Catalog
Yeachan-Heo/refit

Cross-session environment retrospective driven by OMC's own instrumentation — trace timelines, friction reports, logs, and plan notepads. Every finding lands on the surface that owns the fix (deterministic check, steering volume, tooling, or information access); nothing is written without user approval.

NewUpdated Oct 22, 2026

Refit

Refit surveys the environment the agent worked in — not the code it produced (review and verify own that) and not one launch run's lessons (the sediment pass owns that). Every kind of work leaves friction behind: a ralph loop that kept tripping on the same check, an autopilot run that stalled waiting for a signal, a debugging session where the decisive output was never captured anywhere. OMC instruments all of it. Refit reads the instruments and converts recorded friction into environment changes, so the next run starts in a better yard.

Evidence first

The survey starts from OMC's instruments, not from a checklist. Read before asking the user anything:

  • trace_timeline / trace_summary — where turns stalled, repeated, or fanned out wastefully
  • omc session friction report — friction the session recorded about itself
  • .omc/logs/ and session state — errors, retries, swallowed failures
  • plan notepads .omc/notepads/*/issues.md and problems.md — problems that accumulated across a plan
  • the repo's own check surface (package.json scripts, CI workflows) — read before proposing any new check, because a check that exists but sits unwired or silently broken is the finding, not a reinvention

Findings emerge from what the data shows. An empty survey is a valid result and is stated plainly; an invented finding is the same violation as skipping the survey. A repo with no wired automated checks is itself a finding, not a neutral default.

Four fix-owner surfaces

Every finding names the surface that owns its fix. A finding that fits no surface is declined, explicitly and with the reason.

  1. Deterministic check — the failure was mechanical: a fixed syntactic pattern, a banned API, an import shape, a file-location rule. The fix is a lint rule, hook, or CI job — whichever the repo's existing guardrails make cheapest. Default to building the check over writing the rule.
  2. Steering surface — the failure was a judgement call no guardrail substitutes for: cross-file consistency, matches-surrounding-style. The fix lands in the matching docs/standards/ volume, or in CLAUDE.md/AGENTS.md within the thin-entry budget the launch sediment pass defines. Navigation pointers over prose.
  3. Tool surface — the drag was tooling itself: expensive or token-inefficient MCP calls, missing automation, checks too slow to be run when they matter. The fix lands in .mcp.json, scripts/, hooks, or OMC config.
  4. Information surface — the agent needed a signal it could not reach: dev-server logs, service state, third-party readonly access. The fix tees the log, exposes the state, or documents the access path.

The split rests on where enforcement pressure lives: implementation contexts carry the most of it (exploration, writing, and debugging all at once); review contexts receive a diff with none of it. Standards therefore belong to reviewers and checks — never to more instructions loaded onto the implementer.

Proposal and landing

  1. Present findings ranked by severity, each as finding → surface → intended change, with one line of instrument evidence (which instrument, what it showed).
  2. Stop for user approval. Each line is individually approvable or vetoable.
  3. Write approved findings to their surfaces. Before writing any steering prose, call the Skill tool with agent-doc-discipline and pass its verification checklist. A newly installed deterministic check must be proven to bite before it counts as landed: run it clean once, then demonstrate it failing on a deliberately introduced violation, then revert the violation. A check that cannot be shown failing goes back to the proposal — a guardrail nobody has seen fire is decoration, not protection.
  4. Record where each finding landed — one file location per line — so the next refit starts from the record, not from memory. A refit on a launch-run session starts from that run's sediment list; a launch run may defer a lesson to a later refit by naming it.

Headless refit (unattended invocation)

Refit is user-invoked, but its survey does not need the user at the keyboard to run — only to dispose. A host scheduler (cron, CI timer, an automation tool) may start a headless agent session that invokes this skill; the survey runs to its natural boundary under the same contract:

  • The survey reads the same instruments and lands nothing anywhere: headless refit produces the proposal list only — finding → surface → intended change, each with one line of instrument evidence — written to .omc/refit/pending-proposals.md and, when a notification channel is configured (configure-notifications), summarized there so the user learns a survey is waiting.
  • Nothing reaches a surface without the user's approval, exactly as in an interactive refit: proposals wait ranked, each independently vetoable, and the next interactive refit starts from the pending list instead of re-running the survey.
  • An empty survey is stated plainly ("no findings"), never padded — inventing findings is the same violation as skipping the survey.

Output

  • Findings ranked by severity, each with evidence, surface, and intended change — every line independently decidable
  • After approval: what landed and where
  • Declined findings, with reasons
Files1
1 files · 1.0 KB

Select a file to preview

Overall Score

88/100

Grade

A

Excellent

Grades are signals, not a certification. Always review a skill yourself before use.

Safety

92

Quality

86

Clarity

89

Completeness

82

Summary

Refit is a retrospective analysis skill that surveys agent session instrumentation (traces, friction reports, logs, notepads) to identify environmental friction points and propose deterministic checks, steering standards, tool improvements, or information access changes. All proposals are user-approved before landing; the skill is designed to emit findings, not execute changes unilaterally. It supports both interactive and headless (CI/cron-driven) invocation modes.

Detected Capabilities

file read (logs, traces, notepads, config files)session state inspectionstructured output generation (ranked findings)user approval workflowconditional execution (write only after approval)

Trigger Keywords

Phrases that agents use to match this skill to user intent.

session friction analysisretrospective improvement proposalsenvironment check auditcross-session environment surveyagent session instrumentation review

Risk Signals

INFO

Reads .omc/logs and session state files

Evidence first section
INFO

Inspects repo check surface (package.json, CI workflows) before proposing checks

Evidence first section
INFO

User approval gate on all proposals before any write occurs

Proposal and landing section, step 2
INFO

disable-model-invocation flag set to true

Frontmatter

Use Cases

  • Analyze cross-session agent friction to identify reusable improvements
  • Generate ranked proposals for checks, standards, tooling, or information access
  • Run automated friction surveys via CI/cron with manual approval gate
  • Compare environment state across sessions to uncover systematic drag
  • Audit whether existing checks are wired and functional before proposing new ones

Quality Notes

  • Skill explicitly documents its scope boundary: it surveys environment friction, NOT code review or single-session lessons — prevents scope creep
  • Evidence-first principle prevents invented findings: instrument data is required before proposing changes — strong guard against hallucination
  • Four-surface taxonomy (deterministic check, steering, tool, information) is well-defined with clear criteria for each — agents can deterministically route findings
  • User approval gate on every finding before landing is explicit and non-negotiable — prevents autonomous change deployment
  • Verification checklist for deterministic checks (must run clean, must fail on violation, must revert) is concrete and testable — prevents empty guardrails
  • Recording landed findings to .omc/refit/ creates a persistent audit trail — enables iterative improvement without re-surveying
  • Headless mode explicitly defined: proposals to .omc/refit/pending-proposals.md only, no surface writes without user approval — safe CI integration
  • Empty survey is treated as valid result, not padded — indicates strong discipline against false positives
  • Skill deliberately does NOT write code, does NOT modify checks autonomously, does NOT assume user context at keyboard — appropriate for async/CI use
Model: claude-haiku-4-5-20251001Analyzed: Oct 22, 2026

Reviews

Add this skill to your library to leave a review.

No reviews yet

Be the first to share your experience.

Version History

  1. v1.1

    Content updated

    ✦ AINo behavioral changes detected.

    2026-10-22

    LATEST
  2. v1.0

    2026-10-02

    View This VersionInitial version

Use Yeachan-Heo/refit in your dev environment

Command Palette

Search for a command to run...