fix(file-issue,file-pullreq,gh-body-check,gh-body-conventions,research,review-pipeline,review-pipeline-coderabbit): discharge a drafted body's evidence claims against the record

[ultimatile]

Summary

A drafted GitHub body could state that verification was performed with nothing checking the statement against a record of what actually ran. Matching totals do not catch it: a fabricated entry and an omitted one cancel, so an enumeration can be wrong while its count is right.

gh-body-conventions gains an ## Evidence claims section stating what makes a statement one, when a record counts as available, that claims map to records entry by entry rather than by count, that an enumeration need not be complete unless it says it is, that a claim with no record is re-run or dropped rather than hedged, and that text edited after its claims were discharged is discharged again.

Settling a body's statements against that record is called discharging them. The rule is then threaded through the surfaces that draft GitHub text, and gh-body-check records that it does not implement it — the check's subagent is deliberately denied the drafting context this rule compares against, so it cannot.

Closes #138

Changes

• skills/gh-body-conventions/SKILL.md — the new ## Evidence claims section, after ## Exclusions.
• skills/file-pullreq/SKILL.md, skills/file-issue/SKILL.md — a discharge step between drafting and the laundering pass — the existing step that runs gh-body-check — which renumbers the steps after it; both revise-and-re-show loops return through it.
• skills/research/SKILL.md — evidence-claim added to the rule families imported from the conventions; the plan is discharged before it is presented for approval rather than after, and a body derived from the plan rather than posted verbatim is discharged and laundered on its own before filing.
• skills/review-pipeline/SKILL.md, skills/review-pipeline-coderabbit/SKILL.md — the delta section appended to an open PR's body is discharged, and a completeness claim already in that body is re-checked with it; the fix-loop description update re-discharges and re-checks before posting.
• skills/gh-body-check/SKILL.md — its revise loop returns through the discharge, and its scope statement names the rule it does not implement.
• .claude-plugin/marketplace.json — version bump.

Impact

file-pullreq and file-issue renumber their procedure steps 3–6 to 4–7 (5a / 5b become 6a / 6b in file-pullreq). The only references to those numbers from outside the two files are two in research, pointing at file-issue's File step; both were updated.

The rule reaches a body wherever a skill applies the conventions to it, so it binds file-issue and file-pullreq directly, the appended delta section in both pipelines, and — through the widened import — the plan research posts.

Test plan

Sixteen fixtures, each exercising a distinct branch of the rule. Each was handed to a fresh-context agent together with the shipped text of the new file-pullreq step and the ## Evidence claims and ## Exclusions sections, and nothing else. Each agent reported, for every statement in its fixture, the record it mapped that statement to or that it had none, and what it then did with the body; that report was compared against what the rule specifies. All sixteen reached the specified disposition. The fixtures are synthetic scaffolding rather than a surface this repository carries, so they are not committed — this section is their record.

One fixture in full, so the shape is on the record. Its draft body:

Summary

Adds a migration for the new index.

Test plan

pytest -q passes. The migration was also applied and rolled back against a copy of the production schema, with row counts identical before and after.

Its stated record: one command was run and observed, pytest -q reporting 88 passed and 0 failed; the migration apply-and-rollback was performed in an earlier session that has ended, leaving only a note reading "Applied and rolled back the migration against a production-schema copy; row counts matched", with no log, linked run, or committed artifact. The rule requires the migration sentence dropped rather than re-attributed to the earlier session, and the surviving claim restated to name its command and observed result. The agent reached that disposition, and its report named two rewrites it refused — "was previously verified", and attributing the work to the earlier session — as the softening the rule forbids. Neither phrase is in the draft above; both are rewrites the agent considered and rejected.

The other fifteen have the same shape. All sixteen branches, and the disposition each fixture had to reach — the twelfth is the one shown in full above:

• a fabricated entry and an omitted record leaving the totals equal — flag the unbacked entry rather than pass on the count
• an enumeration that omits a record and claims no completeness — pass unedited
• three entries framed as three runs when two runs produced them, one return having reported two results — keep all three entries, correct the count
• a record lost to a summarized stretch with nothing durable holding it — drop, and do not hedge
• an artifact's existence and static properties, plus a report of having run a check on them — drop the report, keep the properties
• a completeness claim falsified by work done after the body was discharged — complete the enumeration
• a body edited after discharge to clear that check, the edit introducing a claim — discharge again and drop it
• a command-backed claim naming neither command nor result, beside a claim for a command never run — restate the first, drop the second
• a record that outlived the drafting context in a committed report and a linked run — keep the claim and cite them
• one run described a second time outside the enumeration — not double-counting
• a claim whose only backing cannot be named under ## Exclusions — drop, since no nameable restatement exists
• work performed in an earlier session that ended, with only a handoff note surviving — drop
• two entries each backed by a different result inside a single return — keep both, the count of runs being right already
• a property knowable by reading or by running, phrased as a coverage run's output — restate it against the code
• inspection reported as having confirmed a code-readable property — not an evidence claim
• bare runtime properties asserting no verification at all — a coverage figure, a suite's pass state, a timing — each caught

Also run: the mdformat hook configured in .pre-commit-config.yaml, against the changed files, reporting Passed with no reformatting; skills/gh-body-check/test-body-math-scan.sh, reporting All checks passed before and after the change; skills/coderabbit-review/scripts/test-completion-state.sh and skills/coderabbit-review/scripts/test-review-skip-reason.sh, each reporting All checks passed; and rg -n 'step [0-9]' skills/, whose surviving hits were each read against the step they name.

Notes

Two posting surfaces the rule reaches are left without a discharge point, and one more is adjacent. skills/review-pipeline/SKILL.md's Phase 4b posts a closing comment on the sub-issue and edits the umbrella issue body; skills/file-issue/SKILL.md's umbrella sub-issue variant edits the umbrella's Phases table when a sub-issue is filed; and both pipelines reply to review comments inline, which are not issue or PR bodies and so fall outside what the conventions currently govern. Covering them means deciding which posting points carry a gate, which this change does not settle; they are tracked in #142.

The rule guarantees that a claim was answerable to a record its author held, not that a later reader can re-derive it.

@coderabbitai ignore

[Copilot]

Pull request overview

Adds an explicit “discharge evidence claims” gate to ensure drafted GitHub bodies’ verification statements are checked against the drafting-session record (per new gh-body-conventions guidance), and threads that gate through the primary drafting and PR-update pipelines.

Changes:

• Introduces gh-body-conventions “## Evidence claims” rules and clarifies that discharging must happen in main context (not in gh-body-check’s subagent).
• Inserts a discharge step before laundering (/gh-body-check) across PR/issue drafting (file-pullreq, file-issue) and plan posting (research), including re-discharge behavior in revise loops.
• Updates review pipelines to discharge evidence claims for appended PR-body deltas and to re-check completeness claims when appending/editing bodies; bumps marketplace version.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
skills/review-pipeline/SKILL.md	Adds discharge requirements to gate-mode PR drafting and to PR-body delta append/edit flow.
skills/review-pipeline-coderabbit/SKILL.md	Mirrors discharge requirements for the CodeRabbit-based pipeline.
skills/research/SKILL.md	Discharges plan evidence claims before approval/posting and adds discharge+launder requirements for derived bodies.
skills/gh-body-conventions/SKILL.md	Defines what counts as an evidence claim, what records are admissible, and when re-discharge is required.
skills/gh-body-check/SKILL.md	Updates revise-loop wording and explicitly states evidence-claim discharge is out of scope for this check.
skills/file-pullreq/SKILL.md	Inserts a dedicated discharge step before laundering and threads re-discharge through revise loops; renumbers steps.
skills/file-issue/SKILL.md	Inserts a dedicated discharge step before laundering and threads re-discharge through revise loops; renumbers steps.
.claude-plugin/marketplace.json	Bumps plugin CalVer patch version.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.