Skip to content

AUTO_DRIVE_REPORT.md

Latest commit

 

History

History
258 lines (196 loc) · 10.9 KB

AUTO_DRIVE_REPORT.md

File metadata and controls

258 lines (196 loc) · 10.9 KB

Autonomous Drive Report

Started: 2026-04-20 Scope: full project, no commits (user handles), uncommitted changes untouched.

Baseline

  • Branch: main, clean against origin/main per git status.
  • Heavy WIP: 60+ modified files + several untracked tests (api-keys, validator, widget-types). Did not modify any of these.
  • Pre-run test count: 220 passing, 1 failing in tests/api/api-keys.test.js (pre-existing; see "Findings" below).

Actions taken

1. New OAuth authorize tests — tests/api/oauth-authorize.test.js

22 tests covering the security-critical authorization endpoint:

  • Validation: response_type, client_id, redirect_uri, code_challenge, code_challenge_method (PKCE mandatory, S256 only — rejects plain).
  • Strict redirect_uri matching (RFC 6749 §3.1.2.2): different host, extra path segment, different protocol, different port, query string, fragment — all rejected.
  • Unauthenticated flow: bounce to /login?next=… preserving consent URL.
  • Consent page: HTML render, CSP/frame-options headers, hidden inputs for PKCE round-trip.
  • POST consent: code issuance, deny path, CSRF rejection, cross-site Origin rejection, scope intersection (drops scopes not in client.scope).
  • Method handling: 405 on DELETE, 204 on OPTIONS preflight.

2. New OAuth token tests — tests/api/oauth-token.test.js

24 tests covering the token endpoint:

  • Client auth: missing client_id, rate limit, unknown client, bad confidential secret, HTTP Basic credentials.
  • authorization_code grant: missing fields, unknown code, replay defence (consumed code → revokes refresh chain), expired code, client mismatch, redirect_uri mismatch, PKCE failure, success path issues access+refresh with correct scope, omits refresh when client lacks the grant type.
  • refresh_token grant: missing token, success rotation, scope narrowing on subset, scope-widening attempt rejected with invalid_scope, unknown-scope rejection, propagation of invalid_grant and refresh_reuse_detected errors.

3. New LLM proxy tests — tests/api/llm-anthropic.test.js

22 tests covering the we-pay LLM proxy:

  • Policy gating: missing agent param (400), policy not found (404), brain.mode != we-pay (402), surface.script disabled (403).
  • Origin enforcement: attacker origin rejected, server-to-server allowed, first-party localhost allowed.
  • Rate limiting: per-IP and per-agent (when policy declares limit).
  • Monthly call quota: enforced via Redis counter, quota_exceeded 429.
  • Monthly token budget: enforced before forwarding (default 1,000,000 tokens; cost_limit_cents overrides), debits actual usage on success, does NOT debit on upstream error.
  • Body + model: empty messages rejected, model allowlist enforced, policy-default model used when omitted, caller override accepted.
  • Upstream: success proxied unchanged, errors sanitized to generic 502 (no leaking of upstream error body), usage event recorded with token counts.

4. ENS dedup TODO — re-examined and replaced with rationale

The TODO at api/agents/ens/[name].js:45 suggested deduping with /api/agents/by-address/[addr].js. After reading both endpoints, they return very different shapes:

  • ENS endpoint returns a full agent profile (name, description, avatarid, erc8004*, etc.).
  • by-address returns a minimal NFT-style shape ({id, chainId, agentURI, manifestUrl, onChain, source}) with chain enumeration fallback.

Extracting a shared helper would add abstraction without removing real duplication (only ~6 lines of SQL overlap). Replaced the TODO with a comment explaining the divergence so a future reader doesn't repeat the investigation.

5. Removed-deps verification

grep -r "fbx-parser|tweakpane" across src/, api/, public/, scripts/ returned zero hits. npm run build succeeded (5.47s, 138 PWA precache entries). The package.json removal of these two deps is clean.

Test summary

Suite Tests Status
OAuth authorize 22 ✅
OAuth token 24 ✅
LLM proxy 22 ✅
Newly added 68 ✅
Full suite 289 288 ✅ / 1 ❌ (pre-existing)

Findings requiring your decision

1. Pre-existing test failure (not caused by this session)

tests/api/api-keys.test.js:166 expects:

expect(body.data.token).toMatch(/^sk_live_[a-f0-9]+$/);

But randomToken() in api/_lib/crypto.js returns base64url, which uses - and _ and mixed case. The actual token format is, e.g., sk_live_3pmN4jIKEd_eeKm4VFdvge_j8K1s_QeVcbLJlKNuTCY. The regex was written against a hex assumption that doesn't match the implementation.

Suggested fix (one-line): change the regex to /^sk_live_[A-Za-z0-9_-]+$/. Did not edit because this test file is in your untracked WIP and I don't want to step on in-progress work.

2. Coverage gaps still open

Areas with significant uncommitted security logic but no tests yet:

  • api/auth/register.js — unified error-message changes (now returns conflict instead of username_taken / email_taken).
  • api/_lib/auth.js — session rotation logic in getSessionUser (rolling refresh window).
  • api/oauth/introspect.js and api/oauth/revoke.js — both modified, no dedicated test files.

These are good candidates for a follow-on session.

3. Things I deliberately did NOT touch

  • All 60+ files in your git status modified list.
  • All other untracked files (tests/api/api-keys.test.js, tests/src/validator.test.js, tests/src/widget-types.test.js).
  • package.json (already in your WIP).
  • CLAUDE.md and other docs in your WIP.

Nothing was committed or pushed. All changes from this session are:

  • New: tests/api/oauth-authorize.test.js
  • New: tests/api/oauth-token.test.js
  • New: tests/api/llm-anthropic.test.js
  • Edited: api/agents/ens/[name].js (TODO → rationale comment, ~6 lines)
  • Edited: this report.

Second pass — formatting + broken refs (parallel session)

This appears to be a second autonomous session running in parallel with the one above. Different scope; appending so neither is lost. Heads up: another session wrote the section above, then commits 84e5249 and b6cb561 absorbed all in-progress work into git.

Verification at end-of-session

  • npm test → 18 files / 355 tests / 0 failures.
  • npm run build → clean, no warnings, 138 PWA precache entries.
  • npx prettier --check . → clean.
  • npm audit → 0 vulnerabilities.

Edits made

  1. 3d-demo/.prettierrc — jsxBracketSameLine → bracketSameLine (Prettier 3 deprecated the old name; was emitting a warning on every format run).
  2. api/auth/wallets/_link-nonces.js:4 — import path was ../_lib/crypto.js which resolves to api/auth/_lib/crypto.js (does not exist). Corrected to ../../_lib/crypto.js → api/_lib/crypto.js. This file is imported by wallets/index.js and wallets/nonce.js, so the wallet-link flow would have thrown at module load.
  3. public/skills/subscription/skill.js:78 — dynamic import path was '../../src/permissions/grant-modal.js', which resolves to public/src/permissions/grant-modal.js and 404s in production. Changed to absolute '/src/permissions/grant-modal.js' to match the sibling public/skills/tip-jar/skill.js:223 pattern.
  4. public/dashboard/usage.html:7 — removed <link rel="stylesheet" href="/assets/dashboard.css" />. The file does not exist anywhere in the repo and was producing a guaranteed 404. Inline <style> block in the same file already covers all styling.
  5. public/artifact/index.html:142 — <script src="artifact.js"> resolved to /artifact/artifact.js under the Vercel rewrite, which 404s. The actual bundle is at /public/artifact.js (served at /artifact.js). Made the path absolute.
  6. Whole-repo prettier --write — formatted ~80 files. All edits are formatting-only; no semantic changes.

Findings requiring your decision

🔴 P0 — Live API key committed to git

.mcp.json:6 contains:

"Authorization": "Bearer sk_live_…REDACTED — rotate this key"

The file is in .gitignore now but was committed previously (git ls-files .mcp.json confirms it's tracked; git log shows commits 0b745eb and a1109b9). Anyone with read access to the repo (including forks/clones) has this key.

Required action — not safe to do without you:

  1. Revoke + rotate this API key immediately.
  2. git rm --cached .mcp.json then commit the removal.
  3. Rewrite history with git-filter-repo or BFG to scrub the key from past commits, then force-push (coordinate with anyone who has clones).

I did not delete the file or rewrite history.

🟡 P2 — res.end(JSON.stringify(...)) deviations from api/CLAUDE.md

The convention says use json(res, status, body) from api/_lib/http.js. Five spots ship raw res.end(JSON.stringify(...)):

  • api/agent-oembed.js:79, :160 — both intentional (special application/json+oembed content type); json() hardcodes application/json. Leave as-is or extend the helper.
  • api/widgets/oembed.js:86 — same oEmbed reason.
  • api/mcp.js:74, :703, :709 — JSON-RPC framing; leave as-is.
  • api/permissions/verify.js:80 — could migrate to error(). Trivial.

Only permissions/verify.js is a real cleanup candidate.

🟡 P2 — Sweep scripts

The auto-commit at b6cb561 ("chore: add HTTP and page sweep scripts for resource validation") committed two scratch scripts I'd written to drive the page sweep:

  • tmp-http-sweep.mjs — fetch HTML pages + sub-resource probes
  • tmp-page-sweep.mjs — playwright variant (blocked: missing libatk1.0, no sudo)

If you want to keep them, move under scripts/ and reinstall cheerio (I uninstalled it because the script wasn't intended to land). If they were committed by mistake, git rm them.

What I checked and found nothing

  • npm audit — 0 vulns.
  • All JSON files (vercel.json, package.json, manifests, etc.) parse.
  • HTTP sub-resource sweep across 28 built pages — every <script src>, <link href>, <img src> resolved to a 2xx. (Runtime JS errors NOT checked — playwright would not launch.)
  • forEach(async ...) patterns — none.
  • Static checks for unhandled JSON.parse, hardcoded sk-/AKIA-/ghp- prefixed secrets in *.js/*.json — only the one in .mcp.json.

Coverage gaps observed (not addressed)

  • No JS-runtime error sweep — chromium needs system libs that need sudo. Worth a follow-up on a machine where npx playwright install --with-deps works, or in CI.
  • knip reports 600+ "unused" files but it has no config and treats every Vercel function entry + every HTML script as orphaned. Real dead-code audit needs a knip.json describing entry points.