Retro

Retro — Three Questions, Then Intents

Why this exists

The session that did the work is the worst judge of its own blind spots and the last to notice its own drift. Verification gates check what the session claimed; the retro asks what it did not claim, and whether the work still points at the slice it started on. It is a grooming step, cheap enough to run whenever drift is suspected, not a closing ceremony: it fixes what the hot context can fix, files intents for what it cannot, never a story, and it never touches the primer. The session close is /handoff, which is gated on a retro newer than the primer.

Precedent: a mid-session retro surfaced that a just-shipped containment check was looser than its intent and that a fail-closed gate needed a second edge kind. No gate would have caught either, because no claim covered them.

Prerequisites

• WIKI_SCOPE: <scope> declared in the project instructions. If missing, suggest /wiki.
• The intent kind served: wiki://template/project/intent via MCP, or kmd resource wiki://template/project/intent where the harness exposes no MCP resources.
• The scope's notes/ folder for the dated retro note (projects/<scope>/notes/retro-YYYY-MM-DD.md).

When it runs

• Any time drift is suspected: a bug-hunting detour, an hour without touching the slice in progress, a compaction, a topic change.
• Before a release cut, and before /handoff — the handoff gate blocks a primer write until a retro note is newer than the primer.
• The operator invokes it. The context clock may prompt it; the agent never runs it unasked.

The protocol

Step 1 — Re-read ground truth

Before answering anything, re-read the story and slice in progress from disk, and list the scope's open intents:

• MCP: search(query, scope=<scope>, kind="intent"); CLI: kmd search "<terms>" --scope <scope> --kind intent.

The re-read is the point. A degraded context cannot diagnose its own drift by introspection; reading the contract back is the cheapest refresh there is.

Step 2 — Answer three 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?
3. Where are we relative to the slice we started? State the delta between the slice's acceptance and what is in the tree, citing the slice by name and its tick state. Then name every detour since the last retro and classify each: decision (recorded where), drift (dropped, or worth an intent), or done.

Answering rules:

• Answers are about the WORK, not the worker. Name code, artifacts, and measurements — never feelings, apologies, or process narrative.
• Every answer to questions 1 and 2 carries a location (file, story, intent, 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 filed intents or stories as discoveries — reference them instead.
• An answer that only references existing intents and stories is a valid answer. A retro that files nothing new is a passing retro.

Step 3 — Route each answer

Nothing stays only in chat. Route by shape:

Rules of the table:

• Cap: at most three new intents per retro. With more findings, file the three with the sharpest falsification paths and name the rest as unfiled in the retro note. If one matters it recurs, and the second sighting is what earns it a file.
• Two strikes before elaboration. The retro never writes a story. Promotion — two sightings, a confirmed falsification, or an operator call — is /triage's outcome, and the story is written there.
• Never a new story, a session log in the plan, or a primer edit. The plan carries only its Story Index; the primer is /handoff's.

Step 4 — The retro note

One note per day at projects/<scope>/notes/retro-YYYY-MM-DD.md, no kind field, updated taken from date -u +%Y-%m-%dT%H:%M:%SZ. Each invocation appends a section headed by the UTC time and a label:

## 14:30Z — mid-session, resolver

- Intents: [[intent-<slug>]] filed; [[intent-<other>]] sightings → 2
- Decisions: [[story-<slug>]] § Decisions, sequencing of the engine publish
- Detours: <one line each, classified>
- Unfiled: <one line each, if any>

Ten lines or fewer, pointers only. The substance lives in the intents and stories; the note is the freshness signal the handoff and tag gates read, so its clock must move every time.

Step 5 — Acting on a finding: fix now, or file

The intent is a cold-start capture: its six sections hold what a future session cannot recover from the code. When the context is hot and the fix is small, the interview re-captures what the conversation already holds and charges the next session for work this one could have finished. Fixing is the default and filing the fallback. Fix now when all three hold:

1. Small and testable. The change is bounded — on the order of twenty lines — and a test named for the finding goes red before it, in a test file already in play. The test is the contract; the record is one Decisions line in the owning story, or fixed_by on an intent already filed.
2. In the current flow. It ships in the commit the session is already building — never its own slice, ceremony, or "next run" — and the operator hears what shipped in the same turn: a report, not a permission gate. Slice ticks, spec corrections, and the retro note ride the same session.
3. Unlocked territory. Anything the story's Decisions mark as ADR-gated routes to its amendment round. When the fix shape IS an amendment, acting now means holding that conversation now.

Anything larger files as an intent; when a filed intent's fix turns out that cheap, it archives with fixed_by naming the test's repo-relative path. Promotion is for findings that need a story's shape.

Output shape

Visible response: the three answers, each with location and falsification path; the artifact list (file → what changed); one line naming what the session is re-anchored on (story-<slug>, slice N). No separate report — the artifacts are the output.

Anti-patterns

• Skipping the re-read. Answering question 3 from memory is the drift the question exists to catch.
• Story from a retro. A finding is an intent until triage promotes it; a story written at filing is the backlog growing by ceremony.
• Primer from a retro. The primer is written once, at /handoff, from the corrected state. A retro that edits it inherits the blind spots it was meant to surface.
• Filing as resolution. A draft intent is a routed finding, not a handled one; the second sighting, not the first file, is the signal.
• Filing as deferral. An intent for a fix the hot context could make is ceremony charged to the next session, which starts cold and pays the interview twice.
• Confession register. Wiki artifacts store project state, never behavioral postmortems.
• Note as substance. A retro note longer than ten lines is a plan session log by another name.