RESEARCH AND DOCUMENTATION / Anthropic
math-proof: siege
Runs an hours-long, multi-round attack on one hard mathematics problem with judge and worker sub-agents and a bookkeeping script, ending in a proof.md whose Status section says what is and is not proved.
“Work on one hard mathematics problem in rounds: each round a judge writes a few self-contained questions, independent workers answer them, and the judge keeps a ledger of what is proved, refuted and open, until the judge concludes or the rounds run out, and then a proof.md says plainly what is and is not proved. A run takes hours and dozens of worker runs. Usage: /math-proof:siege [NAME=value settings] <the problem, stated in full, or the path of a file holding it>.”
01 / THE REASONING
Why this made the selection.
- The orchestrating session does no mathematics: each step goes to a fresh math-proof-judge subagent and each query to a fresh worker that sees only that query and the problem, and the judge keeps a ledger of claims marked PROVED, REFUTED or OPEN.
- Mechanical decisions are made by scripts/ledger.py, a standard-library Python script that validates the judge's output, applies the early-conclusion gate and decides which worker answers are finished; a complete proof or disproof of the goal is checked line by line by two more workers before the judge concludes.
02 / THE REVIEW RECORD
What we actually inspected.
Source review has boundaries.
A clear record is more useful than a “safe” badge.
Material inspected
- plugins/math-proof/skills/siege/SKILL.md (frontmatter, approach, arguments and settings)
- plugins/math-proof/skills/siege/scripts/ledger.py (module docstring and imports)
- plugins/math-proof/agents/math-proof-judge.md
- plugins/math-proof/agents/math-proof-worker.md
- plugins/math-proof/agents/math-proof-worker-deep.md (frontmatter and tool sentence)
- plugins/math-proof/README.md
- plugins/math-proof/LICENSE
Our findings
- Default settings: ESC_ROUND 4, WAVE 4, WAVE_DEEP 10, MAX_ROUNDS 14, MIN_ROUNDS 4, REFINE_STEPS 2 and MAX_COMMIT 9; DIR=path is also accepted, and any other NAME=value token (two or more capitals, a whole-number value) at the start of the message is refused.
- From round ESC_ROUND on, every worker is a math-proof-worker-deep and a round may hold up to WAVE_DEEP queries.
- Worker agents have Read, Write and Edit only and are told they have no web access and no code execution; the judge agent has Read, Write, Edit, Bash, Glob and Grep and may run Python 3 checks.
- ledger.py imports only re, sys and pathlib (besides from __future__ import annotations), needs Python 3.7 or later, and prints a fail-closed verdict instead of a traceback on internal errors.
Not established by this review
- No siege run was started; ledger.py was not executed.
- The SKILL.md was read through its settings section and grepped for state.md, result-so-far.md and the Status section; the round loop and proof-tail steps were not read line by line.
- Not executed: it was not installed or run.
The review applies to the material and revision named here. A newer upstream release can change its behavior.
03 / PUT IT TO WORK
Add a skill to your workflow.
- The math-proof README's install line, `/plugin install math-proof@claude-plugins-official`, adds the whole math-proof plugin (siege, solo and the three agents siege launches), so it is not siege's own command and install.command is null.
- Before the first run the README asks for the "env" block shown in configExample, in ~/.claude/settings.json or in .claude/settings.json of the directory you run from.
- The README also prints an unattended form, claude -p "/math-proof:siege problem.md" with --model claude-opus-5-5, --effort high and --dangerously-skip-permissions, and says adding --max-budget-usd <amount> stops the run once that much has been spent on an API key.
- No line was run for this review.
Before you start
- Claude Code 2.1.280 or later
- Python 3.7 or later on PATH as python3 or python, with no extra packages
Documented configuration example
Use the client and authentication context described in the setup guidance. Configuration formats vary between hosts.
{
"env": {
"CLAUDE_CODE_MAX_OUTPUT_TOKENS": "128000",
"CLAUDE_CODE_DISABLE_BACKGROUND_TASKS": "1",
"CLAUDE_ASYNC_AGENT_STALL_TIMEOUT_MS": "14400000"
}
} 04 / FOLLOW THE EVIDENCE
The source trail.
Our notes are separate from the original resource.
Check upstream before adopting a new version.
- math-proof siege SKILL.md ↗Checked
https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/math-proof/skills/siege/SKILL.md
Supports: upstreamDescription, summary, whySelected[0], whySelected[1], bestFor[0], bestFor[1], bestFor[2], limitations[3], review.findings[0], review.findings[1], access.permissions[0], access.permissions[1], access.permissions[3]
https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/math-proof/skills/siege/scripts/ledger.py
Supports: whySelected[1], review.findings[3], install.prerequisites[1]
- math-proof-judge agent ↗Checked
https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/math-proof/agents/math-proof-judge.md
Supports: review.findings[2], access.permissions[2], limitations[1]
- math-proof-worker agent ↗Checked
https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/math-proof/agents/math-proof-worker.md
Supports: review.findings[2], access.permissions[2]
https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/math-proof/agents/math-proof-worker-deep.md
Supports: review.findings[2], access.permissions[2], review.findings[1]
- math-proof README ↗Checked
https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/math-proof/README.md
Supports: install.sourceUrl, install.steps[0], install.steps[1], install.steps[2], install.prerequisites[0], install.prerequisites[1], configExample, limitations[1], limitations[2], limitations[4], limitations[5], access.cost, compatibility, bestFor[1], bestFor[2], summary
- math-proof LICENSE ↗Checked
https://raw.githubusercontent.com/anthropics/claude-plugins-official/41f32e407b450055044f167db7ded5e8ff67d203/plugins/math-proof/LICENSE
Supports: license, access.cost