voyage

open/voyage

Fork 0

13 releases 21 tags

RSS feed

v6.5.0 1b613421eb

v6.5.0 Stable
ktg released this 2026-10-18 18:46:04 +00:00 | 66 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
Wave 3 adds three skills, each one source for a discipline that was copied
or missing before: voyage:interviewing for the brief's question round,
voyage:writing-criteria for criteria that can run and go red, and
voyage:reviewing-adversarially for every reviewer's stance. The test repo's
trial 5 ran all three end to end. The product now shows English where it
showed Norwegian (decision 56), and three findings from trial 5 are fixed
(decision 59), each with a test that failed on the code before it. Suite
2659 (2657 pass, 0 fail, 2 skipped), measured by npm test on main at
849b242, the tree this release review covered.

Nothing here asks anything of users. No brief, plan or review format changes.

Doors
- two-way: skill reviewing-adversarially and its skills: lines (order 7d) — revert f906ec9; the review contract (JSON, verdict) is unchanged.
- two-way: skill writing-criteria and its pointers (order 7c) — revert 837e925; the criteria runner and the validators are unchanged.
- two-way: skill interviewing and the /trekbrief Phase 3 pointer (order 7e) — revert f6e18a1 restores the inline Phase 3 text; no brief format change (Handover 1 untouched).
- two-way: English UI labels (decision 56) — the stored annotation keys are unchanged, so a revert brings back the Norwegian labels and every saved annotation loads either way.
- two-way: the brief validator warns on a criterion whose first backticked span is not a command while a later one is (trial-5 finding 1) — a warning, never an error; a revert stops the warning.
- two-way: a re-planned step's earlier new files are parked before the fresh build (trial-5 finding 4) — a revert leaves them in the work tree again; files already parked stay in the git directory, never deleted.
- two-way: one reviewer's two findings on one file+line+rule keep a row each (trial-5 finding 7) — a revert merges them again; earlier review.md files are not re-read.
Added
- Skill voyage:reviewing-adversarially (order 7d). One stance for
  every reviewer: no praise, a source for every finding (file:line or a
  quote), no downgrade without a reason, a Declined to Judge list, claims
  checked against the diff and not the report. Preloaded by the three
  /trekreview agents and scope-guardian, and named by /trekreview at
  spawn; it replaces the copied "Interaction Awareness" block. Each
  reviewer's Declined to Judge lines now reach review.md.
- Skill voyage:writing-criteria (order 7c). One source for how a
  success criterion or a Verify:, Regression: or End-to-end: line is
  written so the criteria runner can run it and it can go red; anything more
  is bash tests/<name>.sh. Preloaded by brief-reviewer and plan-critic,
  named by /trekplan. Baseline: 24 of 27 lines refused by the runner
  without the skill, 0 of 27 with it.
- Skill voyage:interviewing (order 7e). /trekbrief Phase 3 asks for
  the interview form by name instead of carrying it: at most four questions
  per call, one decision each, the recommended answer first, facts looked up
  and never asked, the rest written down, and a hollow answer or unasked
  scope never written in as the person's decision. Measured with helpers:
  round-1 size 10–11 → 4, questions without a recommendation 15/31 → 0/12,
  a hollow answer kept as a decision 2/3 → 0/3.
- Eval cases for the three skills (evals/skill-interviewing-*,
  evals/skill-writing-criteria-*, evals/skill-reviewing-adversarially-*:
  three that should fire, two that should stay quiet, one effect case each).
  Written, not run.
Fixed
- English where the product showed Norwegian (decision 56). The
  annotation page's intent buttons, sidebar badges and copied prompt read
  Fix / Change / Question; the stored keys stay fiks / endre /
  spørsmål, so annotations saved earlier still load. /trekreview's
  project-memory approval offers "Add the line" / "Not now". README,
  CLAUDE.md, docs/architecture.md, package.json and plugin.json
  describe what the product shows. The Norwegian hollow-answer list in
  approval-log.mjs and /trekbrief stays: it matches what a user types.
- The brief validator sees a criterion the runner cannot run (trial-5
  finding 1). When a success criterion's first backticked span is not a
  command but a later span is, the validator now warns
  BRIEF_SC_NOT_RUNNABLE, as the criteria runner already reported it
  unrunnable. A criterion that only names a flag or a path stays prose and
  unflagged (Handover 1 says so).
- A re-planned step no longer builds on its earlier attempt's files
  (trial-5 finding 4). On failure: revert leaves a step's new files on
  disk; --resume after a new plan now runs
  lib/util/park-attempt-files.mjs before the fresh build helper. It moves
  every still-untracked file the attempt created into the git directory
  (voyage/parked/step-N-attempt-k), never deletes one, never overwrites a
  parked file, and exits 1 while a file the attempt made is still in the
  tree.
- One reviewer's two findings on one triplet keep a row each (trial-5
  finding 7). Pass 1 merged two different /code-review findings on the same
  file, line and rule key, and the second lost its row and its action. A
  reviewer is never merged with itself: its second finding under another
  title stays its own row, with the id over rule_key#2, counted over the
  final list — also in --workflow, engine-findings and fidelity-diff. The
  decision-23 put-back stays one copy per triplet.
Known issues

The findings from the test repo's trial 5 that are not fixed in 6.5.0
(decision 59), grouped by consequence.

Too lax — can let through what should be stopped
- Core code outside the deep-review patterns is never deep-reviewed
  (trial-5 finding 9). /trekreview triage makes every file outside the
  deep-review patterns (for example bin/ and tests/) summary-only, so
  code-correctness-reviewer gives no verdict on a CLI's core code;
  brief-conformance-reviewer still reads the whole diff. Planned for the
  fix round before v7 (decisions 57 and 58).
Too strict — can stop what should go on
- A controller finding on a locked test line the plan gave has no way
  forward in the fix ladder (finding 2). Fix rounds may not change a locked
  test (the one exception is a test not red at base), so rounds 1–3 repeat
  the same diff before the step goes up to the plan.
- The stop guard holds the session while it waits for the fix helper
  (finding 3). Sub-step F sets the test lock to pending before a fix round;
  the green exception (operator decision 2026-10-08) covers only the wait
  for the controller.
- Phase 7.5 counts a re-approval of the plan as a stray commit
  (finding 5). After the upward path the plan is approved again, but 7.5
  lists only approvals from the review loop as process commits; read
  strictly, the run becomes partial and the merge into main stays closed.
- Plan dedup misses one finding reported by both plan reviewers when they
  cite it differently (finding 6): an absolute and a relative path, another
  rule key, or overlapping line ranges, so one major is counted twice.
- done-line.mjs --check --commit with a short sha answers "no DONE line"
  (finding 10), exit 1, though the line holds the full sha; Sub-step G only
  says --commit {hash}.
Inaccurate — wrong or unclear, nothing stopped or let through
- Line numbers in review findings can be off (finding 8). An engine
  finding cited the line after the code; FINDING_LINE_OUT_OF_RANGE catches
  only a line past the end of the file, and a patch line number written in a
  reviewer's prose, or :135 with no file name, is not caught.
- With --plugin-dir, statistics go to the installed plugin's data
  directory (finding 11): stats-append and event-emit write to
  ~/.claude/plugins/data/voyage-…, so a development copy's statistics mix
  with the installed version's.
- /trekplan Phase 9 points the plan reviewers at their schema in
  agents/plan-critic.md and agents/scope-guardian.md (finding 12), a
  path in Voyage outside the repo, in the same prompt that tells them to read
  only inside it.
- /trekbrief's bounded budget is unclear (finding 13): "8, the three
  fixed ones included" does not say whether the Phase 3.5 tier question
  counts, and Phase 3.5 puts the heuristic's proposal first as "(default)"
  even when the operator's own rule is to go up in doubt.
- /security-review diffs against the remote's default branch, not the
  review's range (finding 14), so it can review more files than the review.
  In this release's review the opposite happened: the remote's main was
  ahead of v6.4.0, so the built-in diff missed the decision-56 commits, and
  its standard prompt was run over the whole range instead.
The release review of 6.5.0 (/trekreview over v6.4.0..main, WARN: 0
BLOCKER, 13 MAJOR, 1 MINOR) found the following. None is a BLOCKER, so none
opens a fix round before this release (decision 52). "Confirmed" means the
release session reproduced it; "read" means it was traced in the code and not
run.

Too lax — can let through what should be stopped
- The park step can leave the earlier attempt's files in the tree and
  still exit 0 (the trial-5 finding-4 fix, lib/util/park-attempt-files.mjs).
  Confirmed: a file still registered with git add --intent-to-add (the
  revert was refused by the guard or declined) is read as "tracked since the
  attempt" and left in place; and an upward diff written under
  color.ui=always has no header the tool can read, so nothing is parked.
  Read: from a subdirectory the diff's paths are resolved against the wrong
  directory, and the tool's two git calls never check their exit status.
- On the opt-in --workflow path, the severity floor and the put-back read
  one row per file, line and rule key (scripts/trekreview-armB.workflow.mjs),
  while the trial-5 finding-7 fix can now give two. Read: a finding can then be
  floored against the wrong row, or take an engine row's one legal lowering,
  and an engine row the coordinator did not return is not put back when its
  sibling was; a WARN can become ALLOW. The default path is not affected.
- Which row a finding joins can depend on the order the reviewers'
  payloads are read (lib/review/coordinator-contract.mjs). Read: the same
  findings in another order can give other rows, severities and verdict.
- The coordinator dropped four engine findings as refuted because their
  line numbers lay past the end of the file. The engine had cited positions
  in the diff; only the citation was checked, never the claim, and two of the
  four were confirmed true (the intent-to-add hole above and the brief
  warning below). Same class as trial-5 finding 8.
- No test holds a reviewer's Declined to Judge lines reaching
  review.md. The copy-through in /trekreview and the coordinator is
  prose only; deleting it leaves the suite green.
Too strict — can stop what should go on
- The brief validator's new BRIEF_SC_NOT_RUNNABLE warning fires on
  prose (the trial-5 finding-1 fix). Confirmed: any bare word in a later
  backticked span counts as a command, so "--verbose prints the login
  field" is told to put login first. A warning, never an error.
Inaccurate — wrong or unclear, nothing stopped or let through
- A finding's #k id depends on arrival order (lib/parsers/finding-id.mjs),
  so the same finding can get another id in a second review; and a kept row
  can take the plain id that a set-aside row on the same triplet already
  carries. Read, not run.
- /trekplan tells the session to load voyage:writing-criteria by name
  but does not list the Skill tool in allowed-tools, as /trekbrief
  now does. Whether Claude Code treats that list as a limit is not verified.
- The bake-off's arm A still merges each triplet to one row
  (scripts/bakeoff-armA-merge.mjs, not changed in this release), so a
  fidelity measurement now shows an artificial gap; the grouping rule lives in
  three copies.
- A new eval scaffold has a non-ASCII character
  (evals/skill-writing-criteria-effect/scaffold.sh:42, an em dash).
Downloads
- Source code (ZIP)
  0 downloads
- Source code (TAR.GZ)
  1 download

v6.4.0 a3fa5de42a

v6.4.0 Stable
ktg released this 2026-10-16 22:25:08 +00:00 | 85 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
A new plan version, 1.9, in which every path in a step's Files: and on its
Test first: File: line stands first, in backticks, so the reader no longer
guesses paths from prose (decision 16). The same delivery closes the six
findings from the test repo's trial 4 and stops npm test from pointing a live
session's stop guard at a temp-repo lock. Every fix began with a test that
failed on the code before it. Suite 2517 (2515 pass, 0 fail, 2 skipped),
measured by npm test on the third remediation's Step 2 commit
(829da48), the commit before this section.

The converging critic loop came after that measurement (decisions 43 and
44). From round 2 the plan gate counts a critic round by rule A, and both
reviewers state rule B. Suite 2602 (2600 pass, 0 fail, 2 skipped), measured
by npm test on main at 4e28dc5, the tree this release review covered.

Nothing here asks anything of users. Older plans warn, they do not fail: a plan
below 1.9 is read as in 6.3.1, apart from the fixes under Fixed, and gets one
warning per step whose Files: names a path that is not backticked.

Doors
- two-way: plan_version 1.9 — a 1.9 plan read by 6.3.1 gets PLAN_VERSION_MISMATCH as a warning and the 6.3.1 reader, so a revert rewrites no plan.
- two-way: Files reader (three 6.3.0 answers restored) — a revert brings back 6.3.1's answers; no file format changes.
- two-way: the test files on the Test first File line are the backticked paths right after File: (operator decision 20) — a revert reads every path-shaped span again.
- two-way: red proof checks --tests against the plan — a revert accepts any list again.
- two-way: brief-conflict mark on review findings — an older verdict gate ignores the field.
- two-way: review Pass 1 ranks by catalogue tier and a dropped duplicate never lowers its group — a revert restores the asserted-severity ranking.
- two-way: the workflow review puts back every engine finding the coordinator did not return (operator decision 23) — opt-in path only; a revert restores the old floor.
- two-way: approval commits require a Co-Authored-By trailer — only Voyage's own commands call the CLI, and they pass it.
- two-way: explorer prompts, stamp order, test env — prose and test-side only.
- two-way: a 1.9 wrapped line keeps its paths only after a comma (operator decision 29), and a File: line with nothing after it takes its first wrapped line — a revert joins every wrapped line again.
- two-way: below 1.9 a test file after a note in brackets, and, plus or + is read when it is clearly a test (operator decision 28 and operator decision 31) — a revert reads only the run right after File: again.
- two-way: red proof reads the step's test files from the plan, prints them, and refuses a plan it cannot read (operator decision 30) — the older prose that passes --tests still runs against it.
- two-way: original_severity names what the kept finding asserted — only tests read the field.
- two-way: the workflow review fails closed when its coordinator returns nothing, and names every reviewer that did not run in allow_blocked_by — opt-in path only.
- two-way: below 1.8, when the plan names no test file, the helper and the executor give red proof the files the helper wrote (operator decision 41) — a revert leaves such a step unproven and unlocked.
- two-way: rule A in the plan gate (decision 43) — an older gate ignores the new verdict fields, and a revert counts every round fully.
- two-way: rule B and the security key in the reviewers (decision 43) — prose only; a revert restores the old rating.
- two-way: plan dedup passes the reviewers' marks and rule-B fields through, and its two fixes — a revert drops the fields and brings back the line-0 collapse and the crash.
- two-way: the post-compact hook test's timeout (decision 44) — test-side only.
- two-way: the critic-loop matcher needs overlapping texts, and a new finding never takes an old id (operator decision 48) — a revert matches by id or line alone again.
- two-way: plan dedup keeps a lowered_from that is not a severity (operator decision 48) — a revert counts it as no lowering again.
- two-way: plan dedup ranks a merged pair by what rule B counts (operator decision 48) — a revert brings back the order-dependent merge.
Added
- plan_version 1.9: the file first on every line. Each entry of Files:
  and each Test first: File: line opens with its backticked path or paths;
  everything from the first other text on is description and is never read.
  A 1.9 plan that breaks this gets PLAN_STEP_PATH_NOT_FIRST or
  PLAN_STEP_PATH_NOT_A_PATH (errors under --strict, warnings without). The
  parser, the validator, the tests-only check, the plan-file sweep (K6) and red
  proof read a 1.9 plan by this grammar. 1.9 is the current target: Handover
  4, the plan template, /trekplan, the planning orchestrator and CLAUDE.md
  say so, and the rules introduced at 1.8 say "1.8 or later".
- A warning for older plans. Below 1.9, a path in Files: that is not
  backticked gets the warning PLAN_STEP_PATH_UNTICKED, under --strict too.
- The brief-conflict mark. A /code-review finding at the MAJOR tier whose
  recommended action goes against a brief Non-Goal or another stated choice
  drops to MINOR, marked contradicts_brief and brief_lowered_from: MAJOR;
  any other engine finding only gets the mark, and a BLOCKER is never lowered.
  /trekreview passes the brief's choices to the coordinator, which writes the
  mark into review.md; the opt-in workflow review applies the same rule and
  floors an illegal lowering. On an ALLOW or WARN review the verdict gate
  answers exit 2 for a finding lowered or marked outside that one shape:
  brief_lowered_from on any other finding, a mark on a reviewer finding, a
  mark or an engine finding below its catalogue tier, or a mark with a missing
  or unknown rule key.
- The critic loop converges by rule A, counted by the plan gate
  (decision 43). Round 1 is unchanged. From round 2 on,
  plan-gate.mjs --diff prints what changed since the text the last round
  reviewed. Both reviewers get it and mark each previous finding by id. Then
  --record counts the round: a new major that cites a plan line outside
  that diff counts as minor. A blocker, a security finding and a finding
  with no plan line count at full severity. A previous finding stays counted
  until it is marked resolved. The gate keeps each recorded round's text in
  the git directory, so a snapshot is never committed. A plan outside any git
  repository keeps none, and its next round counts fully. Every new verdict
  field is optional, and absent means round-1 behaviour, so an older verdict
  block gives the same exit codes. Phase 9, Handover 4, the planning
  orchestrator, CLAUDE.md and the README say so.
- Rule B in both reviewers (decision 43). plan-critic and scope-guardian
  state it: a major that a named test, Regression: command or gate in the
  plan will fail on during execution is minor, never a blocker and never a
  security finding. The gate records each such lowering from round 1 on.
  Plan dedup passes the reviewers' marks and rule-B fields through. The
  operator answered three questions on 2026-10-15: rule B applies from round
  1 and rule A from round 2; both reviewers tag a security finding with the
  review catalogue's security key; a lowered finding that is re-reported as
  major stays minor, and only a re-report as blocker counts as blocker.
Fixed
- The suite no longer hands the session id to the lock CLI. npm test
  preloads a helper that removes CLAUDE_CODE_SESSION_ID for the whole run,
  and every test that spreads the process environment into a child uses a
  shared helper without it. A live session's stop guard is no longer pointed
  at a temp-repo lock. A test file run on its own with node --test still
  passes the id to a child it starts without an env (release review).
- The three Files: regressions from 6.3.1's Known issues give 6.3.0's
  answer again: a source file after a described test is kept, a path that a
  description only mentions is not read as a file, and a backticked bare name
  such as LICENSE alone on a following line is read as a file.
- The test files on the Test first: File: line are the backticked
  paths that stand right after File:, separated only by commas or spaces,
  and below plan_version 1.9 also each further path after a note in brackets
  ((new), *(new)*, (existing …)) or after and, plus or +, as long
  as it is clearly a test (operator decision 28, 13.10, and
  operator decision 31, 14.10; review bd2520b0). Every one of them is read,
  locked and copied for red proof, not only the first. A path in the
  description after them is no test file, at every plan_version
  (operator decision 20, 12.10; review 6c31212c), so a step whose File:
  line names its source in the description is no longer read as tests-only.
- A 1.9 Test first: names its test only on a line that opens with
  File:, the way the executor, K6 and red proof read it. Below 1.9 the
  check reads as before.
- Engine provenance in the review. Ingest takes engine from the payload,
  never from a field a reviewer wrote on its own finding, and dedup ranks by
  catalogue tier, so a reviewer-agent finding is kept over an engine finding,
  and the kept finding carries the group's highest severity, in the
  coordinator's Pass 1 and in the workflow dedup. A brief-conflict mark can no
  longer lower a reviewer's MAJOR (operator decision 2026-10-12). On the
  workflow path, floor (b) and the put-back of every engine finding the
  coordinator did not return (both below) hold the same line after the
  coordinator answers.
- Explorers and reviewers stay in the repo. /trekplan tells its
  brief-reviewer, the eight Phase 5 explorers, the Phase 6 deep-dives, the
  plan-critic and the scope-guardian to read local files only inside the repo
  under exploration plus the paths their prompt gives, and every explorer gets
  the brief.
- The brief review stamp covers only text the reviewer saw. For a small
  brief, the reviewer's gaps are written after the stamp, so the stamp reads
  stale and /trekplan reviews the brief once more.
- Approval commits carry the caller's Co-Authored-By trailer.
  artifact-commit.mjs takes --trailer and refuses (exit 1, before anything
  is staged or committed) a missing, multi-line or malformed trailer, an unfilled {model}, and
  one that names a session link. /trekbrief, /trekplan and /trekexecute
  pass it; /trekexecute prints the reason and goes on when the approval is
  logged but not committed.
- A list joiner no longer ends the read of a plain path below 1.9. In
  Files:, a +, &, and or or between backticked paths is no word of
  description, so a plain path after them is read as a file again, as 6.3.1
  read it, and a step of two tests joined by + and a plain source is not
  tests-only (review 182ceb93).
- A 1.9 entry keeps the paths of a line wrapped after a comma, and a
  numbered File: line names its test. A line that is no list item keeps
  its paths only after a comma, when the entry so far ends with one
  (operator decision 29, 14.10; reviews d21ff81e, eaeb398e), so a
  Files: list wrapped after a comma names every file: the step is not
  tests-only, and K6 plans the second file too. Otherwise that line, and
  every later wrapped line of the entry, is description and never read. A
  File: line with nothing after it takes its first wrapped line. The
  Test first: File: line may open with -, *, +, 1. or 1), and
  any list item ends it (reviews b4fa1927, 4b6460e5).
- A 1.9 plan refuses placeholders and mixed-case bare names as files
  (operator decision 22, 13.10; review 33f37f53). A leading backticked span
  that is a placeholder (none, TBD, TODO, TBA, N/A, NA, ..., in
  any case) or a bare name that is neither in capitals nor a build file, such
  as the mixed-case Parser, is PLAN_STEP_PATH_NOT_A_PATH, and its hint
  says to write ./Name for such a file. LICENSE, README and Makefile
  stay files; below 1.9 nothing changes.
- The engine of a payload is read from an own property. Ingest no longer
  gives the findings of a payload whose reviewer is named toString,
  constructor, __proto__ or another name Object.prototype carries an
  engine (it gave them a function or an object), and toReviewerPayload
  refuses an engine named toString with its "engine must be one of" error
  (review 3799ea6c).
- Pass 1 ranks by catalogue tier, and a duplicate never lowers its group.
  Of findings with the same (file, line, rule_key), the coordinator's Pass 1
  and dedupByTriplet keep the one with the higher catalogue tier, then a
  reviewer-agent finding over an engine finding, then the higher asserted
  severity, then the brief-conformance-reviewer finding, then the first
  seen, and write the kept finding at the highest severity its group asserted.
  Before, a marked code-review MAJOR beat a reviewer's MINOR on the same
  triplet and the review came out ALLOW; now it is WARN (review bc7f4e23;
  the coordinator's prose, operator decision 24).
- The workflow review puts back every engine finding the coordinator did not
  return (operator decision 23, 13.10; reviews ab1b5d9f, bc7f4e23). On
  the opt-in workflow path, dedup ranks as Pass 1 does, and every engine
  finding with no returned finding at its exact (file, line, rule_key) is
  put back at its own severity, one per triplet (the higher-severity copy when
  two share it), so dropping, moving or relabelling an engine finding no
  longer lowers the verdict. Beside it, floor (b): a returned finding is never
  below the lower of its reviewer counterpart's severity and its catalogue
  tier (for a key outside the catalogue, the counterpart's severity). The
  workflow path is now stricter than the default path (see Known issues).
- An approval trailer with a blank or padded name is refused.
  artifact-commit.mjs refuses a trailer whose name is empty or only spaces,
  or that holds two spaces in a row, so an empty {model} ("Claude" and two
  spaces) fails as the module header promises. A name with a parenthesis,
  such as "Claude Opus 5.5 (1M context)", is still accepted (review
  8420915b).
- The plan validator parses each step once. validatePlanContent reads a
  step's fields once and every check, --steps-from included, reads them from
  there; its errors and warnings are unchanged (review 93382464).
- Red proof reads the step's test files from the plan and prints them, and
  the executor defines them once (review cf241526;
  operator decision 30, 14.10; reviews 793fc1aa, 09ae46d8, d01df508,
  11ee78c5). With --plan and --step, red-proof.mjs reads the step's
  test files from the plan as decision 20 and, below 1.9, decision 31 set
  them, and prints the list it ran as tests (a tests: line in text mode).
  The helper and the executor pass no --tests, and the lock's --files is
  that printed list. A --tests that names another set is exit 2 with a
  reason that starts tests-mismatch and the plan's list in plan_tests; a
  plan or step it cannot read is exit 2 with a reason that starts
  plan-unreadable, and the executor corrects its own call; a step whose
  plan names no test file reads as before. commands/trekexecute.md defines
  the step's test files in one sentence the other places point to.
- The /trekbrief pass branch names no ## Brief Quality, and the stats
  rows span their blocks (operator decision 21, 12.10). That section belongs
  to the exhausted and force-stopped branches only. In
  tests/fixtures/jsonl-schemas.md the line_ref of the trekbrief, trekplan
  and trekreview stats rows runs from the phase heading to the closing fence of
  its stats-append block, and a test pins all four command rows (trekexecute's
  too) by exact equality.
- original_severity names what the kept finding asserted (review
  08deae68). When Pass 1 raises a finding to its duplicate group's
  severity, Pass 3 writes original_severity from the finding's own
  severity, and only when that differs from its catalogue tier; the carrier
  field pass1_raised_from never leaves the contract.
- The workflow review fails closed when its coordinator returns nothing
  (review edf2ec56). A coordinator that returns no verdict and findings
  counts as a missing reviewer: the merged findings are read at their
  catalogue tier or higher (never lowered), the verdict starts at WARN (a
  BLOCK stays BLOCK), and every reviewer, engine and coordinator that did not
  deliver is named in allow_blocked_by as missing-reviewer:<name>, so the
  verdict gate exits 2 on such a WARN.
- K6 keeps its own normaliser and LIST_ITEM is one literal again
  (operator decision 27, 13.10; review 7961a700). The two /simplify edits
  that crossed the first remediation's fence around K6 are reverted.
- The control-run test leaves headroom over a slow start (review
  703d87e1). Under full-suite load the start-point run of red proof's
  control-run test was killed at its 2 s timeout. The test now gives red proof
  a 15 s timeout and starts every run but the control slowly on purpose, so a
  timeout without headroom fails on demand; the numbers live in
  tests/helpers/timing.mjs.
- Below 1.8 a plan that names no test file gets the files the helper
  wrote (review 7b3b59ed; operator decision 41, 14.10). A File: line
  below 1.8 may name its test without backticks, so red proof reads no test
  file from the plan and answers no test files given. The helper then runs
  the same red proof again with --tests naming the files it wrote, and the
  executor gives red proof that same list, so such a step is proven and locked.
- The post-compact hook test survives a slow start (decision 44). Its
  hook child had a fixed 5 s timeout and was killed under full-suite load.
  The timeout now comes from tests/helpers/timing.mjs, and a killed hook
  child reads as a failure instead of exit 0.
- Plan dedup keeps distinct file-level findings apart, and no longer
  crashes on a bad field (decision 43). Findings of one rule key on line 0
  or with no line no longer collapse into one. An empty file or a line
  that is not a whole number no longer crashes the finding id; the line
  becomes null and the finding stands.
- The critic loop no longer folds a new finding into an old one by its id
  or line alone (reviews 4ab24b07, d03835c4; operator decision 48,
  16.10).
  - What happened: an equal id or the old finding's line was enough. The new
    text was lost, and a new file-level major could count as an old rule-B
    minor.
  - Now the texts must overlap, a finding with a plan line is never matched
    by id, and a new finding never takes an old finding's id.
- A lowered_from that is not a severity now counts as blocker (reviews
  8782602b, ebe164e2; operator decision 48, 16.10). It was read as no
  lowering. A blank one is still no lowering.
- Plan dedup no longer lets reviewer order lower a blocker (review
  69c9cebd; operator decision 48, 16.10).
  - A merge keeps the copy rule B counts highest, on the merged keys.
  - This also covers a blocker claim merged into a plain major, found during
    this round.
Known issues
- The session split reads a 1.9 plan as an older plan (operator answer 3).
- Marked MINOR findings still feed recurrence / project-memory proposals.
- The default review path's verdict gate does not compare review.md with the
  engine payloads, so only the opt-in workflow path puts an engine finding the
  coordinator did not return back (operator decision 23). The operator made
  that check a separate delivery later (operator decision 25, 13.10).
- Below 1.9 the plan-file sweep (K6) keeps its own wider reader of the File:
  line (operator 2026-10-11): it plans every backticked path on that line, also
  one in the description that decision 20 no longer counts as a test file.
- Red proof accepts a --tests that matches the plan's list and prints it in
  the spelling it was given, not the plan's (review 2d624b54; deferred by
  operator decision 42, 14.10).
- ingest keeps a pass1_raised_from and an original_severity that a
  payload wrote on its own finding (reviews c06d5385, e031a480; deferred
  by operator decision 42, 14.10).
- When its coordinator returns nothing, the workflow review raises the merged
  findings in place; this has no effect on its output today (review
  6ff82e75; deferred by operator decision 42, 14.10).
- A whitespace-only line ends a 1.9 Test first: File: entry but not a
  Files: entry, to be skipped in both by operator decision 40 (review
  4b6460e5; deferred by operator decision 42, 14.10). The same id stands
  in a Fixed bullet above, which fixed the list-item case of that review; the
  whitespace-only case came back open.
- A 1.9 wrapped line read as description is dropped without a warning, even
  when it opens with a backticked path, to be warned on by
  operator decision 39 (review 4e1607cb; deferred by
  operator decision 42, 14.10).
- Handover 4's Files: bullet words its empty-entry rule as for a File:
  line (review 4f22b101; deferred by operator decision 42, 14.10).
- Accepted deviation: the 1.9 File: selector repeats the list-marker set of
  LIST_ITEM (review 4dce7c74; operator decision 38, 14.10, decision 27
  stands).
- The line diff's LCS table has no size bound. A rewritten gap of tens of
  thousands of lines takes seconds to minutes and gigabytes, and can end
  outside exit 0/1/2. Real plans measure about 1,200 lines (review
  8a9b5a17; deferred by operator decision 48, 16.10).
- A previous finding re-reported on its own edited line is counted twice.
  Such a re-report of a lowered finding counts at its full severity (review
  8004ea85; deferred by operator decision 48, 16.10).
- A line sent as a numeric string ("42") is read as no line and counted
  at full severity. The agent templates give line as a number (review
  c257625e; deferred by operator decision 48, 16.10).
- npm test does not run tests/e2e-critic-diff-mode.sh. It runs under
  ## Verification and in /trekreview's success-criteria check (review
  818b1346; deferred by operator decision 48, 16.10).
- The matcher's residual (review 4ab24b07; operator decision 50, 16.10): a
  previous finding marked resolved by every raiser and re-reported in other
  words (content overlap below 0.7) on an unchanged line is a new finding,
  which rule A can lower to minor. The same holds for a re-report in the same
  short words (fewer than 6 content words) away from the previous finding's
  place.
- A short new finding at a previous finding's place is matched to it when
  all its content words appear in the old text: the overlap is measured
  against the shorter text, and the place clause has no word floor. The new
  text is lost, and the old finding's severity is counted, a rule-B minor
  included (review ea8c1752; deferred by operator decision 51, 16.10).
- In a merge of three or more copies, a copy's rule-B claim that ties on the
  keys merged so far is dropped. When the security key arrives with a later
  copy, reviewer order still decides between minor and major (reviews
  436ade33, f9ba85e1; deferred by operator decision 51, 16.10).
- A lowered_from such as "none", "n/a", false or 0 counts as
  blocker, by design (fail-closed). A reviewer that writes one on a finding
  it did not lower turns a minor into a BLOCK (review 79d1dc97; deferred by
  operator decision 51, 16.10).
- The rule that a blank lowered_from is no lowering lives only in plan
  dedup. A --record input that bypasses dedup with lowered_from: "" counts
  as blocker (review 2e7afd61; deferred by operator decision 51, 16.10).
- When the plan has not changed since the previous round, the gate counts
  the round fully, and a previous finding that no reviewer reports again
  leaves the count without being marked resolved. This predates the critic
  loop's diff mode (review 4c340d4a; deferred by operator decision 51,
  16.10).
- A near-duplicate merge in plan dedup keeps only the survivor's place, so
  rule A can lower a new major in the changed text to minor when its partner
  cites an unchanged line, and the round can record PASS. 6.3.1 counted that
  merge major (review ef1633b6; deferred by operator decision 52, 16.10).
- Below 1.9, regression A's restored reading also takes a plain slash path
  that a description only mentions after a comma (`tests/a.test.mjs` — covers the parser, lib/parse.mjs stays as is). 6.3.1 read only the test
  file; now the step is no longer tests-only, and the validator warns
  PLAN_STEP_PATH_UNTICKED on the mentioned path (review 38a8ad34;
  deferred by operator decision 52, 16.10).
- The verdict gate's brief-lowering check compares severities in upper case
  only, so an engine finding written "severity": "major" reads as an
  illegal lowering and the gate exits 2 (fail-closed) (review 20b626e2;
  deferred by operator decision 52, 16.10).
- The opt-in workflow review's floor raises an illegally lowered or marked
  reviewer finding but keeps its brief_lowered_from and contradicts_brief,
  so the verdict gate still exits 2 on that review (fail-closed) (review
  c614baad; deferred by operator decision 52, 16.10).
Downloads
- Source code (ZIP)
  1 download
- Source code (TAR.GZ)
  1 download

v6.3.1 a97efb6989

v6.3.1 Stable
ktg released this 2026-10-11 16:05:57 +00:00 | 151 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
Rounds 7 to 10 of the release review of 6.3.0. Round 7 is the WARN 6.3.0 was
released with, round 8 was BLOCK (1 BLOCKER), rounds 9 and 10 were WARN. Every
fix began with a test that failed on the code before it. Suite 2236 (2234
pass, 0 fail, 2 skipped).

Nothing here asks anything of users. Measured against 6.3.0 over the
marketplace and the test repo: the plan validator reports the same errors and
warnings for all 233 plan files, the verdict gate gives the same exit for all
123 review files, and tests-only reads the same for all 1408 plan steps.

Doors
- two-way: tests-only is decided by the plan's Files: alone, and red proof no longer reads the step's diff (operator decision 2026-10-11) — a revert brings the diff check back; plans stay valid.
- two-way: the Files: reader keeps a plain file beside a backticked path, reads numbered entries, and keeps a description together across commas and lines — it feeds only the PLAN_STEP_TESTS_ONLY warning and red proof's tests_only; a revert restores the 6.3.0 reading.
- two-way: a trailing # comment is dropped from a backticked check too, so such a check now runs; one whose comment holds a double quote, $, a backtick or a backslash is refused out loud — 6.3.0 refused every backticked check with a comment, so nothing that ran stops running.
- two-way: the verdict gate reads the last fenced block once and answers exit 2 unless its info string is exactly json, or when it is never closed; a ~~~ fence is a fence — a review written by /trekreview reads as before; a revert reopens the two holes under Fixed.
- two-way: stops and timeouts, the release check's --review, isMain in every CLI, the test lock's command comparison, one fence reader, and cleanups — fixes; a revert restores the old behaviour.
Fixed
- Two verdict-gate holes, present in 6.3.0. A review that ended with a
  clean json block and then a block holding a BLOCK verdict, written as
  ~~~json or opened with more text after json on its line, passed with
  exit 0: the gate did not see the second block. Both are exit 2 now.
- Tests-only steps. The plan's Files: alone decides, and red proof,
  /trekexecute, step-implementer and the handover contract say so. A change
  already in the tree when the step began no longer counts against it, and a
  plan outside .claude/ reads the same.
- The Files: reader keeps a plain file beside a backticked path
  (", Dockerfile", "and Makefile"), reads a numbered entry (1. or 1)) as an
  entry, keeps a description together past a comma and onto the next line, and
  starts a description at a plain file followed by a colon. Lowercase
  makefile, rakefile and justfile are build files.
- Checks with a comment. A trailing # comment is dropped in a backticked
  check and in a command passed as written, so the words screened are the
  words run. A comment with a double quote, $, a backtick or a backslash is
  refused instead, because the executor hands the Verify: span to red proof,
  the test lock and the done line inside double quotes. Every reader refuses it
  out loud (PLAN_VERIFICATION_NOT_RUNNABLE, BRIEF_SC_NOT_RUNNABLE). A span
  that is only a comment is not-a-command.
- Stops. Red proof reports a command killed by its timeout as killed after
  the timeout, not "could not start". The release check tells a timeout from
  output past its buffer and from another signal. One reading of a stop
  (stoppedBy) serves red proof, the criteria runner, the test lock, the done
  line and the release check.
- Release check. A relative --review is read in --repo, or from the
  current directory when only that one exists. Two different files are exit 2,
  never guessed, and one file reached through a symlinked --repo is one file.
- Every CLI uses one isMain helper. Before, some did nothing and exited 0
  when run through a symlink or a path with a space.
- Test lock. It compares --cmd by its words and without its trailing
  comment.
- One fence reader for the gate and every reader of a trailing json block.
  A fence closes only on a run at least as long as its opener, as in Markdown,
  and CRLF and an indented fence read alike everywhere. A fence indented four
  spaces or more still counts, on purpose: Markdown's rule made a BLOCKER block
  inside a list item invisible.
- Cleanups. The 15-minute timeout is defined once, the findings check reads
  each cited file once per validation, and the guard writes its no-one-to-ask
  refusal once.
Known issues

Round 11 of the release review (0f400d3..10ac9c5) is WARN: 0 BLOCKER, 10
MAJOR, 0 MINOR. All ten come from Claude Code's /code-review, which gives no
category, so each enters as MAJOR. Voyage's two reviewers and
/security-review found none. The operator chose to release 6.3.1 with it
(2026-10-11). None of the ten lets a gate pass. The regressions below were
checked by running them against 6.3.0. None of them occurs in the 1408 plan
steps measured above.
- The Files: reader still guesses paths from prose and misreads in both
  directions. Two misreads are regressions from 6.3.0:
  - A source file after a described test is dropped
    ("tests/a.test.mjs - new test, src/a.mjs - the fix"). The step that
    builds code then reads as tests-only and goes to the plan.
  - A path that a description only mentions is read as a file ("covers the
    parser in src/parse.mjs"). A step of only tests is then not caught and
    goes through fix rounds (at most five, then the operator).
  Red proof has no diff check behind the reader. The structural fix is to
  require every path in Files: to be backticked. It changes the plan
  contract and is left for a planned release.
- A json block closed on the same line as its JSON is no longer read (a
  regression from 6.3.0). The gate answers exit 2, and a reviewer is asked
  again. An opener with a blank before json is not json either, as in 6.3.0.
- A trailing comment with a double quote, $, a backtick or a backslash is
  refused in every backticked check. That includes checks that never reach a
  shell: a ## Verification bullet, Regression:, End-to-end: and a brief
  criterion. 6.3.0 refused these too.
- Cleanups: a dead branch for a span that is only a comment, the comment
  dropped up to four times per command, a re-export of the timeout that
  nothing imports, and a test helper that copies a PATH lookup.
- An engine run that was stopped and not counted in the verdict reported three
  more, each checked by running:
  - A backticked bare name such as LICENSE, alone on a line after a
    description, is no longer read as a file. 6.3.0 read it.
  - One unclosed fence earlier in a review hides the review's own json block
    from the review validator. 6.3.0 found the block, and the gate answers
    exit 2 there.
  - An indented json example after a review's own block is read as the last
    block, as in 6.3.0.
- The example plan's steps 1-2 keep their tests-only warning (operator
  decision 2026-10-09).
Downloads
- Source code (ZIP)
  0 downloads
- Source code (TAR.GZ)
  2 downloads

v6.3.0 0f400d3461

v6.3.0 Stable
ktg released this 2026-10-11 06:33:07 +00:00 | 157 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
Wave 2 of the practice build (five orders, merged 2026-10-09) and the fix round
after it, which closes finding 1 of the test repo's third trial: a step of only
tests can never be proven red. Every fix began with a test that failed on the
old code. Suite 2194 (2192 pass, 0 fail, 2 skipped).

Two changes ask something of users (see Doors): /trekexecute now refuses a
plan that has no logged approval for its current content, and a check in a
brief or plan runs with no shell, as plain words with a runner's read-only
options. Plan 1.7/1.8, brief
2.0–2.2 and review 1.0/1.1 files still validate; the new step fields are
optional.

Doors
- one-way: /trekexecute refuses a plan whose logged approval does not cover its current sha256 (K2) — a plan written before 6.2.0, written by hand, or edited after its approval stops at Phase 1; run /trekplan --project <dir> and answer "Approve" (from 6.3.0).
- two-way: /trekexecute reviews by itself in the closing run (/simplify → /trekreview → verdict gate, K7) and keeps the main merge closed on BLOCK or an undecided verdict; step check (K9), Regression: (K10), architecture rules (K5) and plan-file sweep (K6) run at step end and after the last step — a revert restores the old executor; nothing to do for users.
- two-way: the approval guard refuses a merge into main, a push, a tag or publish and destructive git in the permission mode that shows no prompts — a revert lets them through again; nothing to migrate.
- two-way: progress-validator rejects the runner's exitCode in plan_verification.failed_criteria (C15) — the executor's own prose always wrote exit_code.
- two-way: skills workshop — form check, model-free routing test, test-first and debugging preloaded into step-implementer — new files and one skills: line; a revert removes them.
- two-way: researching-sources skill, preloaded by the five web-search agents and named by /trekresearch — the agents' copied rule text moved into the skill; a revert puts it back.
- two-way: Regression: (K10) and On failure: action words (K11) — an optional field and a warning; plans without them validate as before.
- two-way: trial 2's minor fixes (B8, B11, B12, C13, C14, C16, C17) — checks, prose and agent text; a revert restores the old behaviour.
- two-way: release check and runbook — new files; a revert removes them, and nothing else reads them.
- two-way: tests-only steps (PLAN_STEP_TESTS_ONLY, red-proof --plan/--step → tests_only, which also reads the step's own diff, executor routes them to the plan) — a warning and an optional flag; a revert removes them, plans stay valid.
- two-way: timing tests under load (hook helper kill limit 30 s, SC #14 fastest of three against 800 ms) — test-only.
- two-way: example plan On failure: lines open with a K11 word, and its verbose-content test moves into the step that wires the output (6 steps, progress.json follows) — text in an example.
- two-way: /trekreview saves the reviewer's whole answer for the findings check and names FINDING_PROSE_LINE_OUT_OF_RANGE — command text.
- two-way: the criteria runner's commands run without NODE_TEST_CONTEXT — a false green under node --test is now red; nothing to do for users.
- two-way: the fixes from the release review of 6.3.0 (rounds 1-6) — checks and prose; progress.json may now say simplify: "revert-refused" (an optional value); a revert restores the old behaviour.
- one-way: a check in a brief or plan (criteria, Regression:, proof of red, the done line, the test lock) runs with no shell: its words reach the runner as written, so only plain words run (no quote, backslash, $, brace, or ~/@ at a word's start; a glob only for node --test on Node 21 and later; no path that climbs out with .. or starts at /, git included; a trailing # comment in a shell fence is dropped) with a short list of read-only options per runner — git without --output or --no-index, node --test with a built-in --test-reporter only, nothing after make test, only paths after npm test --; an option that writes a file or loads code is refused — a criterion that used one (git log --grep="x", pytest -k foo, pytest tests/test_*.py, jest --coverage, go test -timeout 30s) is now NOT RUN at review and not runnable on a 1.8 plan; wrap it in bash tests/<script>.sh (from 6.3.0).
- two-way: the phase-model resolver text says effort comes only from the brief's phase signal — prose only, no code change.
Added
- /trekexecute is coupled to the gates (order 6).
  - Plan approval (K2): approval-log.mjs verify --kind plan before step 1;
    not-logged, mismatch, expired, broken and diverged halt with what
    to do. Exit 2 is never a pass.
  - Step end: the step check (K9; exit 2 is "not measured") and the step's
    Regression: command (K10, new lib/verification/regression-check.mjs
    through the criteria runner) run green before the checkpoint.
  - After the last step: architecture rules (K5) and the plan-file sweep
    (K6); a violation or a file no step names turns completed into
    partial.
  - The closing run reviews by itself (Phase 7.2): run the thing,
    /simplify, /trekreview, verdict gate. A BLOCK adds the review's
    remediation steps after the plan validator accepts them and the operator
    re-approves the plan, at most TREKEXECUTE_MAX_REVIEW_ROUNDS (3) times.
  - The main-merge gate is a lock on the verdict, and where the guard asks,
    the executor does not ask again (one question per merge).
  - progress.json gains optional coupling fields (Handover 5).
- Skills (orders 7a and 7b), the first three: voyage:test-first and
  voyage:debugging, preloaded by step-implementer, and
  voyage:researching-sources, preloaded by the five web-search agents and
  named by /trekresearch at spawn. lib/skills/skill-lint.mjs checks the
  form; lib/skills/routing.mjs is a model-free routing test over the
  skills, the 7 commands and the 21 spawnable agents (collision at 0.75 fails;
  thresholds pinned). Eval cases evals/skill-* are written, not run.
  Authoring method: docs/skills-authoring.md.
- Release and rollback (order 8). docs/release-runbook.md (before, the
  release, both rollback paths drilled in a sandbox) and
  scripts/release-check.mjs: clean tree, npm test at HEAD, review verdict
  not BLOCK and covering HEAD, a ### Doors list. 0 ready, 1 not ready, 2
  could not run.
- Plan step rules (order 9). Regression: — one allowlisted command,
  unpiped (PLAN_STEP_REGRESSION_NOT_RUNNABLE); On failure: opens with
  revert | retry | skip | escalate, or PLAN_STEP_ON_FAILURE_UNKNOWN warns.
  Minors open after the last critic round are listed for the operator, and
  --extra-round records the round they pick.
- Tests-only steps (order 10). PLAN_STEP_TESTS_ONLY warns on a step
  whose files are all tests; red-proof --plan/--step reports tests_only,
  and /trekexecute sends such a step to the plan instead of into fix rounds
  that cannot succeed. /trekplan, the plan template and plan-critic say
  the test belongs in the step that builds the behaviour.
Fixed
- The approval guard refused only a merge into main and a push in the
  permission mode that shows no prompts; a tag, a publish and destructive git
  went through unattended. Now every call it would ask about is refused there.
- Plan-review dedup merges two findings on the same file:line with the same
  content; the Jaccard threshold for every other pair is unchanged.
- /trekbrief Step 4c says the brief reviewer runs in the background and is
  waited for.
- The Phase 4 size count includes .mjs, .cjs, .mts, .cts.
- criteria-runner --evidence shows the first output line on PASS rows.
- The findings schema checks path:line citations in finding prose against
  the file; /trekreview checks the reviewer's whole answer and names
  FINDING_PROSE_LINE_OUT_OF_RANGE.
- git-historian never proposes changing a file the brief froze, and names
  the remote from git remote -v.
- The criteria runner's commands run without NODE_TEST_CONTEXT, so a test
  run inside node --test can no longer pass falsely.
- Timing tests tolerate load (fastest of three; hook helper kill limit 30 s),
  and the K2 hook tests use unique session ids per run.
- The example plan's On failure: lines use the K11 words, and its
  verbose-content test moves into the step that wires the output. Steps 1-2
  (golden capture, stability harness) keep their tests-only warning; a step
  type for them is a later question, and step 2's heading no longer says it
  must fail.
- The resolver text in four commands says effort comes only from the
  brief's phase signal.
- Release review of 6.3.0 (seven rounds; rounds 1-3 BLOCK, rounds 4-7 WARN;
  every finding of rounds 1-6 fixed with a test that failed first, round 7's
  under Known issues): the guard refuses where
  no one can be asked also when its rules will not load; review steps appended
  by /trekexecute are held to the step rules on any plan version and need a
  Verify: (plan-validator --steps-from); a restore the guard refuses is
  recorded as revert-refused and stops the run before review, never as
  reverted; every check Voyage runs runs with no shell (round 5, the
  operator's structural fix after the patching of the shell screen did not
  converge), as plain words and each runner's read-only options; the release
  check compares version files without their version, lets the README change
  only in the badge's version, reads scope_sha_end as a commit sha, sees a
  file moved after the review, ignores Voyage's own artifacts committed after
  the review, runs the cheap checks first, stops a suite that never ends,
  works for a plugin in a subdirectory and never reports a crash as "not
  ready"; the plan-file sweep reads Files: as v6.2.0 did, and the
  tests-only check has its own reader, and red proof also reads the step's
  diff (see Known issues); skill frontmatter and routing read YAML values, comments and tags
  as YAML does; regression-check prints the progress.json status and runs
  through the criteria runner itself. /trekexecute says plainly that review
  steps in the template's own form stop with a re-plan.
Known issues — for 6.3.1

The seventh round of the release review is WARN: 0 BLOCKER, 12 MAJOR, 0
MINOR. The operator chose to release 6.3.0 with it (2026-10-11); the findings
go into 6.3.1.
- Tests-only steps: when red proof decides, the step's diff in the documented
  flow holds only the test, so it cannot overrule the plan's Files: reading;
  a change already in the working tree when the step started counts against
  tests-only; an unbackticked file next to a backticked path in Files: is
  not read.
- Red proof reports a command killed by its timeout as "could not start". The
  release check calls a suite killed for too much output a timeout, and reads
  --review from the current directory instead of --repo.
- A trailing # comment is dropped in a fenced check but not in a backticked
  one.
- The example plan's steps 1-2 keep their tests-only warning (operator
  decision 2026-10-09; a step type for them is a later question).
- Four cleanups: the 15-minute timeout defined in three places, a line count
  read once per citation, isMain copied into five more CLIs, and a
  roundabout condition in the guard.
- Found while writing the review, and present in 6.2.0 too: the findings check
  ends a reviewer's json block at the first triple backtick, even one inside a
  JSON string.
Downloads
- Source code (ZIP)
  0 downloads
- Source code (TAR.GZ)
  3 downloads

v6.2.0 4ee711a6ac

v6.2.0 Stable
ktg released this 2026-10-08 18:53:31 +00:00 | 188 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
Wave 1 of the practice build (five parallel orders, merged 2026-10-06), the
fixes from two end-to-end trials in a separate test repo, and the defects an
independent reader found in those fixes. Every fix began with a test that
failed on the old code. Suite 1972 (1970 pass, 0 fail, 2 skipped).

Nothing that was valid becomes invalid: plan 1.7, brief 2.0–2.2 and review 1.0
files still validate. The new fields are optional, and the new rules apply only
from plan 1.8 and review 1.1.

Added
- Intent (/trekbrief). A size class
  (ceremony: small|bounded|architectural; absent = architectural) sets the
  question budget, and the brief asks one round of questions with the
  recommendation first. "høres bra ut", "gjør som du vil" and an empty answer
  are not approval. lib/util/approval-log.mjs keeps an append-only,
  hash-chained approval log in the target repo, bound to the approved
  content's sha256. lib/util/artifact-commit.mjs commits it.
  /trekplan Phase 1 halts on a broken, missing or mismatched entry.
- Plan 1.8 and the plan gate (/trekplan).
  - Plan 1.8 requires Test first: on every step and one End-to-end: step.
    New optional fields: Interface in:, Interface out:, Review focus:.
  - An architectural plan gets a ## Design part.
  - The critic loop has at most 3 rounds, and the plan records them.
    lib/validators/plan-gate.mjs decides: 0 go, 1 stop, 2 could not run.
    An open MAJOR stops the plan, and after the last round the operator
    decides.
  - plan-critic now checks 12 dimensions.
  - New lib/verification/arch-rules.mjs runs the target repo's own
    dependency-cruiser or import-linter.
  - New lib/verification/plan-scope-sweep.mjs lists changed files that no
    step names.
- Step helpers and proof by command (/trekexecute).
  - One fresh step-implementer helper per plan step (new agent: opus, effort
    high, no Agent tool).
  - One controller reads the diff and gives two verdicts: does the step meet
    the plan, and does it hold the quality.
  - Fix rounds go up, at most 5: helper, then plan, then operator.
  - New lib/verification/red-proof.mjs: the step's test must fail at the
    start point and pass at the result.
  - New lib/verification/test-integrity.mjs (the floor guard) catches
    removed asserts, skipped tests, stubs and loosened thresholds.
  - New lib/verification/done-line.mjs writes the done line only when the
    test command exits 0.
  - A closed stop list and a decision log appear in the final report.
- Guards (command hooks only, none runs a model).
  - pre-bash-executor answers "ask" before a merge into main, a push, a
    tag or publish, and destructive git. post-approval-log records the
    answer.
  - pre-plan-gate runs the plan gate when a Voyage plan leaves plan mode.
  - The test lock (lib/util/test-lock.mjs, pre-write-executor,
    stop-test-lock) refuses edits to a locked test file, and refuses "done"
    until the step's own test command is seen green.
  - post-edit-check runs the target repo's own formatter or linter. It runs
    only during a Voyage plan run or after opt-in.
  - New lib/verification/step-check.mjs checks each step when it ends.
- The review is a gate (/trekreview, review 1.1).
  - verdict: ALLOW | WARN | BLOCK is required.
    lib/review/verdict-gate.mjs reads it: 0 ALLOW/WARN, 1 BLOCK, 2 could not
    decide (never green). A control that did not run gives exit 2.
  - Claude Code's /code-review and /security-review run as the engine.
    Their findings enter through lib/review/engine-findings.mjs under three
    rule keys. CODE_REVIEW_IMPROVEMENT (MINOR) covers simplification,
    efficiency, reuse, style and readability. The rule catalogue now has 15
    keys.
  - No finding is struck for length. Only MINOR and SUGGESTION text is
    trimmed, and the trims are counted.
  - The coordinator lists what it declined to judge.
  - Every BLOCKER becomes a numbered remediation step.
  - A finding class seen in two reviews is proposed as a project-memory line
    (lib/review/recurrence.mjs). The operator approves each line.
- Checks when a brief or plan is written. A success criterion or Verify
  line that the criteria runner could never run is caught when the brief or
  plan is written, not when the review runs. The brief validator warns
  (BRIEF_SC_NOT_RUNNABLE, BRIEF_SC_SUBPOINTS, BRIEF_SC_BASH_NOT_PINNED).
  For a 1.8 plan, the plan validator reports PLAN_VERIFICATION_NOT_RUNNABLE,
  PLAN_STEP_VERIFY_NOT_RUNNABLE, PLAN_END_TO_END_NOT_RUNNABLE and
  PLAN_END_TO_END_NOT_VERIFIED: as errors under --strict, otherwise as
  warnings.
Fixed
- The write guard covers Edit and NotebookEdit, not only Write, and
  fails closed.
- The review keeps findings that are right:
  - an engine finding is never set aside for Actionability;
  - recommended_action is the engine's own fix, or otherwise a test-first
    step, never the failure scenario;
  - a cited line past the end of a file is caught;
  - finding ids are computed, not typed by a model.
- A MINOR /code-review finding outside the reviewed range no longer holds
  ALLOW back. A MAJOR one still does.
- The plan template's checkpoint stages the step's files
  (git add … && git commit). The pre-compact hook reads that form.
- The brief template's examples are commands the runner accepts.
  /trekcontinue and /trekexecute call lib/ through the plugin root.
- Manifest must_contain uses grep -E -e … --, so a pattern such as
  --sep is not read as an option.
- /trekreview, /trekresearch and /trekcontinue write stats through
  stats-append and report a failed write.
- annotate prints an absolute path.
- A test lock marked green may stop while it waits for its controller.
- /trekplan runs every exploration agent, and reviews the brief again only
  when it has changed since /trekbrief reviewed it.
Downloads
- Source code (ZIP)
  1 download
- Source code (TAR.GZ)
  1 download

v6.1.1 8f707aab0f

v6.1.1 Stable
ktg released this 2026-10-05 16:43:28 +00:00 | 215 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
A patch release. /trekexecute no longer loses its telemetry when the harness
does not export CLAUDE_PLUGIN_DATA, its progress-file prose now matches its
own validator, and no agent is pinned below reasoning effort high.

Fixed
- lib/stats/event-emit.mjs falls back to the same data directory as
  stats-append and the yardstick when CLAUDE_PLUGIN_DATA is unset, instead
  of skipping silently. Every written:false (failed write, usage error,
  payload-not-json) exits 1 with a reason. The CLI also runs when it is
  invoked through a symlink or a path with a space.
- main-merge-approved, main-merge-declined and gates_mode_flag_override
  are known events.
- /trekexecute Phase 8, Phase 9 and the gates_mode_flag_override record all
  write through a writer that resolves the directory. A failed write is printed
  in one line and the run continues.
- Phase 8 passes each event payload on stdin through a quoted heredoc
  (event-emit.mjs --payload -), so the shell expands nothing in it. Before,
  the plan path went into a single-quoted argument: a path with a quote
  character broke the shell parse, and a quoted pair could run a command
  substitution.
- The progress-file schema and status rules in /trekexecute use the
  validator's vocabulary (in_progress, completed, plan_version).
  progress-validator now accepts the run status stopped and the step
  status blocked, both of which the command writes. A legacy passed step
  is still treated as completed on resume.
- architecture-mapper, dependency-tracer, git-historian and task-finder
  pin effort: high (was medium). No agent is pinned below high; the rest
  follow the session.
Downloads
- Source code (ZIP)
  0 downloads
- Source code (TAR.GZ)
  3 downloads

v6.1.0 5680591ddf

v6.1.0 Stable
ktg released this 2026-10-05 10:43:29 +00:00 | 222 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
The use row in scripts/yardstick.mjs now counts an order as used only when
a Voyage record names the order's id. Counting itself stays off
(USE_COUNTING_SINCE = null) until the operator declares the work done; the
tool ships first, the measurement comes after.

Added
- lib/stats/claimed-order.mjs finds the one coord-queue order the session
  has claimed (repo-mailbox's claim contract, read-only). Exactly one claimed
  order is an attribution; zero or several is null plus a reason, never a
  guess.
- /trekbrief stats records carry order_id, order_claimed_at and
  order_reason, added by the writer (stats-append), never typed by the
  model. The fields are on all five stats surfaces and are not exported.
- facts.ordersNamedByARecord in the yardstick output: a known-positive
  diagnostic for the use row.
Changed
- Yardstick use: the slug join is replaced by an order-id join. YES = a
  trekbrief record names the order's id; declared NO = the order's own
  arbeidsform: ordre line; no record = not measured, never a no.
- /trekbrief --approve now runs Phase 7 after a successful stamp, so the
  approval leaves a trekbrief record (mode: "approve").
Fixed
- The CLI main-guard in stats-append and yardstick compares real paths,
  so it no longer silently does nothing under a symlink or a space in the
  path.
Downloads
- Source code (ZIP)
  0 downloads
- Source code (TAR.GZ)
  5 downloads

v6.0.0 7439dac927

v6.0.0 Stable
ktg released this 2026-10-02 17:54:22 +00:00 | 239 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
Operator decision 2026-09-27: nothing Voyage ships starts a Claude Code
session. Work that needs another session becomes an order plus a starter
line that the user runs in a session they open themselves; parallel work
inside one session uses subagents. This is a breaking change for anyone who
relied on Voyage launching sessions.

Changed (breaking)
- /trekexecute multi-session path (Phase 2.6) no longer launches child
  sessions and no longer finishes in one call. It prepares one wave at a
  time: a worktree per session, one session order per session
  (.claude/trekplan-sessions/{slug}/orders/, written by the new
  lib/util/session-orders.mjs) with a /trekexecute --session N <plan>
  starter, and stops. Running the same command again reads each session's
  progress file, merges the wave sequentially and prepares the next one. A
  plan with W waves takes W + 1 parent runs plus one session per order.
  New Hard rule 21: never start Claude Code.
- Sessions now run under the user's own permission mode and tool set: the
  per-session tool allowlist and the permission-bypass mode are gone, and
  brief.md / architecture/overview.md are no longer injected as an
  appended system prompt.
- Phase 7.6 recovery writes a bounded recovery order for a session you open,
  and the run stops as partial; nothing dispatches it automatically.
- /trekplan --decompose (and its alias --export headless): the
  session-decomposer writes orders.md instead of a launch.sh, and the
  "launch" reply is gone.
- /trekresearch --engine deep-research: the pre-gate no longer runs the
  Claude Code binary to read its version; the post-hoc check decides. The
  version-below-floor reason token is gone (that case now reports
  no-cited-report).
- No {session-dir}/logs/session-N.log is written any more; progress files
  are the source of truth, and each wave's progress files are archived to
  orders/wave-W-session-N.progress.json before its worktrees are removed.
- --dry-run no longer prints a billing line.
Removed
- The ANTHROPIC_API_KEY billing check, the hooks-fire version precheck, the
  headless safety preamble and the per-child tool, turn and budget flags —
  they existed only for launched children. The TREKEXECUTE_MAX_TURNS and
  TREKEXECUTE_MAX_BUDGET_USD environment variables are ignored.
- templates/headless-launch-template.md and
  scripts/q3-cache-prefix-experiment.mjs (a finished experiment that
  started the binary; its results stay in docs/).
- README: the billing-safety and headless MCP-tuning sections, and the
  fourth security layer (scoped per-child tool access), which went with the
  launched children.
Fixed
- Every VOYAGE_ROOT resolution in /trekexecute (multi-session Step 0 and
  plan verification), /trekreview and /trekresearch now finds the
  installed plugin: the literal ${CLAUDE_PLUGIN_ROOT} that Claude Code
  substitutes, else the highest installed version in the plugin cache. Before,
  these blocks always exited 2.
- The multi-session path never force-deletes. Worktrees are removed only
  when git agrees they are clean and session branches are deleted only once
  merged; anything git refuses is kept and listed in the report with what to
  do. Before, a failed session, a merge conflict or a refused merge deleted
  every trek/{slug}/* branch of the wave, finished work included.
- Collecting a wave never tears down a session that is still running (an
  unknown status counts as running); it stops when the main checkout has
  uncommitted changes or has switched branch, and tells a refused merge from
  a real conflict. The parent run must start in the main checkout, never in a
  session's worktree or a subdirectory's view of it.
- A merged-waves record, bound to the plan's path and content hash, keeps a
  re-run from preparing an already merged wave again; a different plan or an
  edited plan is refused with the way out named. The record is cleared only
  when the final merge is confirmed.
Added
- tests/lib/no-claude-code-launch.test.mjs: fails when any tracked file
  outside a reasoned, exact-count exemption list carries a way to start
  Claude Code; proven on a known-positive fixture for every pattern. The one
  operator-run exception is the fixed test set (docs/proevesett.md).
- tests/lib/session-orders.test.mjs: runs the order writer on a fixture
  plan and proves the orders land, statuses classify, progress is archived,
  merged waves are remembered, and nothing is started.
- tests/lib/plugin-root-resolution.test.mjs: runs every VOYAGE_ROOT
  block, substituted and unsubstituted.
Downloads
- Source code (ZIP)
  0 downloads
- Source code (TAR.GZ)
  5 downloads

v5.11.1 57e620b028

v5.11.1 Stable
ktg released this 2026-09-23 11:47:52 +00:00 | 257 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
Changed
- The annotation surface's provenance is described as the pattern it follows
  (pencil toggle, intent buttons, popover form, structured export) instead of
  by naming a repository that no longer exists: .claude-plugin/plugin.json
  and package.json descriptions, README.md, docs/architecture.md, the
  v5.0.3 entry below, and two comments in scripts/annotate.mjs and
  tests/scripts/annotate.test.mjs. No behavior change.
Added
- tests/lib/retired-references.test.mjs fails when any tracked text file or
  path contains a term from the gitignored local list
  tests/.excluded-terms.local.txt. The scanner is validated against a planted
  known-positive, an empty scan fails closed, and the named exemption list is
  empty and rot-checked. Without the local list the scan is skipped with an
  explicit message.
Downloads
- Source code (ZIP)
  1 download
- Source code (TAR.GZ)
  4 downloads

v5.11.0 0be152ad29

v5.11.0 Stable
ktg released this 2026-09-23 10:33:44 +00:00 | 260 commits to main since this release

Signed by ktg

SSH key fingerprint: SHA256:JakMjO6FTBBzN0Bhfj9saOoEjaFxlSdYuZQQpM/lF9Q
Added
- Intent approval gate. /trekplan (both --brief and --project, also
  --quick, no override flag) runs lib/validators/intent-approval.mjs --check
  before it reads the brief. It halts on BRIEF_INTENT_NOT_APPROVED,
  BRIEF_INTENT_APPROVAL_STALE or BRIEF_INTENT_APPROVAL_INVALID, and an
  approval check that cannot run halts too. The marker is intent_approved_hash
  (sha256 over the normalized ## Intent + ## Goal) plus intent_approved_at.
  Editing either section after approval makes it stale. /trekbrief Phase 4h
  shows Intent and Goal verbatim and stamps only on "Approve".
  /trekbrief --approve <project-dir | brief-file> does that step alone for a brief
  from any producer. A ## line inside a code block or HTML comment is not a
  heading, and each section must appear exactly once. The marker is a trace that the
  step ran against this text, not a signature: any session on the same machine can
  stamp.
- scripts/yardstick.mjs prints Voyage's three numbers (use, trust,
  interrupts) with the denominator and the source each one was read from. It
  exits non-zero while any number cannot be counted, and today it is RED.
- scripts/end-state-gate.mjs counts the distance to "Voyage is finished"
  (defects, experiments, dormant agents, open decisions). It is fail-closed: an
  entry whose check cannot run counts as open.
- Plugin eval suite (evals/, docs/proevesett.md): eight cases run as
  real claude -p children with only Voyage loaded, deterministic graders,
  and each expectation committed before its first run. Measured stability: every
  case 5 of 5 over three full runs, and the sequencing-gate case 15 of 15.
- lib/stats/stats-append.mjs: /trekbrief and /trekplan stats records
  land in the data dir the yardstick reads when CLAUDE_PLUGIN_DATA is empty,
  and a failed write is reported, not skipped silently. intent_approved
  reaches the OTel field allowlist.
Changed
- Handover 1 (PUBLIC CONTRACT): a brief from any producer now stops at
  /trekplan until its intent is approved. The schema change is additive (two
  optional fields, no brief_version bump). The requirement sits in the
  consumer. See docs/HANDOVER-CONTRACTS.md §Handover 1.
- The arg parser has a CLI. /trekreview runs it with $ARGUMENTS and relays
  Error: --project <dir> is required. from code (it passed "$@", which the
  Bash tool never has, so the parser never ran).
Fixed
- /trekexecute runs the plan's ## Verification on the single-session path
  too (D-03), and /trekreview runs the brief's success-criteria commands and
  gives the reviewer the result (D-04), through lib/verification/criteria-runner.mjs.
- The criteria runner screens with an allowlist of test runners, not a
  denylist, and refuses writes a brief may not perform. A backticked span runs only
  when it is a command, and the runner reads the repo's own plan format.
- The end-state gate can no longer be turned green by editing its own ledger,
  and a behaviour probe must fell a stub, not just name a test.
- The review gold-scorer key separator is written as an escape, not a raw
  NUL byte. The /trekresearch swarm engine points only at the External agents block.
- Docs no longer claim an opus orchestrator: the commands follow the session model.
Removed
- /trekplan's "execute with team" path: Claude Code 2.1.178 removed
  TeamCreate/TeamDelete, so it always fell back to sequential execution.
- TaskCreate/TaskUpdate from /trekplan's tool lists. Claude Code 2.1.233
  stopped offering them on current models, and no prose called them.
Downloads
- Source code (ZIP)
  0 downloads
- Source code (TAR.GZ)
  4 downloads

1 2

v6.5.0 Stable

Doors

Added

Fixed

Known issues

v6.4.0 Stable

Doors

Added

Fixed

Known issues

v6.3.1 Stable

Doors

Fixed

Known issues

v6.3.0 Stable

Doors

Added

Fixed

Known issues — for 6.3.1

v6.2.0 Stable

Added

Fixed

v6.1.1 Stable

Fixed

v6.1.0 Stable

Added

Changed

Fixed

v6.0.0 Stable

Changed (breaking)

Removed

Fixed

Added

v5.11.1 Stable

Changed

Added

v5.11.0 Stable

Added

Changed

Fixed

Removed