Four memory docs (2 skills, 2 wrapper scripts) + MEMORY.md index lines; ask-minimax-usage.md gains both delegation targets. Records the gac v0.0.2.10 token-discovery gap (needs GITEA_TOKEN_FILE pointed at ai_admin/token; the wrapper defaults it). Claude-Session: https://claude.ai/code/session_01Ku87DDWwbJkTx6SuV3sQDc
2.5 KiB
/review-minimax Skill
Purpose: Delegate a first-pass code review of a git branch (agent-produced worktrees from mm-worktree-task, ralph, container agents) to a MiniMax M3 sub-session. The diff and context reads cost MiniMax tokens; only a findings file returns, which the operator adjudicates. A pre-filter before /code-review or an in-session review — never the final authority.
Location: custom-claude-skills/skills/review-minimax/ — thin wrapper over review-minimax-run (small-scripts), which owns all mechanics.
Usage
/review-minimax <worktree-path> [--profile agent-runtimes] → runs:
review-minimax-run --worktree <path> [--base main] [--profile agent-runtimes]
Findings land in ${XDG_RUNTIME_DIR:-/tmp}/review-minimax/<repo>/<branch>-findings.md.
Key design points
- The wrapper materialises the diff itself (
base...HEAD, three-dot, with:(exclude)pathspecs for secrets/keys); the sub-session getsRead,Glob,Grep,Writeonly — no Bash, no Edit — so it can't run git, inherit the wrong branch, or modify the tree. A post-rungit status --porcelainassertion catches any unexpected write (exit 2). - Credential gate: a diff matching credential patterns (
ghp_,AKIA…, private-key headers,x-access-token:…) aborts with exit 1 — review in-session instead. - Result gate: findings need a
### [BLOCKER|SHOULD-FIX|NIT]block plus at least one citation resolving against the worktree, or the literalNO FINDINGS; otherwise exit 3. - Checklist is explicit patterns (llm-code-security §10/§12 distilled), not "find problems" — MiniMax reasoning-loops on open-ended analysis. The 5 recurring MiniMax test-file bugs load only under
--profile agent-runtimes. - Exit contract 0/1/2/3 mirrors explore-minimax-run; fallback on 2/3 is a Claude Sonnet subagent review.
- Adjudication is mandatory: findings are untrusted data; spot-check 1–2 citations;
NO FINDINGSdoes not certify the branch.
Which modes use it
deep/deep-fable post-agent-merge review, orch for reviewing container-agent branches before merge. Related: script-review-minimax-run, skill-explore-minimax, skill-ci-triage-minimax.
Gotchas
- Synchronous with a 540s internal timeout (harness Bash 600s cap) — the diff budget (
--max-diff-lines, default 6000) keeps runs inside that; oversize diffs exit 1 rather than half-review. Writemust stay in the sub-session allowlist: without it MiniMax exits 0 and the findings file silently never appears (gotchas-minimax).