A MAM script often runs inside one herdr server's pane (the orchestrator's)
while targeting another (the isolated MAM server). The pane exports
HERDR_WORKSPACE_ID, HERDR_PANE_ID and HERDR_TAB_ID, and those ids only mean
something on the server that issued them. They reached the scope filters as a
foreign workspace, whose fail-closed miss (ISSUE-3 / F-2b) reports a live
session as absent: has-session, kill-session, agent prompt and the delegate-job
preflight all answered "no such session". Reproduced on Linux and macOS in the
#23 comments, worked around until now with `env -u` at every call site.
mam_herdr drops the three ids for a call aimed at a named server other than the
caller's own. Scope then falls back to .mam/herdr_workspace_id and finally to
server-global, which is what _herdr_ws_scope already does. `default` is not a
crossing: it is the sentinel for "pass no session flag", so herdr answers on
the ambient server, which for a caller in a pane is that caller's own. A caller
outside any pane keeps an exported scope, so H-18/H-21/H-23 keep their meaning.
The fix lives in mam_herdr, not in the shim heredoc: that heredoc is
materialised into $WORKSPACE_ROOT/.mam/shim/herdr, one file shared by every
checkout bound to a control plane and rewritten on every herdr call, so editing
it would put unreviewed code in the control root and let two checkouts
overwrite each other. S11 fixes that invariant.
lib_py/loop/watcher.py builds its own herdr argv and is the one path that does
not go through mam_herdr, so it applies the same rule via scoped_env.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Issue #25: a promoted plan is reused only with --plan-file; the implicit
newest-plan load is gone in both runners, and an empty planner report no
longer passes silently. VERSIONS.md carries the entry; the four merge-side
review reports and their identity sidecars (base 6fc9194 -> head 1e82651,
evidence review-tree) are archived under .agents/reports/<session>/.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VCgu9Fa9iT7TbJUSzdq4Ff
Review R1/N1 (reviewer-grok-01, planner-reviewer-codex-01), reproduced before
fixing. The usage line and SKILL.md both promise "relative to the control
root", but parse_args did Path(...).expanduser().resolve(), which is relative
to the caller's cwd. The runner never chdirs to workspace_root, so the two are
routinely different - the documented recipe runs the loop from the control root
while the source root is a worktree.
Two ways it went wrong: a plan that exists only under the control root was not
found at all (exit 1), and a same-named file in the caller's cwd was loaded
instead, quietly putting an unrelated plan into the first implementation
prompt. That second one is the very failure #25 is about, reintroduced through
the option meant to replace it.
Resolve against workspace_root the way jobs_dir already does (Issue #21 R2);
pathlib's `/` leaves an absolute or expanded-home path untouched. Every earlier
test passed an absolute tmp_path, which is why none of them caught this: P3c
covers both failure modes with the control root away from the caller's cwd, and
P3d pins that absolute and ~ paths are not rewritten.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The --plan-file check landed on the one path a human is watching. The three
--plan-mode loads read a report written by another agent with nobody watching,
and had no content check at all: find_report only tests that the path exists.
An empty or whitespace-only report therefore blanked current_plan in silence.
Phase 2's `if self.current_plan` then flipped the prompt to self-planning and
the run finished rc 0, having paid for a planner job, a Creator challenge and a
refine job and thrown the result away. The escalation refine was worse: it has
no guard, so the correction prompt asserted "follow the updated plan below" and
carried nothing.
All four loads share _plan_text() now - a plan is a readable regular file with
text in it - and each caller says what None means. An empty initial plan stops
the run instead of demoting it, and an empty refine warns and keeps the plan it
already has, which is still good. Those two together leave the escalation
prompt unable to claim a plan it does not carry, so it needs no new branch.
Found by the Creator's challenge to the plan (job 0f63ac1d), sustained by the
Planner as revision 1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Without --plan, Phase 1 resolved a planner session automatically and globbed
<workspace_root>/.agents/reports/<planner>/plan-*.md for the newest file by
mtime, handing it to the Creator as "the plan already agreed on". Both keys are
wrong: the report directory belongs to a session, not to this task, so other
issues' plans pile up in it, and right after a clone or checkout mtime is
checkout order. 47 such plans are committed in this repository, so an ordinary
run could silently implement against an unrelated design.
Reuse is explicit now. --plan-file <path> is the only way to load an existing
plan; it is resolved against the control root, checked before a single job
exists, and refused if the file is missing, not a regular file, unreadable or
empty (an empty file would quietly turn into a "no plan" prompt). It cannot be
combined with --plan, and --review-only warns and ignores it because no
implementation job is delegated there. The implementation prompt itself is
unchanged, so the Creator's contract does not move.
Phase 1 now says which plan it loaded and where from, or that it loaded none.
The legacy shell loop (MAM_LOOP_PY=0) loses the same auto-load and rejects
--plan-file rather than duplicating the check on a path that is on its way out.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- worktree_prepare.sh installs the same files deploy/install.sh does
(.agents/skills/multi-agent-mux-delegate-job/requirements.txt: paho-mqtt,
PyYAML) plus requirements*.txt. Without paho the loop wrapper, stop_session
and the subscriber tests fail in the worktree - seen as 4 suite failures
in a worktree without its own venv, all passing with one.
- review D1: a failed pip install removes the venv it half made, so the next
run recreates it instead of reporting VENV=kept (test w1c).
- review D3: the [worktree context] block's uncommitted-changes flag uses
--untracked-files=all like the guards.
- review D2: the create skill doc says the install needs the package index.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VCgu9Fa9iT7TbJUSzdq4Ff
R1/R2: `git status --porcelain` honours status.showUntrackedFiles, so a
`no` setting in a submodule (or the worktree) hid untracked files from the
checks that justify worktree_remove.sh's single --force and from the review
snapshot guard (_dirty_submodules), which then approved content it never
captured. All three checks (lib_py/worktree.py, worktree_remove.sh,
runner._dirty_submodules) now pass --untracked-files=all. The source tree's
own snapshot was already safe: `git add -A` ignores that setting.
R3: worktree_prepare.sh no longer links .venv to the control root's. It
creates the worktree's own venv with the control root's interpreter and
installs requirements*.txt; a leftover link is refused (exit 1), a control
root without an interpreter is a warning (VENV=skipped); VENV= is printed.
Tests: w1 (real venv, kept on re-run), w1b (link refused / skipped), w6e
parametrised over the hidden setting for the post-stop recheck, w6e2
(hidden untracked file in the submodule and in the worktree blocks the
removal), l17 untracked-hidden / during-hidden (guard refuses before the
first review; a mid-round hidden file makes the round stale). Each new
case fails without the fix.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VCgu9Fa9iT7TbJUSzdq4Ff
- review jobs' [워크트리 컨텍스트] points at the [리뷰 identity] path only (B-2);
remote re-adjudication gets a matching sentence (N-4)
- dirty/untracked submodule content fails the snapshot and makes a round stale,
regardless of submodule ignore settings (B-R1)
- with --worktree/--base and no reviewer, the creator self-review takes the
reviewer path: snapshot diff, evidence check, pre-promotion re-check and the
identity sidecar; NOT PASS still re-reviews without a correction job (B-R2/B-R3)
- snapshot diff failures keep the UNAVAILABLE marker (N-1); prune only when
worktree remove fails (N-3); the empty per-loop review dir is removed (N-5)
- tests L17-L21, B-2 assertions in L6/L11; SKILL.md env stale note; import order
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017R3Eg7MEpSWnm1oB9ZPP6G
Active with --worktree/--base only; without them prompts, flow and promotion
are unchanged (a loop_id is recorded in the checkpoint either way).
- Every review round is pinned (B3): the source's HEAD and its working state
as a tree built through a temporary index (the real index/status/refs are
untouched); a dirty state becomes a ref-less snapshot commit on HEAD
(committed as mam-loop, not as the user). An untracked nested repository
cannot be pinned and is refused. identity = (loop_id, base, head, tree).
- Review isolation (B4): a detached review worktree of the snapshot under
TMPDIR (default), or `--review-isolation env` (the source root with env and
temp-dir instructions). Failing to open the tree refuses the round - the
isolation is never lowered silently. Trees are removed after the round and
on any exit (run() wraps the body; release_guard too).
- The reviewers' diff comes from the snapshot (B13):
diff_collect.sh takes an optional head commit for a commit-to-commit diff;
its two-argument form is byte-identical.
- Stale rounds (B7/B8): after the round the source is measured again; if it
changed, the round's approvals are discarded and the next iteration
re-reviews without a correction job (last iteration: rc 1). Only the last
non-stale round's PASSes are promoted.
- Records (B9/B15): review_rounds in the checkpoint with per-reviewer
transport/evidence, and report-<jid>.identity.json next to each promoted
report: evidence review-tree | source-tree | diff-only.
- Pre-promotion re-check (B10): a source changed after approval is not
promoted (rc 1).
- Remote reviewers (B14, Creator challenge 91c81095): they cannot reach this
host's review tree. Checked before anything is delegated: refused by
default; `--remote-review diff-only` admits them with the snapshot SHAs and
diff as their only evidence (no local path in their prompt), and refuses a
round whose diff is truncated or unavailable.
- run_loop.sh accepts both options (usage identical to the runner's); the
legacy shell loop refuses them before the lock.
Tests tests/test_issue21_loop_b.py L1-L16 (22).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017R3Eg7MEpSWnm1oB9ZPP6G
Issue #21 phase-1 prerequisites (PR #22) add loop runner options and the
control-plane module: MINOR. VERSIONS.md gains the v4.6.0 entry, the
archived design reviews lose their markdown trailing whitespace so
`git diff --check` is clean, and the three merge-side PASS reports
(claude, agy, codex @ 4a81d15) are archived under .agents/reports/.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VCgu9Fa9iT7TbJUSzdq4Ff
Review 9a1adb46 (planner-reviewer-codex-01):
- R1 [P1] lib.sh new-session: a forwarded binding reached the agent as
arguments. With env flags the idle-pane branch prefixed `env K=V ...` to
the command that `herdr agent start ... --kind <agent> -- <args>` passes to
the agent executable. An idle pane is now reused only when no environment
has to be set; otherwise the pane is created by pane split / workspace
create, which take --env. The dead `env` prefix code is removed (its
send-text fallback would have typed the same line).
- R2 [P2] runner: a relative MAM_JOBS_DIR is resolved once against the
control root, the way the wrapper (cwd=workspace_root) reads it, and that
absolute dir is used for delegation, wait, report lookup and cleanup.
- R3 [P2] stop: the row's control_plane.delegate_job_logs_dir is restored
(exported unless the caller set DELEGATE_JOB_LOGS_DIR), so the
cancellation history lands in the row's audit-log root.
Tests: T3-1 and T3-3 now assert that the binding arrives as --env on the
pane-creation calls and that no env/K=V token follows `--` in agent start
(fresh and reused pane); T3-4c (audit root on stop) and T2-15 (relative
MAM_JOBS_DIR) are new. The design doc's compatibility table and a create
comment no longer describe the removed idle-pane env path.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017R3Eg7MEpSWnm1oB9ZPP6G
- --worktree <path> (CLI over MAM_LOOP_WORKTREE): the source root. Only the
diff, base, head and the empty-set check use it; the wrapper, registry,
checkpoint, loop guard, promotion and cleanup stay on the control root. The
collector is the control root's diff_collect.sh (never the worktree's
unreviewed copy), its arguments shlex-quoted. Validation: an existing
directory, the top of its working tree, the same git common dir as the
control root (".git" is resolved against the directory it came from).
- --base <ref> (CLI only, no env - a stale sha would silently widen the next
range): resolved to a full sha, must be an ancestor of the source HEAD.
- --review-only: requires --base, refuses --plan; Phase 2 is skipped and
logged; correction and comply jobs still go to the creator.
- With --worktree or --base every delegated prompt starts with a context
block (path, branch, base and its origin, head, base..head commit count,
uncommitted changes); without them prompts are byte-identical. All call
sites in run() go through one _dispatch helper.
- Fail-closed guard (only with --worktree/--base): before each review round,
self-review included, an empty change set or one git cannot determine
(exit codes, never output parsing) ends the loop with rc 1 instead of
asking for reviews. This closes the fail-open where a loop started after
an out-of-loop commit sent reviewers "(no changes since base commit)".
- The runner honours MAM_JOBS_DIR and passes --registry-dir to the wrapper.
- run_loop.sh accepts the three options; the legacy shell loop
(MAM_LOOP_PY=0) refuses them before the freeze and the loop lock.
Tests T2-1..T2-14.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017R3Eg7MEpSWnm1oB9ZPP6G
Contract (no new event types): heartbeat every N s (automatic timer),
unresponsive after K x N s without an event, `started` within T s of the
injection, retained per-job state, and a remote transport.
- registry: heartbeat_sec / heartbeat_misses / ack_timeout_sec / transport
fields; idle_timeout_sec derives K x N unless explicit (defaults 30/4/60
keep the historical 120 s idle window).
- heartbeat.py + publish_event --heartbeat N: detached fixed-cadence timer
publishing signed `progress` heartbeats (threaded wire publish, local
record first); stopped by the terminal event via pidfile, by re-dispatch
and by lifetime. Terminal events refuse --heartbeat.
- publish_event: retained `<prefix>/state` (status, last_event, seq, HMAC)
sent on the same connection after every event (--no-state to skip);
publisher.publish_raw takes extra messages.
- job_subscriber: subscribes the state topic, recovers terminal status from
retained state, resets idle on any accepted message, records seq_gap and
reconnected in the audit log, idle budget before the first event is
T + K x N (--ack-timeout), --remote mirrors started/terminal events into
the registry through the transition boundary.
- lib.sh send_keys_safe: an accepted-but-unconfirmed submission is confirmed
by agent_status working or the job's own `started` (_sks_job_started);
pane change and busy token are no longer delivery signals.
- wrapper: --heartbeat/--heartbeat-misses/--ack-timeout (MAM_* env),
injected start command arms the timer, brief states the contract,
_wait_started (registry / events.log / subscriber log, bounded by T and
the wall clock) replaces the 10 s inject grace, R1 grace uses only the
on-disk event age (herdr/pane logged once as a diagnostic),
remote:<name> sessions skip preflight/idle/pane and dispatch over the
broker (remote_dispatch).
- remote.py (+ scripts shim): signed retained dispatch (MAM_DISPATCH_SECRET)
and a runner that mirrors the job into its own registry with the
orchestrator's identity, publishes started --heartbeat, runs --exec and
publishes the terminal event.
- loop runner: passes MAM_HEARTBEAT_SEC / MAM_HEARTBEAT_MISSES /
MAM_ACK_TIMEOUT_S through; YAML rows with transport: remote delegate as
remote:<name>.
- docs: MESSAGING.md 2.5 contract + remote transport, delegate-job SKILL.md,
registry.md, loop SKILL.md.
- tests: tests/test_issue18_liveness_contract.py (16); issue 14 pane-signal
tests flipped to started-based; issue 6 working-agent tests acknowledge
with started; CLI snapshot/help captures refreshed.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VCgu9Fa9iT7TbJUSzdq4Ff
PATCH release for the fix/issue-12-8-7 branch (Issue #12 broker password
kept out of job records, #8 fresh install stamps kept on update, #7
same-cwd workspaces packed). No code logic changes.
- lib.sh MAM_VERSION, lib_py/loop/eager.py FRAMEWORK_VERSION and the eight
SKILL.md frontmatters (create, stop, resume, status, monitor, delegate-job,
loop, orc-onboard) move from 4.4.3 to 4.4.4
- VERSIONS.md: current release v4.4.4 (2026-09-13 KST), the eight-row skill
matrix, and a v4.4.4 changelog section in the v4.4.3 format covering the
password-free registry block restored from MQTT_PASSWORD at connect time,
the scrub on every record write, `registry.py scrub-secrets [--dry-run]`,
why auth_token/dispatch_token stay, the update.sh stamp and install_state
merge rules, the sibling-workspace walk in new-session, and the 8 new tests
- VERSIONS.md v4.4.3 ledger correction: its new-test count predated fae70c1,
which added two purge tests; now 26 (19 + 7) with both listed
- README, README.ko, BOOTSTRAP and BOOTSTRAP.ko badges and the MESSAGING.md
"Applies to" header follow the current version; historical 4.4.3 mentions
(v4.4.3 changelog, reviewer reports) stay as history
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HCGpJXhSZasJLU38ksDghe
PATCH release for the fix/issue-10-11-13-15 branch (Issue #10 agy purge
sidecars, #11 orc_onboard message, #13 orchestrator conversation purge,
#15 workspace-anchored job registry). No code logic changes.
- lib.sh MAM_VERSION, lib_py/loop/eager.py FRAMEWORK_VERSION and the eight
SKILL.md frontmatters (create, stop, resume, status, monitor, delegate-job,
loop, orc-onboard) move from 4.4.2 to 4.4.3
- VERSIONS.md: current release v4.4.3 (2026-09-13 KST), the eight-row skill
matrix, and a v4.4.3 changelog section in the v4.4.2 format covering the
agy purge sidecars/annotation/presence lock and residual_artifacts,
`lib_py.agents purge` and `orc_onboard.sh --remove --purge-conversation`
with their exit codes, the past-tense message, the MAM_JOBS_DIR ->
WORKSPACE_ROOT -> cwd registry/log resolution and the 20 new tests
- README, README.ko, BOOTSTRAP and BOOTSTRAP.ko badges and the MESSAGING.md
"Applies to" header follow the current version; historical 4.4.2 mentions
(v4.4.2 changelog, reviewer reports) stay as history
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HCGpJXhSZasJLU38ksDghe
PATCH release for the fix/issue-9-14 branch (Issue #9 herdr server env
isolation, Issue #14 accepted-prompt delivery confirmation). No code logic
changes.
- lib.sh MAM_VERSION, lib_py/loop/eager.py FRAMEWORK_VERSION and the eight
SKILL.md frontmatters (create, stop, resume, status, monitor, delegate-job,
loop, orc-onboard) move from 4.4.1 to 4.4.2
- VERSIONS.md: current release v4.4.2 (2026-09-12 KST), the eight-row skill
matrix, and a v4.4.2 changelog section in the v4.4.1 format covering the
server env marker denylist, MAM_NO_AUTOSTART, the shim ls "no server
running" translation, SKS_DELIVERY_GRACE_S, MAM_INJECT_ACK_GRACE_S, the R1
pane busy-token grace, the T1-7 trigger change and the 13 new tests
- README, README.ko, BOOTSTRAP and BOOTSTRAP.ko badges and the MESSAGING.md
"Applies to" header follow the current version; historical 4.4.1 mentions
(v4.4.1 changelog, IMPROVEMENTS.md ledger, docs/contracts/issue-6, reviewer
reports) stay as history
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HCGpJXhSZasJLU38ksDghe
Issue #5 (Python core migration, CAS state transitions, shadow watcher,
auto-completion gate, eager-import immutable runtime, B-13 freeze removal)
is complete and merged to main. This stamps the release.
SemVer: MINOR, not MAJOR. The change is large but breaks nothing. Across the
whole Issue #5 range there are zero file deletions and zero renames, the event
wire schema stays at SCHEMA_VERSION 1, all four literal script paths survive as
compatibility shims, the 968-line legacy shell loop is still in run_loop.sh,
and every default change reverts through a single environment variable
(MAM_DELEGATION_PY / MAM_ENGINE_CAS / MAM_LOOP_PY / MAM_LOOP_FREEZE). The
riskiest new feature, auto-completion, ships default-off. No operator has
anything to migrate, which is the decisive reason not to call it 5.0.0. The
reasoning is recorded in the VERSIONS.md release notes, not just here.
Close a gap in the lockstep guard. lib.sh exports MAM_VERSION and eager_load()
refuses to start the loop when it disagrees with FRAMEWORK_VERSION, so eager.py
is a load-bearing fourth leg — but test_version_consistency.py only checked
three, and the sole check on eager.py was a hardcoded literal in the phase 4
test that a release bump has to edit anyway. The lockstep test now covers
eager.py, a new test pins the fail-closed gate itself, and the phase 4 test no
longer carries a version literal.
Fix a latent test-isolation defect the bump exposed. The first full run failed
four tests with "MAM_VERSION mismatch: expected 4.4.0, got 4.3.0 in
environment" — 4.3.0 was not in the tree, it was inherited from the shell that
launched pytest and had sourced the pre-bump lib.sh. Those tests were measuring
the runner's shell rather than the tree under test, and only passed until now
because the two values always happened to agree. A session-scoped fixture drops
the ambient value. Production is unaffected: run_loop.sh:17 sources lib.sh
before anything else and lib.sh re-exports MAM_VERSION unconditionally, so a
stale inherited value is always overwritten before Python starts.
Full suite: 750 passed in 912.50s, 0 failed, 0 skipped.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
Two findings from round 2, both reproduced before being accepted.
R2 -- attribution was compared before the terminal write and nothing tied the
two together. A job could complete, be re-dispatched and claimed again in that
window, and evidence from the first attempt then satisfied a status check and
finished the second one. Reproduced with real transitions: committed, with the
registry already on iteration 2.
This corrects my own earlier reasoning, and I want to be plain about it. Last
round I argued that the attempt guard could live entirely in the evidence layer
because that is where attempt identity lives. That holds for comparing evidence
to a dispatch, and it does not hold for atomicity: a precondition can only be
atomic with the commit if the transaction can read it under its own lock. So the
identity now lives in the record.
It is an opaque dispatch token, not the iteration counter. That distinction is
the part of the earlier rebuttal that still stands: iteration already means the
loop round, and comparing a dispatch attempt against it produces a guard that is
vacuous when they agree and wrong when they diverge. A token has no competing
meaning. It is minted when an attempt begins -- claim and redispatch -- and the
gate passes the token it decided against, so any intervening re-dispatch refuses
the commit.
R4 -- the intent identity was being emitted but nothing consumed it. A replayed
notification is re-published with a fresh wire sequence, so the monotonic guard
cannot recognise it; to that guard a retry looks new. The subscriber now
deduplicates on the stable intent identity, leaving the sequence guard as the
rule for ordinary events. Across a restart the disk fallback resolves terminal
status and a duplicate terminal is refused as an idempotent ack, so a replay
after restart cannot double-apply either.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
The important correction is an honesty one, raised independently by two
reviewers. The note claimed startup recovery runs in the subscriber. It does
not: nothing in production imports the gate or the recovery function, and there
is no per-job opt-in path, so the flag is currently a no-op in the live loop.
Phase 2b delivers a verified library and gate; wiring it in belongs to phase 3,
and the note now says so and lists what wiring requires.
Also records the checks that are now mandatory rather than caller-dependent,
that the rollback switch disables auto-completion too, that backend keys match
exactly, that bookkeeping may touch one field, that settling is conditional on
the intent actually sent, and why the intent's identity travels in the payload
rather than relying on the transport sequence.
Fixes three stale references: task.md named the pre-move path for finalize, the
loop package description still described itself as shadow-only, and a comment in
mqtt_common still pointed at the engine's original location.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
Seven blocking items across three reviews, every one reproduced here before
being accepted. Two themes run through them: a gate whose strictness depended on
what the caller remembered to pass, and a boundary that trusted its only current
caller instead of enforcing its own rule.
Independent verification is no longer optional. Omitting the report directory
skipped the disk check entirely and handed the decision back to the collector's
own flags -- the self-certification this design exists to prevent. It is now a
refusal, as is an empty directory or a job with no recorded dispatch.
Evidence is now bound to its target. The policy compares a scenario against the
dispatch context carried inside that same scenario, which is self-consistent by
construction, so one job's evidence could complete another. The gate now checks
the job, attempt, session and pane generation against the recorded dispatch.
A reviewer with no verdict token is no longer completed. Phase 0 made this a
safety contract and run_loop reads a missing token as NOT PASS, so inferring
completion first would start a correction round against a review that was never
finished. The requirement applies to reviewers only.
Bookkeeping may write one field. A status-free request was an unrestricted merge
on a terminal record: it could rewrite the prompt and the session. Settling is
also conditional now on the exact intent that was sent, because publishing
happens outside the lock and a newer intent can legitimately land mid-flight;
clearing unconditionally dropped a notification nobody had sent. The intent's
own identity travels in the payload, since the transport sequence cannot
identify a retry of the same intent.
Side-effect evidence must be evidence. A kind with no reference names something
without evidencing it, and the CLI parser could not be the only check because
the API and hand-written manifests bypass it.
Also honours MAM_ENGINE_CAS=0, which the operations note already promised
disables auto-completion, matches backend keys exactly rather than by prefix so
a future name cannot inherit evidence never collected for it, records an
absolute report path, and imports the notifier at module scope.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
Two failures the full suite caught, both mine.
finalize.py started in lib_py/loop, which made publish_event import a sibling
package. That is the rule phase 1 wrote down and phase 2a already made me move
the engine for: under a frozen loop, delegation is bound to one tree while the
parent may be another, so reaching sideways mixes them. The manifest is written
into the job's own report directory and the CLI that declares it is a delegation
entry point, so finalize belongs there. Second time on the same rule, and the
guard caught it both times.
The publish_event --help golden no longer matched, because phase 2b
intentionally added the finalization options. The diff is exactly those three
options and nothing else, so the capture is refreshed rather than the assertion
weakened, and the test now states when a refresh is legitimate: an intended
extension, never drift nobody meant.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
Three scenarios join the contract set, so the fifteen vectors now cover an
unproven side effect, an undecidable attribution, and an agent whose state goes
unreadable in the window between finalization and notification. All fifteen
match their expected decision.
The side-effect case needed a decision about layering. Authoritative enforcement
belongs in manifest verification, which has the job record and therefore knows
what was declared; the policy honours an explicit negative from the collector so
the case is expressible as a vector. Absence is not read as proof there, because
the manifest check is the layer that fails closed on missing evidence.
The operations note records why the agent declares and the reader re-derives,
the three gates and why the backend allowlist can only be narrowed, the
notification transaction, and the rollback.
It also records a deviation from my own plan. §D2 asked for reviewer sign-off
before promoting an attempt id into the record, since that changes a contract
phase 1 froze. Measurement during implementation showed the change was not
needed: the attempt guard already lives in the evidence layer, where attempt
identity actually is, so auto-completion passes the engine a status precondition
instead. The schema is untouched, the frozen contract holds, and a
characterization test fails the moment a record starts carrying an attempt id,
which is when to revisit it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
The gate does not re-decide anything. Phase 2a already decides, fail-closed and
against contract-pinned vectors; this supplies real evidence to that decision and
commits only if three independent gates agree: the global switch is on, the
backend has observed evidence that it reaches a quiescent state, and the job
itself opted in. Default is off.
The backend allowlist is evidence-based rather than optimistic. Inferring
"finished" from stillness requires stillness to have been observed, and only
three of six backends have that. An operator may narrow the list; a name that
was never observed is dropped rather than trusted, so it cannot be widened by
configuration.
Everything else holds. An unreadable manifest, a report rewritten after
finalization, missing evidence for a declared side effect, a policy that raises:
all refuse. "Could not decide" is never "done". A late explicit terminal wins,
because the boundary refuses the inferred one -- that is the boundary working,
not an error in the gate.
Notification intent is committed in the same atomic write as the state, so a
crash between committing and publishing leaves a record that still says the
world has not been told. Recovery re-sends; duplicates are already dropped by
sequence. A failed send deliberately leaves the intent pending, because an
unsent notification that looks sent is worse than one that is retried.
Two fixes this surfaced. Settling used a request that terminal records refuse,
which made an owed notification impossible to ever settle; settling is now a
status-free bookkeeping request, valid in any state. And the phase 2a guard
against shell references was matching hashlib.sha256 as a substring, so it now
requires a word boundary.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
The agent declares that its work is finished; the reader re-derives the hash
from the file rather than trusting the number in the declaration. A manifest the
watcher wrote itself would attest only that the watcher saw a stable file, which
is the self-certifying evidence shape rejected earlier in this issue.
Finalizing promotes a staged report by atomic rename before writing the
manifest, so a reader cannot observe a half-written report as final. Rewriting
the report afterwards invalidates the declaration immediately, because the
recomputed hash no longer matches -- a new body needs a new finalization.
A job that declared side effects must show evidence for every one of them.
Partial evidence is treated as none: the work is not known to be complete.
publish_event gains --finalize-report, --side-effect and --attempt-id, and does
the finalization before publishing. A `completed` event whose report is still
being written would invite exactly the partial adoption the contract forbids, so
a failure there fails the publish.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
Both reviewers sustained the round 1 fixes and found four more. Each was
reproduced here before being accepted; none were rebutted.
F-C (claude, blocking) -- my own round 1 fix caused this. Making refused writes
exit non-zero also classified an idempotent ACK as a refusal, and the wrapper
finalises an approved job with `status --set completed` on a record that is
already completed. It runs under `set -e`, so the *success* path aborted with
exit 4 before its validation hook and its final stdout line. Reproduced at the
shell. An idempotent result is now distinguishable from a refusal and exits 0; a
genuine conflict still exits 4.
F-D (claude) -- routing the operator write around update_status dropped the
status vocabulary check. Under the rollback flag an invalid status was written
to the record with exit 0, where before it was refused with exit 1 and the record
left alone. The check now runs first, in both modes.
R1 (codex, blocking) -- the previous fix compared identifiers only when they
happened to be present, so evidence carrying none at all was still approved on
the dispatch record's own flags. That is the self-assertion the check exists to
replace. Every identifier must now be supplied and agree, and a report manifest
must associate with the attempt being judged. Identifiers are read from a
collector envelope beside the payload rather than from inside it: a native Herdr
payload may not carry MAM fields, so one appearing there is either fabricated or
a contract violation, and either way it is not evidence.
R4 (codex) -- a sequence number orders nothing unless the compare and the write
are atomic. Two flushes could both read the old value, the newer land first, and
the slower overwrite it. Reproduced deterministically. The mirror now serialises
read-compare-write under a per-job audit lock, taken outside the registry
transaction so a slow mirror still cannot lengthen it, and each write uses its
own temp file instead of a shared name.
All eight new regressions fail against the pre-fix code.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
Two reviewers raised five defects. Every one was reproduced here before being
accepted; none were rebutted.
F-A (claude, blocking) -- the CAS table had no row for two transitions the
wrapper legitimately makes, so `submit --type loop|discuss` silently stopped
working by default: the per-iteration re-dispatch was refused and its metadata
discarded, the exhausted-budget finalisation was refused, and the CLI returned 0
both times. Reproduced against the real modules. Fixed by naming those
transitions rather than carving an exception: `redispatch` and `finalize` are
requests gated on operator authority, so the wrapper can move a finished record
while an agent still cannot touch its own terminal. A re-dispatch must advance
the iteration, so a replay cannot resurrect a job. The operator CLI now exits
non-zero with a reason when a write is refused -- the silence is what let this
hide.
F-B (claude) -- moving the audit mirror out of the lock made status.json
last-writer-wins, so a slow flush could leave it behind the registry. It now
carries transition_seq and skips a write that would go backwards, restoring the
sequential consistency the lock used to provide for free.
R1 (codex, blocking) -- the policy called attribution with only the dispatch
record, so it proved the record was well-formed and never confronted the
identifiers the evidence carried. A finalized scenario with a different session,
a different attempt and a manifest from another attempt was still approved.
Observed identifiers are now collected from the timeline and every one must
agree; a manifest belonging to another attempt holds. Collecting every
occurrence rather than the first matters, because taking the first would drop
exactly the contradicting value this is meant to catch.
R2 (codex, blocking) -- a supplied expected_attempt was ignored when the record
had no attempt field, so a compare-and-swap with nothing to compare committed.
It now fails closed, which matters because 2b plans to gate inferred completion
on precisely that precondition.
R3 (codex) -- a status could ride in as an ordinary field, so `progress` could
commit a terminal from inside the boundary that exists to prevent it. Status is
now settable only by the request that owns it, unknown requests are refused, and
pending has an owning request instead of being an untyped merge.
Each of the eleven new regressions was checked against the pre-fix code first.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
policy.decide is a pure function so the phase-0 scenario fixtures can be fed to
it verbatim. Those seven decisions survived three-reviewer scrutiny, so they are
the contract rather than a fresh set of rules, and the test asserts against
expected_policy.decision directly.
Order inside the policy matters. Transport doubt and staleness invalidate the
evidence itself, so they are answered before anything is inferred from it; a
blocked agent is waiting for a human and can never be completed; attribution
including the execution anchor comes before quiescence, because idle is not
evidence of finishing when a session is reused; and a report must be finalized,
never merely paused, since a debounce is a candidate filter and not an authority.
The shadow runner is never handed the engine, so it has no route to a write.
That is the structural half of P2A-03. The measured half hashes every file under
the registry directory around a full run and asserts it is unchanged, which is a
stronger claim than checking a status field.
Adds the five scenarios P2A-04 requires: a raced double dispatch, a reused
attempt with no anchor, a mid-write report, an unknown status, and a pane
restarted mid-flight. They pass phase 0's own fixture validator, which is what
keeps them honest. The phase-0 test that pinned exactly seven scenarios now
requires the seven to still exist rather than forbidding growth.
Also pins the two phase 1 handoffs as dormant: nothing here spawns a daemon, so
preexec_fn stays unreachable, and nothing calls a frozen shell helper.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
watcher wraps one `herdr agent wait` barrier and observes only; it never
concludes. The live contract is encoded rather than described: the timeout
parameter is named timeout_ms because herdr counts milliseconds, --until is
always passed explicitly since omitting it silently drops `working`, and a
non-zero exit is read from stderr because stdout is empty by contract.
A timeout counts as healthy. The socket answered and the state simply did not
change; treating that as a failure would make an idle agent look unreachable.
Only a matched snapshot becomes evidence, so a timeout can never be mistaken for
an observation of agent state.
Backoff doubles from 1s to a 16s ceiling and drops the cached snapshot on
failure. Reasoning from state captured before a gap is exactly what would make a
disconnect look like a completion, which is why the contract answer there is
hold_reconnect.
store keeps dispatch and attribution evidence outside the registry. The job
record carries none of attempt_id, agent_session_uuid or
pane_process_generation, phase 1 froze its schema, and this phase may not write
the registry at all. Keeping the evidence here lets shadow mode carry full
attribution while the registry stays untouched; promoting these fields is a 2b
decision. The execution anchor starts false because acknowledging a dispatch is
not evidence that the attempt ran.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
engine.transition is now the only writer of job status. It implements the target
table in state-transitions.md §3: claim, non-terminal metadata, explicit
terminals, idempotent re-receipt, refusal of conflicting or late signals with an
audit entry, stale-attempt rejection, and deadline timeout.
Two write paths become one. update_job_status merged fields unconditionally;
pick_pending hand-rolled a correct pending->running swap plus its own audit
mirror. Neither was wrong alone, but a rule added to one silently missed the
other, which is the shape that produced the phase 1 transport bugs. Both keep
their signatures and their preserved legacy bodies behind MAM_ENGINE_CAS=0.
A merged `status` value is routed to the request that owns it, so a raw field
write can no longer slip past the table. Getting that wrong first was caught by
the characterization tests: a metadata merge was overwriting the caller's status
with the pre-merge value.
Lock discipline is now enforced, not described. The lock covers a decision and
one record write; audit mirroring is queued and flushed after release, carrying a
monotonic transition_seq so ordering survives the move. A test watches for
subprocess, socket and sleep calls while the lock is held and asserts none.
Structural correction to my own plan, caught by a phase 1 guard: the plan put the
engine in lib_py/loop, but delegation would then have had to import a sibling,
which is exactly what mixes trees under a frozen loop. The engine belongs with
the registry writers, so it lives in lib_py/delegation and loop consumes it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UrDNJkHmm18uu7n7JwV5Qd
GodopuandClaude Opus 5