tool-arena-attack-problem

name

tool-arena-attack-problem

description

Autonomous end-to-end attack pipeline for any Einstein Arena problem. Composes recon -- parallel 5-agent research spawn -- hypothesis graph synthesis -- OVERVIEW.md -- user gate -- attack -- polish -- tri-verify -- submit gate -- retrospective. Triggers on "attack arena problem", "investigate arena challenge", "work on challenge", "new arena problem", "autonomous arena attack", "bootstrap arena attack", "arena investigate", "attack this problem". Not for tactical polish only (use tool-arena-runner) or raw API ops (use tool-einstein-arena).

tool-arena-attack-problem -- Autonomous arena challenge attack pipeline

One command, one slug, end-to-end autonomous campaign. Pre-integrates every Organon tool the framework has for arena work: the full arena-framework orchestrator, all 5 specialist recon agents, sci-literature-research parallel fanout, tool-paperclip biomedical grounding, tool-firecrawl-scraper web fallback, sci-council, sci-optimization-recipes, ops-ulp-polish, ops-parallel-tempering-sa, tool-einstein-arena API. The human only enters at gates.

When to use

User says: "attack prime-number-theorem", "work on the new heilbronn problem", "investigate the uncertainty principle challenge", "take this arena problem on autonomously". Skip when the user only wants a tactical polish of an existing solution (route to tool-arena-runner polish) or raw API traffic (route to tool-einstein-arena).

Methodology -- 7 stages

The skill is a playbook. Each stage is one CLI call + (sometimes) parallel Agent spawns. Claude drives the agent spawns directly via the Agent tool -- the Python scripts handle all non-agent work.

Stage 1 -- RECON (autonomous)

python3 .claude/skills/tool-arena-attack-problem/scripts/attack.py recon \
    --slug {SLUG} --workspace projects/tool-arena/{SLUG}

This fetches problem spec + leaderboard + top-K solutions + discussions via tool-einstein-arena, runs the rigor scan, extracts reference papers from problem.json (arXiv IDs, DOIs, named works), writes:

• {workspace}/problem.json, leaderboard.json, best_solutions.json, discussions.json
• {workspace}/recon/SUMMARY.md
• {workspace}/literature/REFERENCES.md (priority-1 context for the literature agent)
• {workspace}/PLAYBOOK.md (bootstrapped from tool-einstein-arena/assets/playbook-template.md)

Stage 2 -- RESEARCH FAN-OUT (Claude spawns 5 agents in parallel)

Claude MUST spawn these agents simultaneously in a single message with 5 Agent tool calls. Each agent receives the workspace path and writes exactly one file. See references/spawn-agents.md for the exact prompts.

Agent	Writes	Purpose
arena-literature-agent	{workspace}/literature/LITERATURE.md	Published bounds, SOTA, BibTeX. Uses sci-literature-research fanout + paperclip + WebSearch fallback.
arena-historian-agent	{workspace}/recon/COMPETITOR_FORENSICS.md	Per-rank structural diffs, discussion mining
arena-pattern-scout-agent	{workspace}/recon/APPLICABLE_PATTERNS.md	Matches problem shape against arena-framework/arena-patterns/
arena-rigor-agent	{workspace}/recon/RIGOR_REPORT.md	Classifies top-K as rigorous vs exploit
arena-router-agent (optional)	{workspace}/recon/ROUTING.md	Problem class + ranked primitive stack

All 5 share the same inputs from Stage 1 and do not see each other. Adversarial diversity is the design.

Stage 3 -- HYPOTHESIZE (autonomous)

python3 .claude/skills/tool-arena-attack-problem/scripts/attack.py hypothesize \
    --workspace projects/tool-arena/{SLUG}

Composes the 4 agent artifacts into a hypothesis graph via arena_framework.hypothesize.synthesize. Writes HYPOTHESES_DRAFT.md.

Stage 4 -- CRITIC REVIEW (Claude spawns one agent)

Spawn arena-critic-agent on the draft graph. It writes {workspace}/recon/CRITIQUE.md with FATAL/MAJOR/MINOR findings + missing hypotheses.

Stage 5 -- OVERVIEW + USER GATE (autonomous)

python3 .claude/skills/tool-arena-attack-problem/scripts/attack.py overview \
    --workspace projects/tool-arena/{SLUG}

Re-synthesises with CRITIQUE.md included, writes:

• {workspace}/HYPOTHESES.md (final graph)
• {workspace}/OVERVIEW.md -- the rich human-readable briefing: problem statement, published bounds table, competitor timeline, top-5 hypotheses with P(BEAT) + kill criteria, 3 proposed attack directions, open questions.

Claude prints a concise summary of OVERVIEW.md and asks the user whether to proceed (AskUserQuestion). The user approves, modifies, or kills the campaign here.

Stage 6 -- ATTACK + POLISH + TRI-VERIFY (autonomous on approval)

On approval, invoke AttackOrchestrator with the hypothesis graph + routing decision. Integrates ops-parallel-tempering-sa, sci-optimization-recipes, ops-ulp-polish, and arena-runner's tri-verify. Writes solutions/best.json, solutions/polished.json, attack_candidate.json.

Stage 7 -- SUBMIT GATE + RETROSPECTIVE (user-gated)

Presents the final candidate + tri-verify report to the user. On approval, tool-einstein-arena submit fires. Regardless of submit outcome, retrospective.py runs and proposes new patterns / fixtures for the next campaign (see arena-patterns/ additions).

API

from attack import run_stage, main

# Library entrypoint for tests and programmatic use
run_stage("recon", slug="prime-number-theorem", workspace=Path("/tmp/w"))

# Full-campaign driver (for scripts)
main(argv=["recon", "--slug", "x", "--workspace", "/tmp/w"])

Dependencies

Dependency	Required	Provides	Fallback
arena-framework library (plugins/arena/arena-framework/)	Yes	Recon, AttackOrchestrator, synthesize	None -- skill fails loudly
.claude/agents/arena-*.md	Yes (5 specialist agents)	Research fan-out	None -- cannot proceed without them
tool-einstein-arena	Yes	API ops (fetch, submit)	Offline mode with cached JSON
sci-literature-research	Optional	Parallel fanout (paperclip + arXiv + PubMed + S2)	Sequential search via paper-search MCP
tool-paperclip	Optional	Biomedical full-text grounding	Federated search only
tool-firecrawl-scraper	Optional	Web fallback when WebFetch blocked	Skip scrape, note in LITERATURE.md

Run bash .claude/skills/tool-arena-attack-problem/scripts/setup.sh once on a fresh clone.

References

• references/workflow.md -- step-by-step playbook for Claude (stage timing, what to do if an agent fails, how to interpret gate prompts)
• references/spawn-agents.md -- exact prompts for each of the 5 agents (Stage 2 + Stage 4)
• references/overview-schema.md -- what sections OVERVIEW.md must contain and why

Tests

python3 -m pytest .claude/skills/tool-arena-attack-problem/tests/ -v

Covers: stage dispatch, recon -> reference extraction round-trip, hypothesize with partial agent outputs (graceful degradation), overview rendering with all 5 sections, missing-workspace error handling, idempotent re-run.

Routing

Use when user says: "attack arena problem X", "investigate challenge X", "work on new arena problem", "autonomous arena campaign for X", "take on Einstein Arena challenge", "bootstrap arena attack on X", "full attack pipeline for X".

Do NOT use for: tactical polish only (tool-arena-runner polish), leaderboard inspection (tool-einstein-arena analyze), post-hoc submission retry (tool-einstein-arena submit).