Execution Log: an event on a PR routes to the PR's linked issue first

Append-only log of progress for the user's visibility. Checked in alongside the spec at docs/specs/issue-93/.

Phase transitions

Phase	Entered	Reviewed/approved by	Notes
requirements-definition	2026-07-25		Issue #93: a PR event must resolve the linked GH issue first and reuse that issue's existing tmux session.
design	2026-07-25		One shared seam (webhook/router.py: extract_work_items) + one provider enrichment (poller/github.py: closingIssuesReferences in the existing gh pr list). Ordering linked-issues-before-PR is the whole mechanism.
tasks-breakdown	2026-07-25		12-task DAG
implementation	2026-07-25		Implemented on claude/github-issue-93-uy0xtc
needs-review	2026-07-25		PR opened; awaiting human review (tier-3, human-approves-pr)
complete

Progress entries

2026-07-25 — spec drafted

• Phase: requirements → design → tasks
• Did: Traced both ingress paths end to end (Router.route → Dispatcher.handle / _on_unmatched, and Poller._process_item → GitHubPollProvider.refs → extract_work_items). Confirmed three cooperating defects rather than the one in the title:
  1. GitHub delivers a PR conversation comment as issue_comment with a pull_request key on the issue object; extract_work_items treated it as a plain issue event, so it resolved to the PR's number alone — with the auto-execute label on the PR that is an unmatched labelled event, i.e. a spawn.
  2. PR→issue linkage was heuristic-only (head branch + a bare Closes #N), so GitHub's own Development-panel linkage, OWNER/REPO#N, GH-N and issue URLs were all invisible.
  3. _on_unmatched spawns for work_items[0], and the PR number was first — so even a resolved linkage still keyed a new session to the PR. Drafted the 3-phase spec; the design settles on reordering at extraction (rather than teaching the dispatcher a second notion of "the ticket") because the poller's own has_session check calls provider.refs, not the dispatcher.
• Next: implement T1–T12.
• Blockers: none. Human approval of the spec + code happens at the PR (tier-3, human-approves-pr).

2026-07-25 — implemented (T1–T12)

• Phase: implementation → needs-review
• Did:
  • webhook/router.py: _pr_entity() (a pull_request* event, or an issues/issue_comment event whose issue carries pull_request); linked_issue_numbers() resolving closingIssuesReferences → head branch → body keywords, deduplicated, never the entity itself; a widened _CLOSING_KEYWORD_RE (#N, OWNER/REPO#N, GH-N, issue URL, Fixes: ) with a cross-repo qualifier guard; extract_work_items emitting linked issues before the PR's own number.
  • poller/github.py: closingIssuesReferences appended to the existing gh pr list --json field list (no extra API call), a latched one-time downgrade + warning when the installed gh rejects the field (any other gh failure still propagates), GhItem.linked_issues, WorkItem.raw["linkedIssues"], and injection into the synthesised webhook-shaped payload so the router keeps one code path.
  • Tests: 8 router unit cases, 4 provider/gh-client cases, and one integration test per ingress path (test_webhook_routing_integration.py, test_poller_integration.py) with Gherkin docstrings + requirement links.
• Evidence: both integration regressions fail before the change (git stash of the two source files → 2 failed) and pass after; full gate green from the repo root: ruff check + markdownlint (209 files), ruff format --check, pyright (0 errors), pytest 332 passed, validate_config.
• Docs: docs/capabilities/webhook-triggers.md and interactive-sessions.md (current behaviour + history rows), skills/the-loop/reference/automation.md, new docs/decisions/decision-036.md.
• Next: PR + reviewer briefing; await human approval (tier 3).
• Blockers: none.

Reviews

Round	Type	Findings	Resolution
1	self	GH- alternative reachable only if the # alternative fails first; confirmed by the parametrised test. Cross-repo references must be dropped, not renumbered — added the qualifier guard + tests.	resolved in the implementation
2	self	Ordering change must not alter PR-close cleanup: _is_pr_close closes all matched sessions and the issue ref was already present for pull_request events — unchanged.	no change needed
3	self	The _no_link_field latch is per-GhClient, so a hot reload rebuilds it and pays one failed call again; self-healing and cheaper than global state. Downgrade is gated on the error naming the field so an auth/network fault still surfaces.	accepted as designed

Security review

Per security.review (tier 3 → no named human sign-off required; humanSignOffMinTier: 4). No new attack surface: the change only decides which work-item ref an already-authorized event resolves to. All newly-read payload fields are parsed into integers, formatted into a WorkItemRef whose slug is already regex-sanitised before it reaches a filesystem path or tmux target. The self-reply marker guard and the routing.authorizedUsers guard still run before dispatch, unchanged and in the same order. No new subprocess, network call or file; the gh invocation is the existing one with one more --json field, run as an argv list (no shell). One narrowing: a closing reference naming another repository is now dropped instead of being read as this repository's issue number. Degradation on an old gh falls back to today's conventions with a warning — never wider routing, never a failed cycle.