Execution Log: one work item may be delivered by several PRs

Append-only log of progress for the user's visibility. Checked in alongside the spec at docs/specs/issue-101/.

Phase transitions

Phase	Entered	Reviewed/approved by	Notes
requirements-definition	2026-07-25		Issue #101 asks for an audit first ("check if it's already the case"). Audit → requirements.md Requirement 0: the event→work-item direction is already list-shaped (issue-93) and the poll path is already per-ref (issue-94); the webhook close path and the loop's own guidance are not.
design	2026-07-25		One payload-only predicate (_closing_refs) in the dispatcher; workspace/tmux behaviour follows from it. No config key, no new stored data, no API call. decision-039.
tasks-breakdown	2026-07-25		10-task DAG (T1–T10)
implementation	2026-07-25		Implemented on claude/github-issue-101-l8j95o
needs-review	2026-07-25		PR opened with the reviewer briefing; make check green; awaiting human review (tier-3, human-approves-pr)
complete

Pull requests

PR	Scope / tasks	Status
the-loop#102	Whole work item — T1–T10	open

Progress entries

2026-07-25 — audit + spec drafted

• Phase: requirements-definition → design → tasks-breakdown
• Did: Walked every site that relates a work item to a PR (router, session registry, poller core, GitHub poll provider, dispatcher, workspace cleanup, guidance docs, templates) and recorded the result as Requirement 0. Five of nine sites were already list-shaped — linked_issue_numbers/extract_work_items emit every linked issue (issue-93), the registry is per-ref, _poll_provider unions all of a listed item's refs, _reconcile_closures asks about the session's own ref, and closure_event carries work_items=[ref]. The gap is the webhook close branch in Dispatcher.handle, which closed every matched session on a pull_requestclosed event: the first PR of a series merging closed the linked issue's session, killed its tmux harness and deleted its checkout while the ticket was still open (_cleanup_workspace runs per closed session, so it inherited the same assumption). The existing test test_dispatcher_auto_closes_session_on_pr_close asserted exactly that behaviour, which is why it had never surfaced.
• Design choice that shaped everything: the webhook path is deliberately payload-only, and a PR-close payload cannot say whether the linked issue is still open — so the rule had to be decidable from the refs alone. "Only the object that closed ends" satisfies that, needs no config knob, and makes the lifecycle-close authz bypass strictly safer.
• Checkpoint/tests: none yet (spec phase).
• Next: implement T1–T10, TDD.
• Blockers: none. Tier 3 (human-approves-pr): spec + code approved together at the PR.

2026-07-25 — implemented (T1–T10)

• Phase: implementation → needs-review
• Did:
  • webhook/dispatcher.py: new _closing_refs(routed) — pure, payload-only, provider-agnostic; matches the number the payload's own issue/pull_request entity carries against routed.work_items. The close branch now skips a matched session whose ref is not in that set, logging it and emitting a new session.kept_open event. Nothing else in the branch moved (never spawn, never prompt, tmux + workspace per closed session).
  • eventlog.py: session.kept_open registered in EVENT_TYPES (the drift test enforces this; the-loop events --types documents it).
  • Tests, red before green (4 new/changed assertions failed against the old dispatcher, all pass after): test_routing.py — the linked issue's session stays active, a PR-ref session still closes, a second PR closing ends nothing, a malformed close payload closes nothing; test_workspace.py — the linked work item's worktree survives, the PR's own session still cleans up (the three existing cleanup tests re-pointed at the PR ref, which is the case they actually describe); test_tmux_runner_integration.py — a linked PR's close leaves the tmux session running even with keepSessionOnClose: false; integration — the work item's session survives its first PR merging (webhook) and a merged PR does not close its still-open work item (poll, guards the already-correct path against regression).
  • Guidance: execution-log template gained the Pull requests list; work-on/execute-tasks label and record every PR; finish-tasks verifies all of them and notes that closing the ticket is what ends the session; reference/automation.md, docs/capabilities/webhook-triggers.md, docs/capabilities/spec-workflow.md and cli/README.md carry the corrected end condition and the operational consequence (a PR merged without closing its ticket leaves the session active); decision-039 + index.
• Checkpoint/tests: make check — ruff, ruff format --check, pyright, scripts/validate_config.py, pytest (463 tests) and markdownlint, all green. Red→green evidence recorded in the PR briefing.
• Next: human review of the PR (tier 3).
• Blockers: none.

Review cycles

Cycle	Type (self/critic/security)	Reviewer	Outcome	Link
1	self	the-loop	Confirmed the poll path needed no change (reconciliation is registry- and ref-driven) and that closure_event's work_items=[ref] keeps the polled close inside the new predicate; added the malformed-payload test after re-reading _closing_refs for its failure mode	this log
2	self	the-loop	Checked the fallout for narrowed webhooks.ghWebhook.events lists: dropping issues now leaks GitHub-ticketed sessions; the receiver's existing startup/hot-reload warning already names that omission, so no new warning was added — documented in decision-039 § Consequences	this log
3	self	the-loop	Re-read the three existing workspace/tmux close tests: they described "the work item's PR" but registered the session against the issue; re-pointed them at the PR ref (the case they actually assert) and added the linked-PR counterparts rather than deleting coverage	this log

Security review (gate)

Required before ready-to-ship (security.review.required). See reference/security.md.

• Mechanism: the-loop checklist (security.review.mechanism: auto; the change is a removal of an auto-close path with no new ingress, credential, command or parsed field).
• Outcome: pass — strictly narrowing. The lifecycle-close bypass of the authorized-actor guard (issue-94) now only ever ends the session for the object that closed, so a hostile Closes #N in a PR body can no longer end #N's session (abuse case 1 of requirements.md). Close events are still never rendered into a harness prompt. The fail-closed direction is keep state: an unparseable close payload closes nothing.
• Human sign-off: n/a — risk tier 3, below security.review.humanSignOffMinTier (4).

Final validation evidence

• Red→green: with the dispatcher change stashed, the new assertions fail (test_pr_close_leaves_the_linked_issues_session_active, test_a_second_pr_closing_ends_nothing_while_the_work_item_is_open, test_a_malformed_close_payload_closes_nothing, test_dispatcher_pr_close_keeps_the_linked_work_items_worktree); with it applied the whole suite passes.
• AC coverage: AC1 (unit + webhook/poll integration), AC2 (existing issue-close tests, unchanged), AC3 (calls() == [] on the close), AC4 (PR-as-work-item close), AC5 (session.kept_open in EVENT_TYPES + the drift test), AC6/AC7 (workspace tests), AC8–AC10 (templates, commands, capability docs, references, decision-039).
• Gates: make check green (see the progress entry above).