fix(server): show Claude project skills before command discovery completes#23114

Problem

After a Claude thread's workspace snapshot is evicted by scans of 16 other folders, its $ picker temporarily shows only personal skills. Project skills wait for a fresh Claude capability probe even though filesystem skill discovery has already finished.

Change

Publish discovered Claude workspace skills before the capability probe, with slash commands pending. Partial and final results use the same registry ownership and freshness guards. The final result replaces its own partial publication; the 16-workspace cap remains.

Other providers keep their existing one-result discovery path. No wire schema or client rendering changes.

Scope and approval

Fixes #22429 GitCafe: https://git.cafe/pingdotgg/t3code/issues/22429 Triaged bug via the automated issue flow. One reproduced cache-eviction timing defect; no authentication, permission, mode, sandbox, or MCP changes.

Verification

• Main reproduction: 17 real worktrees of one repository, each with a project-only Claude skill. Returning to the first after scanning the other 16 temporarily showed personal skills only; repeated with a filtered skill query.
• Red test committed before the fix: the registry had no project skills while a Deferred held command discovery after eviction. After the fix, the same test observes skills with commands pending, then final commands and cleared pending state.
• Post-merge vp run check:changed: formatting, targeted lint, Knip, generated GraphQL verification, affected-package typechecks, and the full registry test file (38/38) passed.
• Additional implementation checks: Codex/Antigravity discovery test files (29/29), Pi driver tests (5/5), server and provider-core typechecks passed.
• Two isolated code reviews returned MERGE. The actual PR diff after merging current main is byte-identical to the reviewed diff.

Evidence matrix

Before: beec3c8b30ce. After: 26c41f4c7fa8.

Formal static render matrix: empty ([]). The changed paths are server/provider internals; the four rows below are supplemental interaction recordings. No still-image/pixel-diff claim is made. The pass-1 matrix and audit are in the linked support bundle; independent pass-2 proof audit passed all six required evidence rows.

Script and observed ordering

Same 17 stopped fixture threads and project skill on both isolated servers. A is folder 02. Prime A, type $triage, visit 03–17 then 01 sequentially and await each real settled snapshot, prove A absent with 16 cached folders, return A within 15–21 seconds, keep $triage visible through discovery, then show / and query /compact. No turns sent. The actual /compact query suggests /autocompact on both sides; the unchanged final provider snapshot contains 70 command names.

Before: the project row is missing while the personal skill remains, then appears only after final command publication. After: the project row becomes visible with slashCommandsPending:true, 448–744 ms before final publication in these runs. A brief filesystem-discovery gap remains; this eliminates the slow CLI interval, not every millisecond of absence. All recorded snapshot lists stay at or below 16.

Rebuild, transcript and audit metadata

Rebuild guide · Support bundle: scripts, fixture creation data, pass-1 audit/matrix, provenance, probe spans, red/green logs · Ordered live runtime transcript · Capture audit metadata

The transcript subscribes to the real appAtomRegistry / serverEnvironment.providersValueAtom and observes actual DOM mutations. It records cwd, skill/pending/count milestones and the final real command names. Server workspace-probe spans are correlated by timestamp; they lack cwd attributes and an independent CLI subprocess span, so no exact CLI launch/completion claim is made. Parent-provided green log labels the final head and reports 38 focused tests passing. Historical red output lacks an in-band invocation/SHA; its provenance limit is retained in metadata.

What else changed and rendering

The intended visible difference is the transient project row during discovery. Final command names match across all eight runs. Live age badges differ with capture time, and caret/text selection plus recording frame sampling vary while typing; no pixel-identity claim is made. The same responsive route navigation, viewport, theme, fixture, provider and scroll position are used per pair. Sidebar filtering hides unrelated threads; update notifications were dismissed. A02 avoids a separate active A01 client refreshing its own evicted cwd.

Tool/browser: T3 collaborative Browser, Chrome 154 on Linux. No device emulation; freeform 1440×900 or 390×900 CSS viewport, DPR 2, visual viewport scale 1, root font 16px, text-size-adjust 100%. Complete native viewport footage is 2880×1800 or 780×1800. A 240-native-pixel caption band sits outside the app; the original recording is uncut and retains its speed, with a labelled four-second final-frame hold. Captions narrate script and browser-clock milestones; compositor/video frames are not perfectly synchronized to DOM subscription timestamps. No artificial provider delays, fake snapshots or synthetic component fixtures were used.

Served Vite source-map originals for five relevant runtime/client files per checkout match the exact git refs. Server cwd/start times and observed partial/final behavior support runtime provenance; no in-process build hash endpoint exists. The before server predates the merge-base checkout, but all four changed runtime files are byte-identical between its original base and the final base.

Limits and publication

Native iOS/Android, desktop shell, remote/relay and multi-device runtime timing remain unverified. Clean-room reseeding was not executed: rebuild scripts were syntax/source checked and captures reused the parent’s fixture. Personal skill source files and unrelated data are excluded. Independent pass-2 proof audit passed; this is evidence for review, not merge approval.

Before upload, both required cafe issue reads succeeded with zero comments and complete pagination; no recent MEMBER/OWNER comment existed. No repository edits, commits, pushes, issue/PR writes, new dev servers or agent turns were made by this capture task.

Unverified: Electron shell, native iOS/Android, remote/relay/tunnel, multi-device, cancellation and instance replacement at runtime. Their code paths are unchanged. Filesystem discovery still takes its normal time; this removes the wait on the slow command probe.

Implementation: Claude Opus 5.5 via Claude Agent SDK; triage, integration and evidence: GPT-6.1-Sol via Codex; independent code/proof reviews: GPT-6.1-Sol via Codex and Claude Opus 5.5 via Claude Agent SDK.