Repository navigation
Releases
## v1.6.0 (2026-11-18)
v1.6.0 (2026-11-18)
The comparative-design and marketing wave: one new read-only seat, the
fleet's first in-repo authored skill, four charter patches from the
comparative seat-design report, the evals convention, and a README
rebuilt for readability. Drafted by charter-architect from the
campaign's second research leg (the 37-source repo-marketing report;
the first leg's 69-source pair fed v1.5.0, and a 32-source comparative
seat-design report drove the patches). Roster thirty-eight:
twenty-seven read-only, ten write seats, one multimodal judge.
- repo-marketer (GLM-5.3-Flash): the launch desk, read-only with
web tools (the assay tool class). Given a repository it returns a
placement analysis across the earned channels (Show HN, language
subreddits, Lobsters, curated newsletters, Discord and Matrix,
dev.to cross-posts, awesome lists) judged against each channel's
live rules, the repo's actual maturity, and the operator's stated
standing; an explicit where-not-to-post list; channel-native drafts
built from the repo's own artifacts on the demonstrate-don't-claim
skeleton (the trurl shape: analogy, problem, copy-pasteable
examples, honest-maturity section, roughly 400 words);
similar-projects research on the live web; a paid-advertising
worth-it verdict with the arithmetic shown (the honest baseline is
that most pure OSS gains nothing; three defensible conditions and
four burn patterns carry their numbers); and a measurability plan
(HN's spike shape, the roughly 20:1 HN-to-Reddit volume ratio,
GitHub's 14-day referrer panel, UTM conventions with the lossiness
caveat). The fences are absolute and first: it never posts, never
comments, never votes, never opens a browser session to any
platform, never authenticates, never touches a social account;
drafts land in the final message only and the human posts. The
enforcement reason is documented, not theorized: the February 2026
matplotlib incident, where an autonomous agent retaliated against a
maintainer over a rejected PR with a fabricated-quote hit piece,
makes drafts-only a structural requirement. Anti-hype guardrails are
charter law (the Stenberg standard): every claim traces to a repo
artifact, no invented traction, no star-count flexing, every launch
draft carries an honest-maturity section, every draft carries its
channel's compliance note. Live probing shaped four rungs: HN's
rules pages and lobste.rs fetch clean keyless while Reddit's rules
routes 302 to a login wall under curl (the charter pins the WebFetch
route and the "rules unverified" marking instead of promising
fetches); hn.algolia.com's search API verified keyless and instantly
returned a direct comparable for the dry-run repo; the CalibreQuarry
dry run surfaced the maturity gate's exact finding class, the
README's own TODO admitting the missing TUI screenshot. - skills/herald: the Reddit presence buffer, authored by the main
thread in this same campaign and landing under skills/ beside the
seven vendored Obsidian skills (validate.py grows the expected set
to eight). Drafts posts and replies in the operator's voice, triages
reply threads, and keeps hostile replies at a sanitized distance;
like repo-marketer it drafts only and the human posts. repo-marketer
is its research engine: placement and first drafts from the agent,
ongoing presence from the skill. - Comparative charter patches: code-reviewer's fix-sketch and
linter-folding rungs, latro's pinpoint-citation rule, and the
untrusted-corpus lines on latro and issue-census; four of the
comparative-design report's five STEALs, landed in-tree before this
release (the report: 32 sources, 25 delta rows, 9 CONSIDERs and 11
SKIPs also recorded). The fifth STEAL is the evals convention below. - evals/ convention: evals//prompt.md holds a realistic
dispatch prompt and evals//graders.md asserts the Output
contract's shapes; seeding is incremental and an empty evals/
passes while the convention fills. validate.py carries the light
parse check. - README rebuilt for readability: the single-mega-paragraph intro
and one thirty-eight-row table became grouped seat tables by lane, a
fleet-at-a-glance block, a skills table, and a condensed model
guidance section pointing at the design guide. Same facts, parseable
layout.
validate.py grows the roster to thirty-eight, adds the web-research
reader class with its four pins (never posts, never authenticate,
final message only, compliance note), expects the eighth skill
(herald), and carries the v1.6.0 comparative pins and the evals check
landed earlier in the tree.
## v1.5.0 (2026-11-18)
v1.5.0 (2026-11-18)
The life-domains wave plus the fleet's first vendored skills: two new
read-only seats, one shared charter line on every charter, and skills/
in the repo. Drafted by charter-architect from the campaign's research
pair (a 37-source stealable-seats report and a 32-source
skills-ecosystem report, 69 sources total). Roster thirty-seven:
twenty-six read-only, ten write seats, one multimodal judge.
- vault-archivist (GLM-5.3-Flash): read-only hygiene audit of an
Obsidian vault, the az9713 /vault-health pattern proven on a
10,918-file vault (baseline ~4,979 unresolved links): unresolved
wikilinks and dangling targets, orphan and dead-end notes judged
against the vault's own linking conventions, frontmatter coverage
and key consistency, tag-taxonomy drift against the vault's declared
roots, duplicate and near-duplicate notes, unreferenced attachments,
plus PROPOSED connections between related notes (the buildwithclaude
connection-agent lane folded in as a proposals-only rung,
modes-before-agents: a suggestion list, never an edit). Vault prose
is never rewritten, and the vault's own guidance file outranks the
charter (the librarian law-corpus rule). Live probing on the
reference vault (730 notes by its own walk) pinned three charter
rungs: the audit universe is reconciled across enumerators because
the vault's .gitignore deliberately hides the daily journal and
eight personal notes (fd 626, rg 633, the vault's own script 730;
an audit trusting one enumerator silently skips the journal), the
vault's own graph_audit.py is the alias-aware primary authority
(live: 5 dangling links, 31 orphans under its no-in-and-out
definition, 2 unreferenced assets), and raw rg extraction carries
three traps (multi-file path prefixes without -I, 405 embeds and
136 empty dataview constructs absorbed by a naive pattern, aliases
and anchors needing strips). One live convention finding of the
kind the seat exists for: the core daily-notes config carries a
template path the vault's own law forbids (the Templater race);
reported, never fixed. - subscription-watchdog (GLM-5.3-Flash): the renewal audit over
manually exported bank and card statement CSVs, the offline pattern
community skills already shipped (dvcrn's subscription-audit; the
agentusecases corpus records a $700 recovery from exactly this
audit): merchant-stem clustering, cadence from date gaps,
stable/price-change/missed-billing/trial-to-paid/promo-step flags,
optional budget-line gate. Deterministic awk/sort work with
judgment labeled; the fences are absolute: never a bank API, never
Plaid, never mail (the OpenClaw inbox record is why), never ~/org/
or anything .gpg/.asc. Dormant until data (the ledger-analyst
precedent): no statements named returns a first-class
not-applicable verdict naming the export convention, so the seat
ships before the data exists. Probing on synthetic statements
verified the whole rubric and two parse traps: gawk FPAT keeps its
quotes (a quoted US date silently zeroed every gap until quotes
were stripped first), and a missed month hides from the median (a
31-day series carrying one 61-day gap needs the per-gap check, not
the average). - Language discipline, all thirty-seven charters: answer in the
dispatch's language to the end of the report. The history is the
optimization-guide self-audit's not-applied list: a mid-session
switch to Chinese or another unrequested language is a GLM-lineage
drift that survives context resets, so the line orders the drift
named in the Debug Report and the report continued in the
dispatch's language. Validator-pinned like the v1.3.4 shell line
before it (issue #7's sibling, promoted from the v1.4.1 candidate
slot into this wave), pin written first. - skills/ vendored: the seven official Obsidian skills
(obsidian-cli, obsidian-markdown, obsidian-bases, json-canvas,
obsidian-headless, defuddle, knap) from obsidianmd/obsidian-skills
commit 1c3f6df5, MIT, vendored verbatim with ATTRIBUTION.md
(provenance, commit pins, license chain) and a vet record shaped by
the Snyk ToxicSkills audit: 36.82 percent of 3,984 community skills
carried flaws, 13.4 percent critical, 76 confirmed malicious (91
percent prompt injection), which is why the set is
official-source-only, license-checked, and source-read before
vendoring. The community CLI-audit reference (the 57-scenario
Obsidian CLI gotcha set) is merged per the skills README.
validate.py grows the roster to thirty-seven, pins the
language-discipline rule on every charter, and adds the skills/
structural check: exactly the seven expected skills, every SKILL.md
present with a frontmatter name matching its folder, ATTRIBUTION.md
present. A punctuation-hygiene sweep rode the same release: five
charters (repo-cartographer, doc-drift-auditor, code-reviewer,
spec-compliance-reviewer, claim-verifier) carried em-dashes in their
output templates from before the house prose rule existed; the
templates now use house punctuation, and no verdict vocabulary
changed.
## 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.
## v1.3.4 (2026-11-18)
v1.3.4 (2026-11-18)
One shared charter line across all thirty agents, from the standing
debug-report duty (issue #7): zsh shell discipline. Two independent
dispatches on 2026-11-17 lost the tails of compound Bash commands to
zsh's =word expansion: an unquoted echo === separator makes zsh
resolve == as a command name against PATH, the expansion fails with
== not found, and everything after it in the compound command never
runs. Every charter now carries the rule in its constraints: never
write an unquoted token beginning with an equals sign; quote separator
text (echo '---') or use printf. charter-architect's drafting
requirements name the line so future charters inherit it, and
validate.py pins the rule on every charter (pin written first: thirty
failures against the old charters, thirty passes after), so a future
charter edit cannot silently drop it.
## v1.3.3 (2026-11-17)
v1.3.3 (2026-11-17)
Two latro charter patches from the standing debug-report duty, batched
under the charter-patch policy. The write fence (#5): "Bash never
writes the report" now states the exhaustive rule, any Bash mutation of
the report file in place or redirected, with the in-place editors named
(sd, sed -i, perl -i, awk redirects, sponge, heredocs,
mv/cp onto the file) so the old verb enumeration cannot read as the
complete list; and a self-caught draft defect now has a sanctioned fix
path, one corrective Write pass of the full file before the first
return, never Edit, never Bash, because Method step 6 said "fix what
you find" while the fence left no legal tool to fix with, which is the
gap an sd call slipped through. The degraded line (#6): an absent web
tool is now named in the marker (DEGRADED: WebSearch absent,
DEGRADED: WebFetch absent, or both gone), the toolset is documented
as fixed at dispatch so the agent plans around the absence instead of
re-probing, and the absence is carried into the Debug Report, making
future partial-drop observations precise enough to answer the open
platform question of whether background launches shed WebSearch.
validate.py pins all three lines (pins written first: all three failed
against the old charter, all pass after), and the AGENTS.md Validation
and Non-negotiables sentences and the README Development check
inventory name them.
## v1.3.2 (2026-11-16)
v1.3.2 (2026-11-16)
Two charter patches from dispatch Debug Reports, batched under the
charter-patch policy. ci-triage (issue #3): an empty --log-failed with
exit 0 on a failed run is not "log unavailable"; the method now checks
the jobs API first, where zero steps, no runner, and a BlobNotFound log
endpoint are the signature of a run that never executed (a queue
teardown), and the teardown, not code, is the root cause.
release-auditor (issue #4): the tag-versus-entry cross-check now
set-compares with one identical sort on both sides, an awk set-diff or
comm only over inputs pre-sorted identically, because comm over
version-sorted lists compares lexicographically, silently disagrees,
and buries its warning; its contradictory output is a tool artifact,
not a finding. validate.py pins both lines so a future charter edit
cannot silently drop them.
## v1.3.1 (2026-11-13)
v1.3.1 (2026-11-13)
workspace-sentinel charter patch from a dispatch Debug Report (issue
#2): a latest-tag pick that disagrees with the top patchnotes entry is
re-verified against the entry's version line before reporting, because
a retagged catch-up tag carries a newer tag-object date than a later
release and can win the creatordate sort; and tag findings are
reconciled at sweep end so a tag minted mid-sweep is never reported
missing.
v1.3.0 (2026-11-11)
v1.3.0 (2026-11-11)
Every agent now closes every dispatch with a structured debug report,
and the dispatcher files actionable findings as GitHub issues in this
repo. Brandon's request: a two-week capture window (2026-11-11 to
2026-11-25) whose triage decides charter, validator, and tooling
improvements per agent and per model.
- The duty is always-on and baked into all thirty charters as a
## Debug reportingsection: the final message ends with a
## Debug Reportblock, one bullet per friction point actually hit,
never speculation, in six categories (tool, environment, model,
charter, dispatch, idea), or exactly- noneon a clean run. The
block is meta: it never replaces mission findings and never enters a
deliverable file. - The separation is explicit: agents report, the dispatcher files. No
agent creates, comments on, or labels a GitHub issue in any form, so
imprint's forbidden-mutations fence and the declined issues-triage
ledger entry stand unchanged. - The dispatcher's filing contract lives in this repo's AGENTS.md: one
issue per actionable finding, deduped against opendebug-labeled
issues, titled[<agent>] <summary>, the body carrying the dispatch
date, the category, the evidence, and the model pin the agent ran
under, so the triage can group improvements per agent and per model.
A one-line pointer joined the global dispatch rules so dispatches
made from other workspaces still file. - validate.py enforces the section on every charter (thirty new
checks); the roster stays thirty and no tool set changed.
## v1.2.0 (2026-11-11)
v1.2.0 (2026-11-11)
Thirtieth agent: schedule-analyst, the school-and-schedule lane,
promoted from the watch list by the suite-completion gap survey. The
survey's baseline named weekly school work (td as the task system,
gcalcli the calendar, the Sunday D2L sweep) as the one recurring domain
with no seat; the same survey found the write axis saturated, and its
other verdicts are recorded in the roadmap ledger and watch list.
- td 5.10.0 is the verified backbone (every rung probed live
2026-11-11): today/upcoming/inbox, stats (live but absent from the
top-level help listing), task list's filter surface with the
results/nextCursor JSON pagination, the project/section/label
structure rungs, and completed list. Two quirks carried in the
charter: the JSON priority field is the API scale, inverted against
the UI (4 is p1), and today and upcoming open with the same overdue
block. - gcalcli 4.5.1 is chartered in its verified state:
listanswers from
local cache, but event queries currently fail with
RefreshError: invalid_grant(the stored OAuth session cannot
refresh; verified live). Until Brandon re-auths, the digest runs
td-only and carries a calendar-unavailable line with the re-auth
direction, the ledger-analyst precedent (chartered ready, the dormant
dependency stated plainly). Re-authentication is never the agent's
act. - Read-only by charter: td's and gcalcli's read verbs only,
recommendations never executed, recurrences kept visible in any
suggestion because they are load-bearing. The modes-before-agents
call is recorded: not data-analyst's named-endpoint lane (local
datasets, not CLI state reads) and not syshealth-auditor's (machine
state shares no corpus or allowlist with task and calendar APIs). - Model pin: GLM-5.3-Flash, the rubric-driven evidence-gathering tier
(syshealth-auditor precedent); the dispatcher can raise it if
digests run thin. - Packaging: roster check extends to thirty; README, AGENTS.md (whose
agent list carried a stale twenty-eight count over twenty-nine
names; fixed here), both manifests, and the roadmap synced. The
roadmap records the survey's verdicts: catalog reconciler declined
(no git baseline under the write target for the fence model's
re-verify leg, verified live), dependency apply mode and kindle-sync
deferred with entry conditions, emulation and ML seats declined as
job-title candidates, and ctf-recon and log-miner keep their unfired
conditions.
## v1.1.0 (2026-11-11)
v1.1.0 (2026-11-11)
Twenty-ninth agent: imprint, the GitHub-metadata seat (Brandon's
request). It tends a repository's public identity, the description and
GitHub Topics, through the gh CLI: a new validator class for Bash-verb
metadata mutators whose tools stay [Read, Bash] (no Write, no Edit).
- "GitHub tags" in the request's phrasing is read and stated in the
charter as GitHub Topics, the repository's discovery labels, never git
tags; tags, releases, and the release flow are forbidden surfaces. - The write lane is exactly two mutations,
gh repo edit --description
and--add-topic/--remove-topic(plus--homepageonly when a
dispatch names it; a topics PUT only for wholesale replacement, which
gh repo editcannot express). Forbidden absolutely: visibility,
archive, rename, default branch, collaborators, branch protection,
repository creation or deletion, feature toggles, gists, issues and
PRs, discussions, webhooks, secrets, actions runs; and never commit,
stage, push, or tag (release-auditor and workspace-sentinel own tag
and release health; doc fixes route to doc-drift-auditor). - These are the roster's first writes that are external: they publish to
github.com instantly with no local diff to review. The fence answer
transposes the v0.9.0 default: propose mode is the charter's default,
apply dispatches name their repos explicitly (a fleet apply must
enumerate; "all repos" is a REFUSED), every mutation is followed by a
read-back, and the itemized before/after report is the only audit
trail, its before-state doubling as the rollback record. Fleet propose
sweeps are chartered from the start; the fences are per-repo
identical, and deferring would only re-charter the same lines later. - Grounding is mandated before a word is proposed: README, spec,
patchnotes, roadmap, git log and tag census. A description that
outpaces the repo is metadata drift, the named failure class. - Probed live before chartering (gh 2.97.0): the read side exercised
(gh repo view --jsonfields, the topics endpoint, cwd repo
resolution); the write side verified from the tool's help and
GitHub's own docs, never exercised during design. Topic rules verified
from the docs: lowercase letters, numbers, hyphens; 50 characters or
fewer; at most 20 topics; names saved as lowercase; always public even
from private repos; 422 means stop, not retry. The description field's
platform cap is not published anywhere reachable read-only; the
charter sets house discipline at one line, 200 characters or fewer. - The docs-prose-writer decline stands: this seat writes two bounded
metadata fields grounded in the repo's own documents, not commissioned
prose. Where the judgment is wanted harder, an apply dispatch can
carry Brandon's approved text and the charter writes it verbatim or
refuses. - Model pin per the anything-that-writes rule: GLM-5.3.
- Packaging: validate.py learns the GH_META_AGENTS class (metadata-lane,
two-mutations, never-commit, forbidden-mutations fences) and moves the
roster check to twenty-nine; README, AGENTS.md, both manifests, and
the roadmap synced.