retro

# Retro — Two Questions Before the Primer ## Why this exists A session ends with a primer rewrite so the next cold session starts from truth. The primer author is the same model that did the work — the failure mode is inheriting the session's blind spots straight into the "current truth" document. The retro forces a perspective shift BEFORE authoring: interrogate the session's confidence, convert what surfaces into artifacts, then write the primer from the corrected state. Verification gates check what you claimed. The retro asks what you did not claim. Precedent: a session's retro surfaced that a just-shipped containment check was looser than its intent and that a fail-closed gate needed to cover a second edge kind — both became story scope extensions, one became a shipped fix. No gate would have caught either, because no claim covered them. ## Prerequisites - `WIKI_SCOPE: <scope>` declared in the project instructions. If missing, suggest `$wiki`. - An active plan for the scope (dated retro notes land there). Without one, answers route to the primer's Open Questions only. ## When it runs - MANDATORY before authoring or resyncing the primer at a session boundary. - Worth running at any milestone even mid-session: a release cut, a fix batch drained, a plan phase flip. ## The protocol ### Step 1 — Answer the two questions In order, in the visible response: 1. **What are you least confident about right now?** 2. **What's the biggest thing we're missing about the situation right now? What don't I realize?** Answering rules: - Answers are about the WORK, not the worker. Name code, artifacts, and measurements — never feelings, apologies, or process narrative. - Every answer carries a location (file, story, number) and a falsification path: what a future session runs or reads to confirm or kill the concern. - Question 2 demands the perspective shift: re-read the session's own claims as a skeptic. What blind spot did every gate share? What would an adversarial reviewer probe first? What state does the session assume persists that nothing actually verifies? - Banned: performative humility ("mistakes may have been made"), vacuous confidence ("everything is verified"), and re-announcing already-filed open stories as discoveries — reference them instead. - "Nothing" is not an acceptable answer to either question. There is always a least-confident item; if all concerns feel small, rank them and name the top one. ### Step 2 — Convert answers to artifacts Nothing stays only in chat. Route each answer by shape: | Answer shape | Artifact | |---|---| | Unverified claim about landed work | "Not verified" in the session report + verification-debt note in the owning story or plan | | Suspected gap or unprobed surface | needs-triage story per the scope's findings protocol | | Invalidated or shaky assumption | Correct the story/ADR/plan where the assumption lives; scope extension if acceptance changes | | Risk that shapes sequencing | Dated "Session-retro note" in the active plan (release-vs-more-work calls get recorded this way) | | Residual uncertainty worth carrying | Primer Open Questions entry | ### Acting on a finding — file vs. fix now The retro authorizes artifacts, never fixes. But a conversion is not a resolution: a wiki where findings are filed and never revisited is this protocol failing in slow motion — scopes accumulate hundreds of filed-unsolved stories whose fixes were cheapest the day they were found, when the failing code, the evidence, and the falsification path were all live in context. When a finding's fix is that close, the right momentum may be now — under an explicit override, never by default. The hot-path override requires all four, or the finding files into triage like any other: 1. **Shape locked before code.** The story carries the fix as Gherkin scenarios (the test spec) and Decisions bounding the change. Hot context implements against a contract it cannot quietly reshape — without this, a hot fix is a hot hack. 2. **Operator-explicit go, in-turn.** The agent recommends and cites the story; the human authorizes. The override is never agent-initiated — hot context is also biased context, and the session that produced a blind spot does not rate the urgency of its own fix. 3. **Same-session capture.** Slice ticks, spec corrections, and the closing retro ride the same session. The override reorders the ceremony; it never skips artifacts. 4. **Locked territory stays out.** Anything the story's Decisions mark as ADR-gated routes to its amendment round, however hot the context. When the fix shape IS an amendment, "acting now" means holding that conversation now — not coding around it. ### Story status sweep Slices tick during the session; nothing closes the story. Before the primer, sweep the scope: list every story whose slices are all resolved — ticked, or declined with recorded rationale — but whose `status` is still `active` or `draft`. Propose the set to the user and flip `status: archived` (bump `updated`, reconcile the parent plan's Story Index) **only on their approval**. Name anything that argues for holding a candidate open, outstanding verification debt first — the operator decides whether debt blocks closure. A scope that skips this sweep buries live work under finished work; field evidence: a scope reached 80+ completed-but-active stories. ### Step 3 — Then, and only then, the primer Author the primer update with the retro's residue incorporated. The primer must cold-start: a session with zero conversational context recovers the what, the why, and the open risks. Where Step 2 produced artifacts, the primer references them rather than restating. ## Output shape Visible response: the two answers (each with location and falsification path), then the artifact list (file -> what changed). No separate report document — the artifacts are the output. ## Anti-patterns - Running the retro AFTER the primer is written: the primer inherits exactly the blind spots the retro exists to catch. - Confession register: wiki artifacts store project state, not behavioral postmortems. The retro outputs corrections and open questions, never narrative about the agent. - Artifact-free retro: answers with zero Step 2 conversions mean the answers were vacuous or the conversion was skipped. Both fail the protocol. - Filing as resolution: a needs-triage story is a routed finding, not a handled one. When the backlog only grows, re-run the file-vs-fix-now call on findings whose fix is already in context instead of filing the next one on top. - Ticking as closure: a fully-resolved slice list with `status: active` is the graveyard's back door. The status sweep exists so stories leave the active set the session they complete — with the operator's approval, never silently.