chore: end-session — log, gac token correction in CLAUDE.md, FUTURE offload targets

Claude-Session: https://claude.ai/code/session_01Ku87DDWwbJkTx6SuV3sQDc

memory/log/2026-09-11.135548.md

@@ -0,0 +1,25 @@
+# Session Log -- 2026-09-11
+
+## Summary
+Built, verified live, and shipped two new MiniMax offloads: `/review-minimax` (first-pass code review of agent branches → adjudicated findings file) and `/ci-triage-minimax` (CI failure triage over gac logs). Wrapper scripts + specs + offline tests in small-scripts, skills in custom-claude-skills, memory docs here; all three repos pushed.
+
+## Decisions
+- Decision: wrapper materialises all inputs (diff/log) itself; MiniMax sub-session gets `Read,Glob,Grep,Write` only — no Bash, no Edit -- Rationale: kills four review blockers at once (wrong-branch inheritance, two-dot diff, tree clobbering, uncontrolled git); post-run `git status --porcelain` assertion backstops Write.
+- Decision: 540s synchronous timeout + `--max-diff-lines` budget instead of detached launch -- Rationale: harness Bash 600s cap kills longer synchronous runs; with no test execution and pre-materialised inputs, review fits.
+- Decision: credential-shaped diff content aborts review (exit 1) but CI logs are redacted-and-continued -- Rationale: a credential in a branch diff is itself a finding for in-session review; logs are the whole payload and routinely carry unmasked clone URLs.
+- Decision: `--run` required for the triage script; picking "newest failure" from gac's human table is the skill's job -- Rationale: unversioned table parse doesn't belong in the script.
+- Decision: Opus plan review (deep-fable named workflow) before implementation -- Rationale: it caught 4 genuine blockers and ~14 should-fixes in the first draft; cost well spent.
+
+## Gotchas Discovered
+- **[gotchas-gitea]** Symptom: `gac` (v0.0.2.10) errors "no token found" despite root CLAUDE.md saying it falls back to the secrets path; also `GITEA_TOKEN_FILE` pointed at `~/dev/claude/secrets/gitea/ai_admin` fails ("is a directory") -- Fix: the token is `ai_admin/token` (file inside the dir); ci-triage-minimax-run now defaults `GITEA_TOKEN_FILE` to that path when no token env is set (env vars override). Root CLAUDE.md's gac token-discovery description should be corrected.
+- **[gotchas-minimax]** Symptom: offline exit-0 testing impossible when the script `rm -f`s the out-file before each attempt (pre-seeded files get cleared) -- Fix: companion fake seams (`REVIEW_MINIMAX_FAKE_FINDINGS`, `CI_TRIAGE_FAKE_TRIAGE`) that play the sub-session's Write after the JSON fake; gates still run.
+
+## Key Context
+- Live verification results: planted-issue branch → 5 confirmed findings, all citations valid; clean branch → `NO FINDINGS` (no invented findings); failed gac-scratch run 3 → correct step + honest low-confidence on a truncated log. Telemetry confirmed `contextWindow=1000000` via `MiniMax-M3[1m]`.
+- Review run cost/latency: ~23k in / 4.6k out tokens, 73s (planted); 37s (clean). Triage similar scale.
+- small-scripts `tests/run-all.sh` has 2 pre-existing failures (test-claude-profile, test-wait-for-version) from another session's uncommitted WIP (`wait-for-version.spec.md` etc. untracked) — not touched.
+- Next offload candidates from the original analysis: `/reflect-logs` draft mode, mechanical test generation from specs (`mm-test-task`), `/end-session`/housekeeping doc summarisation, transcript extraction.
+
+## Process Notes
+- The materialise-inputs-yourself pattern (wrapper runs git/gac; sub-session reads files in cwd) is the strongest MiniMax hardening shape yet — reuse it for future offload wrappers instead of granting `Bash(git *)` globs.
+- MiniMax review checklist must be explicit grep-able patterns; repo-specific items (the 5 agent-runtimes test bugs) go behind `--profile` or they manufacture findings on unrelated stacks.