Ticket
The self-review loop reviews a change faithfully against its INTENT block, and its finders leave alone any "out of scope on purpose" cut that only narrows the ask — they flag a cut only when it contradicts the request or a governing document — so the author's own scoping is largely exempt from review, and the premise is questioned only after the author is invested — and a correct implementation of a wrong ticket survives any number of correct review rounds. A team's ticket is independent because someone else groomed it. This skill is that someone: one fresh reader, before the first edit, who may question exactly that line. It costs one Sonnet agent reading a ten-line block and the request — not a diff — and three tool calls here: write the block, spawn the reader, wait.
When, and the skip
Before the first edit of a task that will write code. After the code exists a validator is reading a ticket the code has already answered, and the Stop gate counts that as no validation at all (§4).
The gate never refuses a skip, and a skip is always written down: every
counted convergence marker must say --intent validated, author or skipped, so
there is no silent default. Skip when the request is its own ticket — a fix
whose premise you have checked rather than assumed. Work that implements an
approved plan, spec or ruling is never its own ticket: the plan is what was
asked, the chat only points at it, and a block written from memory of the plan
is the one that drifts from it — do not skip. Two things audit the call
afterwards: the record says skipped, and the loop's round 1 computes a tier —
a skipped ticket that tiers M or L is a mismatch you name in the final report,
the same way a forced tier is named. A skip you cannot state there in one
clause was not a decision.
1 · Write the ticket
The ticket is the INTENT block, at <reviews>/intent.md — the file the
self-review skill's §0 otherwise writes, and every brief carries. There is no
second file on purpose: two would drift, and the finders would review against
whichever one got copied.
<reviews> is <scratchpad>/self-review/ — the session scratchpad from your
system prompt, else what ${CLAUDE_PLUGIN_ROOT}/scripts/reviews-dir.sh prints — the
same path on every call in the session (with the same $TMPDIR) — outside the
repository, for the loop's own reason (§0). Fill the template from
${CLAUDE_PLUGIN_ROOT}/skills/self-review/references/briefs.md, "The intent
statement": seven to ten lines, the governing documents cited as line ranges (or
none: with what you checked), the premise a checked fact or marked unchecked,
the invariant written so a test could fail it. You are writing what the reader
will attack, not a summary of what you are about to do.
2 · Spawn the reader, then wait in one call
One Agent call, subagent_type: "self-review-ticket-validator", named
self-review-ticket-validator-<suffix> with a suffix no earlier validator in
this session used (a short slug of the task). The type is what the Stop gate
recognises, the name is what wait.mjs finds the transcript by, and both
spellings keep it under tree-guard. Its prompt carries three things and
nothing else:
- the absolute path to
<reviews>/intent.md - the absolute path to
<reviews>/user-words.md— the user's own words, the thing the ticket claims to serve, written by one Bash call before the spawn:"${CLAUDE_PLUGIN_ROOT}/scripts/user-words.mjs" --out <reviews>/user-words.md. It reads them from the session transcript, so no paraphrase of yours stands in for them; the review reads them again at round 1, so a correction typed after the ticket still reaches the finders. - the repository root, so it can grep the premise
Then, as the very next call with nothing in between, Bash timeout 600000:
"${CLAUDE_PLUGIN_ROOT}/scripts/wait.mjs" --work <reviews>/ticket --round 1 <name> \
&& "${CLAUDE_PLUGIN_ROOT}/scripts/salvage.mjs" <session-id> <name>
wait.mjs blocks on the validator's own transcript and exits 0 once it is no
longer active; salvage.mjs then prints its last message, which is the report
— nothing else carries it (a report sent through SendMessage may never arrive).
The session id
is $CLAUDE_CODE_SESSION_ID, or the UUID in your scratchpad path without it. The ticket waits in a directory of its own
because wait.mjs keeps its state under round-<n>/ and numbers from 1, and
the ticket is not round 1 of anything. Exit 1 means call it again, now; exit 3
means it went dead — salvage by hand, and re-spawn once under a new suffix only
if the transcript holds no verdict; a row marked stalled is resumed, not
re-spawned (both in self-review §2f, and in its references/recovery.md). Do not end the turn to wait, and do not check on
it any other way (self-review rule 1).
If the type is not offered, a general-purpose agent with the agent file
(${CLAUDE_PLUGIN_ROOT}/agents/self-review-ticket-validator.md) pasted at the
top of its prompt and the same prefixed name still reads the ticket and is
still guarded — but the gate reads the launched type, so its reading is
recorded as author with a --note saying who read it.
3 · Act on the verdict
sound— build it.revise— overwrite<reviews>/intent.mdwith theintentblock the validator returned, in full, and build against that. The finders review the fileround.shis given, so the corrected block has to be that file; your version is not lost — it sits in the validator's transcript beside its reasons — and a second file is what "no second file" forbids. Read the failed checks before building: arevisethat moved the premise or the scope is what this skill exists for, and one that only moved wording is the signal it should not.do-not-build— do not make the first edit. Report the validator'sreason, quoted, with what already does the job: that is the deliverable, and a turn that changed no code needs no marker. If the user then says to build anyway, or told you in advance not to check in, rewrite the Premise line to carry the objection and the override, build, and say in the report that the verdict was overridden. An objection carried in the ticket is one the finders can attack; an objection you dropped is a ticket nobody groomed.
4 · At the end of the task
The self-review skill's §0 finds <reviews>/intent.md already written and runs
round.sh --new-review --work <reviews> --intent <reviews>/intent.md, which moves
intent.md and ticket/ into the review dir it allocates — do not write a
second block.
The marker then says who read the ticket, in either form the self-review skill's §4 gives:
"${CLAUDE_PLUGIN_ROOT}/scripts/converged.sh" --converged --rounds 2 --fixed 3 --dismissed 1 --open 0 --intent validated --tier M
or "intent": "validated" in CONVERGED.json. The field is required on every
converged and not-converged marker, because an absent field and "nobody
read it" would otherwise be the same state, and it takes exactly validated,
author or skipped.
validated is checked, and only its order. The Stop gate refuses it unless
an agent launched as self-review-ticket-validator completed — its idle notice
in the transcript — before this task's first code change, counting your own
edits, any applier's, and a shell command whose write target the gate cannot
name — an unnamed write is still a write, and it is ordered like any other. It never reads the verdict: a revise you acted on
counts exactly as a sound does. It is scoped to this task: back to the last accepted marker, or further when a
human prompt interrupted a still-running agent. A validator from an earlier,
already-marked task does not count — only one that finished after that boundary
and before your first edit. Refused, write
author or skipped, whichever is true, in a message of its own — never spawn
a validator at that point, because it would read a ticket the code has already
answered, which is the one thing the field exists to make visible. If it
refuses although the validator finished before your first edit, the idle notice
landed late: author with a --note saying so is the truthful record, since
the note is the only free text the marker has. It refuses twice by default
(gate.maxReminders); a third attempt is released with a log line saying the
intent may never have been read, and that line is then the record.
The final report's outcome line carries the answer: Intent: validated before coding (sound) — or (revise: premise, scope) naming the failed checks, or
(do-not-build, overridden) — Intent: author-written, unvalidated, or
Intent: validation skipped, with the tier beside it when you skipped. That
line is the measurement, and its limit: convergence is a claim about the code
against the stated intent, never that the intent was right. This skill records
whether anyone independent read it, which is a smaller thing.