wrapper around a specific remote BGPT MCP service, whose access terms, host configuration and evidence-quality output were not verified here. Too little standalone research methodology for the selected definition downloads. Source ↗
- github-search-ai-optimization-expert: The broad SEO/GEO checklist mixes implementation and optimization claims without a concrete evidence or audit-output contract. Named commercial analysis tools are not bundled or configured; stronger bounded research and technical roles were selected. Source ↗
- voltagent-docs-drift-editor: The otherwise narrow editing workflow incorrectly says an HTML comment preserves an old fragment anchor. It also permits a TODO mutation while describing the page as skipped. Do not promote the unchanged definition until these instructions are corrected. Source ↗
- voltagent-dependency-manager: Broad checklist persona includes fixed, apparently completed vulnerability, package and performance results as a delivery notification, without a sufficiently specific evidence/output contract. A narrower verified workflow is preferable. Source ↗
- voltagent-incident-responder: Includes operationally destructive containment actions and an unrestricted Bash/write tool set without an explicit action-authority gate; fixed successful incident metrics also appear in the delivery template. Not selected as a ready reusable response role. Source ↗
- voltagent-data-engineer: Generic broad implementation checklist asserts zero-data-loss guarantees and supplies fixed successful delivery metrics. Insufficient grounding and role-specific output constraints for this curated selection. Source ↗
- voltagent-assumption-mapping: The full upstream role has invalid YAML frontmatter: its unquoted description contains a colon followed by a space. Excluded from screened downloads rather than requiring users to repair a known-broken native definition. Source ↗
- voltagent-ab-test-analysis: The full upstream file fails YAML parsing because its description contains unquoted colon-space syntax. Its fixed-threshold decision framework also conflates minimum detectable effect with a shipping threshold and lacks a specified sequential design for extending tests. Source ↗
- voltagent-scientific-literature-researcher: Relies on a BGPT service and an unspecified context-manager integration, includes example completed research metrics, and repeats vendor evidence-quality claims without independent validation. The selected Research Synthesist provides a clearer source-assessment method without that dependency. Source ↗
- agency-design-ux-researcher: Includes fixed adoption and satisfaction targets while deferring important methodological detail to the model’s training. Research recruitment, consent, analysis and output verification are insufficiently specified for this screened collection. Source ↗
- agency-design-brand-guardian: Broad brand and trademark strategy persona with unmeasured consistency/equity targets and methodology deferred to core training. Insufficiently bounded deliverables and validation for the selected collection. Source ↗
- agency-product-sprint-prioritizer: Supplies fixed velocity, timeline and feature-success targets without an evidence workflow or configured planning integration. Excluded rather than present the persona’s targets as demonstrated planning performance. Source ↗
- agency-testing-reality-checker: Predetermines a negative assessment and grade range, treats any zero-issue report as an automatic failure, and assumes a particular local capture script and app layout. That is not a neutral evidence gate. Source ↗
- agency-project-management-experiment-tracker: Treats reaching statistical significance for nearly every experiment as a success target and supplies an illustrative revenue claim without a measurement procedure. Excluded because that framing can bias experimental interpretation. Source ↗
- wshobson-observability-engineer: An extensive prospective capability list spanning telemetry, cloud cost, compliance and chaos engineering, with little bounded output or operational authority protocol. Excluded in favor of more focused definitions. Source ↗
- wshobson-context-manager: Broad context-engineering capability lists do not provide a sufficiently concrete handoff schema or standalone memory implementation. Excluded to favor narrowly actionable roles and avoid implying that a prompt installs persistent memory. Source ↗
- ecc-build-error-resolver: Its quick-recovery section deletes the dependency lockfile and reinstalls packages, conflicting with a minimal, reproducible repair role. Excluded from the unchanged downloadable selection. Source ↗
- ecc-e2e-runner: Contains misleading Playwright guidance that raw page.click lacks auto-waiting, recommends networkidle for animation timing, and offers test quarantine plus artifact upload without enough workflow boundaries. A corrected definition would be preferable. Source ↗
- ecc-doc-updater: Assumes companion code-map generation scripts and tooling that are not included in the standalone file. It overlaps the selected documentation roles, whose repository inspection and output contracts are more reusable. Source ↗
- voltagent-content-quality-editor: Delegates its central editing step to a globally installed external unslop package that was not independently reviewed, while using fixed readability targets as a quality gate. The standalone definition is not a self-contained substantive editing method. Source ↗
- voltagent-ai-writing-auditor: The role treats unsupported stylometric frequency thresholds and blanket punctuation/hedging rules as an audit method, then rewrites the entire text. It also incorporates named external writing systems whose applicable upstream permissions were not independently verified in this pass. Source ↗
- github-refine-issue: A short generic issue-refinement checklist requests delete_issue alongside update tools without explaining or gating that unrelated destructive capability. It lacks a concrete change-review contract and was not selected. Source ↗
- github-agent-governance-reviewer: The broad governance role prescribes a Python-style policy decorator and speculative trust scoring across multiple frameworks without a concrete output contract, verification protocol or referenced implementation. It overlaps better-bounded security and threat-modeling roles and was excluded rather than presented as an implemented enforcement layer. Source ↗
- github-task-researcher: Requires deleting non-selected alternatives from the final research record and assumes a large host-specific tool collection. It substantially overlaps the selected technical-spike researcher while preserving less of the decision trail. Source ↗
- sentry-senpai: The full original fails strict YAML parsing: an unquoted description contains colon-space syntax at Examples:. Excluded from unchanged native-agent downloads. Source ↗
- sentry-code-simplifier: This vendored copy puts an attribution comment before its YAML block and duplicates the Anthropic role. Selected the valid original Anthropic definition with its applicable licence instead. Source ↗
- awslabs-aidlc-aws-platform: The persona covers live provisioning, drift remediation and teardown through inherited shell tools. The single file supplies neither an action-specific approval contract nor a fixed output schema; framework controls would need separate operational verification. Not selected for this reusable-definition expansion. Source ↗
- awslabs-aidlc-operations: Includes production chaos experiments and restart, failover and scaling runbooks without defining action-specific authorization, abort conditions or a bounded response artifact in the persona. Prefer a separately reviewed operational workflow over promoting the unchanged role as a general-purpose download. Source ↗
- trailofbits-dimension-validator: The definition assigns severity from mismatch type and broadly excludes commented conversions, test/mock code and documented edge cases from findings without requiring the same path-level verification. It also depends on prior unit annotations and host-specific task tools. The narrower evidence-driven workers were selected instead. Source ↗
- microsoft-coreai-infrastructure: This is a CoreAI DIY project runbook with fixed repository paths, Bicep modules and Azure deployment/teardown commands. It is not a portable infrastructure definition, and the unchanged file cannot supply another project’s deployment boundaries. Source ↗
- microsoft-wiki-researcher: Substantially overlaps the selected code exploration and research roles. It depends on a resolved repository/citation context and has no tools allowlist; the expansion favors more bounded workers with explicit analysis outputs. Source ↗
- AWS Incident Triage: Withheld after inspecting its required aws-cloudwatch-investigation companion: the throttling example treats every API Gateway 4XXError as a throttling event, and the previous-week comparison duplicates the same metric inside one request without shifting its time window. Current AWS API documentation does not support these interpretations. The investigation outline is useful, but the unchanged dependency can produce misleading incident evidence. Source ↗
- aws-principal-architect: Broad AWS expertise inventory lacks a concrete required deliverable or verification contract. The selected serverless architect gives a more specific design output and fills the AWS architecture gap without another overlapping generalist. Source ↗
- aws-serverless-architect: Withheld in favour of more specific, verifiable definitions. Its advice mixes an HTTP API preference with REST-only API Gateway features without explaining the boundary, asks for rough monthly costs without a pricing-source or calculator requirement, and requests working infrastructure examples without a verification gate. Source ↗
- react19-test-guardian: Contains a false React 19 Strict Mode claim and tells the agent to change spy assertions to observed call counts. Current React documentation still describes extra development effect cycles. Its passWithNoTests/forceExit-driven completion loop can also obscure missing coverage or unresolved handles. Source ↗
- arm-migration-agent: Directs automatic architecture and dependency changes while treating builds and integration validation as optional. It also requires learning_path_server, which is absent from the declared MCP tools. The unchanged workflow lacks a sufficiently reliable validation contract for architecture-specific changes. Source ↗
- Oracle-to-PostgreSQL Migration Expert: Withheld after inspecting required companion skills: migrating-oracle-to-postgres-data-access-code instructs retaining CommandType.StoredProcedure for function calls, contrary to the default behaviour documented for Npgsql 7 and later. Its risk-review skill also needs a separate references tree. The phase structure is useful, but the unchanged workflow includes a material migration error. Source ↗
- OpenAPI to Application Generator: Withheld because the procedure asks for a complete deployable application but does not define concrete OpenAPI validation, unresolved-reference handling, architecture decisions or executed acceptance checks; it largely repeats the existing scaffolding role. Source ↗
- backend-development-performance-engineer: Assigns generic latency severity thresholds and asks for expected improvement estimates without a concrete measurement protocol. The selected database and browser performance roles make stronger distinctions between observed measurements and recommendations. Source ↗
- backend-development-test-automator: Substantially overlaps the selected Test Engineer and the expanded QA and Playwright roles. Its general capability list and test-file organization add less distinctive method than the bounded alternatives. Source ↗
- backend-development-graphql-architect: Mostly a broad catalogue of GraphQL capabilities and technologies, with no specific output schema, evidence contract or failure handling. Not selected solely to fill an API-design label. Source ↗
- monorepo-architect: A compact setup checklist with no concrete validation, migration boundaries or deliverable contract. The architecture and build-related roles already cover its general planning advice. Source ↗
- eval-orchestrator: Combines static checks and model-supplied rubric/F1 scores into badges such as production ready without a defensible runtime validation contract. The workflow depends on a CLI and judge absent from the standalone definition and overstates what the composite can establish. Source ↗
- framework-migration-legacy-modernizer: A short general migration checklist substantially overlaps the existing modernization roles. The selected Uplift Migrator adds a more concrete pilot-playbook gate, unit ownership and executed-build reporting contract. Source ↗
- api-designer: The delivery template contains fixed successful endpoint and SDK counts, while the role assumes an unspecified context-manager protocol. The extensive topic lists do not establish an evidence-backed validation or completion contract. Source ↗
- a11y-architect: Misidentifies WCAG 2.2 criterion 2.4.11 as focus contrast and blends target-size/spacing advice into purported universal requirements. Current W3C documentation defines 2.4.11 as Focus Not Obscured (Minimum). The unchanged reference should be corrected before recommendation. Source ↗
- Software Architect: The bounded-context and architecture-pattern guidance is useful, but overlaps the selected Code Architect, Architecture Critic and ADR roles. This expansion prioritizes missing methods rather than another general system-design persona. Source ↗
- Test Automation Engineer: The CI example relies on an unconfigured PLAYWRIGHT_TRACE environment variable for tracing and its worker fixture is not safe across concurrent shard jobs sharing a backend. The role also prescribes fixed reliability targets and automatic test quarantine. Prefer the narrower exploration-first test role. Source ↗
- Spatial Data Scientist: Overstates that non-spatial models on spatial data necessarily produce invalid inference, and lists many spatial methods without a precise data/validation contract. The selected statistician provides clearer design and assumption boundaries without that blanket claim. Source ↗
- Language Translator: A Spanish-English travel and conversation persona rather than a software-localization workflow. Broad cultural and emergency guidance and phonetic approximations would need specialist verification; this pass does not promote it as a general translation or localisation solution. Source ↗
- Lingo.dev Internationalization: Delegates the entire procedure to a remote i18n_checklist tool whose current workflow and service requirements were not independently reviewed. The file supplies no substantive standalone localization method beyond obeying that service. Source ↗
- GitHub Community Elasticsearch Observability Agent: Withheld at this revision: the embedded MCP configuration combines a remote type with local command/args and no remote URL field. It needs a host-specific configuration correction before the advertised Elastic tools can be treated as an installable integration. Source ↗
- Agency Experiment Tracker: Withheld for experiment analysis: its success criteria reward most experiments reaching statistical significance. That is not a sound generic measure of experiment quality and can bias interpretation away from valid null results. Source ↗
- SAST and SCA analyzer: The pinned profile assigns an overall compliance PASS when no Critical or High findings are found. A bounded SAST/SCA scan cannot establish compliance with the named frameworks, and the self-assessed quality score supplies no independent validation. Source ↗
- ARM Cortex expert: The pinned profile both recommends static mut DMA buffers and later calls all static mut undefined behavior. Its critical-section example unconditionally enables interrupts after disabling them, failing to preserve a previously disabled state. Hardware-specific code needs correction and device-bound verification. Source ↗
- Healthcare reviewer: The profile emits SAFE TO DEPLOY for clinical software from an LLM code review without a validated clinical-safety acceptance contract or qualified sign-off. That verdict exceeds this source-review evidence; retain as a private candidate pending a bounded contract. Source ↗
- SE Security Reviewer at the reviewed revision: Withheld: its prompt-injection example labels a simple input-sanitization approach secure, which overstates the protection offered by string filtering. Source ↗
- Technical Content Evaluator at the reviewed revision: Withheld: the instructions depend on an undefined /ultra-think invocation and mandate numerical content scores without a validated measurement procedure. Source ↗
- replay-minimizer: The opened definition is bound to microsoft/TypeScript’s Go TestReplay harness, flags and replay-file format. It is a compiler-repo procedure, not a portable crash minimizer. Source ↗
- Demonstrate: The opened Copilot agent demonstrates a microsoft/vscode pull request through vscode-playwright-mcp and can edit settings. It is not a portable UI check and it does not produce a pass or fail result. Source ↗
- go-sdk-tool-migrator: The opened definition migrates one toolset in github/github-mcp-server from mark3labs/mcp-go to modelcontextprotocol/go-sdk. It is not a general Go or MCP migration guide. Source ↗
- .NET Notebook Migration Agent: The opened definition converts notebooks in microsoft/ai-agents-for-beginners and allows runCommands while the body says not to execute code. It is a course-repo transform, not a general agent. Source ↗
- Claude Security plugin agents: plugins/claude-security/LICENSE at this commit defines the Plugin as "This software, including its prompts, agent and skill definitions, workflows, server code, and documentation" and says "you may not: (a) distribute, publish, sublicense, sell, or otherwise make the Plugin or any modified version of it available to any third party". Its agents patch-verifier and scan-verifier are therefore not listed, not even with a short description quote. The licence is byte-identical at c447c3207a425bc4e2a0d068435f64b0477ae981. Source ↗
- anthropic-pr-review-code-reviewer: Same review responsibilities and confidence threshold as the selected feature-dev code reviewer, but this copy declares no tools, so in Claude Code it inherits every subagent tool, including Write and Bash. Source ↗
- anthropic-pr-review-code-simplifier: The body after the frontmatter is byte-identical to plugins/code-simplifier/agents/code-simplifier.md, already listed as anthropic-code-simplifier. Only the description differs. Source ↗
- anthropic-math-proof-worker-deep: The higher-effort variant of the listed math-proof worker. Its description says it is launched only by the math-proof plugin's siege skill, from the escalation round on. Source ↗
- anthropic-skill-eval-comparator: Assigns 1 to 5 rubric scores and a 1 to 10 overall score, and is told to be decisive because ties should be rare. The pass-or-fail grader from the same skill is listed instead. Source ↗
- anthropic-skill-eval-analyzer: Reads the blind comparator's winner and scores to explain why the winner won, so it depends on the comparator excluded above. Source ↗
- anthropic-valuation-reviewer: The agent computes fund NAV, carried interest and LP allocations itself, with no cite-every-number rule and no independent re-verification step of the kind the GL reconciler has. Source ↗
- vercel-ai-architect: Its model-selection tree names fixed models for lowest latency and for speed with low cost, and a fallback order, with no measurement, source or date behind those comparisons. Source ↗
- openai-codex-awaiter: A built-in Codex helper role that polls a command until it finishes. It is compiled into Codex, not a definition users load, and it reviews nothing. Source ↗
- openai-codex-explorer: The built-in explorer.toml is an empty file (0 bytes) at this commit. Source ↗
- figma-design-parity-review-agent: The Figma plugin's agents are governed by plugins/figma/LICENSE.txt, which points to the Figma Developer Terms and says the beta materials may be modified or discontinued at any time. Those terms were not reviewed. The design-system-rules agent in the same folder was excluded for the same reason. Source ↗
- aws-llm2bedrock-agents: The five llm2bedrock agents (code-analyzer, log-ingestor, prompt-evaluator, code-rewriter, report-generator) are phases of AWS Startup Migrate Track 2 and consume each other's results. The report generator prices source models from a static table to be updated periodically, and estimates a missing model from the closest listed one. Source ↗
- aws-generic-phase-workers: generic-phase-worker-rw and -rwx are dispatched only by the plugin's DSL interpreter. Their descriptions say not to dispatch them directly for user conversation. Source ↗
- stripe-company-researcher: Researches a third-party company's website to pre-fill Stripe Connect discovery answers. Not a code, security, testing or evaluation role, and its confidence levels are inferred by the model. Source ↗
- sentry-for-ai skill-creator and skill-updater: Maintainer agents that create or update Sentry SDK skill bundles in this repository and open a pull request. Source ↗
- github-dynatrace-expert: The embedded MCP server URL is a specific dynatracelabs.com development host (pia1134d.dev.apps.dynatracelabs.com) and sends the user's Dynatrace token to it as a bearer header. Loaded unchanged, the agent would route credentials to that host. Source ↗
- github-amplitude-experiment-implementation: Creates a live experiment in the user's Amplitude project with create_experiment and wraps the feature in it, with no approval step before the remote write. No tools or MCP server are declared in the file. Source ↗
- github-postgresql-dba: A broad DBA persona with pgsql_modifyDatabase, bulk-load and runCommands tools, listing backups and restores among its tasks, with no confirmation gate or output contract. Source ↗
- github-ms-sql-dba: A broad DBA persona with runCommands and query tools covering backups, restores, upgrades and security changes, with no confirmation gate or output contract. Source ↗
- github-se-system-architecture-reviewer: Chooses architectures from fixed user-count and monthly hosting-budget bands (for example under $100 a month means serverless) with no source, and creates an ADR file for every decision. Source ↗
- github-quality-playbook: Its description claims it finds the 35% of real defects that structural review cannot catch, and the body says iterations typically add 40-60% more confirmed bugs. Neither figure is sourced in the file. Source ↗
- github-gem-reviewer: A hidden subagent (user-invocable false) that needs the gem-team orchestrator's plan_id and handoff fields, and returns a model-supplied confidence number such as 0.95. Source ↗
- github-insiders-a11y-tracker: Tracks accessibility issues released to VS Code Insiders in microsoft/vscode only. It does not review the user's own code. Source ↗
- github-defender-scout-kql: Tells the agent to always filter on Timestamp, yet lists DeviceTvmSoftwareVulnerabilities among its tables. Microsoft's schema page for that table (learn.microsoft.com/en-us/defender-xdr/advanced-hunting-devicetvmsoftwarevulnerabilities-table, checked 2026-10-13) has no Timestamp column, so the rule produces failing queries. Source ↗
- github-wg-code-sentinel: A security persona with a broad edit and runCommands tool set and no evidence or output contract beyond a severity label. Much of the file is a speaking style. Source ↗
- github-tech-debt-remediation-plan: Says analysis only, no code modifications, but declares editFiles, runCommands, runTests and github tools, and asks it to create issues from a template. Source ↗
- github-critical-thinking: A questioning mode that overlaps the listed devils-advocate agent and produces no artefact to review. Source ↗
- github-se-system-architecture-reviewer: Generic user-count/budget thresholds and one-line architecture decision trees are unsupported heuristics; overlap with existing architecture reviewers. Source ↗
- voltagent-cohort-analysis: Unquoted description contains Triggers on: and fails YAML parsing; not distributed as a working native definition. Source ↗
- voltagent-backlog-grooming: Unquoted description contains Triggers on: and fails YAML parsing; not distributed as a working native definition. Source ↗
- voltagent-first-principles-thinking: Unquoted description contains Triggers on: and fails YAML parsing; not distributed as a working native definition. Source ↗
- ecc-django-build-resolver: Troubleshooting commands can print database credentials, kill unrelated port owners and fake/reset migrations. Not admitted as a general build helper. Source ↗
- github-technical-content-evaluator: Conflicting grade ceilings, absent /ultra-think dependency and simulated learner-success estimates weaken the claimed evaluation contract. Source ↗
- voltagent-license-engineer: Pseudo-legal engineering terminology and example 100% clearance claims do not establish usable legal review or licence clearance. Source ↗
- wshobson-payment-integration: Instructs acknowledging webhooks before database writes/durable acceptance and groups PayPal under generic HMAC handling; insufficient payment correctness. Source ↗
- wshobson-c4-container: Conflates C4 containers with environment deployment topology and mandates OpenAPI across non-HTTP interfaces. Official C4 documentation separates deployment diagrams. Source ↗
- github-electron-angular-native: Names ipcMain.handleSync, absent from the inspected official ipcMain API, and includes older project-specific assumptions. Replaced with a clearer role. Source ↗
- github-winforms-expert: Partial candidate review found conflicting never-dock versus Dock=Fill directions and questionable floating-range guidance. Full technical validation not completed; not admitted. Source ↗
- agency-engineering-privacy-engineer: Overbroad special-category and legal-basis statements plus unqualified anonymization claims make the profile unsuitable for a general technical privacy recommendation. Source ↗
- agency-engineering-search-relevance-engineer: Says search-time-only synonyms break matching, then recommends exactly that in its example; official synonym_graph documentation explicitly supports search-only use. Source ↗
- agency-engineering-mobile-release-engineer: Android lane calls the iOS dSYM upload_symbols_to_crashlytics action as though it uploads mapping.txt. Current official action docs list iOS support. Source ↗
- Zeroize preflight: The same file says the orchestrator inputs are a closed table that includes languages such as rust, and that table does not name cargo_manifest. The Rust mode section then runs only when cargo_manifest is provided and tells the agent to verify that path. Those two instructions do not match, so the file was not selected. Source ↗
- Zeroize PoC generator: The same file says PoC generation is mandatory and that every finding gets a PoC regardless of confidence, then says Rust PoC generation is enabled for only three categories and that every other Rust category records poc_supported false. That contradiction is why it was not selected. The program-writing steps are not repeated here. Source ↗
- Claude Security orchestrator: The root README says to use each plugin's LICENSE file. plugins/claude-security/LICENSE says this plugin, including its agent definitions, may not be distributed or published to any third party. The root LICENSE is Apache-2.0. Those terms disagree, so this file is not selected. Source ↗
- Scan researcher: The root README says to use each plugin's LICENSE file. plugins/claude-security/LICENSE says this plugin, including its agent definitions, may not be distributed or published to any third party. The root LICENSE is Apache-2.0. Those terms disagree, so this file is not selected. Source ↗
- Code simplifier: plugins/pr-review-toolkit/README.md states the licence as MIT. The plugin LICENSE at this pin is Apache License Version 2.0 and is byte-identical to the repository root LICENSE. The README and LICENSE disagree, so this file is not selected. Source ↗
- Arithmetic scanner: Pass 0 says the inventory that supplies total_files_scanned already applies the Pass 1 path exclusions. The output section calls that same Pass 0 baseline pre-exclusion and sets total_files_scanned equal to files_excluded plus the post-filter arithmetic and non-match counts. Those definitions of the same count cannot both be true. Source ↗
- Dimension annotator: The preserve-annotations section says never to overwrite an existing dimensional annotation, and the next bullet says to update an annotation when it is clearly incorrect. The same section says to add to existing NatSpec and not to replace it. Those instructions disagree about whether an existing annotation may be changed. Source ↗
- PoC builder: Its primary instructions are to build proof-of-concept exploits for a vulnerability that earlier phases already treated as verified, including executable exploit code and tests meant to show impact. This review does not publish that procedure. Source ↗
- Rust compiler analyzer: The input table does not name enable_asm, but assembly analysis is skipped when enable_asm is false and the assembly output is included only when enable_asm is true. An O0 IR emission failure skips Step 4 in the process and skips Steps 4 and 5 in the error section. The opening says Step 7 surveys the crate and writes coverage-gaps.json, while the survey heading and the error section assign that write to Step 6 and assign Step 7 to cleanup. Those instructions do not match, so the file was not selected. Source ↗
- Report assembler: The same file says each finding's severity is high or medium, then requires a high, medium, and low breakdown and tells final mode to downgrade one result's severity to low. Final mode reads the existing findings file and does not rerun collection, but the error section says a missing source-findings file is fatal. Those instructions do not match, so the file was not selected. Source ↗
- PoC validator: The file's purpose is to compile and run every proof of concept in the manifest and record whether the run is exploitable. That walks through an exploit proof of concept, so it is excluded. The commands and result mapping are not repeated here. Source ↗
- PoC verifier: The file's purpose is to walk through each proof of concept and judge whether it demonstrates the vulnerability it claims. That walks through an exploit proof of concept, so it is excluded. The checks and examples are not repeated here. Source ↗
04 / PRACTICAL ANSWERS
Before you add it.
Is an agent definition a running AI service?
No. These are reusable role and workflow instructions for an agent host. You still need a compatible tool, model access and appropriate permissions. The host and model determine what the definition can actually do.
Can I copy the complete definition?
When the applicable upstream licence permits redistribution, the detail page includes the exact reviewed definition, a download, attribution and licence text. Preserve those notices when reusing or redistributing it. Other resources link to the original instead.
Will every definition work in every AI tool?
No. Frontmatter, model aliases, tool names, installation paths and delegation behavior vary by host. Each review lists documented compatibility and relevant limitations. Adaptations need their own review; a downloaded role is not a universal installation package.
How were these roles selected?
We look for a specific job, an actionable workflow, useful output expectations and explicit checks or boundaries. Each definition has original notes explaining its inclusion. We have not benchmarked these roles against every alternative or executed them across all hosts.