Unattended /autoplan on Claude Code says its phase guard is on, and a report hidden in reasoning gets a clear fix (GST-3062) (#4481)

An unattended /autoplan run on Claude Code now says its phase guard is
on, instead of saying it is off while the guard blocks it. A phase
report written only in reasoning now gets a denial that says so and
names the exact command to send.

Why: in the day-one eval, two runs were blocked by a guard their own
record called off, and one run ended blocked after four publish attempts
whose report was invisible.

Nothing to do.

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>
This commit is contained in:
Garry Tanandcapy-ai[bot] authored and GitHub committed 2026-11-18 08:47:07 +00:00
1 parent f87d4325a3
commit 44170d6791
19 files changed
+244 -34

No files matched your search

+16 -1
View File
@@ -214,8 +214,23 @@ be verified, so Claude Code is not writing it to the journal in time.
<a id="publication-missing"></a>
### `publication_missing`
No filled report for the previous phase appears after its close packet Read.
Publish it as your own assistant text in a message whose only tool call is
Publish it as your own visible assistant text, starting with the line
`**Phase <N> complete.**`, in a message whose only tool call is
`true autoplan-published <phase>`, then enter the next phase in a later message.
The close packet, the active plan or any saved report file does not publish it:
the guard's job is to see the report in the conversation, and a file the model
can write without sending the report would let a phase start unpublished.
<a id="publication-not-visible"></a>
### `publication_not_visible`
A `true autoplan-published` message came after the close packet Read, but
none of those messages carried visible text, so nothing was published. Text
written in reasoning (thinking) is visible to neither the user nor the guard;
an unattended run that keeps putting its report there gets this denial on
every attempt instead of a fallback. Send one message whose visible text is the
filled report, starting with `**Phase <N> complete.**`, with
`true autoplan-published <phase>` as its only tool call, then enter the next
phase in a later message. The denial names the exact command.
<a id="publication-unflushed"></a>
### `publication_unflushed`
+23 -6
View File
@@ -3495,14 +3495,31 @@ keep the line in the review record.
### `autoplan guard: not enforced by this host; publication order is unverified (GUARD_NOT_INSTALLED)`
**Meaning.** The autoplan publication hook only executes inside Claude Code.
On every other host, and in every `unattended` session, the hook is present
but never runs, so phase publication order rests on the snapshot hashes
recorded per phase rather than on the guard.
On every other host (Capy, Codex and the other env-var hosts), or when the
installed `autoplan/SKILL.md` does not declare the hook, it never runs, so
phase publication order rests on the snapshot hashes recorded per phase
rather than on the guard.
**Fix.** Nothing on this host. The line is printed once per run and belongs
in the review record; `gstack-artifact validate` requires it for an
unattended run. On Claude Code a *missing* hook file still denies every
guarded call (a broken install must not pass as an absent one).
in the review record; `gstack-artifact validate` requires it (or
`GUARD_ENFORCED`) for an unattended run. On Claude Code a *missing* hook file
still denies every guarded call (a broken install must not pass as an absent one).
<a id="guard-enforced"></a>
### `autoplan guard: enforced by Claude Code's /autoplan hook ... (GUARD_ENFORCED)`
**Meaning.** `gstack-skill-start` runs inside Claude Code (`CLAUDECODE` is set
and no other host is active) and the installed `autoplan/SKILL.md` declares
the publication hook. Claude Code runs that hook whenever `/autoplan` was
started as a skill (the slash command or the Skill tool), in an unattended
session too: a phase-entry Read or reviewer dispatch is denied until the
previous phase's report is published. A session that only reads SKILL.md
with the Read tool does not register skill hooks.
**Fix.** Nothing to change. Publish each phase report as visible message text
(not reasoning) in a message whose only tool call is
`true autoplan-published <phase>`; the denial codes are in
[autoplan-guard-troubleshooting.md](autoplan-guard-troubleshooting.md).
<a id="review-status-mismatch"></a>
### `review-log: status mismatch ... (REVIEW_STATUS_MISMATCH)`
+9 -6
View File
@@ -47,12 +47,15 @@ it, so `headless` keeps its BLOCK semantics.
`timing: analytics skipped (...)` instead of writing to a disk that vanishes.
Point `GSTACK_STATE_ROOT` at a durable directory when you want cross-run
history (see [state-root.md](state-root.md)).
- **The autoplan publication guard is named, not faked.** On hosts that do not
execute Claude Code hooks, and in every unattended session, `gstack-skill-start`
prints `autoplan guard: not enforced by this host; publication order is
unverified (GUARD_NOT_INSTALLED)` once. `gstack-artifact validate` requires
that line in an unattended run's `review-record.md`. Claude Code's
deny-on-absent behavior is unchanged.
- **The autoplan publication guard is named, not faked.** `gstack-skill-start`
prints one guard line for /autoplan, chosen by host, not by session kind. On
Claude Code with the hook declared in `autoplan/SKILL.md` it prints
`autoplan guard: enforced by Claude Code's /autoplan hook ... (GUARD_ENFORCED)`,
because the hook denies phase entry there in unattended sessions too. On hosts
that do not execute Claude Code hooks it prints `autoplan guard: not enforced
by this host; publication order is unverified (GUARD_NOT_INSTALLED)`.
`gstack-artifact validate` requires one of the two in an unattended run's
`review-record.md`. Claude Code's deny-on-absent behavior is unchanged.
An unknown `GSTACK_SESSION_KIND` value coerces to `interactive` and says so:
`SESSION_KIND: interactive (unknown kind '<x>')`.