feat(ai-project-board-view): add AI Project Space Status Board

[LukasHirt]

AI-generated · OSPO-70 · Gate: ✅ 1.00

Problem

Project spaces render as a flat grid, so a manager can't tell which deliverables are draft, in review, or final without opening every folder.

Solution

Registers an alternate project-space folder view: a three-lane board (Draft / In Review / Final) built from the same file list already loaded. The LLM classifies each file's lane from its name plus a short excerpt, returned as structured {fileId, lane} pairs. Degrades: no structured output → one lane-per-file text line, pattern-matched; no tool use → single classification pass on load, no re-run button; small context → excerpt dropped to filename only.

Extension points

app.files.folder-views.project-spaces

Why ship this now

Managers today infer status from naming conventions alone; this surfaces it visually for free from data the view already fetches.

What was built

The extension is web-app-ai-project-board-view, which adds an LLM-classified status board as an alternate view mode for project-space folders. Rather than the flat file grid, it groups the same resource list already loaded into the view into three lanes — Draft, In Review, Final — inferred from each file's name and a short content excerpt. It's registered in src/index.ts as a folderView extension; note that despite the spec targeting app.files.folder-views.project-spaces, the implementation actually registers on app.files.folder-views.folder (documented with rationale in the package README).

Classification is driven by composables/useBoardClassification.ts, which reads activeResources from useResourcesStore(), batches up to 60 files per prompt, and asks the BYO-LLM endpoint (via useLLM().complete(), configured through applicationConfig.llm) to return structured {fileId, lane} JSON. composables/useExcerpt.ts supplies a capped 2KB text excerpt per file over WebDAV (Range-header fetch), skipping folders, non-text mime types, and oversized files so filename-only classification kicks in when content isn't available. Presentation is split across BoardCard.vue (icon, name, last-modified), BoardLane.vue (header with icon/label/count plus a scrollable card list and empty state), and ProjectBoardView.vue, which orchestrates loading/empty/error/cross-origin/unconfigured states, a truncation notice when the 60-file cap is hit, and a manual "Re-run classification" button.

The design leans on graceful degradation rather than hard failure: if the model doesn't support structured output, utils/lane.ts falls back to a regex line-parser (laneFromKeyword/parseLaneLines) that reads "<name>: <lane>"-style text, and any file the classifier can't confidently place defaults to Draft. useLLM.ts gained a responseFormat option (wired to response_format in the request body) to request JSON output where the endpoint supports it, reusing the existing same-origin check and bearer-token completion path from the BYO-LLM composable pattern. Unit coverage (55 tests) exercises the excerpt fetch/truncation boundaries, both JSON and malformed-JSON classification paths, the full LLM error-code-to-panelError mapping, and all view states. The end-to-end Playwright test drives the real view-mode switch and lane placement for uploaded fixture files, then re-runs classification after a fourth upload — its LLM response is mocked by intercepting the proxy call and deriving a realistic non-JSON reply from the actual outgoing prompt, deliberately exercising the fallback line-parser instead of depending on real oCIS file IDs.

Gate

Check	Result
Hygiene	✅ ok
Build	✅ ok
Lint	✅ ok
Unit tests	✅ ok
E2E tests	✅ ok
Score	1.00

---

_{Effort: M · 🤖 Generated by extctl}

[kw-security]

✅ Snyk checks have passed. No issues have been found so far.

Status	Scan Engine	Critical	High	Medium	Low	Total (0)
✅	Open Source Security	0	0	0	0	0 issues
✅	Licenses	0	0	0	0	0 issues
✅	Code Security	0	0	0	0	0 issues

💻 Catch issues earlier using the plugins for VS Code, JetBrains IDEs, Visual Studio, and Eclipse.