Repository navigation
## v1.4.0 (2026-11-18)
v1.4.0 (2026-11-18)
The issue-and-propagation wave: five new seats drafted by
charter-architect from the roster-expansion campaign (a 28-source
prior-art report, a full sweep of the workspace's workflow corpus, and
usage mining), riding one release with a release-auditor method patch.
Roster thirty-five: twenty-four read-only, ten write seats, one
multimodal judge.
- issue-census (GLM-5.3-Flash): the retrospective census of OPEN
GitHub issues, one repo or the whole owned fleet, which nothing ships
(first-party Copilot automations label new issues one at a time; the
fleet view is unbuilt). Aging profile, open-vs-open duplicate
clusters, fix-already-landed cross-checks against git log, tags, and
patchnotes, label hygiene, needs-info zombies, then label-only triage
verdicts taken from Anthropic's own triage-issue command design: a
closed vocabulary fetched live from the repo's own labels, duplicates
marked against open issues only, and the false-positive aversion rule
("false positives are worse than missing labels"). It ends with
rule-based SemVer hints (never model-judged: rules already do it
exactly) and a proposed release batch. Absolute fences: never
comments, never closes, never writes a label, never posts anything;
on-demand dispatch only, never a resident watcher. The 2026-11-11
issues-triage decline carried its own revisit condition ("if the
workflow ever moves to issue-driven"); it has fired (the v1.3.0
debug-report duty files issues as a standing flow, CalibreQuarry's
release log is issue-driven end to end, opends gates a release on two
open issues), and the ledger entry records the supersession. Two gh
quirks carried in the charter:--json commentsembeds every full
comment body (verified on opends #13), so comment fields are only
read through a--jqprojection, and the RESTopen_issues_count
counter includes pull requests whilegh issue listdoes not. - fleet-sync-auditor (GLM-5.3-Flash): the four-tier propagation
audit the sync-reference-docs skill has hand-walked after every
release: the repo's own docs (VERSION carriers, patchnotes placement,
roadmap ticks, shields.io badge URLs where the version hides inside
the URL, AGENTS.md status headers, Flatpak runtime pins), the
workspace catalog row, the public surfaces (the profile README's
shipping tags, the site codex status spans), and the Obsidian vault's
identity stubs, where versions are forbidden by design. Propose-only
by charter: the catalog-reconciler decline stands, the public
surfaces are push-gated, and the vault is Brandon's prose. The tree
is the truth (a reference document is never evidence about another
reference document); tier 3 is curated, so absence from the profile
README is not drift while a stale listed version is. Trigger lines
drawn against both neighbors: doc-drift-auditor keeps one repo's docs
versus code, workspace-sentinel the git-state sweep. - upstream-triage (GLM-5.3: its port/skip verdicts gate porting
work): the weekly NetNewsWire window behind Viaduct, generalized to
any dispatch-named reference clone. Reads the last recorded sync
point from the parent repo's dated sync notes, enumerates the
upstream window, clusters it into arcs (windows reach +352 commits),
and returns port / skip / roadmap-candidate verdicts with reasons,
conflict flags against the parent's local ports, an ordered port
worklist, and a draft sync note in the parent's own format. Its one
chartered mutation, pinned in validate.py, is the roster's narrowest
fetch exception: fetch only the dispatch-named reference clone,
because triage needs commit content andgit ls-remotereturns only
the tip (verified live: the clone's origin/main last saw f3a0102c3 on
2026-11-08 while the tip had already moved, and the window's size was
unanswerable without the objects). The sentinel's never-fetch rule
stands everywhere else; the fast-forward, the sync note, and the
ports stay with the main thread. A clone not declared reference
material in the parent's AGENTS.md or the catalog is a REFUSED. - release-executor (GLM-5.3, the write axis's first
post-saturation seat, chartered on Brandon's explicit 2026-11-18
appetite): the mechanical half of cutting a release, taken out of the
main thread after a failure history measured in the hundreds (91
broken tags across 20 repos, 184 missing GitHub Releases backfilled
in one day, whisker-street's empty-notes-file release, deadbeef-cui's
four release-workflow fix cycles). Propose mode (the default) drafts
the exact command plan; apply mode, enumerating repo, version, and
release commit and carrying release-auditor's ready verdict, verifies
CI green on the release commit, extracts the patchnotes entry
verbatim into a scratch file, proofreads it (version, date, heading
style, no em-dashes, no neighboring tails, never the whole file, body
non-empty), cuts the annotated tag with --cleanup=verbatim, verifies
the message with cat-file against the previous tag's detail, pushes
only that tag, creates the GitHub Release from the same file with
--verify-tag, reads back every mutation, and closes with the
for-each-ref subject sweep. Its writes are the roster's second
instantly-public class (after imprint's metadata): three mutations
plus the tag push, never force-push, never re-tag, never a branch,
issue, or file. release-auditor's description now draws its side of
the line: it gates and sweeps; the executor runs. - pixel-judge (GLM-5.3-Flash, the roster's first multimodal seat
and first vision-dependent charter): non-document visuals (GTK/Qt
windows, Godot game screens, the bar, Hyprland, terminals, theming
changes) judged from handed-in PNG captures plus state text against
the dispatch's acceptance bar, one JSON verdict line per capture with
evidence-backed issues and recapture guidance; self-capture is
chartered only as region-scoped grim to /tmp (verified live on this
Hyprland session; Catagotchi's own gotcha table documents grim dead
under KWin, where the repo's capture script is the dispatcher's
path), and UI driving via ydotool is forbidden absolutely. Document
deliverables stay with the four plugin visual-judge seats. - release-auditor truth patch (v1.4.0): the pre-release commit
cross-check now runs both ways. Step 4 caught commits the entry fails
to mention; it now also catches any claim in the entry that no commit
since the previous tag backs (an invented or aspirational entry is a
finding), the notes-versus-diff truth audit prior art found shipped
nowhere, folded in as a method extension rather than a seat
(modes-before-agents). The patch grew one more leg during this very
release's gate: the auditor's Bash allowlist never had gh, so it
could not verify CI itself and reported it dispatcher-asserted; the
allowlist now carriesgh run list/gh run viewas read-only CI
queries, and a new pre-release step makes green CI its own check
(red, pending, or absent isnot ready; unverifiable is reported,
never assumed).
validate.py grows the roster to thirty-five and pins the wave's
dangerous lines: issue-census's false-positive aversion, closed
vocabulary, and never-post fences; fleet-sync-auditor's propose-only
and vault no-versions fences; upstream-triage's named-clone fetch fence
and fast-forward rule; release-executor's clean-tree,
three-mutations-allowlist, verbatim-tag, and force-push/re-tag fences;
pixel-judge's multimodal-description, UI-driving, and code-edit
requirements; and release-auditor's claims-without-commits truth check.