Skip to content

Business operating system for Claude Code — 57 skills, 21 agents, smart daemon. Unified inbox (WhatsApp/Email/Slack/Telegram), autonomous PR merge, full-AWS monitoring, revenue (Stripe+RevenueCat), e-commerce (Shopify), marketing (Klaviyo/Meta/GA4), voice (Bland/ElevenLabs), APM (Datadog/NewRelic/OTEL), YOLO mode.

16.7 KB
Raw Blame

PII audit — public claude-ops repo (2026-07-22)

Owner rule: no owner-specific data in this public repo. This tracks what was found, what this branch scrubbed, and what remains for a dedicated follow-up pass.

Scrubbed on this branch (chore/scrub-pii-to-local-prefs)

Highest-sensitivity: real people + a real phone number + a real WhatsApp JID, all in skills/ops-inbox/SKILL.md example / standing-rule prose (illustrative only — safe to genericize, no logic depended on the literals):

This section deliberately does not quote what it removed. An audit that reprints the phone number, JID, and contact names it scrubbed leaves the repo exactly as exposed as before, just with the data moved to a different file. The original version of this document did precisely that; see "Round 3" below. Describe the class of value removed, never the value.

  • Five real contact first names in dedup examples → generic <contact> / <contact-A> / <contact-B>.
  • One real E.164 phone number → <number>.
  • Real deal and issue references in the task-tracker example (four internal issue keys plus three counterparty/deal code names) → generic wording ("resolve counterparty→issue mappings from your local prefs/board").
  • A real WhatsApp LID used as a format example in a script docstring → a synthetic all-zeros LID.

Mechanism added:

  • ~/.claude/ops-prefs.json (OUTSIDE the repo, gitignored by location) — central prefs holding the real owner/company/issue-prefix/channel values.
  • .gitignore now blocks ops.local.json, *.local.json, .ops-prefs.json.
  • docs/LOCAL-PREFS.md documents the pattern.

Remaining — follow-up pass (NOT done here; needs care/review)

Raw grep counts (many are already example placeholders like a[at]x[.]com, 123[at]s.whatsapp[.]net, <...> — real count is much lower):

Category Raw hits Real (est.) Notes
The operator's primary company name 149 ~load-bearing Integration target across Slack scoping, dashboards, skills. Genericizing to a config-driven company key is a real refactor — do not blind-scrub.
Emails 155 ~4 real The public maintainer contact ×5, a product support alias ×3, one vendor contact ×1; rest are examples.
WhatsApp JIDs 79 ~0 real left Overwhelmingly format examples.
Issue keys 15 ~11 Internal tracker keys in traceability comments in credit-rotation code + CHANGELOG. Low-harm, numerous.
Phone E.164 7 1 (scrubbed) Rest are 123456… examples.
Aurora 4 0 All AWS RDS Aurora, not the company — false positive.
One counterparty name 1 scrubbed —

Recommended follow-up

  1. Issue keys in code comments + CHANGELOG → drop the ticket refs or replace with <ISSUE>; mechanical, low-risk, ~11 spots.
  2. Real emails (the maintainer contact, a support alias, one vendor address) → placeholders / read from prefs; ~9 spots.
  3. Company name (×149) → a real refactor: introduce a company-key indirection (<company> in prose; runtime value from prefs), keeping the product-specific Slack/dashboard integration working. Do as its own reviewed PR.

Nothing here is pushed. Branch: chore/scrub-pii-to-local-prefs.

Round 2 (2026-07-22) — mechanical scrub done

  • Internal issue keys in code comments / tests / install scripts / plist templates / one runbook → removed or genericized (<ISSUE> / <TEAM>-123), plus the runbook Owner: line degenericized.
  • A person's example email in agents/memory-extractor.md → example.user@example[.]com.

Deliberately kept (NOT PII to scrub)

  • The public maintainer contact address — intentional, published in marketplace.json / SECURITY.md and allowlisted in tests/test-no-secrets.sh.
  • A product support alias — functional config default, overridable by env; left to avoid breaking the launch-gate default.

Round 3 (2026-08-16) — this document was itself a leak

Two failures found by enabling the operator identity denylist, which had never been configured and therefore reported PASS while checking nothing:

  1. This audit re-published what it scrubbed. Rounds 1 and 2 quoted every removed literal verbatim: the phone number, the WhatsApp LID, contact first names, deal code names, and issue keys. One commit removed the phone number from a skill and added it to this file in the same change, so net exposure never dropped. All such literals are now described by class instead of quoted.
  2. Operator identifiers the earlier rounds missed — real email addresses in script docstrings and a test assertion, hardcoded /Users/<user>/… paths, a username fallback, a Tailscale tailnet name, an EC2 instance id, a private Tailscale IP, a PEM filename, and account-pool labels naming the owner's brands.

Both are now guarded rather than merely documented: tests/test-no-secrets.sh gained a tracked-preferences check and a prefs-write-target check, and the identity denylist must be configured per machine (see CLAUDE.md Rule 0) or it verifies nothing.

Scope limit, stated plainly: everything above fixes the working tree only. Values committed earlier remain readable in this repo's public git history, and the maintainer identity is present in the author/committer trailers of most commits. Neither is addressed by editing files. Treat any address or hostname that ever appeared here as public.

Round 4 (2026-09-08) — the gates were not running

Round 3 said the identity denylist "must be configured per machine ... or it verifies nothing". That was true and it stayed unconfigured everywhere it mattered. Three findings, each verified by running the thing rather than reading it:

  1. pii-gate in CI passed while checking zero identity terms. The job runs tests/test-no-secrets.sh on every PR, but no denylist exists on a CI runner, so identity_denylist_check took its "none configured" branch and counted it as PASS. Proven locally by running the suite with the denylist hidden: the summary read 27 passed, 0 failed, identical to a run that really did check 25 terms. Fixed: that branch now reports SKIP, the summary names the skip count, and OPS_PII_DENYLIST_REQUIRED=1 turns it into a failure.
  2. .githooks/pre-commit was dead code. Present and executable, but git only runs hooks from core.hooksPath or .git/hooks, and nothing set either — in the primary checkout or in any worktree. Every commit for the life of this repo bypassed it. Fixed: bin/ops-install-git-hooks wires it and warns when no denylist is configured. Verified by staging a line containing a real denylist term: the commit was refused and no commit object was created.
  3. main has no branch protection at all, so pii-gate was advisory even when it did run. Not fixed here — that is a repository-settings change for the owner, not a code change.

This document was still leaking. Round 3 removed the quoted literals but left the operator's primary company name in the follow-up table, in the row explaining that it is load-bearing. It is now described by class, like every other entry.

Working-tree state at this round: 3 denylist matches across 930 tracked files. One is a detector pattern inside a test (legitimate — the test must name what it searches for). One was this document. One is a private hostname in a test comment. Zero real secrets: every sk_live_/ghp_/AKIA/xox* hit is a detector pattern or documentation, none with a credential-shaped tail. No tracked prefs-shaped files. One hardcoded /Users/<name> path, and it is the literal word username in a Rule 0 example. All phone numbers are reserved-range placeholders.

History, stated plainly again. 14 of 25 denylist terms appear somewhere in the 173-commit history, across 34 distinct paths. Round 3's conclusion stands and is now stronger: this repo is public, has 21 forks and 187 stars, and forks were created both before and after every scrub. A history rewrite would not reach the forks, would break every open PR and clone, and cannot un-publish what was already served. Treat every value that ever appeared here as public. Rotate anything sensitive instead of rewriting; rewriting is theatre once a fork exists.

Round 5 (2026-09-09) — third-party identifiers, and what the agent got wrong

Round 4 closed the gates. This round used them, and the first thing they found was not the operator's own data but other organisations' identifiers, which is the worse category: the operator can decide to publish his own name, and cannot decide that for a client.

Fixed.

  1. Client organisation UUIDs in a Linear bridge. TEAM_TO_COMPANY and COMPANY_TO_TEAM in scripts/hermes-linear/linear_paperclip_delegate_bridge.py held ten UUIDs belonging to six client organisations. One of the six was already read from the environment three lines above the hardcoded five, so the pattern existed and had simply not been finished. All six now come from LINEAR_TEAM_MAP_JSON; an absent map is valid and yields an empty mapping.
  2. A client's Linear team key, hardcoded across several files including a filename. It is now CLIENT_TEAM_KEY from LINEAR_CLIENT_TEAM_KEY, defaulting to the neutral TEAM. The fix-all script was renamed to linear_fix_all.py and its environment variables to LINEAR_FIX_ALL_*.
  3. A client's real Linear issue ids as fallback "thrash canon" dictionaries in two files. The canons already loaded from an out-of-repo canon module; the fallbacks are now empty, which is the correct meaning for anyone who does not have that module. The public import name is client_thrash_canons.
  4. The operator's timezone, hardcoded in ten places. The two functional sites read OPS_TZ and fall back to the host zone; the rest are documentation and now describe the host zone rather than one city.
  5. A brand name in a tracked filename — an unreferenced logo under assets/, renamed to ops-pixel-logo.svg. Its contents were clean. The old name is not repeated here: the denylist would flag this document too, and correctly so.

Two agent findings that were false, and the checks that falsified them.

  • "Private MCP hostnames in ops-mcp-reauth.py, ops-mcp-watchdog.py, ops-cron-pocket-watcher.py." No such hostname is in the tree. The only .ts.net string anywhere is example.ts.net, in two test files.
  • "A tailnet CGNAT IP in local-failover/config.example.json." The value is 100.64.0.10 — the first host of the CGNAT block, the conventional example address, and the same value the sanitizer's own tests use as a placeholder.

Both were reported against history, where the terms do appear, and read as working-tree findings. All 25 configured denylist terms now return zero hits in the tree, which is why the agent's own numbers (49 hits, 228 hits) could not be reproduced against it.

Deliberately not changed: paperclip. The deprecated task system's name appears 272 times, and docs/public-template-contract.md says to replace it with legacy task system. It stays, because 45 of those occurrences are wire format — [Paperclip <id>] title prefixes and paperclip-export: / paperclip-comment: comment footers that this code writes into Linear and later parses back out. Renaming the string in the reader without rewriting every marker already sitting in a live Linear workspace would make the bridge silently stop recognising its own pairs, and "silently stops pairing" is the exact failure these files carry the most defensive code against. The name is a product name, not personal data, and no PAPERCLIP_* variable is set in any environment file on this machine. Renaming it is a protocol migration, and it needs its own change with a compatibility window that reads both spellings — not a search-and-replace in a PII pass.

Unchanged from Round 4: history is not rewritten, and main still has no branch protection. Both remain owner decisions.

Round 6 (2026-09-09) — the gate now fails closed

Round 5 scrubbed a client's identifiers by hand. That fixes one tree, once. This round changes what the build refuses, so the same class of leak cannot come back quietly.

The structural gap. Every identity check here was denylist-driven, and a denylist can only ever hold the operator's own terms — a hardcoded list of anyone else's would itself be the leak. So it structurally cannot hold a client's workspace UUIDs, their team key, or their issue ids: nobody can enumerate a third party's identifiers in advance. That is why ten client UUIDs, a team key in about twenty-five places plus a filename, and a set of real issue ids sat in a public repo while the scanner reported PASS. The scanner was not broken. It was answering a different question.

The inversion. For identifier-shaped literals the default is now refusal. Every UUID and every <KEY>-<number> in the tree fails unless it appears in tests/known-public-constants.txt with a stated reason, and an IANA timezone in code or config fails outright. Pasting a client identifier now costs a line in a reviewable file, argued for in a diff — which is the only kind of protection that does not depend on somebody remembering.

Three properties make it hold. The checks need no configuration, so they run on the CI path where every denylist check SKIPs. They sweep every tracked file including tests/, which EXCLUDE_DIRS drops — the scanner's own directory was the one place a client id could sit unseen. And they cannot report SKIP: a check that could not run is counted as a failure, because counting it as a pass is how a gate silently stops gating.

The negative control. tests/test-pii-gate-fires.sh plants the exact shapes that leaked in a throwaway repo and asserts each gate refuses them, with a clean-tree control so a scanner that failed on everything could not pass either. A gate nobody has watched fail is not a gate. Its first run proved the point by finding four defects in the checks written minutes earlier:

  • grep -o prints the filename only when given more than one file. With a single tracked file the prefix vanished and every filter anchored on : silently stopped filtering — the allowlist admitted nothing and refused nothing.
  • The issue-key regex capped the prefix at six characters, so <CLIENT>TEAM-<n> matched nothing at all. The anchoring that was supposed to stop a long key riding in on a short allowlisted one was never reached.
  • Both checks skipped rather than failed when the file list came back empty.
  • PLUGIN_ROOT was resolved with pwd, but git rev-parse --show-toplevel returns a physical path. Under a symlinked checkout the two never matched, the sweep ran over an empty list, and the suite passed.

Every one of those would have read as PASS forever.

Found by the new checks, missed by the hand scrub: Europe/Amsterdam in two daemon cron notes and one script header, now stated in UTC.

The hook was loud and inert. Committing the work above tripped three BLOCKED lines in the pre-commit hook — and the commit landed anyway. The hook applied an amnesty to its own failure flag after all checks had run: if the only email hits were example domains, it reset the flag to 0 and took every other failure with it. So a UUID and an issue key were both detected, both announced, and both waved through.

That is a worse shape than silence. Output that reads like enforcement is why nobody looked. The fix is structural in the same way as the inversion above: the example-domain filter now applies inside the email check, at the point of the check, and nothing resets the flag afterwards. A late amnesty can only ever be broader than the check it was written for.

tests/test-pre-commit-hook-blocks.sh closes it for good. Eight cases run a real git commit against an installed copy of the hook and assert the exit status, not the output — the exit status being the only part of a hook that stops anything. Case 1 is a clean commit, so a hook that refused everything cannot pass. Case 5 keeps the amnesty's legitimate purpose (an @example.com address alone is fine). Case 6 is this regression itself. Cases 7 and 8 pin the exemption to four exact paths — the scanner, its allowlist, and the two negative controls — rather than to the tests/ directory, so the scanner's own folder stays scanned.

The lesson generalises past this hook: a gate must be tested through the interface that enforces it. test-pii-gate-fires.sh proved the scanner refuses the leaked shapes and said nothing whatsoever about the hook, because the hook is a separate program with its own copy of the patterns and its own exit path.

Unchanged: history is not rewritten, main still has no branch protection, and paperclip still awaits its own protocol migration. All three remain owner decisions.