13 releases 21 tags

RSS feed

• v6.5.0 1b613421eb

  Compare

  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