feat(ai-smart-collections-nav): add AI Smart Collections app menu item

[LukasHirt]

AI-generated · OSPO-71 · Gate: ✅ 1.00

Problem

Files land wherever a user drops them; there's no way to browse "all my invoices" or "all my contracts" across spaces without manual tagging.

Solution

Adds a "Collections" entry to the Application Switcher, listing AI-inferred thematic groups (e.g. "Invoices", "Contracts", "Meeting notes") built by clustering recent files' names and short excerpts via the LLM's structured {fileId, collection} output; clicking a group filters to those files. Degrades: no structured output → one collection label per line, parsed leniently; no tool use → a single batch pass, no iterative re-clustering; small context → smaller batches merged client-side.

Extension points

appMenuItem

Why ship this now

Purely read-only discovery — unlike a folder-reorganizing action, nothing moves, so it's low-risk to ship for users with flat, untagged storage.

What was built

Correction (post-review): this PR was originally drafted claiming a Files-app left-nav placement (app.files.navItems/sidebarNav), which an earlier version of this description flagged as an unresolved internal contradiction. That's now resolved: only an Application Switcher entry ships. A sidebarNav/app.files.navItems registration was tried during development and dropped — a live gate run confirmed the installed @ownclouders/web-pkg never renders that extension point, and sidebarNav isn't one of this repo's four documented extension types (see src/index.ts's comment and CLAUDE.md). The rest of this description has been corrected accordingly.

---

What was built

This adds web-app-ai-smart-collections-nav, a Files-app extension that surfaces AI-inferred thematic groupings of a user's recent files — e.g. "Invoices," "Contracts," "Meeting notes" — as a "Collections" entry in the Application Switcher (registered as an appMenuItem-typed extension), letting users browse by inferred topic instead of manually tagging or reorganizing anything. Clicking a collection filters down to just the files in that group; nothing is moved or modified, so the feature is purely a read-only discovery layer over existing storage.

The data path starts in useRecentFiles.ts, which fans out one WebDAV REPORT request per space, merges and sorts the results by modification date, caps the set to 100 files, and pulls capped text excerpts (1MB limit, text-extension allowlist only) for clustering input. useCollections.ts orchestrates the actual clustering: it batches files 30 at a time through sequential complete() calls against useLLM.ts (a non-streaming, per-package LLM wrapper matching the convention already used by the insights/brief sidebar extensions), then merges the resulting {fileId, collection} assignments across batches by fileId. Prompting and parsing are split into clustering-prompt.ts and parse-collections.ts, which support a strict JSON path (response_format: json_object, wrapped in an assignments array since the API requires a top-level object) with a lenient line-based fallback parser for models that don't return structured output.

On the UI side, CollectionsView.vue is the orchestrating component: it drives useRecentFiles/useCollections, handles loading/error/empty/consent-declined/grid/detail states, and gates the first clustering call behind a session-scoped consent flag (mirroring the existing useInsights.ts consent pattern) before any file content is sent to the LLM. It composes CollectionCard.vue (clickable summary card with file count and an icon badge), CollectionFileList.vue (a filterable, sortable file table modeled on the existing search results table), and ConsentDialog.vue (the disclosure prompt shown before the first LLM call).

Test coverage spans five unit spec files (86+ tests) covering the WebDAV REPORT/XML parsing, both clustering parser paths, batching/merge logic, the LLM wrapper's ready/unconfigured/cross-origin states, and the view's state transitions — plus four Playwright acceptance tests exercising the nav entry, end-to-end clustering with dynamically-derived mocked LLM responses (rather than hardcoded fileIds, since those are server-assigned), click-to-filter, and the lenient-parsing degrade path. Because the target extension point's real DOM rendering couldn't be inspected during development, the e2e nav-item locator matches by accessible name across either a link or button role rather than assuming a fixed element type. Batch size, recent-file cap, and file-size limits are all named constants rather than hardcoded magic numbers, and the README documents configuration, output format, and UI states for the extension.

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.

[dj4oC]

Adds web-app-ai-smart-collections-nav, a new opt-in extension that clusters a user's recent files into AI-inferred thematic groups via the existing LLM proxy, gated behind a session consent dialog.

Verified independently (not just from the diff): checked out the branch, ran pnpm install --frozen-lockfile (clean), pnpm --filter web-app-ai-smart-collections-nav check:types (clean), and pnpm --filter web-app-ai-smart-collections-nav test:unit (86/86 pass) — matches the PR's own gate table.

Findings

Security. No exploitable issue found. The client-side Origin check in useLLM.ts (blocks with status: 'cross-origin' before any request) is a reasonable belt-and-suspenders layer on top of the required server-side check in ai-llm-proxy, which this PR doesn't touch. Traced the consent gate end-to-end — no path calls the LLM before consent or bypasses denial. No secrets, no v-html/innerHTML anywhere (grepped the full diff), so even a fully adversarial LLM response can't produce more than a mislabeled collection card. Two nits: clustering-prompt.ts interpolates file excerpts into the prompt without escaping quotes (a prompt-injection surface, but not XSS given current rendering); and the new .gitignore carries a leftover comment about "stray debug artifacts" from the generation process — verified no such files actually exist in this checkout, so it's inert, but worth deleting the comment/ignore entries since they reference nothing real.

Stability.

• Resolved an internal contradiction in the PR description: src/index.ts registers only an appMenuItem (global App Switcher entry) — there is no app.files.navItems/sidebarNav registration, and per this repo's own CLAUDE.md, sidebarNav isn't even one of the four documented extension types. The code comment at index.ts:42-50 candidly explains that path was tried and dropped because it doesn't render in the installed web-pkg. This is the right call and matches the documented appMenuItem pattern (draw-io/group-management), but the PR title/description ("Nav Item", "Files navigation") oversells what actually shipped — worth correcting before it's used for changelog generation.
• useCollections.ts clusterFiles() (~L119-149): batches are merged into a local Map and only committed to collections.value after the entire sequential batch loop succeeds. If batch 2 of N throws, the catch block discards all already-clustered assignments from batch 1 — a user with >30 files gets a bare error with no partial results, verified in source.
• useRecentFiles.ts searchSpace() (~L171-187) correctly swallows a single space's REPORT failure and returns [] so one bad space doesn't block the rest — but fetchRecentFiles() never distinguishes "all spaces failed" from "genuinely no files," so a systemic WebDAV outage renders as the normal empty state rather than the error state the README promises for "recent-files listing failed."
• package.json pins @ownclouders/extension-sdk/web-client/web-pkg at 12.3.2, while every sibling AI extension (e.g. ai-data-insights-sidebar) is on 12.4.2 — the lockfile is internally consistent with this, so it's not broken, just drifted; worth bumping to match siblings before merge.

Performance. Nothing pathological (single-user, opt-in, read-only, every call individually timeout-guarded and error-isolated), but two real inefficiencies: fetchExcerpt() in useRecentFiles.ts downloads the entire file body (up to 1MB, no Range header) purely to keep the first 200 characters (MAX_EXCERPT_CHARS) — a ranged read would cut this by orders of magnitude. Also, CollectionsView.vue re-runs the full space fan-out + excerpt fetch + up to 4 sequential LLM batches from scratch on every mount, with no session-scoped result cache (unlike the consent flag, which is already cached that way) — a user revisiting the tab a few times pays the full IO+LLM cost each time.

Test coverage: GAP. The claimed 86 unit tests are real and substantive for parsing (both strict-JSON and lenient-fallback paths), the LLM wrapper's state machine, and per-space partial-failure handling — verified by reading the spec files, not just the PR description. However, the consent-gating flow — the actual privacy control deciding whether file names/excerpts reach an LLM — has no test in either direction: CollectionsView.spec.ts explicitly routes around it (own comment: avoids the module-scoped consent flag because it "isn't exposed for resetting"), and the e2e suite only ever exercises the confirm path, with no assertion that zero LLM calls happen before consent or on denial. The consentDeclined UI state also has no test anywhere. Per this pipeline's testing requirement, here's a ready-to-apply e2e addition to close the gap:

test('does not call the LLM before consent, not at all on denial, and only after confirming', async () => {
  await mockRecentFilesResponse(adminPage, SEED_FILES)
  let completionsCalls = 0
  await adminPage.route('**/chat/completions', async (route) => {
    completionsCalls++
    await route.fulfill({
      status: 200,
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({ choices: [{ message: { content: JSON.stringify({ assignments: [] }) } }] })
    })
  })

  const collections = new CollectionsViewPage(adminPage)
  await collections.openViaAppSwitcher()
  await expect(collections.consentDialog).toBeVisible()
  expect(completionsCalls).toBe(0) // nothing sent before consent

  await collections.consentDialog.getByRole('button', { name: 'Cancel' }).click()
  await expect(adminPage.getByText('Grouping was cancelled. No file data was sent to the AI service.')).toBeVisible()
  expect(completionsCalls).toBe(0) // denial must never trigger the call

  await adminPage.getByRole('button', { name: 'Group my files' }).click()
  await expect.poll(() => completionsCalls).toBeGreaterThan(0) // confirming does trigger it
})

Deployment topology check

N/A — client-side web extension bundle only, no service/deployment-topology impact. Root-level config changes (docker-compose.yml, dev/docker/ocis.apps.yaml, support/actions/ocis.apps.yaml, .github/workflows/test.yml) were checked against this repo's two documented (and differing) key conventions and are correctly wired.

Verdict

Commenting — nothing here rises to a security block or data-loss risk (this is a read-only, opt-in, retry-recoverable feature), but the consent-gating test gap, the batch-failure data loss, and the masked-outage empty state are worth addressing before or shortly after merge.

🤖 Automated review by Claude Code (security · stability · performance · coverage)

---

Generated by Claude Code

[LukasHirt]

Addressed the review findings above in four commits on this branch:

• Prompt injection (clustering-prompt.ts): file names/excerpts are now sanitized before interpolation into the LLM prompt, so an embedded " can no longer break out of the quoted field.
• Batch-failure data loss (useCollections.ts): collections.value is now committed incrementally after each batch, so a later batch's failure no longer discards earlier batches' results. CollectionsView.vue's template now renders the grid alongside an inline error notice instead of replacing it entirely.
• Masked outage (useRecentFiles.ts): per-space failures are now tracked via Promise.allSettled; when every space fails and nothing was found, a distinct error surfaces instead of the normal empty state. Per-space resilience when only some spaces fail is unchanged.
• Excerpt over-fetch (useRecentFiles.ts): fetchExcerpt() now sends a Range: bytes=0-1023 header instead of downloading the full file body.
• Consent-gating test gap: extracted the session consent flag into a new useConsent.ts composable (mirroring ai-data-insights-sidebar's useInsights.ts pattern) with test-only reset/give helpers. Added unit and e2e coverage proving zero LLM calls happen before consent or on denial, and exactly one call after confirming (incorporating your suggested e2e test).
• Dependency drift: bumped @ownclouders/web-client/web-pkg/extension-sdk from 12.3.2 to 12.5.0 to match every other package in the monorepo; pnpm install, check:types, and test:unit all pass clean.
• Placeholder description: package.json's "description": "CANDIDATE" is now a real description.
• Stray .gitignore: removed (referenced two paths that don't exist anywhere in the tree; no sibling package has a package-level .gitignore either).
• PR scope: retitled/redescribed this PR and corrected the index.ts comment — only an appMenuItem (Application Switcher entry) ships; the Files-app left-nav placement was tried and dropped, as your review noted.

Deferred: result caching across CollectionsView.vue mounts is real but lower-priority (no correctness/security impact) — filed as #504 for follow-up rather than bundled into this PR.

Verification: check:types, lint, test:unit (99 tests), and test:e2e (15 tests × 3 browsers, including the new consent-denial test) all pass against a live oCIS stack.