The body of _process.yml is read from dev (@dev, #623). After "Перейти на
ветку задачи" the working copy of job prepare is the task branch, and a
show/ship branch with a clean merge is not rebased before review: its
scripts/ may lag dev or be replaced. #749 fixed job integrate; prepare
still ran four control scripts from the material — the issue-body digest
for the anchor, --reuse of a green verdict (#499), validate-gate (#510)
and the spec-change check (#517). The material decided its own admission:
a branch whose review-doc-guard.mjs prints reuse=true merges without the
model. model_review took model-usage.mjs from the material too: a lagging
branch has none, and the publication silently wrote reason=missing.
Now prepare extracts one snapshot right after setup-node, before the
branch switch: `git rev-parse origin/dev` once and `git archive <sha>
scripts .github/workflows/validate.yml`, so the git fetch of the track and
rebase steps cannot mix versions. Every repo script of the job runs from
it via TOOLS — the track step and the rebase guard lose their own
extractions. The SHA goes out as job output tools_sha; the usage step of
model_review archives the same commit inside itself, so a snapshot failure
is a failure of the reporting step (continue-on-error), not of the stage.
The model session runs on that runner, so the usage line stays untrusted
input parsed strictly (#556, #737).
withMaterialAnchors is idempotent: a repeated call drops the separator the
previous call wrote instead of piling up `---` lines.
Tests: test/process-prepare-tools.test.mjs — the job contract (no step
calls scripts/ from the working copy, one pinned archive before the branch
switch, tools_sha reaches model_review) and the steps as they are, on real
bash and git: a branch behind dev without model-usage.mjs and with a
substituted review-doc-guard.mjs; the anchor digest, reuse and the spec
check come from dev, the usage line is data even after dev moved. Red on
the old workflow: all four. Harnesses of process-track, rebase-generated,
review-doc-guard and model-usage take the prepare snapshot. Three registry
mutants (reuse from the material, usage from moving dev, separators).
Canon: PROCESS.md §10.4 «Скрипты конвейера — из dev» covers prepare and
the usage step; «Расход модели» names it a pipeline step, not the reviewer's.
Issue: #765
User-Visible: no
The weekly process metrics weigh tracks and the nightly ship review in the
order quality, speed, tokens (#707), but the third axis had no source: the
claude-code-action step hides usage from the Actions log on purpose, nothing
read its execution_file, and the #728 reader printed "no data" every week.
scripts/model-usage.mjs is the single module that builds and parses the line:
`<!-- hp:usage input_tokens=N output_tokens=N cache_creation_input_tokens=N
cache_read_input_tokens=N num_turns=N -->` (sums over every model in the last
`result` message, `result.usage` when modelUsage is absent) or
`<!-- hp:usage-none reason=<code> -->`. Only the result message is read; the
rest of the file holds tool results, and no byte of it is printed.
A new step right after Review in both model_review jobs (always(),
continue-on-error) hands the line out as the job output `usage`. Usage is a
reporting figure like the stage duration, so it travels as a job output and
not through the sealed artifact: REQUIRED_FILES and the #556 gate are
unchanged. Publication treats the line as untrusted input and writes the
normalized form as the last line of the anchor block (review-doc-guard
--anchor --usage=) or right after the SHIP-REVIEW block; empty becomes
reason=missing, anything off-format reason=invalid.
The #728 reader now takes the line only from the machine block: a reviewer
quoting the previous round in prose no longer doubles its usage, and
"no data" is counted as missing, never as zero. PROCESS.md §10.4 documents
the source, the format and why it is a job output.
Issue: #737
User-Visible: no
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018qZfe7YS4rqEMKoVeS3GKd
claude[bot]