AGENTS.md sends every instruction to a file in internal/prompts, and the
workspace scoping rule, the PR-review adversary mission and the model probe
were the three that still read as Go string literals. The text is unchanged:
the templates render the same bytes the concatenations built.
A step is id/type/settings with when and on_failure; the five kinds become
the types agent.run (inline mission), connector.action, approval.wait,
run.notify and workflow.sleep. Documents declare inputs, repository,
instructions, identity and profile; references are {{ task|inputs|steps }}
checked at save. Any number of approval steps, each asking only about what
was held since; actions still held at the end are put to people before the
run finishes. Triggers are refused for a workflow they can't start, and the
workflows API lists, reads and saves documents beside a step-type catalog.
Copied from internal/domain/workflow: definition.go, step_refs.go,
step_flow.go's checks, control_steps.go's when parser, vocabulary.go and
task/interface.go. stableid moves to internal/platform.
Label to workflow rules and a default workflow are the intake-routes
resource: ordered, case-insensitive, first rule wins, seeded with bug to
tdd, feature to feasibility, default triage. A write naming an undefined
workflow or an empty label is refused. The daemon reads it through the
workflow definitions client when pinning a task and at intake.
The kind indirection only carried this routing, so it is removed: the
per-kind NATS subjects (every consumer already bound archie.task.>) are
one subject, the envelope and playbook triggers lose their kind, and
x-intake-routes is dropped. The dashboard gets an Issue routing page
under Automation; the trigger card reads the same setting.
Each step type declares its title beside its factory and settings; the
control plane serves it in x-step-types, and the dashboard reads it instead
of a hand-written TYPE_TITLES table. Parallel and switch are YAML shapes, not
registered types, so their titles stay with the graph.
Archie runs its own agent loop only. Removes Kit resolution and launch,
the harness runner and adapters, the setup terminal and sandbox view,
harness secrets and OAuth capture (table dropped in migration 0075), the
task-tools contract that only Kit containers reached, and the
sandbox-kit-spec dependency. Agent profiles carry image and tools only.
The egress proxy keeps the model-session path; the sandbox relay code
stays for #1867.
Model usage was recorded per call and never read, so an org could not see
what it spent. Add the read side:
- model_usage carries identity_id. Task calls take it from the task row,
chat and curator calls from the calling principal carried in the usage
record, so an outbox replay keeps its attribution. Existing task rows are
backfilled from tasks.
- StatsService.UsageTotals sums the caller org's rows by UTC day, source,
provider, model and identity. The org comes from the principal, with no
org parameter, so cross-org reads do not exist.
- GET /api/usage prices tokens at the catalog's list rate, netting cache
hits, and reports calls to unpriced models as tokens only. The published
catalog view now carries each model's cost.
- A new org Usage page behind a read_usage grant: admin and owner only,
with a 7, 30 or 90 day range, totals, the daily series, and model, source
and identity breakdowns. The org shell keeps it on the caller's own org.
Drop the bot_user-derived root identity and the empty-identity sentinel. System
polls forge connections and owns instance-level work; a polled task stores the
polling identity, task_queued and work_request_submitted carry task id, actor
and principal, and an empty task identity is refused. Migration 0072 moves
empty tasks.identity, the identity-grants root key and the old root's
connection permits to System so a re-poll dedups.
Issue-level writes could comment and close but not open an issue, so no
curator or workflow stage can file its own work yet. Adds the rpc to
proto/forge/v1, regenerates the contracts, and threads it through the
domain interface, forgerpc client/server (comment permission, like
CloseIssue), forgeext, and the archipelago SDK + gitea/github plugins.
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.
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.
workflow.approve pre-writes the task's workflow column with the workflow
that runs once approved, so task_list named a workflow that had not
started while the task's own log named the one that produced the plan.
Chat now names the compiled workflow through the pinned definition for a
plan wait, matching the log.
A run that parked left tc.Outcome empty, so the taskrun response carried an
empty status and the daemon logged 'taskrun run recorded no outcome' for a run
that had recorded one, then retried the park it could not apply.
Record the outcome once the parked transition lands, so a budget stop and every
other park surface with the reason the run parked.
StagePRReviewAndOpenPR set tc.prReview.ownPR before the decision stages ran,
but intake allocates tc.prReview, so every implement run dereferenced a nil
state at repo.open-pr and the agent container panicked with exit code 2
before opening the pull request.
GitHub refuses REQUEST_CHANGES from a pull request's author, so a blocking
finding on the PR a task opened lost every finding. The retry's identical
refusal is now reported once, and forge errors drop the gRPC framing.
An approval, ask or park can be resolved by an operator, a rule, a
classifier or automatically, so the vocabulary stops naming a person:
- step type human.approve -> workflow.approve
- park class needs_human -> needs_decision (ParkNeedsDecision)
- event kinds human_* -> operator_*
- alert kind task.needs_human -> task.needs_decision
- status grouping "needs you" -> "needs attention"
0061 rewrites the stored values, including the approval step type in the
workflow-definitions resource and each task's pinned definition, with the
digest recomputed so resumed tasks still match.
The prd step returns the PRD as structured fields and human.approve takes
them as its document. The plan text every surface reads and the approval
card's layout are both rendered from that one document, so what the human
approves is what the implement workflow is handed. Replaces the separately
generated OpenUI view, which could drift from the plan.
agent.run gains views: the agent may return an OpenUI Lang view of its
answer through an optional view tool, read as steps.<id>.view. human.approve
stores it as the task's plan_view, and the approval card renders it with
the dashboard's own components, falling back to the plan text when there
is no view or it does not parse. The feasibility prd step uses it.
A test run used to suppress only forge and push writes, which named one
destination among many. Deliveries now have one seam: TaskContext.Deliveries
records each as "would ..." and performs none, and a recording egress
session does the same for a Kit run's HTTP writes and central tool calls,
answering them synthetically. Provider hosts stay live, since model traffic
is how a test run reads. Native containers have no proxy (archie-core-n8d3).