abort

Team-mode lifecycle teardown. Under the host engine the workers are reaped by the engine itself — there is no harness team to TeamDelete. Abort is therefore a recorder, not a reaper: it writes a durable abort-report postmortem, an 'aborted' team_audit_log event (the authoritative resume signal), transitions teams.status, and applies the worktree policy.

When to use

Call abort to deliberately tear down the CURRENT agent-team cycle — the user wants to stop the cycle, or the team has stalled past recovery. This skill is the in-session teardown recorder.

Pre-flight (always first)

Run from scripts.atelier_entrypoint import startup_check; startup_check().

Branch on the returned action:

• proceed-local — Memex is not installed. Continue with the rest of this skill's recipe; all writes go to the project-local .ai/atelier.db.
• proceed-memex — Memex is installed and bootstrapped. Continue; all writes go through Memex.
• prompt-migration — Memex is installed but this project still has a local DB. Read internal/migrate-local-to-memex/SKILL.md and follow its prompt protocol. After the user answers, restart the pre-flight (startup_check() will now return proceed-memex or proceed-local depending on the user's choice).

Procedure

1. Identify the team_pk and team_id

The --team-pk is the run/cycle correlation id that scopes the cycle's records — the same value used when the cycle started. The --team-id is the team's TEXT id; it is required for the 'aborted' audit event (team_audit_log.team_id references teams(team_id)). Pass --team-id whenever you hold it — without it the audit event is SKIPPED and resume detection cannot find the arc (the abort-report is still written).

2. Run the abort script

From the target project root:

PYTHONPATH=. python3 -m scripts.abort --team-pk <pk> [--team-id <id>] --project-id <teams.project_id> --phase <projects.phase> [--hard] [--reason "<why>"]

• soft (default) — graceful teardown: teams.status -> 'shutting_down'.
• --hard — forced teardown: report written FIRST, then teams.status -> 'closed'.
• --reason "<why>" — recorded in the report + the aborted audit event (defaults to operator-initiated abort).
• --db / --bridge-db both default to .ai/atelier.db (atelier's single project-local DB); pass nothing unless overriding.
• --clean-worktree — soft path only; remove the worktree iff it is clean (see step 4).
• --project-id <teams.project_id> and --phase <projects.phase> — the #66 resume hooks. The orchestrator already holds the cycle's textual teams.project_id correlation string and the live projects.phase (the phase the arc is being aborted AT), so pass them BOTH. They are folded into the aborted audit payload (the authoritative resume signal) and the abort-report metadata so the NEXT /atelier:run pre-flight (scripts.resume.find_resumable_arc) can OFFER to continue FROM this abort point (AC3) and force-phase AT it (AC4). Omitting them is silently degenerate: abort_phase/project_id default to None, so the resume prompt renders "aborted arc was found at phase None" and there is no phase to continue at — the offer still fires (never-silent holds) but its CONTENT is unusable. Always thread both on a live abort.

The script is mode-aware: in Local mode it performs the full teardown; in non-local mode the state mutators are skipped (team-state mutation is Local-only), so it WARNs, writes the report where possible, and returns 0.

3. What the script does

In Local mode abort.py performs the durable writes (the shared core, run by BOTH paths):

1. Abort-report — backend.write_document(domain='postmortem', subdomain='abort', ...), a durable markdown postmortem. On --hard this is written FIRST so the report survives even if a later step fails. (The workspace-less Memex write now persists in non-local mode too — #90 part-3 — so the report holds cross-mode.)
2. teams.status — set to shutting_down (soft) or closed (hard).
3. 'aborted' team_audit_log event — backend.write_team_audit(event_type='aborted', ...), carrying the #66 resume hooks (project_id, abort_phase, incomplete_task_ids) threaded from --project-id / --phase (step 2). This payload is the AUTHORITATIVE resume signal scripts.resume.find_resumable_arc reads on the next pre-flight (it joins team_audit_log.team_id -> teams.project_id, since the workspace-less abort doc carries project_id=None at the column level). The event is SKIPPED when --team-id is unresolved (the FK to teams needs it); the report still records the abort.
4. Worktree policy — applied per step 4. NEVER destroys uncommitted work.

There is no team_delete row to enqueue and no TeamDelete handshake to drive — the host engine already reaps the workers. Abort's job is purely the durable record above.

4. Worktree preservation

The script NEVER destroys uncommitted work. If the current worktree is dirty (or its state cannot be read), it is ALWAYS preserved — on --hard it is preserved with a warning. A CLEAN worktree is auto-removed only on --hard, or on the soft path only when you pass --clean-worktree. The actual worktree decision is folded into the report's "what was torn down" section.

5. Graceful stop (soft path) — the TM-005 handshake (optional)

The soft path is graceful. If live teammates are still mid-task, the orchestrator MAY drive the TM-005 shutdown_req / shutdown_resp handshake on the kept bridge_messages wire first (each teammate echoes the request_id within one turn; see internal/team-mode-rules/SKILL.md TM-005) so workers wind down cleanly before the cycle stops. This handshake rides the message WIRE only — it is NOT tied to any dispatch-queue row, and abort no longer enqueues one. The --hard path skips the handshake entirely.

Hard rules

• Abort is a recorder: it writes the postmortem + the 'aborted' audit event + the status transition. The host engine reaps the workers — there is no team-delete step for the live session to service.
• Pass --team-id whenever you hold it; without it the 'aborted' audit event (and thus resume detection) is SKIPPED.
• Never destroy a dirty worktree to force a clean teardown — preservation is intentional.
• --db defaults to .ai/atelier.db; pass no db path unless overriding the target.