Skip to content

Release 1.10.0 - #487

Merged
peg merged 101 commits into
mainfrom
staging
Nov 11, 2026
Merged

peg merged 101 commits into
mainfrom
staging

Conversation

@peg

@peg peg commented Nov 6, 2026 •

Copy link
Copy Markdown
Owner

Summary

Release 1.10.0: promotes staging to main. Full notes are in CHANGELOG.md under 1.10.0.

Changed   Claude Code: a policy allow returns no decision, so Claude Code's own permissions apply
Added     rampart setup claude-code --hook-failure=auto|block|continue
Removed   alpha Python SDK; rampart quickstart is a hidden stub pointing to protect
Security  shell-write protection for Rampart's own files, launcher and wrapper unwrapping,
          audit hash and reader consistency, init --from-audit verification,
          OpenClaw browser URL and reply-provider checks, Codex hook detection,
          doctor setup records, release images built from the tagged commit
Fixed     rampart test labels ask as ASK; doctor and protect guidance
Toolchain Go 1.26.11

Evidence

  • CI on the exact head passes on Linux, macOS and Windows, along with docker smoke, the
    goreleaser snapshot, the Hermes plugin check and the CI gate.
  • go test ./..., go vet ./..., make security-assurance, git diff --check, API reference
    and toolchain sync checks pass.

peg and others added 30 commits October 2, 2026 20:01
Resolve the changelog conflict: the OpenCode entry moves into a new 1.10.0
section above the 1.9.2 section that staging already dated.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Synchronize the release metadata for the experimental OpenCode minor release:
assurance baseline, threat-model applicability, site version markers, and the
current-release summary.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Prepare experimental OpenCode support for 1.10.0

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
The scheduled Hermes gate followed Hermes' rolling latest release, so every
upstream change (a new tag scheme, an unstamped package version, a dropped
dependency) turned the whole Upstream Compatibility workflow red without
telling us anything about Rampart. Add --release-tag so the same gate, with all
of its validation and the editable exact-tag install, runs against one reviewed
release, and pin CI to v2026.9.24 (Hermes 0.21.5).

Declare the integration's upstream_ci as pinned, narrow the support claims to
the reviewed release, and state that newer releases are not claimed compatible
until reviewed.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Pin the Hermes compatibility gate to the reviewed release

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Claude Code lets a command hook that cannot start, times out, exits with an
unexpected code, or prints invalid output proceed through its normal
permission flow. Every Rampart-internal error exits 1, so a failure of Rampart
silently disabled enforcement (an outside user reported exactly this on a
domain-joined Windows machine).

Claude Code 2.1.295 added a per-hook onFailure field. Add
--hook-failure=auto|block|continue to the Claude Code setup command:

- auto (default): a new installation on a detected 2.1.295+ host sets
  onFailure "block" on Rampart's PreToolUse hook only; an existing installation
  is left exactly as it is and told how to opt in.
- block: explicit opt-in; refuses an older or undetected host, because an older
  release may reject a settings file containing an unknown hook field.
- continue: returns to Claude Code's default.

Post hooks keep the default so a scanning failure cannot hide tool output.
Fields a user added to Rampart's own hook entry and other tools' hooks are
preserved. Version detection runs `claude --version` with a scrubbed
environment and a short timeout. verify reports the active mode.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Keep both unreleased changelog entries: the new setup option and the Hermes
gate pin.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Fail closed when the Claude Code PreToolUse hook fails

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Bumps [docker/build-push-action](https://github.com/docker/build-push-action) from 7.3.0 to 7.4.0.
- [Release notes](https://github.com/docker/build-push-action/releases)
- [Commits](docker/build-push-action@53b7df9...c3c9e26)

---
updated-dependencies:
- dependency-name: docker/build-push-action
  dependency-version: 7.4.0
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <[email protected]>
…cker/build-push-action-7.4.0

Merge verified dependency bump

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Every rotated audit file starts with a chain-continuation header linking it to
the previous file. ReadEventsFromOffset decoded it into a zero-valued Event, so
each rotation added a phantom record with a 0001-01-01 timestamp to everything
that reads audit files (API, dashboard, status, inventory, init-from-audit,
reports). Skip records that carry both top-level chain_continue and prev_file,
the same fields the chain verifier requires, while still consuming their bytes.

This also removes the midnight race in a sink test that reads the current file
and expects a single record.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Skip chain-continuation headers when reading audit events

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Bumps [docker/setup-buildx-action](https://github.com/docker/setup-buildx-action) from 4.2.0 to 4.4.1.
- [Release notes](https://github.com/docker/setup-buildx-action/releases)
- [Commits](docker/setup-buildx-action@bb05f3f...f87e599)

---
updated-dependencies:
- dependency-name: docker/setup-buildx-action
  dependency-version: 4.4.1
  dependency-type: direct:production
  update-type: version-update:semver-minor
...

Signed-off-by: dependabot[bot] <[email protected]>
…cker/setup-buildx-action-4.4.1

Merge verified dependency bump

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Bring the download-and-pipe detection fix and the 1.9.3 changelog back into
the integration branch. Keep staging's 1.10.0 release markers (it still carries
the held minor release) and order the changelog 1.10.0, 1.9.3, 1.9.2. The
current-release summary now names 1.9.3.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Merge main into staging after the v1.9.3 release

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Bring the group and substitution handling, its regression tests and the 1.9.4
changelog back into the integration branch. Keep staging's 1.10.0 release
markers (it still carries the held minor release), order the changelog 1.10.0,
1.9.4, 1.9.3, 1.9.2, and name 1.9.4 in the held 1.10.0 summary.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Merge main into staging after the v1.9.4 release

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Revert the OpenCode merge (#443) on staging so the rest of the integration
branch can ship as 1.10.0. The integration is preserved on
codex/opencode-parked until its acceptance items finish.

#443 also opened the dated 1.10.0 changelog section, which held only
OpenCode. Remove it and return the release markers to v1.9.4, the latest
release staging contains; the remaining 1.10.0 notes stay under
Unreleased.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
The Claude Code hook answered an explicit allow for every call policy
permitted, which skipped Claude Code's own permission rules, modes and
prompts and overrode dontAsk. Rampart now returns no decision for a
policy allow and for every decision in monitor and audit modes, so it
only narrows what Claude Code permits. An explicit allow is kept for a
call the user approved through Rampart's approval queue. Other hosts are
unchanged.

An empty hook payload from Claude Code in enforce mode is now refused as
a parse failure instead of being allowed.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Remove sdks/python (never published to PyPI), its CI jobs, Dependabot
entry and docs page; custom agents are pointed to the HTTP API reference.
The CI job that also ran the Hermes tests on Python 3.9 is kept as a
Hermes-only job and remains in the CI gate.

Replace `rampart quickstart` with a hidden stub that prints the
`rampart protect <agent>` guidance to stderr and exits 1, accepting and
ignoring all arguments. Helpers shared with interactive setup move into
setup_interactive.go and setup.go. The zero-driver branch of
runProtectionPlan existed only for `quickstart --agents none` and is
removed with it.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>
Park the experimental OpenCode integration for 1.10.0
peg added 28 commits November 9, 2026 15:09
Setup collapses duplicate Rampart PreToolUse hooks into one entry but only
read onFailure from the first, so re-running setup could drop block mode.
fix: 1.10.0 shell-write protection precision, setup fail-closed duplicates, docs accuracy
Find comments in the quoted command, so a quoted or escaped # operand is
still checked. Read -t/--target-directory values, take only the
destination of cp, install, ln and rsync (every operand for install -d
and rsync --remove-source-files), and check find starting points when an
-exec action changes the found path.
An existing PreToolUse hook marked async, asyncRewake or if enforced
nothing, so setup now plans it as a fresh install. Also drop the
'No policy found' tip: the hook always loads the shipped standard
profile.
The script given inline, or with -e/-f, is no longer read as a file the
edit writes, and perl options such as -Mstrict are no longer mistaken for
-i. Adds a regression row for a comment line before a command, which
crashed the hook on staging.
- Claude Code guides use rampart protect, which installs the Guard that
  answers the canary; setup alone installs only the hooks.
- verify prints the PreToolUse failure mode in its human report.
- doctor accepts no policy files: the hook and serve use the built-in
  standard profile.
- Approval fallback messages and the threat model say a headless_only ask
  is denied when rampart serve cannot take it.
- protect help, allow precedence, and the older-surface onFailure caveat
  are stated precisely.
…ested digest

The container workflow now checks out the release commit, confirms the
tag still points at it before building and again before publishing,
pins base images by digest, and promotes tags only to the digest that
passed the smoke test.
A Rampart handler with extra fields such as async or if no longer counts
as installed, and setup rewrites it. A user config.toml with
allow_managed_hooks_only = true makes doctor and verify fail. Messages
now say /hooks trust is not checked.
…cation

Setup now records each user-level hook install next to the verification
receipts. Doctor treats either record as proof the agent was protected,
so hooks deleted after setup --no-verify or a failed verify are reported
as a failure, not a warning. setup <agent> --remove clears both.
A send whose explicit channel names a different provider than the
originating conversation, or names one when the origin provider is
unknown, is now an external message that needs approval, even when the
target ID matches the origin conversation.
The hash covers the typed event, so a record with an unknown field, a
repeated key, or a schema key in another letter case could change while
its hash still verified. Decoding now marks such records and VerifyHash
fails them in every verify, recovery and report path. Dynamic request
maps stay covered as before.

Measured decode cost: 5.4 to 6.2 us per record, no added allocations.
ReadEventsFromOffset still reads past a complete record that is not
valid event JSON, but now returns a MalformedRecordsError with the
events. audit tail, watch, the dashboard and the HTML report show a
warning; the compliance report and export fail.
init --from-audit now requires a directory to pass the audit verify
checks and a single file to have a valid hash on every record, and
learns only from the exact records that verified. Modified, unhashed or
malformed history generates nothing and fails. --allow-unverified keeps
the old behavior and marks the generated policy. The guide no longer
claims --since accepts a date.
status and inventory dropped a whole file's events from today's counts
when the reader reported a malformed record. They now count the events
it could read. Also corrects the changelog: rampart serve logs malformed
records read through its audit API; the dashboard does not show them.
Moves a misplaced doc comment back to hasCFlag.
fix: 1.10.0 pre-release fixes (shell writes, audit, Codex, doctor, OpenClaw, release)
- policy: a shell started with --rcfile/--init-file, and env -P, are no
  longer unwrapped for allow rules; deny and ask rules still see the command
- policy: option clusters (env -iC, time -ao, time -pf) and find -H/-L/-P/-f
  before starting points are read for writes and payloads
- audit: schema keys must be lowercase ASCII snake case, so a key JSON
  decoding folds onto a schema key no longer verifies
- doctor: upgrade and doctor record agents whose Rampart hooks they find; a
  failed verify of an agent never set up is not a record; removing a Cline
  workspace install keeps the user-level record
- docs: post-tool blocking depends on the host; stray quickstart mention;
  SECURITY.md supported versions; CHANGELOG date and entries
fix: close 1.10.0 review findings before release
govulncheck reports the called standard-library vulnerability GO-2026-6859
(HTTP/1 server crash from a MaxBytesReader header race in net/http) against
Go 1.26.10, fixed in 1.26.11. Update the module baseline, container build
image (pinned by index digest), and public source-build requirement together,
as the toolchain sync check requires. Workflows read the version from go.mod.
chore: update Go toolchain to 1.26.11
- Threat model lists only the obfuscation deny rules standard policy has.
- Architecture describes failure behavior per boundary.
- OWASP mapping uses generic credential-scoping wording.
- Test fixtures and comments use generic names.
- .gitignore drops unused entries and ignores the make linux binary.
docs: align threat model and architecture wording; tidy repo hygiene
@peg
peg merged commit c4e45c6 into main Nov 11, 2026
17 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