On track:ask every spec review round is 10-45 minutes of waiting, and
rule #1 kept the author idle for all of it. Rule 10 (#738) judges a
class A commit by its author date, so code written in the
S4-spec-review epoch was always refused: nothing told a draft written
against the reviewed text from a violation. The owner allowed changing
rule #1 for this (decision 2026-10-01).
A draft commit carries `Spec-Draft: sha256:<issueBodyDigest(body)>`,
the hash the pipeline already writes as "Тело issue" into the review
document anchor. Rule 10 accepts a class A commit written in S4 only
when its S4 epoch (a repeated S4 does not restart it) was closed by
S5-ready, the track at the author date was ask, and the trailer equals
the body of the green, High 0 SPEC-REVIEW added inside that epoch, read
from the range head or origin/dev. The first failing check is the one
finding: trailer format, track, how the epoch ended, the missing
document with a `git fetch origin dev` hint, or both hashes and the
document name. A trailer on a commit written in an allowed epoch is a
warn. Without the document reader rule 10 is exactly #738; main always
passes one, and it reads git only when the range holds a draft.
The task packet tells S4 on ask that a local draft is allowed while the
branch stays closed, prints the trailer line, and in S5/S6 names the
green spec review, whether the body changed since, and the --report
check before push. SPEC-REVIEW documents are a separate input
(specDocs) from the branch and origin/dev, so the previous verdict and
the AC witness keep their source.
PROCESS.md gets §11.8 and the points that refer to it (§1, §2.4-2.6,
§3 item 1, §7.2, §9, §10.2 item 10, §12); AUTHOR.md, REVIEWER.md and
AGENTS.md follow, with three new key rules in process-digests. The
pre-push hook fixture copies the modules process-gate now imports.
Issue: #729
User-Visible: no
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018qZfe7YS4rqEMKoVeS3GKd
Ship tasks merge without a model review and their code was first read by
the batch review right before a beta: one session over the whole range,
ten to forty-five minutes on the release path, days after the merge. The
gate also knew a single document (SHIP-REVIEW-<tag>.md) and covered tasks
by number only, so a commit that landed after the review under the same
trailer still counted as read.
- scripts/ship-review.mjs: the patch set of a task is the sorted
`git patch-id --stable` of its range commits, without `Release:`
commits (the beta candidate carries every Issue: of the line) and
commits touching only docs/reviews/**; the diff options are explicit
so a local git config cannot change it. shipCoverage rates every ship
task from the documents of the same base (candidate and origin/dev,
latest publication wins): clean, high, stale, none; documents without
`patches` cover by number. `tag=nightly` is a reserved mode: the
candidate is required, the document is
SHIP-REVIEW-<base>-dev-<sha12>.md, only none/stale tasks are read and
nothing runs when nothing is uncovered. The beta reads the same delta
(force=true reads everything, as before); the brief names what the
night already read. The gate refuses none/stale with the command and
keeps the High refusal with force=true; all clean passes without a tag
document. The machine block gains `mode` and `patches` at its end.
comment-high writes one line per task of a nightly document with High,
once per document (hp:ship-review-high).
- _ship-review.yml: prepare refuses nightly without a candidate before
defaulting to the dev tip, computes the document from base and SHA and
no longer reads a prepare failure behind `| tee` as "no ship tasks";
publish takes mode and patches from prepare, never from the model
result; a new step comments High at night with HP_PROCESS_TOKEN.
- _nightly.yml: the Validate run SHA is a separate step output before
the wait; a new job dispatches ship-review.yml -f tag=nightly on it
whatever Validate's outcome, waits only for the run to appear and
never colours the night. Thin files in main are unchanged.
- reviews-index/reviews-archive: the nightly name is a ship document
with nightly: true; a beta base archives with its line, a stable base
with the nearest archived line newer than the base, or stays.
- PROCESS.md §11.7, §10.4 and REVIEWER.md describe the nightly mode,
patch set, coverage and beta delta; the digest test pins the key rule.
Tests run the prepare, publish and comment steps and the nightly steps
on real bash with real git in temporary repositories; only push
transport and gh are faked.
Issue: #727
User-Visible: no
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018qZfe7YS4rqEMKoVeS3GKd
A non-green show verdict that found "something to decide" went down the same
path as "fix the code": S6 with a limit of 2. Promoting the task to track:ask
was left to the agent's memory, with no named criterion and no trace, and the
exhausted budget only surfaced on the next S7 - after a fix nobody would read.
The structured verdict now carries `route` (fix | reclassify) and an optional
`criterion` (one of the six show criteria of PROCESS.md section 5). The trust
boundary reads a missing route as fix, rejects one outside the dictionary and
rejects reclassify on a green verdict. `reviewRoute` in process-track.mjs is
the single decision: on a code review of an unconfirmed show it moves the task
to track:ask and S3-spec; on an owner-confirmed show it adds `blocked` and asks
the owner; anywhere else reclassify degrades to fix with a note. The verdict
that spends the last cycle sets review-4 at once; the stage budget is shared
across tracks, so promotion changes the limit (4), not the count.
The "Решение по вердикту" step makes one `process-track.mjs route` call (from
dev, like the track step) and only executes its output: comment from a file,
labels from add/remove lists, status via status-label.mjs as before. The track
step also emits `confirmed` and a `route_note` for the review prompt; the
review document anchor gains a route tail that the old reader still parses;
wait-verdict reports the two new pipeline comments. The guard's own
spent >= limit check stays as the safety net.
Issue: #726
User-Visible: no
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018qZfe7YS4rqEMKoVeS3GKd
The ship limits count lines and files but not what was touched: a
12-line pointerdown handler passed them like a typo and merged unread.
The track rule also lived twice - the guard computed the cycle limit in
bash while process-track.mjs computed the track, and the two disagreed
on multiple track labels. The packet still told authors to rebase
show/ship branches that merge cleanly.
- scripts/change-risk.mjs: one pure classifier over `git diff -U0` from
the merge base. Class A lines only; comments, blank lines and pure
renames give no risk; deletions do. Area and token rules per class
(geometry, touch, migration, devices, perf, ux, visual render/ui),
evidence as path:line, five per class.
- process-track.mjs: owner confirmation is a comment line
"Трек: <x> — решение владельца" by the repo owner (latest wins, only
for the current track); several track labels read as the strictest
with a warning; cycleLimit, guardLimit and rebaseBeforeReview are the
single source. `stage` makes the whole S7 track decision in one call:
ship with risk and no confirmation is raised to show with evidence,
a confirmed ship keeps merging without the model and records the risk
for the batch review; show/ask get a risk note for the reviewer.
- _process.yml: the guard asks process-track.mjs for the limit and keeps
no track logic; the track step calls the script once and only
executes its raise flag and comment file; risk_note reaches the
Review prompt, ship_risk reaches the hp:ship-merge comment (marker
line unchanged).
- task-packet.mjs: track basis, limit and rebase policy; next step
without the stale rebase line; risk with its consequence per track;
required checks with reasons (ci:golden only on render risk);
changelog and visual evidence - from the same exports.
- ship-review.mjs: the batch brief prints the risk line of a ship merge.
- Canon: PROCESS.md §5, §5.1, §10.4, §11.7, both digests, AGENTS.md.
- Registry anchors that watched the moved code are moved, not dropped.
Issue: #707
User-Visible: no
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018qZfe7YS4rqEMKoVeS3GKd
Owner's decision 2026-09-29: mutants check the tests, not the product.
During development they are not run at all — not locally, not in CI,
not by the reviewer. The whole registry is the nightly run
(mutation-gate.yml, #513); a survivor files an issue (#472). The #693
post-mortem: 36 of 57 minutes of a one-line fix went to optional work.
- process-track.mjs: `mutants` is always false (no track, no label).
- classify-changes.mjs: Validate requests no diff mutants on any event;
the `mutants` input stays so old `-f mutants=…` calls do not fail.
- _process.yml: the default for the gate and the merge is false.
- pre-push-gate.mjs: the manual run no longer runs mutants.
- Canon: PROCESS §2.7 (a mutant is written, not run; `--check` keeps the
anchors), §5.1 (`ci:mutants` retired), §8 (ship/show: nothing beyond
gate:small and the spec — one proof per item, no `--smokes` on ship,
a stray flake is an issue, not an investigation), §10.4; AUTHOR,
REVIEWER, AGENTS, TESTING.
- Registry: four mutants of the old request rules replaced by
dev-mutants-requested-again and track-pays-for-mutants-again.
Issue: #709
User-Visible: no
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018qZfe7YS4rqEMKoVeS3GKd
Сверка PROCESS.md, ролевых выжимок, AGENTS.md, TESTING.md, CONTRIBUTING.md
и скриптов по 26 найденным расхождениям (D1–D26): трейлеры по классам
изменений, gate:small как единственный источник состава, пороги ревью,
путь реестра мутантов, golden по ci:golden, порядок чтения промпта ревью.
- scripts/change-classes.mjs: классы A/B/C/D — один модуль для
process-gate и проверки трейлеров.
- commit-msg: коммит только с файлами класса C (документация) трейлеров
не требует; указанные трейлеры по-прежнему проверяются.
- Маршрут автора без docs/STATUS.md: 5345 → 4703 слова.
- Промпт ревью читает SCOPE → AGENTS → REVIEWER, как ROUTES.reviewer.
Issue: #701
User-Visible: no
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018qZfe7YS4rqEMKoVeS3GKd
The analysis of 85 closed tasks #600-#691 showed that the light track
cost as much as the full one (115 min and 12 events vs 102 and 13) and
that the owner had no label to choose the route. The owner accepted the
proposal on 2026-09-28.
- PROCESS §5 is the track table: track:ship (S1 -> S5, one line under
"## ТЗ", <= 30 src lines, batch review before the beta), track:show
(default, S2 -> S5, up to three AC, no spec review, 2 code cycles),
track:ask (full route). The owner's label beats the criteria, which
become a hint; any agent may raise a track, only the owner lowers it.
- §5.1: ci:full / ci:golden / ci:mutants order heavy checks on any track;
small and trivial read as track:show, no label as track:ask, an
infrastructure task as track:show.
- §2, §2.2, §2.4, §2.5, §4, §7.1, §7.2, §9, §11 follow; AUTHOR/REVIEWER
digests and AGENTS.md follow with the digest test and its mutants.
- task-packet.mjs reports the track via trackFromLabels(); the pipeline
reads track:show/track:ship for the cycle limit of 2 and lets an
explicit track:ask win. Pipeline behaviour by track is #696.
Issue: #695
User-Visible: no
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018qZfe7YS4rqEMKoVeS3GKd
Six workflows run from the default branch (issues, schedule, workflow_run):
process, process-resume, process-reconcile, mutation-gate, nightly,
process-metrics. Their bodies move to _<name>.yml (on: workflow_call); the
original files keep only triggers, run-name, permissions, concurrency and one
job `uses: Matysh/houseplan-card/.github/workflows/_<name>.yml@dev` with
`secrets: inherit`. A pipeline change becomes one commit to dev.
- caller job permissions = union of body job permissions (#556 minimum kept
per job inside the body); caller `if` repeats the body guard for process and
process-resume so unrelated events stay skipped;
- dispatch inputs forwarded via workflow_call inputs of the same names;
- _mutation-gate.yml keys evidence/marker on job.workflow_sha (the body SHA):
in a called workflow github.workflow_sha belongs to the caller in main;
- action-pins: narrow exception for this repo's _*.yml at @dev with a reason;
- preflight workflow_sync compares all six thin callers (was 3 of 6);
performance.yml excluded: its schedule judges main with main's own body;
- tests read bodies from _*.yml; new test/default-branch-workflows.test.mjs;
six mutants; PROCESS.md §10.4, AGENTS.md, REVIEWER.md updated.
Issue: #623
User-Visible: no
Вход агента до первого файла кода стоил ≈ 26 700 слов (аудит 22.09).
- docs/process/AUTHOR.md и REVIEWER.md — выжимки PROCESS.md: каждый пункт
ссылается на раздел канона, ключевые формулировки дословные;
test/process-digests.test.mjs сверяет якоря, ссылки и правила.
- scripts/entry-cost.mjs — маршрут чтения по роли и бюджет (автор ≤ 12 000
слов, AC1); AGENTS.md «Read this first» называет те же маршруты.
- docs/STATUS.md: блок Snapshot генерирует scripts/status-snapshot.mjs
(версии — release-contract, счётчики — inventory, теги — git); feature
surface и ранние milestones перенесены дословно в docs/STATUS-FEATURES.md.
- docs/TESTING.md — действующая инструкция (684 строки, AC3); ручные
чек-листы и приложения по issue перенесены дословно в docs/testing-notes/
с индексом и тестом на полноту.
- Промпт ревьюера в process.yml читает конспект вместо пересказа правил;
машинные требования (строка вердикта, REVIEW_DOC, запрет fetch, таблица
«чем краснеет», разделы повторного раунда) сохранены и закреплены тестом.
- PROCESS.md: правила не менялись; добавлены ссылка на конспекты в шапке и
уточнение в §10.4, что ревьюер конвейера читает конспект.
- 7 мутантов в реестре.
Issue: #634
User-Visible: no
Claude