Codex Session Evolution

Codex Session Evolution

Overview

Review historical Codex evidence safely, then turn that evidence into small, auditable improvement candidates for workflows, prompts, and skills. Use this skill as the missing ingress layer before scorecard or optimizer flows, not as a replacement for them.

Read these references only when needed:

• references/route-matrix.md for lane selection and handoff rules
• references/review-packet-contract.md for packet, candidate, and decision shapes
• references/pattern-labels.md for reusable labels and route hints

Boundaries

• Keep raw session logs machine-local.
• Distill sessions into observation and inference packets; do not paste large transcript blocks into durable memory.
• Treat rollout files as event and state-transition evidence, not just message history.
• Prefer candidate snapshots and suggest-first patch proposals over direct canonical skill rewrites.
• Do not promise automatic thread-to-rollout mapping. If mapping is unclear, work from explicit rollout files or user-confirmed batches.

Quick Route

• Use this skill when the ask spans multiple past sessions, archived rollouts, or recurring patterns across history.
• Use $skill-skill-usage-close-loop when the user only wants a current-thread skill retrospective.
• Use $skill-skill-scorecard or $skill-skill-optimizer-darwin after this skill produces enough evidence for one bounded skill change.
• Use $proposal-critique-refine when the candidate improvement is substantial, contested, or still fuzzy after aggregation.

Workflow

1. Inventory the history surface

• Run scripts/inventory_codex_history.py to inspect recent thread index rows, history snippets, and rollout files.
• Build a bounded batch; default to 3-5 sessions.
• Mix success cases and failure-prone cases when possible.

2. Select one review lane

• pilot-batch
  • Use when this is the first pass and the goal is to discover recurring patterns.
• workflow-targeted
  • Use when the likely improvement area is orchestration, verification, closeout, routing, or resume behavior.
• skill-targeted
  • Use when one or two named skills appear implicated across history.
• prompt-targeted
  • Use when the likely issue is route wording, guardrails, or task framing rather than workflow sequence.

Keep one primary lane per batch. Do not let one batch try to fix workflow, prompts, and multiple skills at the same time.

3. Build one review packet per rollout

• Prefer explicit rollout files under ~/.codex/archived_sessions/.
• Run scripts/build_review_packet.py --rollout-file ... --output ....
• Keep task snippets off by default. Add them only when necessary and safe.
• Separate observations from inferences.
• Fill the packet's manual fields:
  • task_summary
  • successes
  • failures
  • pattern_labels
  • candidate_cues
  • route_override
  • next_action

4. Aggregate across packets

• Run scripts/aggregate_review_packets.py packet-1.json packet-2.json ....
• Look for recurring labels, tool patterns, and route cues.
• Promote only repeated or decision-relevant issues. Do not treat one bad session as canonical truth.

5. Generate bounded candidates

• For workflow issues, propose the smallest workflow guardrail or sequence change that addresses the pattern.
• For prompt issues, create candidate versions with representative inputs and explicit evaluation criteria.
• For skill issues, produce a suggest-first patch proposal or route to scorecard and optimizer flows.
• Keep each candidate attributable to packet ids or rollout files.

6. Route the follow-up

• Route workflow candidates to the owning workflow skill or a small decision memo.
• Route prompt candidates to a bounded evaluation pass before keeping them.
• Route skill_patch candidates to $proposal-critique-refine first when the repair shape is still ambiguous, then to $skill-skill-scorecard or $skill-skill-optimizer-darwin.
• Route weak or noisy evidence to defer and stop honestly.

Output Contract

Produce, at minimum:

• one bounded review batch definition
• one sanitized review packet per selected rollout
• one aggregate pattern summary
• the top 1-3 candidates, each with:
  • candidate type
  • evidence refs
  • why it matters
  • proposed route
  • evaluation plan

When working inside PersonalBrain, prefer writing only sanitized source and output pages. Keep raw rollout files outside durable memory.

Script Notes

• scripts/inventory_codex_history.py
  • Inventory the three main Codex history surfaces without assuming they already map cleanly to each other.
• scripts/build_review_packet.py
  • Build a sanitized packet from one rollout file and keep raw snippets optional.
• scripts/aggregate_review_packets.py
  • Count recurring labels and suggest the smallest candidate directions worth advancing.

Quality Checks

• Keep the batch bounded.
• Keep raw logs local.
• Keep observations separate from inferences.
• Keep candidates small and attributable.
• Keep direct canonical rewrites out of this skill unless the user explicitly changes the task.

Common Mistakes

• Treat a current-thread retrospective as a historical batch review.
• Paste raw transcript blocks into durable notes.
• Assume session_index.jsonl ids and rollout ids are the same object.
• Rewrite multiple skills before one candidate has been evaluated.
• Rewrite prompts based on vibes instead of repeated evidence.