[CI] Coordinator revision/remediation re-invocations omit baseline context — prompt-integrity fires on every cycle

[akaszubski]

Problem

When re-invoking an agent for revision (plan-critic returns REVISE) or remediation (reviewer returns BLOCKING), the coordinator constructs the re-invocation prompt as only the new feedback, without the full baseline agent context. The prompt-integrity hook detects this as shrinkage and fires, requiring the coordinator to manually reconstruct the full context before retrying.

This is a distinct failure mode from #1002 (which covers research-skip pipelines where the planner template itself is not reloaded). The revision/remediation case is: agent was invoked correctly the first time (full context), plan-critic/reviewer returned feedback, and the coordinator passed ONLY the feedback to the second invocation instead of baseline + feedback.

Evidence

Issue #991 pipeline run (2026-05-24):

• STEP 5.5 plan-critic returned REVISE with ADR-1 falsified. Coordinator re-invoked planner with 661-word prompt vs 1537-word baseline → prompt-integrity fired
• STEP 11 reviewer BLOCKING fix: re-invoked implementer with 680-word prompt vs 1380-word baseline → prompt-integrity fired
• Both required manual context reconstruction before the re-invocation succeeded

Pattern: every revision/remediation cycle costs 1 extra agent invocation (the blocked attempt) plus coordinator overhead to reconstruct.

Suggested Fix

Add a helper function to plugins/autonomous-dev/lib/prompt_integrity.py:

def construct_revision_prompt(agent_type: str, baseline_context: str, feedback: str) -> str:
    \"\"\"Construct a revision/remediation prompt that passes prompt-integrity checks.
    
    Combines the full baseline context with the feedback suffix, ensuring the
    resulting prompt meets the minimum word count threshold the prompt-integrity
    hook measures against.
    \"\"\"
    return f\"{baseline_context}\\n\\n## REVISION FEEDBACK\\n{feedback}\"

The coordinator implement.md STEP 5.5b (plan revision) and STEP 11 (remediation cycle) should document that re-invocations MUST pass baseline_context + feedback, not just feedback. A COORDINATOR FORBIDDEN note: "Passing only feedback to revision/remediation re-invocations — always include full agent baseline context."

This is analogous to the existing pattern in STEP 4 (research skip) where the coordinator must explicitly reconstruct the planner template.

Repo: autonomous-dev
Session: 2026-05-24T01:17:32
Related: #1002 (research-skip shrinkage — different trigger, same hook)

Plugin Version: 3.50.0 (fcfec4b)

---

Filed automatically by continuous-improvement-analyst

[akaszubski]

Recurrence: 2026-05-28, commit 6d8c97f37 (Spektiv intent_watchdog docker-py refactor)

Pattern: Security-auditor re-invocation (second call, post-first-success) had prompt shrink from 510 words to 308 words (39.6%) — blocked. Coordinator expanded to 677 words for third invocation, which succeeded (998 words, PASS).

Correction to pipeline summary: The session description characterized the first SA invocation as 'output swallowing — returned only Relevant file: line.' Activity log shows this is incorrect. First invocation (11:21:37) completed successfully with 957 words. The blocked attempt (11:24:09) was the coordinator's re-invocation with a compressed prompt, which the prompt-integrity hook correctly denied — the denial looks like a minimal response from the coordinator's perspective. Root cause is this issue, not a serialization bug.

Sequence (from .claude/logs/activity/2026-05-28.jsonl):

• 11:21:37 — coordinator invokes SA (first invocation)
• 11:23:49 — SA SubagentStop, 957 words, success=True, dur=132651ms
• 11:24:09 — coordinator re-invokes SA, prompt-integrity DENY (510→308 words, 39.6% shrink)
• 11:24:38 — coordinator re-invokes SA with expanded prompt (677 words), allowed
• 11:26:12 — SA SubagentStop, 998 words, success=True, dur=123692ms

Cost: 1 wasted SA invocation attempt + ~2 min pipeline delay. Total SA time: ~4:35 instead of ~2:15.

This is the 4th documented occurrence of this specific pattern (revision/post-success-re-invocation with compressed coordinator context). Suggests the coordinator's post-success-re-invocation path (e.g., asking SA to scope-narrow or clarify) has the same baseline-omission bug as the plan-critic REVISE path.

[akaszubski]

Post-fix behavioral replay — PASS

Per the standing rule that unit tests prove the code runs but not that behavior changed in real use, ran a live replay of the original failure modes against prompt_integrity.py at 8c5f751+.

Artifact: scripts/replay_1116_revision_prompt.py

Check 1 — helper contract

construct_revision_prompt(agent_type, baseline_context, feedback) exists, importable from prompt_integrity, returns baseline + "## REVISION FEEDBACK" marker + feedback. ✓

Check 2 — planner / plan-critic REVISE (baseline=1537 words, feedback=661)

Exact numbers from this issue's body.

Path	prompt size	passed	shrinkage
Pre-fix (prompt = feedback only)	661 words	False	57.0%
Post-fix (prompt = construct_revision_prompt(...))	2201 words	True	-43.2%

Pre-fix BLOCK reason: Prompt for planner shrank 57.0% from baseline (1537 -> 661 words, threshold: 30% [relaxed from 15% for remediation]). Matches the issue body verbatim.

Check 3 — implementer / reviewer BLOCKING (baseline=1380, feedback=680)

Path	prompt size	passed	shrinkage
Pre-fix (prompt = feedback only)	680 words	False	50.7%
Post-fix (prompt = construct_revision_prompt(...))	2063 words	True	-49.5%

Notes

• Used the most permissive threshold (invocation_context="remediation" → 30% relaxed from 15%). Pre-fix BLOCKs anyway.
• Post-fix shrinkage is negative because the helper concatenates baseline + feedback, always growing the prompt.

Verdict

PASS — pre-fix path reproduces the original BLOCK on the exact numbers from the issue body. Post-fix path defeats the shrinkage detector via the prevention primitive. The fix is active and effective.

Replayed: 2026-06-01.