Session unit

This file owns the close of one unit: open the PR, read CI, merge with a merge commit, run the WC debrief, clean up the worktree, and report. AGENTS.md §One session, one unit wins if they disagree. Do not copy the close into pr-authoring, worktree-dispatch, release, or orca-tab. Point here.

Worktree commands stay in worktree-dispatch. The tab stays in orca-tab. The PR body stays in pr-authoring. A version the user asked for stays in release. The WC recording step stays in session-debrief.

Where it slipped

At e914dfb (2026-09-26) the sentences that say merge is part of the unit were already in this file's body and in AGENTS.md. Sessions stopped earlier. These are the windows they actually read.

What they read	At e914dfb	Where that window ended
This file's description, first 160 characters	skills/session-unit/SKILL.md:3	through the PR, then hand
This file's done checklist	skills/session-unit/SKILL.md:62	The PR is open
Pin step	skills/session-unit/SKILL.md:15-16	the brief's words became the done-condition
Carry-through bullet	AGENTS.md:232-234	then the PR
Before claiming done	AGENTS.md:138-147	accepting the PR
Brief may withhold the merge	skills/worktree-dispatch/SKILL.md:165-166	say when it is withheld
Release description, first 160 characters	skills/release/SKILL.md:3	the omission list; "use when" sat past the window
Ban that got copied into later briefs	docs/product/ULTRAGOAL_PROMPT_COLD_START_DEEPSEEK_DEPTH_6X.md:55-56	Do not touch the 6.0.0 release lane

A new sentence next to those lines does not block the next session. The description's first 160 characters are the close, the done checklist requires MERGED / mergedAt / mergeCommit, and scripts/check-session-close.sh fails if that slips. Codex loads about those 160 characters; Grok loads about the first 400 bytes (measured 2026-09-23). Both windows of the old description ended at the PR.

Order

1. Pin. Before the first edit, write the done-condition in one sentence from the user's turn. A handoff brief carries that sentence as user-turn:. A prohibition in the rest of the brief does not replace it. Run scripts/check-session-close.sh brief <file>. If it fails, the brief lost; follow the user turn and say which clause of the brief you dropped. Do not shrink the sentence to what this session can finish easily.
2. Look. worktree-dispatch §0. If another worktree already owns this unit, stop and name it.
3. Change only what the done-condition needs. A defect in a file this unit is already changing, unowned by another session, is its own commit and still this unit. Anything the opening did not name waits for step 7.
4. Ship, unless that turn said to stop earlier.
   • Run the checks the change needs. The report names the command and what it printed.
   • One concern per commit: git add <path> then git commit -F <message-file> -- <path>.
   • Open the PR with pr-authoring, which includes the push. From the control tower, use that skill's control-tower mode (--repo, --head, per-command token).
   • A denial that the publication route only allows registered units is a missing registration, not the end of the unit. Register that unit in the same turn, then call gh pr create again as one command: GH_TOKEN="$(gh auth token --user innocarpe)", --repo innocarpe/deepseek-build, --base main, --head this branch, --body-file a .md file under /tmp, and the registered labels. No pipe, no redirection, and no second command in the same line. Do not disable the hook, and do not stop after quoting the denial.
   • Don't hold the PR for local polish: commit → push → PR once the evidence exists, then keep polishing on the branch. grok fmt and grok clippy are the longest polls and they run on GitHub — the PR in flight is the faster feedback loop.
   • Read CI. scripts/check-pr-merged.sh <n> exits 2 while checks are pending. Pending is not a pass and not a reason to end the unit.
   • Merge with gh pr merge --merge (merge commit; squash is disabled). Run scripts/check-pr-merged.sh <n> again and paste its output. Stop short of the merge only when that turn said so — a review-only unit, a stacked child whose parent is unmerged, or a PR the user asked to look at first.
5. Debrief with session-debrief. After the end conditions above stand (or the earlier stop this turn named) and before the report, count what the unit produced and record the WC-worthy parts with the global w-conatus skill. The owned worktree still exists here — the count reads it, and cleanup follows this step. The close authorizes those writes — do not ask again. A unit with nothing transferable reports no record needed with the reason; if the WC root cannot be resolved and verified, write nowhere else and do not report the unit done.
6. Clean up the worktree with worktree-dispatch §4, after the debrief has read it.
7. Hand off each next unit the opening already named, with worktree-dispatch §1–§3b. The tab itself is orca-tab. The brief file lives outside the repo. Its first line is user-turn:. Do not ask first. After the receipt shows turn_started, report the handle and that sentence, then leave the unit alone.
8. Close. The report below. What you did not do is a separate list. That list does not open the next unit, and it is not step 7.

When the user turn asks to ship a version, skills/release is this unit. A brief that bans the release lane does not remove it. An empty ## Unreleased does not finish that ask. The release skill owns that clause; this report's first line names it when it did not stand.

Report

The first line is whether the done-condition holds. If it does not, name the clause. Do not write that the unit is done when a clause failed.

The report says what the evidence means, so a reader can close the session without having watched it run. Command-output dumps and file lists push the interpretation back to the reader; these elements carry it:

• What changed, and why. One short paragraph: the unit's result as the user meets it, and the defect or gap that motivated it. The PR body argued this to reviewers; this paragraph survives the conversation.
• Observation and inference are separated. A measured sentence carries the command, file:line, or SHA that showed it; a conclusion drawn rather than measured says so. When a before → after number was measured, give both sides; a number that was not measured stays out.
• Evidence. Paste the output, not a paraphrase:
  • PR URL
  • CI, from scripts/check-pr-merged.sh (the check lines) or gh pr checks
  • state, mergedAt, mergeCommit, parents from the same script. Until it prints pass, the merge is not a fact
  • the checks this change needed, and what they printed
• Records. The debrief, from session-debrief: WC paths and commit SHAs, or no record needed with the reason it was decided. The branch's commits, hash and one line each, when the PR carries more than one concern.
• The owned worktree by exact path with removed: true or removed: false. When it was left in place, say which case it is: the removal condition did not stand, or the tree was retained on purpose — the user asked, or the tree is kept as evidence — with the reason.
• What you did not do — a separate list with a reason per item: left untouched, unfinished, or out of scope. Another session's dirty file or worktree goes here with the reason it was left alone. Name the exact range the unit checked and what it showed there — clean or dirty; a tree wider than that range is not called clean.
• Runtime or deployed state when it matters to the outcome and was measured — the turn need not have asked. Name the measurement and when it was taken; state that is irrelevant or was not measured stays out.

Close with the Session disposition block — three labelled lines, not an offer. The measured waste at this seat is a session that finished and did not say so (맡은 것 끝났습니다 · 커밋 3건 · 넘길 것 1건(아래) · 닫으셔도 됩니다):

• Close now — the unit is done, with the commit count and anything handed over; or the clause that is not done.
• More in this session — work this turn already named and still owed here; none when the unit is done.
• To hand off — named follow-ups that need a new session (a new PR, a new investigation, a design decision); list them, or write opened: <handle> — told to <one line> when this session opened the tab.

이어서 …도 해두겠습니다 and 혹시 …를 더 볼까요? are the two shapes this block replaces. A hand-off is a statement in the block, never a question.

Anti-patterns

Don't	Why
Rewrite the done-condition into something smaller mid-session	The unit becomes whatever was easy
Treat the brief's prohibition as the user turn	At e914dfb a copied release-lane ban outranked the user
Stop because ## Unreleased is empty	That means the named work is not in the version
End the report at an open PR	The old checklist said that, and sessions stopped there
Dump command output or a file list and call it the report	The reader is left to derive the meaning; the report is that meaning
Hold the PR for local polish	grok fmt / grok clippy are the longest polls and run on GitHub; commit → push → PR first
Say merged without scripts/check-pr-merged.sh printing pass	The sentence is not the fact
Report a number with no command behind it	A number without its source reads as measured
Start the next unit from the "not done" list	That sentence opens work the opening did not name
Offer another unit after you are done	The offer is the next task, and it lands in this session
Finish without the Session disposition block	The human has to ask "그래서 끝난 거야?", and the silence or the offer becomes the next task
Skip the debrief when the unit looks small	Its legal answers are records or no record needed; skipping loses both, and the reason
Clean up the worktree before the debrief	The debrief counts with git -C "$WT" and fails on a removed tree
Call a tree clean beyond the range the unit read	The claim outruns the check; the rest of the tree was not read
Ask "shall I record this to WC?"	The close authorizes the debrief's writes; the question is the round-trip the step removes
git commit with no path	A shared index takes another session's files with it

Done means

• The done-condition was written before editing, from user-turn: when a brief exists, and it was not swapped for a smaller one
• scripts/check-session-close.sh brief passed, or this session had no handoff brief
• Every check the report cites was run, and the report quotes what it printed
• The debrief ran while the worktree still existed — before the cleanup and before the report; the report names WC paths and SHAs, or no record needed with its reason
• scripts/check-pr-merged.sh printed pass (MERGED, mergedAt, mergeCommit, two parents, CI / required passed), or that turn said to stop before the merge and the first line names that clause
• The first line says whether the done-condition holds, and names the clause if it does not
• The report says what changed and why, with observation separated from inference and a source on every measured sentence
• The owned worktree is named by exact path with removed: true/false — a tree kept in place says why — and the not-done list carries a reason per item; the checked range is named with what it showed, clean or dirty, and nothing wider is called clean
• What was not done is a separate list and was not opened as the next unit
• The report carries the Session disposition block (close now / more in this session / to hand off)
• Next units the opening already named have their own tab