kitrun ran the first Kit of a profile's list as the container image, so
every mixin's layers (its tools, skills and staged context) were dropped
without a refusal. A profile's `kit` is now one digest-pinned reference:
a workload Kit, or a published Kit set, which the Kit spec's frontend
merges from a workload and its mixins into one image. archie composes no
Kits itself, and a bare mixin is refused with a pointer to publishing a
set. kit.Compose had no production caller left and is removed.
BREAKING CHANGE: `[containers.profiles.X] kit` and the agent-profiles
document's `kit` are a string, not a list. A stored list fails to decode
and must be rewritten as the one workload or set reference.
Part of archie-core-egkf.6.
The egress proxy now holds a Kit's OAuth tokens on the host. A refresh
carrying the sentinel refresh token goes upstream with the org's stored
one; a successful response is written to the State Store and returned to
the container with sentinels only. Requests to the Kit's resource hosts
carrying the sentinel access token get the stored one.
The org's token set is read or written only for a run whose credential
carries the service, so a Kit declaring a service never reaches the org
secret by itself. The launcher grants a bound OAuth service without a
config secret, and Register refuses a required OAuth credential when no
harness secret store is wired. A provider that does not rotate refresh
tokens keeps the stored one.
Part of archie-core-egkf.7 (#1208). Credential-file rendering and
per-binding concurrency remain.
A worktree terminal cleanup keeps (uncaptured or unreadable work) now has a
reap path: archiving calls Daemon.RemoveWorktree for the owning identity's
trees, and the keep logs name archive as the way out. The no-change cleanup
test reports completed through one direct transition instead of reaching
merged via a faked pr_open hop.
Recover the OAuth storage slice from f6079e9d and the missing workflow persistence error handling from f07c314e, adapted to the current step contract. Require encryption for OAuth token storage and verify RPC round trips and task-grant denial.
The task-list table and NeedsYouCard read a task.stage field the API
stopped emitting once tasks.stage dropped; remove the column and the
fallback rather than let it render an empty cell forever.
stateStoreDeps grew a fourth optional-capability branch (StepReader,
alongside Pi's ExecutionCanceller and the RecoverStale rebase) and
tripped golangci-lint's cyclop gate; split the execution-tree surfaces
into executionDeps, mirroring the existing accessDeps split.
docs/prds/execution-tree-state-machine.md: step_executions (0017) is now the
durable per-step record; the dashboard's run detail (GET /api/tasks/{id}/
attempts) still folded stage_start/stage_finish events into a rail by hand,
and tasks.stage duplicated what StartStep/FinishStep already record properly.
Added the missing read:
- ListSteps(execution_id, attempt) -- attempt 0 lists every attempt of the
execution, oldest first, so the rail renders every attempt from one call.
New storecontract.StepReader (separate from StepRecorder: this is a
dashboard/API/dispatch action per the PRD's wire contract, never
task-scoped-grant callable), postgresdb.ListStepExecutions, staterpc
client/server/values.go plumbing, wired through state_store.go's Deps and
archieui/compose.go's Server.Steps the same optional-capability pattern
every other dashboard read uses.
- internal/webui/api_tasks_attempts.go: stageViewsFromSteps replaces
foldStages/finishStage/stageOutcome. Attempt bounds, event_count and which
attempt numbers exist stay sourced from events (an attempt can carry
events -- agent calls -- with no step), only the stage list itself moves.
The wire JSON contract is unchanged, so no frontend change was needed;
`task ui` confirms ui/dist still matches a fresh build.
- A running step past the task's own current attempt reads interrupted, not
running: RecoverStale (archie-core, still open per #1214) is what corrects
a crashed run's stale "running" row, and until it does this is the honest
read rather than trusting a row that may be stale.
Dropped tasks.stage and every reader, no fallback:
- Migration 0018 drops the column; tasks.sql's INSERT/UPDATE/RETURNING lists
and the sqlc-generated code no longer carry it.
- workflow.Task.Stage is gone; workflow.Run's in-flight "which stage is this"
bookkeeping moves to TaskContext.Stage (execution-time only, never
persisted -- park/finish's event tagging is its only reader).
- pb.Task.stage (wire field 10) is `[deprecated = true]`, not deleted (never
reuse a field number); staterpc's taskProto/taskValue no longer map it.
- taskactions.Task.Stage and the retry event's "previous_stage" data field
are gone: the operator-action package had no reader for the column beyond
echoing it.
- The chat /tasks command's per-task "workflow/stage" display drops to
workflow-only (ChatTaskSummary.Stage removed, workflowStage() deleted):
it was a genuine reader with no replacement in scope for this change.
Read the archie-web-ui skill and the state-store/daemon docs/development
pages first, per instruction. Stayed out of cancel/stop code throughout
(Pi's concurrent CancelExecution work); the only shared files touched
(workflow.go, daemon-adjacent SQL) got the smallest possible diff.
Plan item 5 of docs/prds/execution-tree-state-machine.md (archie-core-1217).
Until now a stop was a context cancel plus a caller-side transition, and the
step rows a worker still held running could never be told: the execution moved
while the work carried on, and the old interrupt-first ordering existed only
because the record could not defend itself. The step table can: the PRD's
ordering -- the store is the record, the context cancel is the delivery --
became safe the moment a worker's next step write fails ErrStaleTransition
against a cancelled step.
- storecontract.ExecutionCanceller and the CancelExecution RPC: one
transaction moves every non-terminal StepExecution of the execution's
current attempt to cancelled (one event row each, in the same
transaction), and the execution itself to the status the operator action
names, under the shared transition table with its audit row. Staleness is
decided before legality, the way guardTransition decides it. Earlier
attempts are never rewritten. Deliberately administrative: the
deny-by-default grant rule covers it, pinned by a test.
- taskactions.Service routes stop, cancel, abandon and reject through it and
the old Transition/CancelTask combinations are deleted; the daemon's
in-memory cancel is demoted to the delivery hook the PRD names, called
after the store record.
- The gateway's chat controller records first and delivers second, the same
order, and records even when no runtime is wired here (a standalone
gateway cannot reach the daemon's goroutines): the record is what unsticks
a lost-cancel task, and the delivery is best effort. StopRunning and
TaskRuntime.CancelRunningTasks are deleted outright -- no production
caller remained once the task_action tool owned stopping, and the daemon's
CancelRunningTasks/runningTasks.stopAll go with them.
- The gateway's chat adapter writes through CancelExecution instead of a
bare Transition, so chat /cancel and the dashboard's action service land
on the same path.
PRD decisions recorded here:
- Each cancelled step's event row is stage_finish with {cancelled: true} and
the reason as its detail -- the exact-one-event-per-step-transition rule
covers the pending -> cancelled transition too, even for a step that never
ran.
- A parked cancel carries the needs_human class, matching the Transition
path this replaces; the transient class stays with parkRunningTask's
unwind parks, which now meet ErrStaleTransition for cancelled runs and are
ignored there.
- Cancelling a wait:true callee is deferred to #1216's agent-call children:
the callee's own tree and the wait flag are what make
"cancelling a caller cancels each callee it waits on" decidable, and the
callee writes it cannot fake are this issue's half of the contract.
- RecoverStale's interrupted sweep is plan item 6, untouched here.
docs/prds/external-agent-harness.md (Credentials) and orgs-and-access.md:
egress.Resolver had no production implementation, and no run carried a
credential naming the org secrets its identity is granted, so a Kit's
required credential@1 was refused at the proxy and no harness run
authenticated (archie-core-egkf.11).
Two things, as scoped (this does not attempt #1026's full run-credential
consolidation replacing WorktreeGrantIssuer/StateStoreGrantIssuer -- that
PRD item is real, separate, and still open; this is deliberately narrower):
1. A run credential naming granted secrets: config.CredentialBinding maps a
Kit credential@1 service to a secret; config.Config/IdentityConfig gain
Org and GrantedCredentials (a per-identity allowlist of service names,
reusing the existing per-identity config-override pattern DiffCapLines
already uses, not a new Cedar/Authorizer integration -- KindSecret policy
objects exist in internal/domain/access but assembling a full Principal
is #1026's job, not a two-line addition here). config.ContainerConfig.
BoundCredentials computes the declared (Kit) ∩ granted (identity) ∩
org-matched intersection Sam's own rule requires: neither side widens
the other, proven by a mutation-checked test flipping each of the three
independently.
2. egress.GrantResolver: the production Resolver, holding only what a
caller explicitly Grant-ed one run, replaced wholesale (not merged) on
a repeated Grant, gone entirely on RevokeGrant. kitrun.Launcher.Launch
computes the intersection per launch and resolves real values through
the existing secret registry (SecretResolver), Grants it before the
proxy session opens, and Release/error-path rollback (a new shared
Launcher.release) always revokes it.
Course-corrected mid-implementation (flagged by Sam): credential bindings
first landed as a boot-time config.ContainerConfig.Credentials field
captured once into Launcher.Bindings -- exactly the "parses but is not a
resource" trap CLAUDE.md warns about, and inconsistent with agent-profiles
(#1207) making a Kit profile apply without a restart. Rebuilt as its own
live control-plane resource, CredentialBindingsKind ("credential-bindings"),
following the agent-profiles/container-runtime-policies split precisely:
config.ContainerConfig.Credentials is json:"-", [containers].credentials
is only its seed, and kitrun.Launcher now holds *config.Holder (not a
captured slice) and reads Containers.Credentials fresh on every Launch.
Org/GrantedCredentials were already read fresh per dispatch (d.configFor
inside runKitTask, never cached), so no equivalent fix was needed there.
Deliberately not built: identities and grants as their own State Store
resource (orgs-and-access.md's "Org owns ... identities and their grants").
Identities are entirely config.toml-defined today with no control-plane
resource at all; moving GrantedCredentials alone into one would put half
an identity in two places. That migration is identities' own scope, not
credential bindings'.
Plan item 3 of docs/prds/execution-tree-state-machine.md (archie-core-1220), on
top of the transition tables and ErrIllegalTransition enforcement (1214/1215,
archie-core-qdbo.5):
- Migration 0017 adds step_executions -- id, org_id, workspace_id, execution_id,
attempt, parent_id, depth, kind, name, called_execution_id, status, detail,
tokens_used, started_at, finished_at, indexed on (execution_id, attempt, id).
Ownership fields are stamped from the execution's row; depth is derived from
the parent, never supplied by a caller. Existing rows gain no backfilled
steps.
- task.Store (the workflow engine's mid-run facade) gains StartStep/FinishStep
(5 methods, within the 8 cap), with storecontract.StepRecorder fronting them
on the server; deadlineStore wraps both with its per-call timeout.
- The store enforces each write as one transaction: the guarded row write
under the shared step transition table plus the transition's own
stage_start/stage_finish event row. Staleness is decided before legality,
the way guardTransition decides it, and the store returns the persisted
event for the caller's post-commit publish -- the EmitDurable convention
(the daemon's event sink skips already-persisted rows by assigned ID), which
is why workflow.Run stops emitting stage events beside the store and no
longer discards step-write errors: a failed start or finish parks the
execution, and the stage does not run unrecorded. An interrupted stage
records interrupted rather than failed; a recording failure on that
shutdown path is logged and left to RecoverStale, which owns the
interrupted-execution sweep in plan item 6.
- The run credential of docs/prds/orgs-and-access.md authorises StartStep and
FinishStep on its own execution only: both requests carry the execution ID
the grant check verifies (FinishStep carries it because the interceptor
never reads the database), and FinishStep verifies handler-side that the
step really is that execution's. authorizesTaskScopedCall's sanctioned
widenings are now held as a lookup table so the next one stops growing a
switch.
- The PRD's wire contract is StartStep(execution, attempt, parent, kind, name)
-> step id; the responses also carry the persisted event (needed for the
post-commit publish) and FinishStep's request carries execution_id (needed
for the grant check) -- both noted here as deliberate additions to the PRD's
shorthand. called_execution_id exists in the table but not yet on the wire:
its named consumer is the workflow-call recording of plan item 4 (#1216).
Conformance drives both adapters; grant tests pin the own-execution scoping;
the wire sentinel set gains ErrInvalidStep (InvalidArgument, "step execution
invalid") for a step kind no consumer names.
docs/prds/external-agent-harness.md ("Selection"): a Kit profile is an org
resource that applies without a restart. Profiles were bundled inside
container-runtime-policies (ApplyMode: restart-required), which is correct
for Image/MaxConcurrency/Network but not for Profiles, which nothing
startup-builds: daemon.configFor reads config.Holder fresh per dispatch
(pinTaskProfile -> Containers.Profile).
Split Profiles into its own resource, AgentProfileKind ("agent-profiles"),
ApplyMode: live, seeded from [containers.profiles] and validated with the
same rules the file document uses (config.ValidateAgentProfiles, factored
out of ContainerConfig.ValidateProfiles so the two cannot disagree).
config.ContainerConfig.Profiles is now json:"-" so container-runtime-policies
never carries it (one authoritative resource per field, not two), and
runtime_config.go layers the new resource in by wholesale map replacement
(matching ModelRoleAssignmentsKind's precedent) rather than merging into the
file's map. Added Profiles to reload.go's Containers reloadable-subfield set
to match this.
Deliberately NOT built here: true per-org multi-tenant storage for this
resource (docs/prds/orgs-and-access.md places agent profiles under org
ownership, granted to workspaces). No existing control-plane resource kind
in this registry does real per-org storage yet either -- schedules,
workflow settings, etc. are all single instance-wide documents -- so
agent-profiles follows that same established, tested pattern rather than
inventing org-scoped storage unilaterally for this one resource. Org/
workspace grants over agent-profiles are a follow-on, not a redo: nothing
about this resource's shape needs to change to add that scoping later.
applyWorkflowExecutionSettings swapped the running config with cfgHolder.Set,
so the dashboard's published projection kept the previous budgets until the
next reload. Route it through publishConfig like every other config change.
Fixes#1148
2fcc32d8 gave archied sd_notify: one READY=1 when the run loop starts serving.
RunGateway and RunStateStore serve contracts and health surfaces of their own --
the State Store even takes -ready-addr for a readiness HTTP listener -- but told
systemd nothing, so run as Type=notify either would hang until TimeoutStartSec
and be killed in a restart loop while perfectly healthy (archie-core-1174).
Both now announce at the moment each declares itself serving, the same notion
archied's runLoop announces beside healthSurface.markServing -- no second
definition of started. The Gateway announces from its serving tail
(serveGatewayListener, extracted from RunGateway), after the listener is bound;
the State Store announces after its optional surfaces have started, because a
readiness surface that fails to bind fails the boot and READY must not precede
it, and before serveStateStore, which begins accepting the moment it is called.
No unit file changes Type=notify for them; the helper already lives in package
archied beside both entry points, so no move was needed.
Tests follow the notify pattern the daemon's announcement already has: the
Gateway test drives its serving tail over a bound loopback listener with a
stub contract and judges READY against a real Snapshot RPC on the same socket;
the State Store test drives the real RunStateStore boot over a pgtest database
and waits for the datagram. Both failed red before the announcement existed --
no datagram within the window -- and pass after.
archie-core-1000's failure mode was silent persistence: provider-settings seeded
from an already-resolved snapshot stored the process-local ARCHIE_PROVIDER_<hash>
env name with no api_key_ref, and once the resource outranked the file no later
boot had a reference left to resolve. The producer was fixed in ff77612b and the
stored shape is refused in 093e677a, but nothing pinned the seed path's own
refusal: ImportConfig over a resolved snapshot must skip the kind, store nothing,
and report why -- the State Store still starts and the file document stays in
force -- rather than wedging on it or storing it.
Pin that contract, and prove the refusal comes from the guard: with
rejectBootDerivedProviderEnv removed the test fails with 'provider-settings was
stored from a resolved snapshot'.
container-runtime-policies seeded and validated config.ContainerConfig directly,
the last resource in the registry whose document was the internal config struct
(archie-core-1171). encoding/json fell back to that struct's Go field names --
Image, MaxConcurrency, LegacyEnabled, Profiles -- the only Go-cased keys the Web
UI still rendered, and case-insensitive matching let a lowercase misspelling
("maxconcurrency") decode as the field while "max_concurrency" was read back
half-applied.
Define the document here instead: containerRuntimePolicies with explicit
snake_case tags and agentProfile at the profile nesting level, the same shape
267d21a2 gave the channel document. The two durations stay config.Duration,
which already writes the string the file accepts; this leak was the keys only.
The mapping to config.ContainerConfig is one function the validator and the
runtime layering share.
Reading a legacy document is deliberate, so it stays strict about everything
else: DecodeRenamingLegacyKeys renames the seven Go-cased keys an existing store
holds (and the four at the profile level), drops LegacyEnabled -- the removed
containers.enabled switch, which had no runtime consumer and no current
counterpart -- and keeps DisallowUnknownFields, because swallowing an unknown
key decodes as unset and an unset image is a broken worker. A misspelling is
refused.
normalizeContainerPolicies, through the Definition.Normalize hook the channel
document already uses, rewrites a legacy document into the current shape on
write, so a stored document converges the next time it is saved. The layering
keeps its profile semantics -- a document that omits profiles (one stored before
they existed) inherits the file's, one that carries them replaces the file's --
and now checks both spellings of the key.
docs/development/control-plane.md's 'adding a database-owned field' still
instructed the untagged practice this class of issue removes; it now names the
wire-shape rule, the legacy-key tolerance and the Normalize hook instead.
The equivalent fix was made in the crew/qc8p worktree but never committed there before the merge, so the merge silently carried the pre-fix version through despite a clean, no-conflict merge. Re-applied directly on main; task check clean.
Readiness probe test asserts each subsystem registers exactly once instead of
an ordered five-item list. Config descriptor test drops the exact-count check
and its stale settings.js rationale. Container defaults test references the
production constant.
Refs #1229#1231#1232
TestExecuteTaskRequestStartsWorkflowCallCallee built its transport with no URL, so it silently fell through to nats.DefaultURL (127.0.0.1:4222). It passed locally against a developer's own NATS server and failed in CI with no servers available for connection. Start an ephemeral server per the sibling test's pattern and dial it explicitly.
Flip provider-settings, model-role-assignments, repository-policies and
scheduling-policy to ApplyMode live, and give the daemon a watch per kind
(keepWatch's reconnection and backoff rules): on a new stored version the
watch re-runs boot.runtimeConfig over the published snapshot and republishes
through config.Holder, keeping last-known-good when the layered document
fails to validate and reporting through apply status.
The stored model-role assignments now own cfg.Models entirely (the map
merge would resurrect a role the store deleted on every re-layering), and a
provider or role update rebuilds the gateway chat runtime's provider set
(swapped runtime, read per turn) and the chat model manager's offered list.
Scores a reviewer's findings on the 38 runnable Martian Code-Review-Bench
PRs: the deterministic core turns findings into posted comments, an
independent judge model matches them to the goldens, and the run reports
golden micro-recall and golden-only precision. The reviewer is a findings
fixture directory until the pr-review workflow lands.
Run with: task bench:prreview -- -findings DIR
The execution transition table refuses most of the pairs shipped fixtures
wrote to build a finished task -- running -> merged, queued -> parked,
queued -> waiting_human, queued -> merged -- so those fixtures fail the
moment the store enforces it.
internal/taskstate/taskstatetest seeds a status by walking the table's
legal edges (shortest route, computed from taskstate.CanTransition), so a
fixture cannot write an illegal pair and fails loudly when no route
exists. The fixtures in internal/daemon, internal/app/archied,
internal/webui and internal/infrastructure/postgres go through it, and
the table's own test names running -> merged as refused.
archie-agent mcp is a stdio MCP server, on the official go-sdk, serving a
stage's capture tools and appending accepted calls to a captures file.
The harness runner writes the tool spec and a standard mcpServers config,
registers it through HarnessSpec.MCPConfig, and reads the captures back
into the stage result. The captures file is writable by the harness user,
so the worker validates every record again with the same acceptCapture
the built-in loop now shares, and forged, undeclared or over-cap records
are dropped. A stage with capture tools on a harness without MCP is
refused before it runs.
A Kit task's worker shares the sandbox network with its harness, so it
reaches NATS and the State Store through the relay like the harness
reaches the proxy. archie-agent relay takes repeated -forward
listen=target pairs, binds them all before serving so a port conflict
fails at start, and stops every forward when one fails. RelayArgs builds
the command from the fixed relay ports.
archie-agent relay listens on the sandbox's isolated network and joins
every connection to one fixed target, the proxy. The relay holds no
secrets and can reach nothing else, so the host needs no address on a
sandbox network at all.
configuredSecretRegistry resolved provider keys in place, and the State Store's
copy of the config shared the providers map with b.cfg, so ImportConfig seeded
the process-local ARCHIE_PROVIDER_<hash> env name. Provider resolution now
runs only in the daemon and gateway boot (openStores).
archied already finds usable providers (key set in the environment, or
configured) through the model catalog; the published config view now
carries that catalog, providers with their model ids. Settings > Models
offers role models only from configured providers, lists the catalog's
unconfigured providers as available, and adding one carries its class, key
variable and endpoint. Roles and providers are two accordions, each ending
in an add row.
Every install serves from Postgres, so the legacyread/legacyimport packages,
the archie-state-store import command, the boot gate that refused to serve
over unimported SQLite files, and db_path (which only located those files)
go. Migration 0010 drops the import_completion table.
Each credential is now only a {engine, key} secret reference. A stale
token_env in a file is reported at startup as an unrecognised key.
nats.token_env stays (no secret-ref form), and providers.api_key_env stays:
it is the carrier the resolver writes each resolved key's env name into.
None had a production consumer: [agent] and containers.enabled were decoded
only so old files kept loading, [oidc] was never read (archie-ui takes OIDC
from flags), and [image] was validated but no provider exists.
docs/development/ holds one page per kind of change (settings, State Store
and persistence, task lifecycle, workflows and steps, event automation,
dashboard, agent execution and containers, processes and wiring, merging
work), each naming every layer the change must reach and pointing at the
document that owns the design. The docs build puts them in the development
set; docs/index.md and CLAUDE.md link them.
Also adds the container-profile case to the validator agreement test, the
check the profiles work skipped.