Service accounts created through the identities API are assigned the
developer role in the request's org, whose shipped policy permits read and
write across the org -- the permit alone did not scope the canary. Install a
matching org-canary-grant-forbid policy (access.CanaryGrantForbidPolicy) that
negates exactly the permit's surface for the canary identity; a forbid wins
within the level, so the role can no longer widen access. Engine tests pin
the surface with the principal carrying the developer role, and the
rehearsal now proves it live: the scoped token gets 403 on /api/access/policies
and 200 on /api/chat/update, and the full canary loop runs on it.
task check passes; task canary:rehearse green (REHEARSE_EXIT=0).
The canary now authenticates as a dedicated service-account identity in the
system org whose only grant is access.CalaryGrantPolicy: read the signal
endpoints (chat update, alerts, logs, tasks), read and write the
update-settings resource, and trigger an update install. CANARY_TOKEN no
longer carries instance admin; DEPLOY_TOKEN stays CI-only.
- authorize: POST /api/chat/update/install and /defer map to
update on Object::"update" (dashboard) instead of run on any workflow,
so the grant can cover installs without granting workflow runs
- accounts: active service-account identities may authenticate with bearer
tokens (CanAct), not just users
- new POST /api/identities/{id}/tokens (behind manage_identities) mints a
token for a named non-user identity; new AccountService RPC
BindIdentityPersonalToken backs it remotely, refusing user-kind targets,
with the caller recorded as audit actor
- engine tests pin the grant's exact allow/deny surface
RELEASING.md documents the one-time setup. task check passes on this tree.
- one shared UpdateStateDir() resolver for daemon and dashboard, instead of
two copies that could drift
- the recovery path for a committed/rolled-back journal writes
installed.json too, so a crash after the outcome phase cannot lose the
install record
- JudgeCanaryNow is the explicit signals-only entry point for promoting
without the window, replacing the zero-install-time sentinel
- canary version comparisons use IssuedAt, matching Promote's replay rule,
so semver/issued-at divergence cannot wedge check red; queuedBehind
reuses the judgeable predicate instead of a looser check
- promote prints the channel/pin; promote.yml cancels superseded runs
- canary card: no duplicated elapsed rendering, channel row waits for the
settings resource instead of showing a default, agent fallback '?'
- rehearsal script bounds its docker build/push/run calls
task check passes on this tree.
The canary window now counts from when a release is installed on the host
(installed.json / installed_at in the update snapshot), not from the
statement's issued_at, and 'release canary check' judges the release the
canary actually runs via its exact-pin tag instead of the moving next tag.
A newer release on next queues behind the one being judged instead of
restarting the clock, so stable moves once per release.
New subcommands: canary status (read-only state, signals, verdict), canary
promote (promote now when hard checks pass), canary unpin (back to next),
and start -force. A held/pinned canary is verified installable every check
and a stranded pin fails the run red instead of looking green. Rollback
pre-flight checks stable's zip and agent image digests before pinning.
Relative signals are judged whenever both windows carry minimum data, so a
breach rolls back mid-window; promote still needs the full window unless
the operator promotes early (new promote.yml workflow_dispatch). The window
is configurable via -window / CANARY_WINDOW.
Also: dashboard canary status card (ui), and task canary:rehearse, a local
end-to-end rehearsal of publish -> canary start -> status -> check ->
promote against a throwaway registry and host.
task check passes; rehearsal run green (REHEARSE_EXIT=0).
Record the model's reasoning parts. The step run view shows reasoning as a
Thought row, shell calls as a terminal, the gate as a test run, web_fetch as a
source, finish as the verdict, and each fan-out branch as a subagent.
The loop runner hooks every model call and tool call and reports ordered,
redacted, clipped entries. Each is persisted as an agent_transcript task
event through InsertEvent, left out of the timeline read, and served per
attempt at /api/tasks/{id}/transcript. The step run view shows it with Trace.
With an existing database_url, setup prefilled the URL, so a blank answer
kept it and the bundled socket database was unreachable without typing its
URL. The re-run database step now presents a choice menu: keep the existing
URL (still the default), switch to the bundled database (socket URL plus a
generated PGPASSWORD written to the env file), or enter a different URL.
Fresh installs and -database-url are unchanged; a blank answer anywhere
still keeps the existing URL. Pinned at the setup API level with fake
prompter/sink: TestRunParamsDatabaseChoice.
The bundled internal/channels/telegram package (17 files) is deleted; the
channel now ships as archipelago/channels/telegram, launched by archiemessaging
through the existing channel-extension machinery. An existing [chat.telegram]
config enables it with the same keys: the token stays a secret reference
(token_ref) the host resolves, allowed_user_ids and show_tool_calls map onto
extension settings. Per-org bots compose the same extension per org from the
org's credential binding. Reload support is now derived from a channels.Reloader
interface instead of a *telegram.Gateway assertion. The five host-only seams
(/update, release announcements, running versions, in-place show_tool_calls,
reload) are dropped from the channel and need a contract home or deletion.
Add [access].denial_retention (default 90 days, matching the audit-retention
seed) wired through config defaults and validation; the State Store sweeps
denial records older than the configured age hourly via the new
access.DenialPruner / PruneDenials store query. The dashboard gains
GET /api/access/denials (read_audit-gated, so org admins/owners see their
org's records; the store scopes by the credential's org) and an
/org/denials page backed by a Pinia store.
The deploy gate's govulncheck step failed on the first run that reached it, for
two real findings: crypto/x509 in the standard library, fixed in go1.27.3, and
golang.org/x/net v0.61.0, reached from webfetch.htmlToText calling html.Parse,
fixed in v0.62.0. govulncheck now reports none.
deploy.yml inlined its own sequence, which drifted weaker than `task check`:
no db:check, news:check or ui:check, so a release could ship stale sqlc
output, a stale ui/dist, or release notes that disagree with CHANGELOG.md,
and the dashboard suite ran outside the gate without its type check.
The release now runs `task check TEST_FLAGS=-count=1`: same definition, the
full uncached suite including the Docker-backed tests -short skips. quality.yml
already invoked the Taskfile, so both pipelines now share one gate.
The pinning section said CI republished ghcr.io on every push to main. CI
publishes git.catlow.cloud, and only for a release tag; ghcr is where the
hand-run tools/publish.sh writes, which is the reference the default names.
A release zip pins a digest on a registry an anonymous pull cannot reach, and
[containers].registry_auth -- the setting that fixes it -- was documented
nowhere. Documented there, beside the pin.
Also corrected: every shipped profile does not pin the image, and the section
link omitted its number so it matched no heading.
setup-node's cache save fails on this runner -- the pnpm store path it computes
does not exist at post-step time, which marks the job failed after the gate has
already passed. Run 19 proved it: 0 lint issues, ui/dist matched, no task failure,
then 'Path Validation Error' and exit 1.
Manager.StepTypes and Providers had no reader; the catalogue they were meant
to feed is not missing -- Registry/Settings serve it as x-step-types in the
workflow-definitions schema, read by the dashboard workflow editor and the
gateway's workflow_step_types tool. Remove the write-only providers field
the accessor deletion left behind, and pin that production reader in
controlplanerpc so a dropped schema extension fails the gate.
A failed state-label change was dropped at three points: the forge.Forge
interface returned nothing, forgeext logged the extension's gRPC error
instead of returning it, and the forgerpc client and server discarded the
response error. SetStateLabel now returns an error on the interface and on
NoopForge, forgeext propagates the extension's error, and the forgerpc
server carries it in the response envelope while the client returns it,
matching the other forgerpc handlers.
The proto is unchanged: a gRPC error rides the status and
SetStateLabelResponse is already an empty message.
TestSetStateLabelFailureReachesTheCaller pins the full chain: an extension
whose SetStateLabel RPC fails, through forgeext and the forgerpc
client/server, to the caller receiving the error.
Refs #280
Registry.filter stripped the LLM capability whenever a manifest declared
no tools, skills, memory or conversation capabilities, but
DefinitionEngine.Pass always calls the model. An instruction-only custom
curator therefore registered healthy and then failed every pass with
'no model runner is bound'. The definition engine is inherently agentic,
so the filter now keeps the model runner for it while every other
capability service stays bound only when declared.
Adds tests reproducing the instruction-only pass through the
registry/engine path and pinning that the registrar view for an
instruction-only manifest exposes the model but no tools, skills,
memory or conversations.
Closes#1132