[martin-conur]

Summary

On jira repos the task-monitor showed the issue key twice (ML-450 ML-450: …), and when a PR had no ## Summary it used the JIRA: <url> line as the description. Now a leading key is dropped from PR titles, and the summary fallback finds the prose after Closes ML-450 and skips paragraphs that are only links or Closes lines.

Closes #299

Changes

• lib/monitor-assemble.jq: new strip_label, applied to in-flight PR titles and done items. It drops KEY:, [KEY], key - and KEY —, ignoring case. Only a whole label counts (ML-4500 is left alone), and a title that is only the label is kept. pr.title stays raw. gh labels are #N, so gh titles are unchanged, and describer/queue titles are not touched.
• lib/monitor-text.sh: the Closes fallback also matches a tracker key (generic [a-z][a-z0-9_]*-[0-9]+, lowercased) and a /browse/KEY URL. Both fallbacks use the new tm_first_prose, which skips paragraphs that are only a URL, a markdown link, Label: <url>, or a single Closes …. With nothing left, the description is null.
• The shared snapshot fixture's PR body was only Closes #5, which now gives a null summary, so the scenario and snapshot-reviewer.json gained a prose line (only types are compared, and pr.summary stays a string).
• Docs: docs/task-monitor-schema.md, README monitor paragraph, CHANGELOG.

Tests

New cases in tests/task_monitor_snapshot.bats: four title forms dedupe against PROJ-7 and PROJ-70 is kept; in-flight jira dedupe with pr.title raw; gh Fix #4: … untouched; JIRA: <url> then prose gives the prose; a link-only body gives null; tm_pr_summary on Closes KEY, Fixes <browse URL>, a link-only body, and ## Summary still winning.

Mutation-checked, and each one turned at least one new test red: no dedupe, a case-sensitive match, no separator boundary, the old #N-only Closes regex, no link filter. Full suite: 1640/1640, plus check-drift.sh and shellcheck -x. The awk was also checked under mawk 1.3.4 (Ubuntu 24.04).

🤖 Generated with Claude Code

[martin-conur]

Spec compliance

All four fix items and the acceptance list are delivered:

• Title dedupe is done once, in lib/monitor-assemble.jq (strip_label), for in-flight and done rows. It covers KEY:, [KEY], key - and KEY —, ignores case, requires a whole label (ML-4500 is kept), and leaves a label-only title alone. pr.title stays raw.
• The Closes fallback now matches Closes ML-450, Fixes PROJ-1 and a /browse/KEY URL.
• Both fallbacks skip link-only, Label: <url> and Closes …-only paragraphs, and give null when nothing is left. ## Summary still wins.
• Tests, mutation-check notes, CHANGELOG, README and schema docs are all in. CI ran on the head SHA (1 run). I ran ./run_tests.sh task_monitor_snapshot locally: 70 tests, 0 failures. tools/check-drift.sh and shellcheck -x lib/monitor-text.sh are clean. The PR body opens with ## Summary.

Code-review findings

No blockers. One nit worth fixing in this PR:

• Over-eager junk() in lib/monitor-text.sh. The Closes … test is ^(close…|fix…|resolve…):?[ \t]+[^ \t]+$, so any one-word-object sentence is dropped as junk. I confirmed Fixes bug as a whole paragraph now yields an empty description, where before it was the description. Tightening it to require a key, #N or URL-shaped token (the same patterns the found matcher uses) avoids discarding genuine short prose. Please add a test line for it.

Checked and fine: Closes the gap in retry handling and Fixes: the crash on start are still kept; ascii_downcase preserves length, so the slice indexes in strip_label are correct; null titles pass through; the fixture change (Closes #5 → prose) follows from the new rule.

Verdict

clean-with-nits: tighten junk()'s Closes pattern (and add a test); nothing else blocks.

[martin-conur]

PM: one round before merge

CI is green. Two small items, both in this PR:

1. Reviewer's nit: junk() is over-eager. ^(close…|fix…|resolve…):?[ \t]+[^ \t]+$ drops any one-word object, so a whole-paragraph Fixes bug becomes no description. Require the object to be a #N, a jira key or a URL (the same shapes the found matcher uses), and add a test that Fixes bug is kept while Closes ML-450 / Closes #5 / Closes https://…/browse/ML-450 are still skipped.
2. PM mutation gap: a title that is only the key. I changed if $r == "" then . else $r end to plain $r in strip_label (so ML-450 / ML-450: / [ML-450] become an empty title) and every test stayed green. Add those three cases to the done-jira title test, asserting each title is kept as it was.

Mutation-check both (revert each fix and watch the new assertion go red), then re-review-requested.