lfg2025/archy
1
0
Fork 0
Code Issues 24 Pull Requests 4 Actions Packages Projects Releases 72 Wiki Activity
Files
main
archy/aiui/loop/hardening-plan.md
T
archipelago 7ba3109b6d Add 'aiui/' from commit 'e30ac1d1069532fb6d652d87e2d4a2fe9d1b4773' 
git-subtree-dir: aiui
git-subtree-mainline: 0c4826f8cc
git-subtree-split: e30ac1d106
2026-08-03 15:07:11 -04:00

14 KiB
Raw Permalink Blame History

AIUI Debug, Fix & Hardening Plan

Execution Rules

  • Run on development branch
  • Each - [ ] task is one agent iteration — complete fully before moving on
  • After each phase: pnpm typecheck && pnpm lint && pnpm test -- --run
  • If checks fail, fix before proceeding
  • Commit at end of each phase with type(scope): description format
  • Do NOT push — human will review and push

PHASE 0: Baseline Verification

  • P0.1 — Baseline check
    • Run: pnpm install && pnpm typecheck && pnpm lint && pnpm test -- --run
    • Record output. Note existing failures. Fix any blockers before proceeding.
    • Commit: chore(app): verify baseline before hardening

PHASE 1: Critical Security Fixes

  • P1.1 — Fix CORS wildcard in claude-proxy.ts

    • File: packages/app/server/claude-proxy.ts
    • Replace all 'Access-Control-Allow-Origin': '*' (lines ~415, 473, 531) with the correct localhost origin or import ALLOWED_ORIGIN from dev-auth.ts. Match the pattern already used on lines 294/303.
    • Verify: grep -n "Allow-Origin.*\*" packages/app/server/claude-proxy.ts returns zero matches.

  • P1.2 — Fix dev auth token bypass

    • File: packages/app/server/dev-auth.ts
    • Line 11: if (!token) return true skips auth entirely when token empty.
    • Fix: Only skip in non-production. if (!token) { if (process.env.NODE_ENV === 'production') { res.writeHead(401); res.end('Unauthorized'); return false; } return true; }
    • Add console.warn when auth disabled.

  • P1.3 — Symlink traversal protection in vite-fs.ts

    • File: packages/app/vite-fs.ts
    • In walk function: after constructing fullPath, add if (lstatSync(fullPath).isSymbolicLink()) continue;
    • In handleRead: before statSync, check lstatSync(filePath).isSymbolicLink() → return 403.
    • Import lstatSync from fs.

  • P1.4 — Body size limit on vite-fs.ts mkdir

    • File: packages/app/vite-fs.ts
    • handleMkdir reads body with no size cap (lines 185-209).
    • Add const MAX_BODY_SIZE = 1024. Track let size = 0 on data events. Return 413 if exceeded.

  • P1.5 — JSON schema validation in vite-dev-chats.ts

    • File: packages/app/vite-dev-chats.ts
    • After JSON.parse(body) on line 66, validate shape: must be object with optional conversations (object) and activeConversationId (string|null). Reject with 400 if invalid.

  • P1.6 — CSP headers in nginx-archy.conf

    • File: packages/app/server/nginx-archy.conf
    • Add inside /aiui/ location block: 
      add_header Content-Security-Policy "default-src 'self'; script-src 'self'; style-src 'self' 'unsafe-inline'; img-src 'self' https: data:; connect-src 'self' https://api.anthropic.com https://openrouter.ai https://wavlake.com https://itunes.apple.com https://openlibrary.org https://covers.openlibrary.org https://en.wikipedia.org https://www.googleapis.com https://image.tmdb.org; frame-ancestors 'self'; base-uri 'self'; form-action 'self';" always;
add_header X-Content-Type-Options "nosniff" always;
add_header X-Frame-Options "SAMEORIGIN" always;
add_header Referrer-Policy "strict-origin-when-cross-origin" always;

  • P1.7 — OpenRouter validation + streaming timeout

    • File: packages/app/server/claude-proxy.ts
    • OpenRouter handler: parse reqBody, validate has model (string), messages (array), stream (boolean). Return 400 if invalid.
    • Both streaming reader loops: add idle timeout (120s). let idleTimer = setTimeout(() => reader.cancel(), 120000). Reset on each chunk. Clear on completion.
    • Verify: pnpm typecheck
    • Commit: fix(app): critical security — CORS, auth, symlink, CSP, body validation

PHASE 2: Proxy & Server Hardening

  • P2.1 — Fix tool use loop in claude-proxy.ts

    • File: packages/app/server/claude-proxy.ts (lines 164-222)
    • Unknown tool names: push error tool_result { type: 'tool_result', tool_use_id: tu.id, content: 'Error: unknown tool', is_error: true }.
    • Add turnMessages size guard: if (JSON.stringify(turnMessages).length > 500_000) break loop.
    • Add AbortSignal.timeout(30000) on each API call within loop.

  • P2.2 — SSRF mitigation in vite-rss.ts

    • File: packages/app/vite-rss.ts
    • After tryParseFeed returns, validate returned article URLs with isPrivateUrl. Filter out any private URLs.
    • Add comment documenting residual TOCTOU risk (dev-only middleware).

  • P2.3 — Sanitize web search context in useAI.ts

    • File: packages/app/src/composables/useAI.ts (lines 399-408)
    • Create helper: function sanitizeSearchText(s: string): string { return s.replace(/[[\]()#*_~ >]/g, '\$&').replace(/\n/g, ' ').slice(0, 500) }`
    • Apply to r.title and r.content in formatWebSearchContext.
    • Commit: fix(app): proxy hardening — tool loop, SSRF, prompt injection

PHASE 3: State & Logic Bugs

  • P3.1 — Fix contentType assignment in useContentPanel.ts

    • File: packages/app/src/composables/useContentPanel.ts (lines 219-225)
    • The contentType ref type is 'film' | 'song' | 'podcast'. Expand to include all content types or use ContentTab. Fix assignments: books→'film' is acceptable if type can't expand, but news/TV should map correctly. If the type is only used for player logic, keep narrow type but fix assignments to be semantically correct.

  • P3.2 — Fix apiFetching race in useBannerFallback.ts

    • File: packages/app/src/composables/useBannerFallback.ts
    • Change let apiFetching = false to const apiFetching = ref(false). Update all references to .value.

  • P3.3 — Fix event listener cleanup in chat.ts

    • File: packages/app/src/stores/chat.ts (lines 111-118)
    • Extract handlers to named functions. Add import.meta.hot?.dispose() cleanup for HMR.

  • P3.4 — Fix news RSS race condition in useContentPanel.ts

    • File: packages/app/src/composables/useContentPanel.ts (lines 134-154)
    • Capture const tabAtStart = activeTab.value before RSS fetch. In .then(), only set activeTab.value = 'news' if activeTab.value === tabAtStart (user hasn't manually switched).

  • P3.5 — Add error logging to silent catch blocks

    • Files: useImageFallback.ts, useAI.ts, stores/chat.ts
    • Replace all catch { } and catch { /* ignore */ } with catch(e) { console.debug('[module] op failed:', e) }.
    • Keep console.debug (not error) for expected failures.
    • Commit: fix(app): state bugs — contentType, race conditions, error logging

PHASE 4: Content Extraction & Filtering

  • P4.1 — Add recipe instructions to system prompt

    • File: packages/app/src/composables/useAI.ts (SYSTEM_PROMPT)
    • Add after Apps section: 
      **Recipes:** When sharing recipes, use the <recipe_ext> XML tag:
<recipe_ext title="Name" servings="4" time="30 min" calories="450">
- ingredient 1
1. Step one
</recipe_ext>

  • P4.2 — Fix app category validation

    • File: packages/app/src/composables/contentExtraction.ts (line ~1330)
    • Add const VALID_CATEGORIES = new Set(['nostr-client','lightning-wallet','bitcoin-wallet','privacy','node','dev-tool','relay']). Validate before cast, fallback to 'dev-tool'.

  • P4.3 — Fix song deduplication

    • File: packages/app/src/composables/contentExtraction.ts (lines 606-638)
    • After combining library + external songs, deduplicate by normalized title|artist key. Library songs (with real IDs) take priority.

  • P4.4 — Fix film/TV tag ambiguity

    • File: packages/app/src/composables/contentExtraction.ts (lines 914-916)
    • Only convert film_ext to TV when: query is TV-like AND no explicit tv_ext tags present AND response reads TV-like. If AI used both film_ext and tv_ext, keep both as-is.

  • P4.5 — Fix isBookLikeResponse false positives

    • File: packages/app/src/composables/contentFiltering.ts (lines 60-62)
    • Increase by [Author] threshold from >= 1 to >= 2.

  • P4.6 — Fix isRecipeLikeResponse for prose recipes

    • File: packages/app/src/composables/contentFiltering.ts (lines 92-94)
    • Add prose detection: recipe keywords + list structure pattern.

  • P4.7 — Fix looksLikeSong false positives

    • File: packages/app/src/composables/contentExtraction.ts (lines 488-506)
    • Add financial terms to blocklist: 'market cap','etf','price','trading','volume','earnings','valuation','stock','portfolio','investment','yield','inflation','interest rate'.
    • Commit: fix(app): extraction — dedup, tag ambiguity, false positives, recipes

PHASE 5: Memory & Cache Hardening

  • P5.1 — Bound caches in useImageFallback.ts

    • File: packages/app/src/composables/useImageFallback.ts
    • Create helper: function boundedSet<K,V>(map: Map<K,V>, key: K, val: V, max=500) { if (map.size >= max) { const first = map.keys().next().value; if (first !== undefined) map.delete(first); } map.set(key, val); }
    • Replace all .set() calls on memory caches with boundedSet().
    • Also bound failedUrls Set to 1000 entries.

  • P5.2 — Bound caches in usePlayer.ts + cleanup

    • File: packages/app/src/composables/usePlayer.ts
    • Bound resultCache and nullCacheTimestamps to 200 entries.
    • Add destroy() method that calls destroyPlayer(), resets DOM refs, cancels active search controller.
    • Commit: fix(app): bound caches, player cleanup

PHASE 6: Accessibility

  • P6.1 — Fix touch targets and aria attributes
    • Audit all *Grid.vue and *Detail.vue for: interactive elements < 44px, images without alt, SVG fallbacks without aria-label.
    • ImageGrid.vue: add role="img" aria-label="Image unavailable" to fallback SVG wrapper.
    • Ensure all interactive genre/tag pills in grids have adequate touch targets (min 44x44 including padding).
    • Non-interactive info pills are exempt.
    • Commit: fix(app): accessibility — touch targets, alt text, aria

PHASE 7: Security Tests

  • P7.1 — Write dev-auth tests

    • New file: packages/app/src/__tests__/dev-auth.test.ts
    • Tests: auth bypass when no token (dev), auth required when token set, rate limiting 429, CORS headers correct.

  • P7.2 — Extend proxy tests

    • File: packages/app/src/__tests__/proxy.test.ts
    • Tests: unknown tool error result, max rounds termination, turnMessages size guard, streaming timeout.

  • P7.3 — Write vite-fs security tests

    • New file: packages/app/src/__tests__/vite-fs.test.ts
    • Tests: path traversal rejected, symlink blocked, sensitive files blocked, body size limit enforced.

  • P7.4 — Write SSRF tests for vite-rss

    • New file: packages/app/src/__tests__/vite-rss.test.ts
    • Tests: private IP detection, localhost blocked, non-http blocked, private URLs filtered from results.
    • Commit: test(app): security tests — auth, proxy, fs, SSRF

PHASE 8: Content Extraction Tests

  • P8.1 — Write content filtering tests

    • New file: packages/app/src/composables/__tests__/contentFiltering.test.ts
    • Tests: isBookLikeResponse false positive fix, isRecipeLikeResponse prose detection, isMusicQuery rejects financial terms, filterTabsByContext ordering.

  • P8.2 — Extend content extraction tests

    • File: packages/app/src/__tests__/contentExtraction.test.ts
    • Tests: song dedup, looksLikeSong rejects financial terms, app category fallback, recipe parsing.

  • P8.3 — Write TV/film resolution tests

    • Same file as P8.2.
    • Tests: film_ext→TV conversion rules, book_ext in news context, explicit tags respected.
    • Commit: test(app): content extraction — filtering, dedup, tag resolution

PHASE 9: State & Composable Tests

  • P9.1 — Write useContentPanel tests

    • File: packages/app/src/composables/__tests__/useContentPanel.test.ts
    • Tests: contentType assignment per content type, RSS tab override prevention, closePanel resets, empty text → no tabs.

  • P9.2 — Write useBannerFallback tests

    • New file: packages/app/src/composables/__tests__/useBannerFallback.test.ts
    • Tests: primary URL fallthrough, API fetch on exhaustion, apiFetching guard, gradient fallback.

  • P9.3 — Write chat store tests

    • New file: packages/app/src/__tests__/chat-store.test.ts
    • Tests: createConversation, addMessage, deleteConversation, branchFromMessage, debouncedIDBSave.
    • Commit: test(app): state tests — contentPanel, bannerFallback, chat store

PHASE 10: AI Integration Tests

  • P10.1 — Extend useAI tests

    • File: packages/app/src/__tests__/useAI.test.ts
    • Tests: system prompt includes recipe instructions, web search results sanitized, editAndResend truncates, sanitizeHistory merges consecutive roles.

  • P10.2 — Write useImageFallback cache tests

    • New file: packages/app/src/composables/__tests__/useImageFallback.test.ts
    • Tests: boundedSet evicts at max, generatePosterFallback valid SVG, escapeXml handles specials.
    • Commit: test(app): AI integration, image fallback cache tests

PHASE 11: Final Verification

  • P11.1 — Full integration check

    • Run: pnpm typecheck && pnpm lint && pnpm test -- --run
    • Run: pnpm build — verify production build succeeds
    • Fix any regressions.

  • P11.2 — Final commit

    • Run coverage report if configured: pnpm test -- --run --coverage
    • Ensure all changes committed on development.
    • Commit: chore(app): hardening complete — all checks pass

Phase Dependencies

P0 → P1 → P2 → P3 → P4 → P5 ─┐
                                ├→ P7 (tests P1,P2)
                          P6 ──┤
                                ├→ P8 (tests P4)
                                ├→ P9 (tests P3,P5)
                                ├→ P10 (tests P2.3,P4.1)
                                └→ P11 (final)

P5 and P6 can run in parallel. P7-P10 can run in any order after their fix phases.

Summary

  • 12 phases, 42 tasks
  • Phases 0-6: fixes (security → proxy → state → extraction → caches → a11y)
  • Phases 7-10: tests (security → extraction → state → AI)
  • Phase 11: final verification
  • Target: all checks pass, 40%+ test coverage on critical paths, secure for Archy deployment
Reference in New Issue View Git Blame Copy Permalink