review-last-run-plan

name: review-last-run-plan description: Review last-run artifacts and build a user-approvable, editable fix plan in plan mode. Use when the user asks to analyze lastRun.txt, validate failures across runs, confirm bug hypotheses, and produce a phased plan block before implementation.

Review Last Run Plan

Overview

Analyze run logs and output artifacts, then produce a decision-complete remediation plan that is explicitly presented for user approval and iteration. Keep this skill planning-focused; do not implement code changes while using it.

Workflow

1. Gather context from:

• lastRun.txt (required)
• Current generated outputs (fib50.txt, pattern_report.txt, users_sorted.txt, analysis_results.txt) when present
• Current orchestration sources if needed for root-cause mapping: src/agentic_workflows/orchestration/langgraph/graph.py, mission_parser.py, mission_tracker.py, mission_auditor.py

2. Identify each run by RUN START run_id=... and extract:

• mission count
• step count
• timeout/fallback events
• planner outputs (MODEL OUTPUT, PLAN QUEUED, PLAN QUEUE POP, PLANNED ACTION)
• mission reports
• audit summary/findings
• cache/memo events

3. Evaluate and classify:

• mission execution correctness (PASS/WARN/FAIL)
• mission attribution correctness
• finish-claim correctness
• cache/memo correctness (including poisoning risk)

4. Validate explicit user hypotheses (example):

• first run failed due to planner timeout vs parser truncation
• rerun output correctness (file contents and acceptance criteria)

5. Build a plan, not code:

• define phases
• define acceptance criteria per phase
• define logging/observability additions required to verify fixes
• include approval gate before implementation/commit

Plan-Mode Behavior

• If plan mode exists in the environment, stay in plan mode and output plan only.
• If plan mode does not exist, emulate plan mode behavior:
  • do analysis
  • produce only a <proposed_plan> block
  • request user edits/approval before implementation
• Treat the first plan as a draft that must remain editable by user feedback.

Output Contract

Always return:

1. LAST RUN REVIEW summary:

• run-by-run findings
• hypothesis confirmations (CONFIRMED / NOT CONFIRMED)
• concrete log evidence (run ids + step numbers)

2. <proposed_plan>:

• title
• phases with sequencing
• exact files/components to change
• test/verification checklist
• approval gate (explicitly "wait for user approval before implementation")
• assumptions/defaults called out explicitly

3. Open Questions / User Controls:

• list only decisions that the user may want to override
• keep plan deterministic if user does not override

Quality Rules

• Do not guess when evidence exists in logs/files; cite exact lines/steps.
• Distinguish "what happened" (evidence) from "why" (inference).
• Prioritize root-cause mapping over broad summaries.
• Keep scope on run review + planning; do not mutate code during this skill's execution.

References

• Reuse the original review rubric at references/review-last-run.md for detailed mission checks and issue verification structure.
• Use references/mission-rubric.md for compact expected mission outputs and acceptance hints.