Skip to content

Release v3.76.0 - #213

Merged
edmondscommerce merged 19 commits into
mainfrom
release/3.76.0
Nov 20, 2026
Merged

edmondscommerce merged 19 commits into
mainfrom
release/3.76.0

Conversation

@edmondscommerce

Copy link
Copy Markdown
Collaborator

Release v3.76.0.

See RELEASES/v3.76.0.md for the notes, upgrade instructions and changelog. Scope: Plans 00626 (remote-delete proofs), 00628 (a long project path no longer breaks the daemon socket), 00623 (secret-guard link index and fail-closed bounded walks) and 00603 (supervisor idle compaction at the orange tier, dry-run by default), plus the branch-pool warning ladder and the empty-commit branch-delete fix already on main. This pull request is the release gate: its CI run tests the exact commit that is merged and tagged.

🤖 Generated with Claude Code

LTSCommerce and others added 19 commits November 20, 2026 09:13
…t paths, harder secret guard

- Updated version to 3.76.0 across all registered files (regenerated HOOKS-DAEMON.md, re-locked uv.lock)
- Added changelog entry
- Generated release notes with every pending callout folded into Highlights

Full changelog: RELEASES/v3.76.0.md

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…es into the release

- git mv release-notes callouts 001-004, 006-008 into
  CLAUDE/UPGRADES/v3/v3.75.0-to-v3.76.0/release-notes/ (all folded into
  RELEASES/v3.76.0.md); fix the CcySupervisor.md relative link in 004 for
  its new depth
- git mv config-changes/v3.76.0.yaml to the live directory and set its
  date to 2026-11-20
- no pre-upgrade, post-upgrade or truth-changes items were staged; the
  UNRELEASED indexes are already empty

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d, ledger N682

- claude-code-versions.yaml: v3.76.0 entry (claude_code_version unknown,
  reviewed_through 2.1.329). contract-status reports the upstream hooks page
  CHANGED since the 2.1.296 audit, so the hooks META is not bumped.
- Stop.json / SubagentStop.json: document background_tasks `type` as an open
  label set including `monitor` (adopt row 1, quoted from the raw hooks page).
- Instruction-loading META: last_audited 2.1.329, hash/bytes/fetch_date moved
  to the re-captured memory page (no fact changed).
- NIGGLES N682: FACTS G5-01 auto-compaction note (adopt row 2).
- Review report copied into the Plan 00474 subagent-reports folder.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- CHANGELOG: restore the lost [3.75.0] heading; drop the SECURITY label
  from the 512-directory usability fix; name the two branch-pool pressure
  handlers
- Release notes: callout 008 cites release note 007 (not the staging
  ordinal 003); fix the CcySupervisor.md link depth; Upgrade Instructions
  name the two recommended handlers

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tern fixes

F1: with no protected pattern (mode: replace, protected_paths: []) the
unresolved and gave-up routes now return no finding, matching
fail_closed_finding's contract, instead of denying.

F2: a word through the calling process's own view is judged by the Bash
scan as a mention, so the grep-PATTERN exemption clears it as it clears a
protected name, and a quoted heredoc body written only to .md files (prose
no command opens) is not judged for views. Every read position still
denies: the consumer and encrypted-target exemptions never clear a command
holding a view, and every caller outside the guard's Bash route keeps the
N668 refusal (ProcessViewRefusedError subclasses ResolutionFailedError).
Near-miss DENY rows added.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…nale

A terminal DENY, not an ALLOW, is what ends the chain.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…kes no directory move

effective_cwds now walks the command with every quoted heredoc body that
nothing runs, now or later, blanked (strip_unrun_heredoc_bodies). A body
fed to an executor, or written to a non-markdown file, is still walked.
In a loose (unbalanced) segment, a move counts only at a command start or
just inside a quote, never mid-prose. The deny reason's cause is computed
from the same judged command text the decision used.

Reproduced against a warmed link index past LINK_SCAN_DIRECTORY_CAP.
Executor heredocs, bash -c, eval, ssh and after-closer moves still deny.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…ride

Short override: events dir beside the override. Over-limit override:
events dir follows fit_socket_path's fallback, not the raw override.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…uns later

cat > n.md <<'EOF' ... then bash n.md (or . n.md) runs the body it
wrote, so its moves count. strip_never_run_heredoc_bodies blanks a .md
body only when .md is named no more often than the bodies it removes.
Also carries the formatter's line joins on the new tests.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…nner umask and /run/user

The fake XDG_RUNTIME_DIR was made with a bare mkdir, so under the CI umask
(022) it was 0755, failed the private-dir check, and the relocation fell
through to the runner's real /run/user/1001. Create it 0700. The tmp
fallback test patched Path.is_dir, which the code no longer consults; point
_RUN_USER_ROOT at a missing directory instead.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…-only graph writable first

git writes objects/info/commit-graph read-only (0444/0400). Root rewrote
it anyway, so the suite passed locally; the CI runner user got EACCES.
chmod only that test-owned file to 0644 before the in-place forgery.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…er fetching, not before

git >= 2.48 (CI runs 2.56.1) records refs/remotes/<remote>/HEAD on every
fetch (followRemoteHEAD=create), so "other has no recorded default" was
false on CI and the proof rightly accepted the delete. Delete the symref
after the fetch so both tests set up the state they claim; no code change.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…e number

test_no_unregistered_file_carries_the_current_version flagged four src
modules whose Step 10 comments named the version; the next bump would
leave them stale. Keep the finding label, drop the number (comments only).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…n private shared tmp

A CI runner can already hold /tmp/hooks-daemon-<uid> that is not private, which the code rightly refuses. The test now points the shared tmp at a fixture directory.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…the command then runs

The F2 process-view exemption blanked a quoted heredoc body written only to a
.md file even when the same command then ran that file (bash n.md, . ./n.md,
sh n.m?), so a /proc/self or /dev/fd view inside it was allowed. The views pass
now reads strip_never_run_heredoc_bodies, which keeps a .md body the command
names again, and that helper now also keeps every body when a glob character
appears outside the blanked bodies (the safe direction).

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
… may respell its file

strip_never_run_heredoc_bodies blanked a .md-only body unless .md recurred
or a glob char appeared. A run step naming the file through a substitution,
expansion, brace, quote or escape (bash $(ls), bash ${f}d, bash n.{md,},
bash n.m''d, bash n.m\d) slipped past, so a process-view path in the body
was allowed. Now any of $ ` { ' " \ or a process substitution outside the
heredoc operators keeps every body for judgement.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
@edmondscommerce
edmondscommerce merged commit de8833b into main Nov 20, 2026
58 checks passed
@edmondscommerce
edmondscommerce deleted the release/3.76.0 branch November 20, 2026 13:34
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.

2 participants