feat: find threads by PR number in the command palette

[jonathanKingston]

Summary

Cmd/Ctrl+Shift+K now finds a thread by the GitHub PR it touched. Searching 2262, #2262, copse-dev/agent-pane#2262, or a pasted PR URL all narrow to the thread that opened #2262.

Why

The palette matched threads on title and project name only, so the thread behind a PR was unfindable by the number everyone actually refers to it by.

The data was already on disk. Thread.prRefs caches every PR a thread touched — scraped from message prose, recorded directly by gh_pr_create, and filled in from remoteAgentLink.prUrl for cloud-agent runs. It just never reached the palette: the palette reads catalog.jsonl, and ThreadCatalogEntry dropped the field.

What changed

Index the refs. ThreadCatalogEntry carries prRefs. catalogEntryFromDisk already reads the meta.json that holds them, so this costs no extra I/O and no transcript fold.

Match without a parser. PR keys are owner/repo#number, so 2262, #2262 and copse-dev/agent-pane#2262 all hit as plain substrings. A pasted PR URL is the one spelling that would not, so queryTerms folds it through the existing parseGithubPrUrl to that same shape — a link copied out of the PR pane and the bare number from the same PR should not disagree.

Show why a row matched. Thread rows render a #2262 chip; without it a bare-number query returns a title with no visible reason for the hit.

Migration, and the hazard it created

The new field is required, not optional, so readCatalog rejects every line of a pre-prRefs catalog. Those lines then look like missing threads to catalogMissingIndexedThreads, which reuses the existing rebuildCatalogFromDisk to backfill. Without that, every thread predating the field would stay silently unsearchable by PR number.

That rejection has a sharp edge worth reviewing closely: upsertCatalogEntry and refreshCatalogLine read the catalog and write the whole map back, so the first message append against an old file would have persisted a one-line index and dropped every other thread. Both now read through ensureCatalogMap, which rebuilds from thread dirs first. There is a regression test that appends a message against a stale two-thread catalog and asserts both survive.

Notes for reviewers

• Threads are still matched by title, so a title containing 2262 hits too. That is existing substring behaviour, not a regression.
• The @-thread mention picker shares loadProjectCatalog and picks up PR matching as well.
• matches() no longer re-splits the query per candidate row; terms are computed once per keystroke.
• refreshCatalogLine now costs one extra readdir per message append (via ensureCatalogMap), alongside the catalog read and write it already did.

Issue numbers are deliberately out of scope

Nothing stores a GitHub issue against a thread today — the only issue refs in the codebase live in roadmap note frontmatter (parseIssueRef), never on thread metadata. Covering issues needs a new persisted field, a scraper widened to /issues/, and a decision on bare #123, which GitHub itself resolves to either a PR or an issue. That is its own piece of work.

Validation

• npm run check — typecheck, lint, format, dead-code, oracle and e2e-syntax all clean. Unit suite: 8240/8262 passed. The 13 failures are the known host-flaky SSH/askpass/socket-path and workspace-indexing set (ssh-workspace/openssh-transport, ssh-workspace/ssh-paths, ssh-workspace/askpass, search/workspace-indexing); none of those files reference anything in this diff, and they reproduce identically in isolation.
• npm test -- thread-store — 83 passed
• npm test -- command-palette mention-picker — 19 passed

🤖 Generated with Claude Code

Co-Authored-By: Copse noreply@copse.dev
Copse-Models: acp:claude-acp#opus[1m]

[copse-release-bot]

Reference screenshots ready for review

Review GitHub’s image diffs in screenshot PR #2339.
Merge it (or enable auto-merge) to apply the accepted PNGs to copse/can-we-make-the-cmd-shift-k-search-also-fi-08afd2.

Rendered for 4a3af65aa29c by CI run 33860133909; the immutable artifact is reference-screenshot-candidates-33860133909.

If this source branch moves, a later successful render closes the stale review PR and replaces this link.

[jonathanKingston]

Review: find threads by PR number

Verdict: changes needed before merge. The migration design holds up under probing, but the branch now conflicts with main and the visible change has no screenshot evidence.

Validation

Check	Result
pnpm run typecheck	pass
oxfmt --check on the 7 changed files	pass
eslint on changed .ts files	pass
pnpm test -- thread-store command-palette mention-picker	102/102 pass
pnpm run check:dead-code	pass
Ad-hoc probe: pre-prRefs catalog with 3 threads, then deleteProjectThread, updateMeta({archivedAt}), updateMeta({title}), a mixed old/new file, and a malformed prRefs line	no entry lost in any case; rebuild persisted once
Dry merge with origin/main	CONFLICT in src/main/services/thread-store.ts and src/shared/types/thread.ts

Blocking

1. Rebase onto main. #2317 moved the store into packages/thread-store: src/main/services/thread-store.ts is now a 5-line re-export and ThreadCatalogEntry lives at packages/thread-store/src/thread-types.ts:296. The existsSync ? readCatalog : rebuild pattern this PR fixes is still present in the package copy (upsertCatalogEntry at packages/thread-store/src/thread-store.ts:876, refreshCatalogLine at :923), so the fix has to be re-applied there rather than in the shim.

2. Visual evidence. The #2262 chip (command-palette.ts:229-237, command-palette.css:72-80) and the new placeholder (command-palette.ts:118) are user-visible, but tests/ is untouched. tests/e2e/command-palette.e2e.ts seeds four threads with no prRefs and types login, so command-palette.png cannot show either. #2296 touches 12 screenshots and none is command-palette.png, so it is unrelated rendering drift, not evidence for this PR. seedThreadDir writes meta.json from the seed object verbatim, so adding prRefs to one seeded thread, searching by number, asserting .command-palette-pr, and saving a screenshot is a small addition.

Migration analysis

Every catalog reader-then-writer on the head:

• upsertCatalogEntry (:904) and refreshCatalogLine (:954) go through ensureCatalogMap. Safe.
• saveProjectThreads (:1331) writes wholesale from its argument by design. Safe.
• deleteProjectThread (:1631-1637) is the only remaining direct readCatalog + writeCatalog. Against an all-stale file readCatalog returns an empty map, the delete returns false and nothing is written; against a mixed file it writes only the new-format lines. Both self-heal because catalogMissingIndexedThreads checks every non-archived thread dir against the map. No data loss, but routing it through ensureCatalogMap would make the invariant uniform.
• A malformed prRefs line drops that line, which triggers rebuild from meta.json. Rebuild cost is meta.json plus parseSpine of events.jsonl for the first user message, pre-existing and one-off.

Required rather than optional is defensible: the catalog is derived and rebuildable, and optional-with-[] would leave every existing thread silently unsearchable by PR with no backfill. Older builds reading the newer catalog ignore the key and drop it on write; the newer build then rebuilds.

Non-blocking

• upsertCatalogEntry and refreshCatalogLine now run catalogMissingIndexedThreads on every append and meta patch: a readdirSync plus a meta.json read per archived thread. Projects with many archived threads pay that per write. Having readCatalog report rejected lines would keep the rebuild trigger precise without the scan.
• Substring matching means 2262 also hits #22620 and any title containing 2262, and rows sort by updatedAt with no preference for the exact PR. Anchoring numeric terms to the end of the key would fix it.
• Main-side loadProjectCatalog (:1899) does not fold PR URLs; only the renderer queryTerms does, so api.threads.catalog(id, url) from the mention picker returns nothing for a pasted link.
• queryTerms does not strip trailing punctuation from a pasted URL the way extractGithubPrUrls does, so .../pull/2262. stays a raw URL and matches nothing.

---

Generated by Claude Code

[jonathanKingston]

Pushed four commits. e19a22a merges main, with the PR's catalog changes re-applied inside packages/thread-store/src/ (prRefs on ThreadCatalogEntry, the parseCatalogPrRefs decoder, ensureCatalogMap routing for upsertCatalogEntry and refreshCatalogLine, the PR-key haystack in loadProjectCatalog); the original diff was saved and checked so nothing was dropped. 912857d routes deleteProjectThread through ensureCatalogMap too, with a stale-catalog delete test. faf1597 makes an all-digit or #digits term match #<n> only at the end of a key (so 2262 no longer hits #22620) in both the palette and main's loadProjectCatalog, and strips trailing punctuation from pasted URLs via a shared stripUrlTrailingPunctuation. 4a3af65 extends tests/e2e/command-palette.e2e.ts with a seeded prRefs thread, a 2262 search asserting the #2262 chip, and a new command-palette-pr-number.png screenshot (CI renders it; no PNG committed). Typecheck, eslint, oxfmt, dead-code, e2e-syntax, oracle and 296 targeted tests pass.

---

Generated by Claude Code

[github-actions]

🖥️ PR preview

• Marketing site
• Browser demo
  Interactive, with a mocked backend — jump to a scenario:
• Builds a cupcake site
• Agent-proposed edits open the Changes panel
• Markdown list indentation
• Responsive composer footer
• Subagent display visual reference
• Subagent stays ungrouped beside reading tools
• Semantic search subagent markdown
• Light-theme approval with a bright accent
• Grouped outside-sandbox command approval
• Remote desktop discovered-port list with one selected
• Settings scroll + sticky footer geometry
• Per-model reasoning / temperature / top-p
• Thread switch hydration notice
• Thread switch hydration failure
• Chat layout styling
  _{Static demo of the renderer against a mock API — the browser pane, terminal, and all main-process IPC are stubbed, so it never proves main-process behaviour. Rebuilt whenever a push changes the demo bundle; these links go live once the queued deploy lands, shortly after this comment. Removed when the PR closes.}