Skip to content

Security review: GitPush tree kill, README facts line, witness-backed verify - #172

Merged
sykosyber merged 2 commits into
mainfrom
claude/security-followups-three
Nov 4, 2026
Merged

sykosyber merged 2 commits into
mainfrom
claude/security-followups-three

Conversation

@sykosyber

@sykosyber sykosyber commented Nov 4, 2026 •

Copy link
Copy Markdown
Member

Before: after #171, three items from the security review were still open. A git_push publish that timed out ended git but left the transport it started (ssh) running. The README described the facts route as if any fact could be asserted. And Verify history checked only the hash chain, so a history the database owner rewrote and rehashed still showed as intact, even when the witness had signed the original head.

After: a timed-out push or status read ends its whole process tree. The README's facts line says when the route refuses a key. Verify history also reads the witness's log when one is configured, so a history rewritten after a head was signed fails verification. Without a log, the response says that the chain alone cannot see a rehash.

Class: P
Agent-platform: claude
Admitted-by: sykosyber (Mateo approved merging on Nov 4)

What changed, by item

  • GitPush process-tree kill. GitPush.apply and its ls-remote status read now run through publish._bounded: their own session (POSIX), output to a file, and a deadline that ends the tree through checks._stop (taskkill /T on Windows), as a publish command and a check already do. A timeout is still an unknown outcome (effect_unknown), as before. ls-remote keeps Repo.git's scrubbed environment and core.fsmonitor=false. Test: GitPushTimeoutEndsTheTree (a stand-in ssh via core.sshCommand that starts its own child; fails on main with the child still running).
  • README facts line (Class P). POST /api/cases/{id}/facts now says it is refused (attestation_denied) for a key the contract reads, unless a required fact names that same source without asking for verification, matching Work._attestable.
  • Hash chain vs. a database owner who rehashes. New Work.verify_history(case_id, log=None), used by GET /api/cases/{id}/verify and syberwork verify. With a witness log it judges the signed heads as witness_status does; valid is false on head_differs, beyond_history or chain_broken. The response now carries chain, witness (witnessed_through, count, conflicts, or null) and a note. valid keeps its meaning without a log. Tests: HistoryVerifiesAgainstTheWitness (Work level, in-process witness) and ConsoleWitness.test_a_rehashed_history_passes_the_chain_but_not_verify (route). README, docs/ARCHITECTURE.md and docs/console/NOTES.md say what this covers and what it still doesn't.

Behavior change to review

  • GET /verify returns more fields than {"valid": ...}. The console reads only valid, so it is unchanged, but it now shows "does NOT verify" when a witness conflict exists. Four test_console asserts were updated from exact-dict to valid/chain.

Still not solved

  • Events after the newest signed head, and an installation with no witness log, are covered by the chain alone. Someone holding the witness seed can still rebuild the log (already stated in ARCHITECTURE).

Evidence (claim ceiling: Implemented and Tested)

  • Unit suite (sharded): 1256 tests, 4 failures, the same 4 that fail on main in this environment (2 Playwright browser missing, 2 hidden-check staging).
  • conformance.run: matched 23 golden traces. G5: 0 protected spans changed. G6: ok.
  • Rebased on main e1bc9bb before merge; suite 1447 tests with the same 4 environment failures, conformance 23 traces, G5 0 changed, G6 ok.

Generated by Claude Code

sykosyber and others added 2 commits November 4, 2026 05:09
…ts line, witness-backed verify

- GitPush: push and ls-remote run in their own session under a deadline that
  ends the whole tree (checks._stop), so a timed-out push no longer leaves ssh
  running. A timeout is still an unknown outcome.
- README: the facts route line now says when it is refused (attestation_denied).
- Verify history (Work.verify_history: GET /verify, syberwork verify) also reads
  the witness's log when configured, so a history rewritten and rehashed after a
  head was signed fails. Without a log it says the chain alone cannot see that.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@sykosyber
sykosyber force-pushed the claude/security-followups-three branch from 5eae162 to e168952 Compare November 4, 2026 05:15
@sykosyber
sykosyber marked this pull request as ready for review November 4, 2026 05:37
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 4, 2026 •

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-04T05:39:14.047475Z e168952 Draft marked ready
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@sykosyber
sykosyber merged commit 9a29642 into main Nov 4, 2026
25 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant