v2.31.0

Added

• unbounded-background-push: a push sent to the background needs a
  deadline. git push runs the pre-push gate inside it — a test suite, a
  lint pass — and with run_in_background no clock ends it and nothing says
  it is still going: on 2026-10-08 a push seeding a fork held a session for
  forty minutes before anyone asked (#86). The rule advises on a background
  git push whose output goes to a file (> push.log 2>&1, the shape a
  backgrounded push takes) and no timeout wrapper, and points at timeout <s> git push …, which push-preview and implementation-review read through, plus
  amont rehearse --wait so the push itself skips the suite. A foreground
  push (the tool's own timeout bounds it) and a dry run stay silent.

---

What's Changed

• feat(rules): bound a background git push; 2.31.0 by @fredericrous in #88

Full Changelog: v2.30.2...v2.31.0

v2.30.2

Fixed

• Pushing commits some remote already has needs no preview and no
  implementation review. Seeding a fork (git remote rename origin upstream, add the fork as origin, push a branch) publishes nothing this
  session wrote: every commit is already on upstream/*. push-preview
  listed the files with --not --remotes=<target remote>, which excludes
  nothing on an empty fork, so upstream's whole history read as an
  unapproved interface change; implementation-review found no default
  branch on the fork and asked for a review of it. Both guards now stop at
  "nothing new to any remote" (git rev-list <src> --not --remotes is
  empty); a commit new to every remote is still judged exactly as before.
  The same mistake in amont's pre-push gate was amont#301 (1.47.2).

---

What's Changed

• fix(push): commits on a remote need no gate; release 2.30.2 by @fredericrous in #87

Full Changelog: v2.30.1...v2.30.2

v2.30.1

Fixed

• push-preview: an unedited template placeholder is not evidence. The
  plan templates carry evidence: <why …> under ## Preview. A plan
  approved with that line left as it was declared evidence and passed the
  push with no question. A reason that starts with < now declares nothing,
  and the push asks as usual.

---

What's Changed

• fix(preview): a placeholder is not evidence (2.30.1) by @fredericrous in #85

Full Changelog: v2.30.0...v2.30.1

v2.30.0

Approve only what needs judging (ADR-0028, which replaces ADR-0023).
Existing installs must re-run amont-agent install --write: the
evidence gate needs the new PostToolUse hook on ExitPlanMode.

Added

• push-preview: a plan can declare the preview's evidence. When the
  plan a branch landed was approved through ExitPlanMode and its body has
  a ## Preview section whose first line is evidence: <reason>, a UI push
  passes with no question (rule work.preview-unless-planned-evidence).
  • The approval record. A new PostToolUse hook on ExitPlanMode
    writes the approved plan's canonical body sha to
    ~/.claude/amont-agent/plan-approved/<sha>, hashed from
    planFilePath when the hook runs. Only the approve shape counts
    (tool_response an object carrying plan).
  • The check. The first plan the branch adds under docs/plans/ since
    the merge base with the default branch, read at the commit that added
    it, must have an approved sha and the declaration. A picked mockup, a
    pointer plan, a plan already on main, a section added later and a
    missing base ref are held as before.
  • Audit. Every pass is journalled evidence, with the plan, the
    commit and the sha; every record plan-approved.
• preview register prints question_prefix: [preview <id>] <label>,
  what the marked question starts with, verbatim. A refusal to bind quotes
  it.
• implementation-review counts a resumed review. A reviewer resumed
  with SendMessage on a new block is a review of that tree: the message's
  to is a reviewer's agentId, its result's resumedAgentId matches, and
  the round's structured notification names that SendMessage and agent.
  The plan-review panel still counts fresh launches alone.

Changed

• A marked question approves by id. Approve on a question whose
  [preview …] marker lists a pending registration approves it; the
  repo@sha7 label is no longer required (the id starts with the sha7).
  The "bound NOTHING" answer and its unlisted journal line are gone.
• A chained register binds. preview register only has to be the last
  clause of its line, joined by && or ;; the JSON is read from the last
  non-empty line of stdout. A PreToolUse stamp
  (previews/registers/<tool_use_id>, swept after a day) and an id, page
  and mtime check keep a line an earlier clause printed from passing as
  the register's. No stamp: not bound. A register that is followed,
  piped, redirected or substituted stays unbound.
• The vendored decisions pack carries ADR-0028.

Fixed

• The person was asked twice for one preview (2026-10-08): a question
  carrying the id but no label bound nothing.

---

What's Changed

• feat: approve only what needs judging (2.30.0) by @fredericrous in #84

Full Changelog: v2.29.1...v2.30.0

v2.29.1

Fixed

• push-preview: an unreadable push is held as unreadable. Under
  deny a push whose shape cannot be read is still held, but it no longer
  says "no approved preview covers" or gives the preview steps. It names
  the shape and its own remedy: push the ref by name, in its own command.
  The most common case is git tag v1 && git push origin v1, judged
  before the tag exists; that shape now says so. A push of an existing tag
  needs no preview, as before. The missing-directory hold no longer gets
  the preview steps either (Confirmed::YesSaying gains remedy).
• preview register on the default branch itself. A commit that is
  its own merge base (tagging main) no longer fails with
  `git diff <c>..<c>` failed; an empty range is no screen.

---

What's Changed

• fix(preview): unreadable holds, register on main by @fredericrous in #82
• chore: release 2.29.1 by @fredericrous in #83

Full Changelog: v2.29.0...v2.29.1

v2.29.0

Added

• publish-without-skill: a release or a merge needs its skill. A new
  rule, shipped as advise (ceiling deny). It fires on three commands:

  • a git push that publishes a v* tag (--tags and --mirror included);
  • gh pr merge;
  • an HTTP client sent to /pulls/<n>/merge.

  The rule looks for a call of the tag-release or merge-when-green skill
  (a Skill call, or the person typing /tag-release) in this turn or in the
  human turn before it. A skill called within that window passes silently.

  • Reading the transcript. It is read backwards from its end, and only
    as far as the second human prompt.
  • Subagents. A subagent's own transcript (agent_transcript_path) is
    read too.
  • When it cannot tell, it never refuses. That covers a missing,
    unreadable or truncated transcript, a turn longer than 16 MB, or a skill
    that is not installed.
  • To refuse: git config --global amont.agent.publish-without-skill.stance deny.
  • Measured on 2026-10-07: 8.4–28.3 uncovered publishes per 1000 Bash
    calls by week, falling. Count with tools/skill-rate.py.

Changed

• release-tag-push also matches +refs/tags/v… and --mirror. It now
  shares its matcher with publish-without-skill. --follow-tags stays
  unmatched.

---

What's Changed

• feat(rules): publish-without-skill gate by @fredericrous in #80
• chore: release 2.29.0 by @fredericrous in #81

Full Changelog: v2.28.0...v2.29.0

v2.28.0

Added

• plan-phases-open: an agent no longer stops between plan phases. A
  new Stop hook, shipped as deny: when the agent ends its turn while the
  active plan its branch carries (docs/plans/, changed since the
  merge-base with the remote default branch) still has an open phase that
  is not a 🧑 decision:, the turn continues with that phase. It lets the
  turn end in plan mode, while background tasks run, on a last-message line
  starting with WAITING: <reason>, and after three continuations on one
  phase, when it tells the person instead. Existing installs must re-run
  amont-agent install --write to add the Stop entry; doctor now
  names the events an install lacks.

---

What's Changed

• feat(rules): plan-phases-open — keep an agent going between plan phases by @fredericrous in #78
• chore: release 2.28.0 by @fredericrous in #79

Full Changelog: v2.27.1...v2.28.0

v2.27.1

Fixed

• A delete-only push is not held under deny (#66). git push origin --delete <branch>…, -d (also inside a cluster such as -ud), and a
  push whose every refspec is :<ref> publish nothing, so push-preview
  and implementation-review now pass them (a delete publishes nothing
  and delete-only in the journal) instead of holding them as unreadable.
  A push that deletes and publishes at once, --delete with a src:dst
  refspec, and --prune are still unreadable, and still held under deny.
• A push from a removed session directory is judged, not waved
  through (#67). Every confirmed rule used to decline as soon as the
  session's cwd was gone, so cd /abs/repo && git push … from a session
  left in a torn-down worktree passed every push gate. The test is now
  whether the matched clause's own directory exists, cds followed. When
  it does not, push-preview and implementation-review hold the push
  under deny, because the shell then runs somewhere the hook cannot name
  (Claude Code resets it to the project root). Other rules still decline.
• A firing hook call no longer spawns git once per key (#70). Every
  stance lookup ran git config twice (--global, then --system), so a
  call firing one rule spawned six git processes and a Bash command firing
  several spawned more. Each scope's amont.agent.* keys are now read once
  per process with --get-regexp, both scopes in parallel. A boolean that
  is actually set is still normalised by git (--type=bool), and the scope
  rule is unchanged: only --global and --system, with their includes.
  Measured on a loaded macOS machine, release build, 100–200 runs each: a
  firing Read went from 111 to 38 ms p50, and a Bash command firing several
  rules from 398 to 38 ms p50 (p99 788 → 69 ms). The silent path is
  unchanged (~10 ms). One git spawn costs ~30 ms here under load, so the
  issue's 30 ms budget is not met on this machine.

---

What's Changed

• fix: delete-only pushes, removed cwd, config spawns by @fredericrous in #75
• chore: release 2.27.1 by @fredericrous in #77

Full Changelog: v2.27.0...v2.27.1

v2.27.0

Added

• amont-agent rules --json. One array, one object per rule and per
  assertion, with id, kind, default_stance (what ships), stance
  (what is in force here), max_stance, per_1000 and measured. The
  keys are an interface: the homebrew tap reads them. rules now refuses
  any other argument (exit 2); until now it ignored them, so rules --json
  printed the text table.

Fixed

• The brew caveat names every rule that refuses by default. It said
  only pipe-to-tail refuses, false since plan-review-panel shipped at
  deny in 2.22.0. The release now writes the list from the published
  binary's rules --json (scripts/bump-tap.py), and from 2.27.0 the
  formula's brew test checks it against the brewed binary.

---

What's Changed

• docs(plan): lint-suppression-added, done by @fredericrous in #73
• feat(rules): rules --json; tap caveat generated from it by @fredericrous in #74
• chore: release 2.27.0 by @fredericrous in #76

Full Changelog: v2.26.0...v2.27.0

v2.26.0

Added

• lint-suppression-added: an edit that adds a lint suppression is
  advised. On an Edit, MultiEdit or Write the hook rebuilds the file
  before and after (a bounded read that never opens a FIFO) and compares
  suppression markers, not lines: # type: ignore, # noqa,
  # pyright: headers, eslint-disable, @ts-ignore, #[allow]/#[expect],
  //nolint, and a lint configuration made looser (tsconfig, eslint,
  pyright, ruff, Cargo [lints], .golangci.yml). Editing a line that
  keeps its suppression, or moving one, is silent. The advice points at
  general.no-disabled-safety; the journal records the marker kind and
  whether the file was read whole or only the fragments were seen.
  Ships advise, ceiling deny. tools/suppression-rate.py prints the
  weekly rate from transcripts.

---

What's Changed

• feat(rules): lint-suppression-added by @fredericrous in #71
• chore: release 2.26.0 by @fredericrous in #72

Full Changelog: v2.25.0...v2.26.0