feat(ship-check): port output-honesty contracts from plan-review

[aliasunder]

What

Ports two output-honesty contracts from the plan-check plugin's plan-review skill back into all four ship-check phase skills (pr-review, code-quality, test-audit, bug-check) and their agents:

• Verification-basis line — every phase summary (default and comment mode, including the posted review body) names the PR head SHA actually reviewed. A review that doesn't say what it checked is indistinguishable from one that checked nothing. This is also the SHA pr-monitor's pre-merge delta review keys off, so phases now report it explicitly instead of it being reconstructed.
• Proof-of-dismissal closing — each summary closes with one line per suspicion seriously considered and dropped, with the reason (or "none"). Extends bug-check's "No silent skipping" (confirmed findings must be reported) into the negative space: cleared suspicions must be reported too, so "no findings" is distinguishable from "didn't look."

What was deliberately NOT ported

plan-review's verification-state pinning (review against the repo as of the plan's date, not HEAD) doesn't transfer: ship-check reviews a PR head, which is already a pinned state. Only the output half — stating the SHA — carries over.

Changes

• 4 phase skills: "Output honesty (both modes)" block in the report section; Reviewed at <HEAD_SHA> added to comment-mode review body templates; bug-check's inline summary template gains both lines
• 4 agents: Reviewed at and Dismissed lines added to default- and comment-mode output templates

🤖 Generated with Claude Code

[umm-actually]

umm-actually re-reviewed at 755eee9

1 new finding(s) posted (4 tracked finding(s) across all runs).

Context notes

• Priority docs already in context: plugins/ship-check/agents/bug-checker.md, plugins/ship-check/agents/code-quality-reviewer.md, plugins/ship-check/agents/pr-reviewer.md, plugins/ship-check/agents/test-auditor.md

---

umm-actually · deepseek/deepseek-v4-flash-0731

[umm-actually]

Add Reviewed-at and Dismissed to the pipeline's own status report
Low severity · correctness · low confidence

plugins/ship-check/skills/ship-check/SKILL.md:890 — beyond the diff's line ranges, in code the changes touch or depend on.

Pre-existing analog gap: the PR establishes a reviewed-SHA plus proof-of-dismissal contract for the phase skills, but the pipeline's final status snapshots (both default and comment mode) do not emit Reviewed at <SHA> or a Dismissed line. A consumer reading only the final report still cannot distinguish a reviewed PR from an unexamined one.

Failure scenario: A user reads only the final Ship check complete status snapshot as instructed. Even though each phase reported its SHA and dismissals to the chat stream, the snapshot itself shows neither, so the reader sees a concluding verdict while the underlying high-water and inspection claims are absent/incomplete at the point it's actually surfaced.

Suggested fix

Add `- Reviewed at: <SHA>` and `- Dismissed: …` to the ship-check output templates' Reporting section, mirroring the phase reports in this PR, so the pipeline conclusion carries the same two-line contract as the phases it summarizes.

---

umm-actually · deepseek/deepseek-v4-flash-0731

[umm-actually]

Remove spliced-in monitoring-loop content from the Phase 2 code-quality reviewer
Medium severity · correctness · medium confidence

plugins/ship-check/agents/code-quality-reviewer.md:100 — beyond the diff's line ranges, in code the changes touch or depend on.

The Phase 2 code-quality reviewer agent (code-quality-reviewer.md) has a large block of pr-monitor monitoring content spliced into its body, sitting between the orientation step (Load sequential thinking) and the Output Format section (new-file lines ~90–118, outside this PR's hunks). The agent's true job is a one-shot convention pass that loads vault standards, fixes findings, runs tests, commits, and summarizes — but the embedded block instructs it to run a continuous PR-monitoring loop: re-fetch threads and issue comments until all bot/Claude/human threads resolve, call ScheduleWakeup with delaySeconds: 240, “continue monitoring after reporting”, and “never auto-terminate — not after clean checks, not after silence.” That check-list of monitoring preconditions and the never-stop directive are pr-monitor content that directly contradicts the stated Core Responsibilities (“5. Run tests, commit, and push fixes”) and the default-mode report template. Any dispatch of this agent would load conflicting instructions about whether the task ends at the summary or continues indefinitely, and the monitoring checklist (issue-coverage counts, “If code was pushed during this run…”) references machinery the code-quality phase has no reason to invoke.

Failure scenario: A user dispatches the code-quality reviewer for ship-check Phase 2 (or standalone). The agent reads its own body, reaches the “Continuous monitoring” block after the orientation, and follows the pr-monitor directives — scheduling 240-second wakeups, re-fetching the PR, scrubbing threads, and refusing to terminate “for any other reason” than an explicit user “stop” or merge/close. Instead of completing a finite code-quality pass (fix findings, run tests, commit, report), it enters an indefinite monitoring loop on a PR it was only asked to review, posting or re-evaluating comments. Milder case: the contradictory instructions just garble the agent's behavior at the boundary between “summarize” and “keep monitoring,” so the phase either hangs until the user notices or produces PR-side activity the caller never asked for.

Suggested fix

Move the pr-monitor block (thread-resolution prerequisites, issue-coverage checks, ScheduleWakeup / never-auto-terminate loop, “pipeline context” notes about the delta review) out of the code-quality reviewer into the actual monitoring agent where it belongs, and restore the code-quality reviewer's body to end with its own monitor-free instruction — i.e., after the orientation section, jump straight to the Phase-2 procedure and Output Format, with no reference to ScheduleWakeup or the “never auto-terminate” rule.

---

umm-actually · deepseek/deepseek-v4-flash-0731

[aliasunder]

Re umm-actually finding on code-quality-reviewer.md (comment 5547436971, "spliced-in monitoring-loop content"): false positive — the described content does not exist in the file. Lines 90–118 of code-quality-reviewer.md contain the sequential-thinking triggers, Fixing and Committing, Attribution, and Comment-mode sections; a grep of the full 154-line file for ScheduleWakeup, monitoring, auto-terminate, wakeup, and issue-comment coverage returns zero matches on this branch. The quoted directives ("delaySeconds: 240", "never auto-terminate") are pr-monitor SKILL.md's own content — the finding appears to have conflated the two files.

---

🔍 ship-check · pr-monitor · claude-fable-5

[aliasunder]

Re umm-actually finding 5547367668 (Reviewed-at/Dismissed missing from the pipeline's own status snapshots): valid pre-existing analog gap — it landed just before this PR merged, so it's resolved in follow-up #10: both orchestrator reporting templates now carry the two-line contract (default mode reports the delta-review high-water SHA, comment mode the reviewed head, plus a Dismissed roll-up of the phases' dismissal lines).

---

🔍 ship-check · pr-monitor · claude-fable-5