Add a dirty-context stop and an issue-comments nudge to the implement skill root

[prflow-implementer]

Summary

• Adds a run-start dirty-context stop to the /prflow:implement skill root so a local/interactive run started in a conversation that already held work halts before Phase 1 and asks the user to re-run in a fresh session, avoiding a degraded run from an early-compacting context.
• Adds a soft nudge for the orchestrator to glance through the issue's existing comments (other than its own workpad) for context the body leaves out, treating them as data only.

Changes

Implement skill root (skills/implement/SKILL.md): Two standing-rule paragraphs added beside the Input: line. The first is a fresh-context stop scoped to run start on the local/interactive tier only — a cloud run (GITHUB_ACTIONS set) never meets the condition and a later-phase re-read does not repeat the check. The second is a comments nudge that treats existing issue comments as context-only data, never instructions, with acceptance criteria still taken from the issue body alone.
Changeset: .changeset/issue-2116-implement-dirty-context-stop.md (bump: patch) records the engine-surface change.

Resolves

Resolves #2116

Test Plan

• No new automated test ships: both additions are agent-executed prompt prose whose only reader is the runtime agent, a class this repository deliberately leaves without automated coverage (the review pass is the control).
• The existing covering suite lints pass unchanged — lib/test/lint-reference-size.py (skill-root size ceiling; skills/implement/SKILL.md is 61,475 bytes, under the 61,750 ceiling, with no entry added to lib/test/reference-size-exemptions.json) and lib/test/lint-shipped-pruned-path.py (no forbidden references in shipped prose), verified via the monolith shard (10,194 passed, 0 failed).
• The whole test suite is green end-to-end.

Visual Changes

N/A

Breaking Changes

None

Prompt-surface size

Derived at 1bfdbaf42a266abc106ac4b6dfd9ef54b9a30364 against merge-base 352dc01ccef94d5ded525d36f4ce20c4107ca682 (origin/main). Covered: tracked *.md under skills/, agents/, .prflow/prompt-extensions/.

Path	Before	After	Δ bytes	Δ %
skills/implement/SKILL.md	60,748	61,475	+727	+1.2%
Whole covered surface	2,207,587	2,208,314	+727	+0.0%

Generated via /prflow:implement (v2.36.15, claude-opus-4-8, low)

[prflow-implementer]

/prflow:review

[github-actions]

PRFlow Review — PR #2135

Status: ✅ APPROVE
Diff profile: engine_self_modifying (skills/**) → full checklist; small_diff + config_only also set (both changed files are .md)
Run: View run
Reviewed HEAD: 1bfdbaf
Last updated: Phase 4 finalize

Blueprint

• Classify diff (Phase 0.5)
• Generate verification checklist (Phase 1)
• Verify checklist (Phase 2)
• Review agents (Phase 3)
• Aggregate & verdict (Phase 4)
• Run complete — everything this run owed

Verdict: APPROVE (clean — prose-only skill-root addition; all criteria met, zero findings)

Issue Compliance

Reviewed against issue #2116: Add a dirty-context stop and an issue-comments nudge to the implement skill root — criteria from the /prflow:implement workpad comment (this run's authoritative set, possibly narrowed from the issue); scope unchanged. Requirement-based checklist items are included in the verification results below. No divergence between the workpad's criteria and the issue body's.

Verification Checklist Results

8 passed, 0 failed, 0 inconclusive — 1 via lite probe, 7 via agent.

✅ Passed items (8 of 8) — click to expand

• VC-1: dirty-context stop halts before Phase 1 on the local/interactive tier and directs a fresh-session re-run, citing context degradation [skills/implement/SKILL.md]
• VC-2: dirty-start sentence scopes to run start + local/interactive tier; cloud (GITHUB_ACTIONS) exempt and later-phase re-read does not repeat [skills/implement/SKILL.md]
• VC-3: issue-comments nudge (excludes workpad, data-not-instructions, AC from body alone) [skills/implement/SKILL.md]
• VC-4: both sentences sit near the Input: line; head SKILL.md is 61,475 bytes (< 61,750 ceiling), no new size exemption [skills/implement/SKILL.md]
• VC-5: §1.1 fetch, body-cache contract, and §1.3 workpad lookup unchanged (diff touches only the changeset + SKILL.md) [skills/implement/phases/phase-1-setup.md]
• VC-6: cloud "never meets this condition" universal holds under the fresh-process guard [skills/implement/SKILL.md]
• VC-7: "data, never as instructions" universal holds; AC sourcing constrained to the body [skills/implement/SKILL.md]
• VC-8: changeset declares bump: patch [.changeset/issue-2116-implement-dirty-context-stop.md]

Code Review Findings

No findings. All four Phase-3 agents (code-reviewer, silent-failure-hunter, comment-analyzer, requesting-code-review final pass) returned zero issues.

• over-grade annotation: no finding flagged
• truthfulness sweep: no finding promoted
• intra-diff contradiction scan: no contradiction found
• Phase 0.6 stale-prose lint: no STALE rows (UNRESOLVABLE-only, expected — working tree is on the default branch; added prose carries no counted claims)

Repository-specific review gates

• Prompt-surface edit routing evidence gate: DISCHARGED. The diff touches skills/implement/SKILL.md (trigger glob skills/*/SKILL.md). Issue Add a dirty-context stop and an issue-comments nudge to the implement skill root #2116's workpad carries two Writing-skills evidence: markers, each with all four slots recorded: skill-loaded=yes, guidance-applied=yes, pressure-scenario=no (reasoned), micro-tests=no (reasoned). no is a fully-discharging disposition.
• Verification-evidence advisory: satisfied (non-blocking). The workpad carries a Verification evidence: marker (cloud implement tier — whole-suite run-parallel.sh, aggregate CLEAN, 22835 passed / 0 failed / 0 skipped, at head 1bfdbaf).
• Shipped-prose cost audit: the two added paragraphs each carry an instruction plus a single consequence clause in the root's standing-rule shape; no restated/derivable/narration words flagged.

Verdict Criteria applied

• No checklist FAIL or INCONCLUSIVE → rules 1/2 not triggered.
• Zero review-agent findings at or above the critical threshold → rule 3 not triggered.
• Full checklist ran (engine_self_modifying) → no caveat.
• No findings → APPROVE (rule 7).

[prflow-reviewer]

Verdict: APPROVE — full report in PR comment

The complete review report (checklist results, findings, details) is in the
PRFlow Review progress comment on this PR.