A new skill, voyage:reviewing-adversarially (contract S4): no praise, no
finding without a source (file:line or a quoted line), no downgrade without
a reason, summary-only said and not judged, a Declined to Judge list, claims
checked against the diff rather than the report, axes kept separate.
Red baseline first: six review runs by helpers without the skill (two
invented fixtures) had no separate Declined to Judge list in 6 of 6, an
unsourced finding in 6 of 6, praise in 2 of 6 and a verdict on something
unmeasured in 4 of 6. With the skill: 6 of 6, 0, 0 and 0.
- brief-conformance-reviewer, code-correctness-reviewer, review-coordinator
and scope-guardian preload it through skills:; the 21-line "Interaction
Awareness - MANDATORY OVERRIDE" block copied into three of them is gone.
Each agent keeps its own role.
- The two reviewers and scope-guardian end with their own Declined to Judge
list; /trekreview names the skill at spawn and passes the reviewers'
lines to the coordinator, which copies them into review.md as prose
(the JSON declined_to_judge array and the verdict are unchanged).
- Eval cases 3 / 2 / 1; pins in tests/lib/skills-review-prose-pins.test.mjs.
- tests/fixtures/jsonl-schemas.md: the trekreview-stats line_ref moves with
Phase 8 (636-686).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Round 8 (0f400d3..29b4ac9, BLOCK: 1 BLOCKER, 12 MAJOR, 1 MINOR). Each fix
began with a test that failed on 29b4ac9 (13 tests, all red there); suite
2221 (2219/0/2).
- BLOCKER: a backticked check whose trailing comment holds a quote, $,
backtick or backslash is refused, not dropped. The comment was never
screened, while the executor hands the Verify span as written to red
proof, the test lock and the done line inside --cmd "...", where a $(...)
runs and a quote ends the argument. trekplan's text and the reviewer's
NOT RUN table say so.
- A span that is only a comment is not-a-command, not a placeholder. The
dropComment text says the allowlist still refuses a word with a # in it.
- The criteria runner, comparison-gate and end-state-gate use isMain: they
did nothing and exited 0 under a symlinked path. The test now refuses any
process.argv[1] comparison outside lib/util/is-main.mjs.
- Files: in a field with backticked paths, a plain word counts only before
the entry's description (a dash, or a colon after the path); a name with
no slash only first in its entry or after and/or/+/&. "mocks
fs.readFileSync", "runs on Node.js 22.x" or "no change to package.json"
no longer make a tests-only step look like it builds something.
Lowercase makefile, rakefile and justfile are build files. Measured on the
520 Files: fields in 48 plans of the marketplace and the test repo:
3 fields read differently (two notes no longer read as files, one
duplicate name dropped), tests-only changed in 0.
- The test lock compares --cmd without its trailing comment, and the test
lock and the done line read a stop with stoppedBy, the one reading.
red-proof drops its second reading, spawnError.
- release-check reads a relative --review in --repo or, when only that one
exists, from the current directory as 6.3.0 did; two different files are
exit 2, never guessed. The runbook says so.
- One trailing-json reader: a closing fence may be longer than three
backticks, in the block reader and the verdict gate alike, and the review
validator and the synthesis digest use it, so three backticks inside a
JSON string no longer hide a BLOCKER from the remediation-step check.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Round 6 (v6.2.0..c3641ff, WARN: 11 MAJOR, 3 MINOR; 0 BLOCKER). The round-5
fixes hold (code-correctness-reviewer); three of the new MAJOR were side
effects of running checks with no shell. Each fixed behind a test that
failed first on c3641ff.
- git checks stay in the repo too: no path that climbs out with .., starts
at / or reads <rev>:../x, and no --no-index (both reviewers).
- A trailing `# comment` on a shell-fence line is dropped: with no shell,
its words made an honest check NOT RUN, among them the first check of the
shipped example plan.
- A glob reaches `node --test` only on Node 21 and later, which expands it
itself; on an older Node it is refused with that reason (engines >=18).
- With no shell, exit 126/127 is the test program's own: red proof reads it
as a red (the control run still tells an environment fault apart), the
done line and the test lock as a failure.
- A runner that could not start (not installed) is NOT RUN in the criteria
runner and not-run (exit 2) in regression-check, as red-proof and
done-line already said; never a red suite.
- Tests-only: a backticked `Makefile` alone in its entry is a path, and red
proof confirms the plan's tests-only reading against the step's own diff
(the altitude finding: the diff is a fact, the Files: prose is not).
- release-check stops a suite that never ends (15 minutes) and reads it as
not measured (exit 2).
- The two lines this release added to pre-bash-executor.mjs are ASCII.
- MINOR: git and the other runners share one plain-word pattern.
Not in this round: MINOR f00a2eb8 (shared norm/ticked-path and isMain
helpers) and MINOR on plan-review-dedup's repeated tokenizing.
Suite 2194 (2192/0/2).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
voyage-lekerepo end-to-end trial (2026-10-06), report points 1, 5, 11:
- engine-findings: recommended_action is "Fix <file:line> so this no longer
happens: <failure scenario>", not "Resolve: <title>" (8 of 8 engine
findings were set aside as actionability:empty); the coordinator never
sets an engine finding aside for Actionability.
- findings-schema --repo-root: a cited line past the end of an existing
file is FINDING_LINE_OUT_OF_RANGE, so Phase 5's bounded re-ask fixes it
before the coordinator sees it. The reviewers' instruction said "the
actual line number from the diff"; it now says the file's own line,
never the patch's.
- fill-finding-ids.mjs: the coordinator (no shell) writes @@ID_<n>@@ and
trekreview Phase 7 computes them over (file, line, rule_key); an
unresolved placeholder or a wrong 40-hex id fails. The 4-field id with
`title` the docs named is gone; the code always used three.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A denylist in front of /bin/sh is whack-a-mole. Measured 2026-09-18, end to
end through both screens: 5 of 11 named evasions ran with real effect - a
`command` prefix reached git, an escaped `rm` inside a shell fence deleted a
directory, `find -delete` deleted a file, `>|` and `tee` wrote outside the
working tree, a python one-liner deleted the whole tree - and 19 of 28 got
past the refusal list on its own. Every quoting, aliasing and indirection form
of the shell is another mole.
So the screen is now an ALLOWLIST. A criterion runs only when its first word
is a known test runner (npm test, npm run <script package.json declares>,
node --test, vitest, jest, pytest, python -m pytest, uv run pytest,
cargo test, go test, make test, bash <script under tests/>, a read-only git
subcommand) AND the command carries no shell operator and no newline.
Everything else is NOT RUN with the reason said out loud: never run, and never
reported as a failure either - an absent measurement is not a finding. That
also closes the smaller hole in the same file: a bare word a sentence merely
names (`whoami`, `login`, `package.json`) is no longer executed, because it is
not a runner.
REFUSED_BY_POLICY is gone with the list that produced it; a command outside
the allowlist is `unrunnable`, which in plan mode still fells the run and in
brief mode is reported to the reviewer as an absent measurement.
What the allowlist deliberately does NOT do, said in the file and in the
reviewer's rubric: it is not a sandbox. `npm test`, `npm run <script>` and
`make test` run whatever the repo's own package.json/Makefile says they run,
including a script that pushes - that is the repo's responsibility. And it
rejects honest commands too: an env prefix, a project's own binary, anything
piped. A check that needs one of those is declared through
`bash tests/<script>.sh`, the documented way in.
Red first: 6 of the new tests fail against the previous runner (measured with
an always-allow shim so the module still loads), including the end-to-end one
where the canary directory was deleted and files were written outside the
tree. The fixtures move from `true`/`false` to two allowlisted shell fixtures,
because `false` is no longer a runner - the fail case must still be a real
non-zero exit, not an unrun criterion.
Suite 1148 (1146/0/2).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
/trekreview now runs the commands a BRIEF declares, and a brief is an artifact
that can arrive from outside the repo. Measured 2026-09-18 on 6cafb4c: the
executor denylist stopped a download piped into a shell, but the remote-writing
git subcommand and a recursive delete of a path both RAN. The denylist screens
catastrophe (root deletion, fork bombs, mkfs); it was never meant to screen an
artifact under review.
A second screen, in the runner and ahead of the denylist, refuses four classes:
- a remote-writing git subcommand. The subcommand is found by walking git's own
options (`-C`, `-c`, `--git-dir`, ... take a value), so `git status` and
`git log` still run and `git -C sub push` does not.
- a recursive delete: any `rm` carrying `-r`/`-rf`/`--recursive`. A plain
`rm build/artifact.txt` still runs.
- a download piped straight into a shell (also caught by the denylist; pinned
here so the runner does not depend on another file for it).
- a write outside the working tree. `/dev/null`-class devices are fine, and so
is anything under the working tree; `~/...`, an absolute path elsewhere, and
a target carrying an unexpanded `$VAR` are refused - the runner cannot know
where a variable points, and guessing is how a screen stops screening.
A refusal is its own outcome, REFUSED_BY_POLICY: the command never reaches a
shell, and `summary.ok` is false in both plan and brief mode. For the reviewer,
REFUSED is like NOT RUN - the absence of a measurement, never on its own a
finding - and the rubric and the evidence block both say so.
Chosen deliberately, and it is stricter than today's habit: writing scratch to
/tmp is refused too. The repo's own example plan does `> /tmp/out`. Verification
output belongs in the working tree; exempting the whole system temp dir would
have made the rule unstatable, since a working tree created under /tmp then
contains its own escape hatch.
NOT covered, stated rather than implied:
- other writing git subcommands (tag, remote, config, gc) - only push is listed
- writes through a wrapper: `sh -c '...'`, `xargs`, `find -exec`, a Makefile
target, a script the criterion invokes. The screen reads the command it is
given, not what that command goes on to do.
- `>` inside a quoted string reads as a redirect, so a criterion echoing a
literal `>` is refused. Fail-closed, on purpose.
- the whole surface still runs with the invoking process's permissions; this is
a refusal list, not a sandbox.
The denylist-layer test now uses a stand-in command with a screen double: the
refusal list catches a recursive delete first, so naming one there would have
stopped exercising the denylist layer at all.
Red first: the 6 new tests failed before this change (`refuseCommand` did not
exist), and the fixture brief's four writes ran.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"The first backtick span is the command" is right for a plan, whose template
puts the command first, and wrong for a brief, whose criterion usually opens
by NAMING the thing under discussion. Measured 2026-09-18 on the repo's own
example brief: 5 of 6 criteria FAILED, 3 of them parse artifacts - `--verbose`
run as a command gave exit 2 ("invalid option"), `tests/` gave exit 126 ("is a
directory"). The rubric reads a FAILED result as decisive, so each one became
a BROKEN_SUCCESS_CRITERION BLOCKER about prose.
looksLikeCommand() screens the span by SHAPE only - no filesystem lookup, so a
span parses the same everywhere. Refused: a leading flag, a directory, a token
carrying quotes/braces/prose, and a lone relative path with a slash (an
explicit ./, ../, / or ~/ still runs, as do env-var prefixes). A refused span
is `unrunnable` with reason `not-a-command` - its own outcome, never FAILED,
and it never reaches a shell.
It deliberately does NOT scan on to a later span. "The first span that LOOKS
like a command" invents commands out of prose: in that same example brief it
would have run `whoami` and `login`, two real binaries a sentence happens to
name. An absent measurement is honest; a guessed one is not.
The shape check applies to prose spans only. Inside a shell-tagged fence the
author has already declared shell, so `[ -f x ] || exit 1` still runs.
The rubric follows: a NOT RUN result is never on its own a finding. The
Partial row now describes half-built DELIVERED CODE, and the reviewer gets a
table of the three reason strings - no-command, placeholder, not-a-command -
with what each says about the sentence rather than about the code.
Not covered, stated for the record: a multi-token span whose first token is a
non-executable file (`tests/golden/login.stdout --check`) still runs, and a
criterion whose command is real but whose binary is absent still reports the
shell's exit 127 - that is a true measurement of a missing binary, not a
parse artifact.
Red first: 4 runner tests + 1 doc-consistency pin failed before this change.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The rubric required `brief-conformance-reviewer` to classify a Success
Criterion as Full only when "its verification command/test exists and passes".
Its tools are `Read`, `Glob`, `Grep`. It cannot run anything, so "passes" was
either guessed from the command's mere existence or quietly downgraded to
"exists" — a BLOCKER-tier rule key resting on an impression.
The reviewer stays read-only — a reviewer that executes the code it reviews is
not an independent reviewer. The command does the running instead:
- `/trekreview` Phase 4.5 runs the brief's `## Success Criteria` commands
through `lib/verification/criteria-runner.mjs --brief --evidence` and captures
the block as `sc_evidence_block`, pasted verbatim into the reviewer prompt in
Phase 5. The exit code does not stop the review — a failing criterion is
exactly what the review exists to find.
- `formatCriteriaEvidence` builds that block in code: one row per criterion with
the command, the exit code and the first output line. Chose a code-built block
over an orchestrator-written summary so the orchestrator cannot narrate a pass
that never happened.
- The rubric now judges the supplied result: `PASS` supports Full, `FAILED` /
`BLOCKED` is `Broken` with the exit code cited, and `NOT RUN` is the absence
of a measurement — never evidence in either direction.
- Phase 4.5 is skipped in `quick` mode: that mode does not launch the
conformance reviewer, so there is nobody to hand the result to.
Red first: seven tests in `tests/lib/criteria-runner.test.mjs` against a
committed brief fixture whose three criteria pass, fail, and are prose-only.
The two doc pins were verified red against the pre-fix files (rubric asked
"exists and passes"; no Phase 4.5; the block reached nobody).
Suite: 1117 (1115/0/2), up 9.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Flip model: sonnet → model: opus across 20 agent files, 4 prose references
in commands (trekplan, trekresearch), trekendsession command frontmatter,
and CLAUDE.md tables. Aligns CLAUDE.md premium-profile row to actual
premium.yaml content (all-opus, which has been the case since v4.1.0 but
the doc was drift). Companion to VOYAGE_PROFILE=premium env-var (set in
~/.zshenv same day) — env-var governs orchestrator phase model; this
commit governs sub-agent models which are frontmatter-pinned and not
reachable by the profile resolver.
npm test: 516 pass, 0 fail, 2 skipped (unchanged from baseline).
Operator rationale: complete Opus coverage across all Voyage activity,
including the 20 sub-agents that the profile system does not control
(architecture-mapper, task-finder, plan-critic, scope-guardian,
brief-reviewer, code-correctness-reviewer, brief-conformance-reviewer,
review-coordinator, session-decomposer, plus the 6 researcher agents,
plus the 5 codebase-analysis agents).
Cost implication: sub-agent runs ~5x more expensive vs sonnet. Accepted.
