Retro

Run a retrospective on the loop itself — not on the code. The loop's roles review every code increment; nobody reviews the harness unless you do. Your output is diagnosis + proposals; with apply, also one bounded change. If .cid/journal.md does not exist or has fewer than ~5 rows, say there is not enough track record yet and stop.

Step 1 — Measure

Run uv run .cid/bin/metrics.py --window 20 --all for the deterministic journal numbers — verdict mix and NEEDS_WORK rate, Advances tally, streaks, same-step bounces, phase trajectory (script or uv unavailable → derive them from journal.md + journal-archive.md by hand). Then add what the script cannot see, from git history (git log --oneline, the cid(): cadence), learnings.md + learnings/, issues.md, and policy.md, for the last ~20 iterations:

• NEEDS_WORK clustering — do the script's failures cluster on a phase, an area, or a role's

handoff?

• Convergence — is the script's polish share drift (criteria open and reachable) or

legitimate (hardening in stable)?

• Step sizing — steps that blew the phase budget or bounced (blocker handoffs)
• Escalation health — STOP/IDLE signals: were they real, or should the loop have handled them?
• Ledger health — learnings.md index and detail files vs. the review role's rotation

budgets (that role is the single source for the numbers); journal.md past the archive threshold its own header states; stale issues

Step 2 — Diagnose

Name the top 1–3 loop-level problems with evidence (iteration numbers, journal rows, commit hashes). Distinguish harness problems (bad policy knob, missing learning, vague Verify criteria in target.md) from project problems (hard code — not yours to fix). No problems found is a valid finding — say so and stop; do not invent tuning for its own sake.

Step 3 — Propose

For each diagnosis, one concrete proposal:

• Policy tuning — a specific edit to a tunable section of policy.md (phase knob, a role

adjustment for define-next/advance; update-state and review adjustments go to the human as proposals)

• Target sharpening — a target.md Verify criterion that proved vague or unverifiable →

propose the runnable form (human applies it; the loop never edits target.md)

• Ratchet — a recurring failure that should become an executable check in the gate
• Hygiene — ledger rotation, journal archiving, stale-issue sweep

File each proposal as an issues.md entry (Source: [retro], priority low for human-directed ones, normal for ones the loop should pick up), and summarize them in your report.

Step 4 — Apply (only if $ARGUMENTS contains apply)

Apply at most one proposal per retro, and only within these guardrails:

• Only tunable policy.md sections — within the phase table, only the Step budget and

Allowed work rows — plus the define-next/advance role adjustments and hygiene (ledger rotation, journal archiving).

• Never the quality gate, the branch model, the update-state or review role adjustments,

the rigor rows of the phase table (Tests required, Review depth, Debt tolerance), target.md, or any source file. The reviewer and the assessor are the loop's quality gate — a self-tuning loop that can soften its own reviewer or gates will Goodhart itself (a softened reviewer improves the NEEDS_WORK metric, so the rollback test would ratify the move instead of reverting it), and a softened assessor is the symmetric case (update-state computes the phase that selects the binding rigor column and judges when a Verify criterion counts as met, so biasing it improves the Convergence metric the revert test also keys on); those changes go to the human as proposals.

• Record the change in journal.md as a row:

| - | | - | retro: | - | - | - | so the next retro can evaluate it against the metrics — and revert it if the NEEDSWORK rate, the step-bounce rate, or the Convergence rate (Step 1) got worse since. Convergence is in the revert test because the tunable Allowed work row steers what work gets picked: widening it toward easy polish work improves the NEEDSWORK rate while Verify criteria stall — only the journal's Advances column catches that.

• Commit as cid(retro): .

These guardrails are prose, not mechanism. Unlike the runner-enforced harnesses this skill distills (which auto-revert guardrail-breaking meta commits and own the re-evaluation schedule), nothing here mechanically reverts a bad retro commit, no role checks a cid(retro): commit against these guardrails (review's integrity scan pattern-matches gate circumvention in the commits since the last review — that includes a retro commit, so it can catch a gate-weakening retro edit, but not an off-limits tuning such as a rigor-row change), and an applied tuning is re-evaluated only when a human runs the next retro. Granting apply means trusting the retro to obey the list above — review cid(retro): commits yourself until that trust is earned.

Without apply, touch nothing except the issues.md proposal entries (commit them as cid(retro): file proposals).

Step 5 — Report

Metrics summary → diagnoses with evidence → proposals (and the one applied change, if any) → the follow-up: when to run the next retro (default: every ~10 iterations, or after a NEEDS_WORK streak).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: titusz
• Source: titusz/skills
• License: MIT

Install and usage instructions live in the source repository linked above.