chore: remove local machine paths from the repo and guard against new ones

[blafourcade]

🎯 What & why

Task documents and their captured evidence held paths and identifiers from one contributor's machine: home directories, Claude scratchpad paths, macOS temp folders, worktree layout, session ids and a device id. Two kanban specs held a third party's personal email address. This replaces them with placeholders in 164 files and adds a check so none comes back.

🛠️ How it works

• Cleanup (adfbfc4, fe85acb, 10dbf9f, 6059541): files under aidd_docs/tasks/, cli/aidd_docs/tasks/ and two kanban/aidd_docs/tasks/ specs. Each value is now a placeholder: <home>, <worktree>, <scratch>, <tmp>, <session>, <device>, <installation>, <user> (owner column of an ls -l capture).
  • Commands in fences were rewritten so they still read as commands: a relative path (cd cli, cd ../test-sdlc-891-codex) or a quoted placeholder (node "<worktree>/cli/dist/cli.js", B='<scratch>/bench').
  • Session ids were replaced where a key or folder names them (session_id, sessionId, session, Codex session id:, Copilot session-state/<id>, Codex rollout-<time>-<id>.jsonl file names, whose timestamps are replaced too because they give away the timezone), along with the other places the same ids appear (thread_id, prompt_cache_key, window ids, copilot --resume=). The device_id fingerprint and Codex installation_id that sit next to them were replaced too. The seven Codex threadIds in the sandbox observations are <session-1> to <session-7>, so uniqueThreadIds: 7 still holds. Random per-run ids stay: context_window_id, turn_id, root_turn_id and msg_ ids. Made-up ids (3333…, 0000…000a) are unchanged.
  • cli/.claude/scheduled_tasks.lock, already listed in cli/.gitignore, was tracked with a live session id; it is now untracked.
  • Two kanban specs named a third party's personal email address as Decider/Owner/Consumer; it is now someone@example.com.
  • Test fixtures are untouched: their placeholder accounts (/home/user, C:\Users\someone, …) stay.
• Guard (scripts/check-no-local-paths.js): it fails on a home path, raw or JSON-escaped (/Users/<name>, /home/<name>, C:\Users\<name>, with or without a trailing separator), a home folded into a project slug (-Users-<name>-, C--Users-<name>-), /private/tmp/claude-<uid>/, /var/folders/, and a UUID that a session or thread key (session, session_id, sessionId, thread_id, threadId, threadIds, including every id of an array under one) or Copilot's session-state/ folder names, and a UUID naming a .jsonl transcript (<project>/<id>.jsonl, D:\work\<id>.jsonl, rollout-<time>-<id>.jsonl). Each hit is reported as file:line: <pattern>; the matched text is never printed.
  • It matches patterns, never the current account, so it works for any contributor.
  • A name only counts at the root of a path and must start with an alphanumeric. /proj/home/x/ and the fake HOME /home/.claude/ therefore pass. A session UUID with fewer than 8 distinct hex digits is treated as made up (00000000-…-00000000000a) and passes.
  • Allowlist: user, u, a, me, dev, someone, alice, tester, test, p, aidd, jane, first, runner, runneradmin, compared case-insensitively, ignoring a ~N 8.3 suffix and a trailing .. Each one is matched in a test or task doc on the tree; removing any one turns --all red. jane/first are the first word of Jane Doe/First Last; runner/runneradmin are the GitHub Actions accounts. No path is excluded; binary files are skipped.
  • Files are read inside a try and skipped on ENOENT/EISDIR, with no existence check first (CodeQL js/file-system-race).
  • Wiring: lefthook pre-commit → no-local-paths on {staged_files}. In CI, validate.yml already runs lefthook run pre-commit --all-files, so it checks every tracked file. --all runs the same scan by hand.
  • {staged_files}, not {files}: with no glob, {files} was skipped with "no files for inspection" on a staged leak.

🧪 How to verify

• Red, the check run on an origin/next checkout before the cleanup:

  $ node scripts/check-no-local-paths.js --all   # cwd: detached worktree of origin/next
  ❌ 1883 local machine path(s) in tracked files
    aidd_docs/tasks/2026_07/2026_07_08_onboard-tutorial-rebuild/headless-runs-codex.txt:15: /private/tmp/claude-<uid>/
    aidd_docs/tasks/2026_07/2026_07_08_onboard-tutorial-rebuild/headless-runs-codex.txt:30: /Users/<name>/
  ...
  

  By pattern: 1034 /var/folders/, 743 /Users/<name>/, 106 /private/tmp/claude-<uid>/, across 110 files.
• Green on this branch: node scripts/check-no-local-paths.js --all → ✅ No local paths: 0 leak(s) in 3154 file(s).
• The final guard on origin/next (3167d690): 2179 hits (1034 /var/folders/, 743 /Users/<name>, 188 session id, 106 -Users-<name>-, 106 /private/tmp/claude-<uid>/, 2 session file). On this branch: 0. On the merge of this branch with origin/next (git merge-tree, exit 0): 0 leak(s) in 3173 file(s).
• Mutation: restore line 54 of aidd_docs/tasks/2026_10/2026_10_15_ledger-by-day/phase-5.md from origin/next →
  phase-5.md:54: -Users-<name>- and phase-5.md:54: /private/tmp/claude-<uid>/, exit 1.
• Tests first: scripts/tests/no-local-paths.test.js was red (Cannot find module), then green; red again on each new shape before its pattern landed. Mutants run against the test file, each red: each new pattern (slug, session key, transcript name, its rollout- and / prefixes), the bare-name lookahead, the trailing-. strip, the ENOENT skip, the made-up-id threshold at < 16 and at < 5, the escaped-key support removed, threadId alone removed, thread_id/threadIds removed, the array pass removed, an array read to the last ] instead of the next, the line sort removed, the skip past an array removed, the line search off by one, < offset - 2 and high = newlines.length - 1 in the line search, a stale ] reused for a later array, return turned into continue on an unclosed array (pinned by a count of ] scans, not a time bound), the array cap at 0, .jsonl turned into .json or .jsonl?, and the backslash branch of the transcript pattern removed. Allowlist names are pinned by --all on the tree, not by the unit tests: removing any one turns it red.
• Cost of the array pass on synthetic input, round-3 code → now (findLocalPaths, one run each): 480 KB of unclosed "threadIds": [ 2127 → 19 ms; 20000 single-id arrays 2588 → 40 ms; 100k ids in 2000 arrays 70312 → 181 ms. --all on the tree: 0.42-0.43 s before, 0.47-0.49 s now. The pass looks up each array's ] once, and stops when no ] is left in the text.
• Gates: node --test 'scripts/__tests__/**/*.test.js' 474 pass, 0 fail (13 of them in the guard's own test file). lefthook run pre-commit --all-files --force is all green. lefthook run pre-push is green: cli-knip, and cli-test with 6748 passing.
• git diff origin/next...HEAD: no added line holds the username, /var/folders/, /tmp/claude-<n> or the worktree layout.

⚠️ Heads-up

• Git history still holds every old path and id. Rewriting history was deliberately not done. This PR cleans the working tree and stops new leaks.
• Some evidence manifests hash capture files whose content was redacted: rule-contracts/evidence/{index,copilot/index,cursor-native/index}.json, autonomous-rule-generation/rule-presentation/evidence/index.json and realistic-sandbox/native-evidence/index.json. Those hashes still describe the original captures, which only history holds. No test verifies them.
• /private/tmp/aidd-* sandbox paths remain on purpose: they name no account.
• Untracking cli/.claude/scheduled_tasks.lock deletes it from other checkouts on their next pull. It is a short-lived scheduler lock, recreated on demand.
• Not covered by the guard, one line each:
  • an array under a session key that is nested, or whose closing bracket is missing;
  • an id more than 64 KB into one array, and a session array that starts more than 64 KB after an unclosed [ (the unclosed [ takes the next ] as its own, and only its first 64 KB is read);
  • a bare username outside a path (an ls -l owner column, whoami or $USER output);
  • a session id under any other key, or not shaped as a UUID;
  • a session id under an unquoted YAML or JS key (session_id: <uuid>);
  • a UUID in a file name other than a .jsonl transcript, and a bare <uuid>.jsonl with no folder before it;
  • a session id passed as a CLI flag (--resume <uuid>);
  • email addresses;
  • Claude stream "uuid" message ids.

🔗 Linked issue

None. Found by the R1 telemetry cleanup audit.

✅ I certify

• I DO CERTIFY I READ EACH LINE OF THE PULL REQUEST BECAUSE I AM A SOFTWARE ENGINEER, NOT A AI PUPPY.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VWNxk63AGKkqE8HRqHLjGp