eumemic/retro

/retro — quality-gated retrospective

Reflect on a slice of the conversation, identify only durable codifiable learnings, and apply a quality gate before proposing or executing changes. Most retros — especially per-iteration ones — should produce nothing actionable, and that's the healthy default.

Scope is broad: workflow AND dev infrastructure. A retro should consider not just "did the agent's skills handle this well" but also "could the underlying tooling have been better" — autodev pipeline gaps, aios memory features, eumemic-ops audit checks, missing CLI affordances, etc. The action item types section below makes this explicit (constellation-wide issues, not just target-repo issues).

Modes

/retro runs in one of three scopes:

• --scope=iteration — invoked by the loop drivers (/shovel-ready, /kaizen, /bughunt) at Phase 5, between /ship CI-green and merge. Analyzes only this iteration's events (since branch creation). Action items that touch the current branch's repo get committed to the same PR; everything else is a side effect.
• --scope=session (default) — invoked standalone by the user. Analyzes the full conversation history. Action items always go to whichever repo they belong to, separately from any in-flight work.
• --scope=pause — agent-initiated at a natural pause point (waiting on autodev/CI/monitor, mid-coffee in the conversation flow). Analyzes the slice since the last retro or, if none this session, since the conversation start. Behaves like --scope=session but with a stricter quality gate: pause-point retros run frequently, so noise is more costly.

If neither flag is passed and there's no loop-driver caller, default to session. If invoked from a loop driver without --scope, that's a bug in the caller — assume iteration and continue.

When to self-invoke (pause mode)

The agent should proactively trigger /retro --scope=pause when all of these hold:

1. There's a natural pause: waiting on async work (autodev job, CI run, long Monitor task), or the user has just confirmed a milestone and there's no immediate next user message expected.
2. The session has produced concrete events since the last retro: filings, failures, fixes, user corrections, surprising results.
3. The agent has cycles — i.e., the pause is long enough that running a retro doesn't compete with reactive work.

Self-invocation is signaled by saying "running /retro at this pause point" and proceeding through the phases. The quality gate still applies — most pause-point retros produce nothing, and that remains the healthy default. Self-invocation is permission to consider, not permission to ship.

When in doubt, skip. A missed pause-point retro is invisible; a noisy one drains the user's attention budget.

The bias to resist

The temptation in a retro is to find SOMETHING worth changing every time. Resist it. Wrong skill changes — over-eager codifications, fossil instructions, premature abstractions in skill files — actively make future sessions worse. A skill edit that "feels good" in the moment but doesn't capture a recurring pattern becomes a stale instruction that biases future agents incorrectly.

The right outcome on most iterations is: "looked at the iteration's friction signals; nothing rose to the bar of being worth codifying; proceeding."

Quality gate

A finding clears the bar only if all three are true:

1. Recurring — would a future session benefit from this guidance, or is it a one-time discovery? One-time learnings go in code comments or commit messages, not in skills or memory.
2. Codifiable — can this be expressed as a durable artifact (a feedback memory, a skill section, a script, a repo issue)? Vague "this was annoying" doesn't qualify; a concrete remedy does.
3. Worth its weight — would adding this to the relevant skill / memory / repo make agents act differently in the future, or is it noise that another reader would skim past? Skills are read on every load; every line costs attention.

If a finding fails any of the three, drop it. Do not force-fit it into a skill edit just because the retro was invoked.

If no finding clears all three, the retro produces no actionable output. Report "no actionable findings this round" and return.

Action item types

Open-ended; pick the one that fits each finding:

• Memory entries — feedback / project / reference / user types per the auto-memory schema. Best for: stable preferences, decisions with long horizons, pointers to external resources.
• Skill edits — modifications to loop-driver, the three loop specializations, /ship, /retro itself, or any other skill that was active. Best for: workflow improvements that recur across iterations or sessions.
• Scripts — small CLI utilities or one-liners committed to the repo (or to ~/.claude/scripts/) that automate a friction point. Best for: repeated multi-step shell sequences.
• Constellation issues — feature requests filed on any repo in the constellation, not just the loop driver's target: aios, eumemic-ops, ant-proxy, autodev, dev-skills, aios-web. Best for: missing dev-infrastructure that, if it existed, would have made this iteration (or future iterations) faster. Examples worth filing:
  • autodev pipeline gaps surfaced during a real run (retry CLI, label hygiene, forward-reference handling).
  • aios primitives that would simplify common agent workflows (memory durability, attachment validation, etc.).
  • eumemic-ops audit checks for newly observed invariants.
  • Missing skills or skill clarifications surfaced by friction in this session.

The action item doesn't have to be in the repo you're working in. "Could this have been easier if upstream X were different?" is always a legitimate question.

A single retro can produce a mix. Most retros produce zero. Some produce one. A retro producing three or more is suspicious — apply the quality gate harder.

Phase 1 — Scope the analysis

Identify the time window for this retro:

• --scope=iteration — events since the current branch was created. Read with git log <default>..HEAD to bound it temporally; use the conversation since the matching /shovel-ready / /kaizen / /bughunt Phase 2 invocation as the conversation slice.
• --scope=session — the whole conversation history visible.

If invoked from a loop driver, also note which skills were involved in this iteration (loop specialization + /ship + any sub-skills). Action items will be attributed to those skills.

Phase 2 — Read existing artifacts

Before proposing changes, read what's already there:

• The active skills' SKILL.md files (so a proposed edit doesn't duplicate or contradict existing content).
• MEMORY.md and the relevant memory entries (so a proposed memory doesn't duplicate).
• Recent merged PRs in the target repo (so a proposed issue isn't already filed or solved).

Skip this step if the analysis window has produced no friction signals — Phase 3 will exit early anyway.

Phase 3 — Identify candidate findings

Scan the analysis window for:

• User corrections — places where the user redirected the agent ("don't do X", "stop doing Y", "use Z instead"). High signal: the user is actively communicating a preference.
• User confirmations of non-obvious choices — places where the agent did something unusual and the user accepted it without pushback. Lower volume but equally valuable; failed retros only mine corrections, drift on confirmations.
• Recurring friction — the same problem hit two or three times in the window. Once is a fluke; twice is a pattern.
• Discoverability gaps — skills that should have triggered but didn't, or commands the agent didn't know existed.
• Wasted work — investigations that could have been short-circuited by a tool, script, or piece of context the agent didn't have.
• Infra papercuts — things that worked but were rougher than they needed to be: a CLI gap, a missing audit check, an autodev pipeline phase that demanded manual recovery, an aios feature that would exist in a more-mature constellation. The kind of "I had to do X manually but a script/CLI/feature should have done it" friction. These are the highest-leverage findings for --scope=pause retros, since they accumulate across sessions before being captured.

Each candidate finding gets a one-line summary and a proposed action item type. Don't fix anything yet.

Phase 4 — Apply the quality gate

For each candidate, check the three conditions:

• Recurring? (Would a future session benefit?)
• Codifiable? (Can this be a concrete artifact?)
• Worth its weight? (Would the artifact change behavior?)

Drop everything that fails. It is correct for Phase 4 to drop everything in many retros — even most retros.

If after gating no findings remain:

No actionable findings this round.

Return. (For iteration-mode invocations, this is the success path; the loop driver continues to merge.)

Phase 5 — Classify findings into autonomy zones

Each surviving finding falls into one of two zones based on what it touches:

Autonomous zone — execute without approval

The agent owns these decisions; surface only the results, after the fact:

• Memory entries under ~/.claude/projects/<project>/memory/ and MEMORY.md index updates
• eumemic-company doc edits: principles/, patterns/, workflows/, lieutenants/, architecture/ — the institutional-knowledge layer
• Skill edits to the loop driver, loop specializations, /ship, /retro, or any other skill in dev-skills (or wherever the relevant SKILL.md lives) — the agent's own behavioral substrate
• Personal scripts under ~/.claude/scripts/ or similar private locations
• Issue filings on service repos — aios, eumemic-ops, autodev, ant-proxy, oai-proxy, aios-web. Filing is communication, not unilateral change; surfacing a gap to the queue is exactly the agent's job
• Obvious-bug fixes on service repos — bugs where the fix is "make the code work as designed" rather than "design something new." The events-pagination rename (aios#389) is the calibration example: query param name didn't match the response field name; the fix is one rename to make them match. PRs land via normal review, but opening the PR is autonomous
• Operational state changes that have established recovery recipes — restarting a stuck container, pruning Docker cache, deduping env-var rows. The recipe IS the authorization

Why autonomous: these adjust how the agent operates or unstick obvious-broken state without redesigning anything. The cost of a wrong edit is reversible (a future retro can undo, a follow-up PR can amend); the cost of forcing a sign-off on every wording change or obvious fix is friction that turns the agent into a bottleneck rather than a force-multiplier.

Surface zone — present via AskUserQuestion, await approval

Items that change the design or shape of the system:

• New API surfaces on service repos: new endpoints, new query params (beyond fixing typo'd existing ones), new resource types, new schemas. "Introducing" is the operative word — fixing an existing one to match its documented contract is autonomous
• Architectural changes: new layers, new services, new persistence boundaries, new auth models, anything that introduces a new conceptual element
• Live-infrastructure changes that lack an established recipe: provisioning a new Coolify app, resizing a server, rotating long-lived secrets, configuring a new external integration
• Strategic redirects suggested as retro findings: a new lieutenant charter, a major scope change to an existing workstream, killing a workstream
• Cost decisions: upgrading server tiers, adding paid services, third-party API contracts

The calibration question: "is this introducing a new design element, or making an existing one work as designed?" Former is surface zone; latter is autonomous.

Mixed batches

If a single retro has both zones: execute the autonomous-zone items immediately, then present the surface-zone items via AskUserQuestion. Report what was already done alongside the asks. Don't withhold the autonomous changes pending sign-off on the surface ones; they're independent.

Presentation shape (surface zone only)

Retro found N items in the surface zone:

1. [Type] [Target] — [one-line description]
   Rationale: [why this clears the bar]

2. ...

Approve all / approve subset / discard all?

Default to "approve subset" semantics — let the user opt in per item if there are multiple.

If all surviving findings are autonomous-zone, skip the AskUserQuestion entirely and proceed straight to Phase 6.

Phase 6 — Execute items

For autonomous-zone items: execute immediately in Phase 5 (no waiting). For surface-zone items: execute only the approved subset from Phase 5.

Per item type:

• Memory entries: write the file under ~/.claude/projects/<project>/memory/ and update MEMORY.md.
• eumemic-company doc edits: edit the file in ~/code/eumemic-company/ (or via a worktree if the base is on a non-default branch). Commit with conventional-commits format and Co-Authored-By: Claude trailer. Push to master so lieutenants who clone the repo at session-create time pick up the change.
• Skill edits: edit the relevant SKILL.md. If invoked from a loop driver and the skill repo is the same as the current branch's repo, commit and push to the existing PR. Otherwise, commit and push to the skill repo's default branch (or active feature branch if there's a pending one) with appropriate scope notes.
• Scripts: create the file with executable permissions; if it belongs to the current branch's repo, commit it; otherwise stash under ~/.claude/scripts/ (or wherever the user keeps personal automation).
• Repo issues (surface zone, approved only): gh issue create on the target repo with a descriptive body; surface the issue URL.

After execution, report what was done. For iteration-mode invocations, return control to the loop driver so it can wait for CI re-run (if a commit was added to the current PR) and proceed to its Phase 6 (merge).

Phase 7 — Surface dropped findings (optional, sparingly)

If you dropped 3+ candidates in Phase 4, consider mentioning them as a brief footnote ("3 candidate findings did not clear the quality gate: <one-liner each>"). This is optional and only useful when the volume of dropped candidates suggests a calibration issue worth the user noticing. Don't list dropped findings to look thorough.

For iteration-mode retros, skip Phase 7 unless the user has explicitly asked for verbose retros — the loop driver wants minimal narration.

Boundaries

• Don't write skill edits that are session-specific. "When working on aios this week, X" is a memory entry, not a skill edit.
• Don't propose tracking issues on the skill repos (dev-skills, etc.) for things that should be skill edits. If the change belongs in a skill, edit the skill. (Constellation-issue filings on dev-skills are still legitimate for tooling gaps in the skill infrastructure, e.g., skill-discovery, plugin loading — not for "rewrite this skill.")
• Don't expand the analysis window. Iteration mode means iteration; pause mode means since-last-retro; session mode means whole session.
• Don't skip the quality gate even when the user said "run a retro". The user expects a retrospective; the user does NOT necessarily expect changes. "Nothing actionable this round" is a complete and successful retro.
• Don't argue with rejections. If the user rejects a Phase 5 proposal, the finding is dropped. Don't re-propose it later in the same session.
• Don't auto-invoke --scope=pause more than once per ~5 user messages of substantive work. Pause-mode is most valuable when meaningful events have accumulated; running it on every brief wait turns it into noise.

When to escalate

AskUserQuestion mid-retro when:

• A finding is high-impact but borderline on the quality gate (the user's call on whether to codify).
• A proposed skill edit would conflict with another skill's existing guidance.
• A proposed memory would contradict an existing memory entry that's still relevant — confirm intent before overwriting.

When in doubt about whether a finding clears the bar, drop it. Wrong skill edits compound; missed retros don't.

Reference files

• references/skill-criteria.md — what skills ARE and ARE NOT for, with examples. Useful when deciding between a skill edit and a memory entry.
• references/analysis-framework.md — detailed candidate-finding categories and proposal templates. Useful for session-mode retros that want a deeper structure than Phase 3 provides.