Review-first automation. One hard rule is encoded everywhere and must never be relaxed, plus one policy that keeps the queue honest:
- The bot never merges. Landing requires the manual protocol
(merge gate -> issue repro vs main ->
cargo test --workspace->corpus_harness.py->differential_loop.pyFULL) plus a credit commit. - The queue drains itself (doctrine change, 2026-10-24). The old rule —
"the bot never closes" — was written for a repo receiving a handful of
PRs; at ~65 arrivals/day it produced a 400+ PR review-debt pile that no
session could honestly work through. Now
tools/queue_sweeper.py(daily via.github/workflows/queue-sweeper.yml) closes unprotected PRs older thanSWEEPER_MIN_AGE_DAYS(default 5) as not-planned, with a written reply on every close (the CONTRIBUTING "no silent closes" promise is unchanged). A PR survives only with a protected label —triage:wildcard-hold,triage:pareto, ortriage:keep— or by being listed intriage_config.json(wildcard holds, parked drafts). The daily triage board is what earns that protection: if a PR deserves review time, it must surface within its window. Closes are reopen-friendly by design; merges stay human.
| Piece | Where | Cadence | What it does |
|---|---|---|---|
tools/triage_bot.py scan |
local / agent sessions | on demand | Classifies every open PR: wildcard hold, draft, conflicts, stale-base, duplicate cluster, tests-included, new arrival, overlap-risk. Writes TRIAGE.md + TRIAGE.json. --verify K runs the worktree merge gate on the top-K land-candidates. |
tools/triage_bot.py approve |
local / agent sessions | on demand | Approves action_required CI runs for open PR heads (was ~190 manual clicks per session). Quota-guarded: --max N. --dry previews. |
tools/triage_bot.py labels |
local / agent sessions | on demand | Ensures the triage:* palette and applies per-PR labels (protects wildcard holds from staleness, flags conflicts/dups). |
tools/triage_bot.py board --post |
.github/workflows/triage-board.yml |
daily 04:30 UTC + manual | Rewrites the Triage Board issue: Pareto NOW list, land-candidates, conflicts, stale bases, duplicate clusters, new arrivals, wildcard holds. |
actions/stale |
.github/workflows/stale.yml |
daily 04:17 UTC | triage:stale label after 21 idle days. Never closes. Exempts triage:wildcard-hold, triage:pareto, triage:keep. |
ci.yml |
every PR | per push | Already runs the full gate suite: census, cargo tests, corpus verdict agreement, differential loop, Go runtime hygiene, examples. |
tools/queue_sweeper.py |
.github/workflows/queue-sweeper.yml |
daily 05:15 UTC + manual | Closes unprotected PRs older than SWEEPER_MIN_AGE_DAYS (default 5) as not-planned, each with a reopen-path reply. Protected: the triage:wildcard-hold / triage:pareto / triage:keep labels plus config-listed wildcard/parked numbers. |
tools/queue_sweeper.py --dry |
any run | before changes | Preview exactly what the next sweep would close — same policy, zero writes. |
Board posting uses board_issue in tools/triage_config.json when configured;
this repository pins #940. The target must remain an open issue with the exact
board marker. Without that setting, discovery requires the exact board title
and marker plus a repository owner/member/collaborator or GitHub Actions author.
An incomplete search or multiple eligible boards stops without writing. With
no eligible board, the existing first-run workflow creates one.
python3 tools/triage_bot.py scan --out /tmp/triage # fresh queue
python3 tools/triage_bot.py approve --max 40 # unblock CIThen read /tmp/triage/TRIAGE.md (or the Triage Board issue — same content,
updated daily). Work the Pareto NOW list top-down; everything else is
pre-classified so recon is only needed where the bot says RECON.
The bot classifies with cheap signals (API mergeability, commit distance, file-set overlap). It cannot detect silent hunk deletion, semantic breakage after clean API merges, or subsumption — those stay on the verify protocol. Treat every bot verdict as a prior, not a decision.
- GitHub native merge queue (free for public repos, needs branch protection + required checks): would replace hand-merging for PRs that pass everything. Deferred — credit messages must stay hand-written.
- GitHub native auto-merge: same blockers as above.
tools/triage_config.json: wildcard holds (M1 experiment, verdicts due
2026-11-01), parked drafts, thresholds, label palette. Keep
wildcard_holds sorted; the loader asserts it.
tools/shepherd.py + the shepherd-* workflows (W0–W7) absorb the
repetitive steps: live board refresh (W1), CAS-guarded rebase (W2), cheap
preflight gates (W3), duplicate-cluster proposals (W4), verify-and-land
mechanics with push-and-assert (W5), stale-premise probe (W6), structured
JSONL session log (W7). Kill switch: the SHEPHERD_MODE repo variable —
unset means dry-run (read/comment/label only); mutations unlock at
live. The two judgment calls stay human by design: stale-premise closes
and real-conflict extraction. Full manual: SHEPHERD.md.
tools/autoland.py + the autoland workflow close the loop for the landing
queue: a fix bundle at automation/land-queue/<slug>/ (meta.json + fix.patch
- repro.hy) is verified end-to-end and landed with no human awake — baseline
ACCEPT → patch → patched REJECT with the expected code → full battery →
commit (
Fixes #N) → push-and-assert LAND-CONFIRMED → bundle consumed → shepherd log. Any gate failure records BLOCKED with evidence and lands nothing. Same kill switch as the shepherd:SHEPHERD_MODE=liveunlocks landing; the defaultdry-runruns baseline checks only. Format, gates, and trust boundaries:automation/land-queue/README.md. The judgment work (authoring the fix + its tests + the corpus registration) stays with the agent sessions; everything mechanical after that is the machine's.
tools/author_stage.py + the author-stage workflow produce a land-queue
bundle from an issue (brain = any OpenAI-compatible endpoint, e.g. an
Ornith-1.0-9B GGUF llama.cpp server on a Kaggle GPU kernel, or a cloud
model). It never pushes code, never lands, never closes: a READY bundle
enters automation/land-queue/ and is gated by the SAME autoland
pipeline; BLOCKED bundles never enter the queue (artifacts only, full
trace). Scope-guarded diffs (compiler/ + tests/ only), accept
controls read from HEAD so a patch cannot weaken its own tripwire, PRs
refused outright. Kill switches stack in front of autoland's:
AUTHOR_ENABLED arms the workflow, AUTHOR_MODE=dry-run (default)
keeps everything to artifacts + evidence comments, SHEPHERD_MODE
(unchanged) still guards the landing. Brain/harness decisions, the
problem taxonomy (63.4% of open bugs = const-eval + bounds), and the
pre-declared bake-off protocol: AUTHOR-STAGE.md.