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.
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.
* feat(archied): read-only run-scoped sandbox view over the harness contract
Serve RunTarget: the daemon resolves a running task and attempt to its Kit
container and follows its output, refusing a finished or unknown task with a
stated reason. The dashboard adds a Sandbox tab behind the task log's read
authority. Takeover stays out of scope until the sod1 authority is wired.
* fix(archied): scope and close sandbox views correctly
Forward harness activity into container output, bind containers and dashboard views to their attempt, enforce the asserted org, and report unavailable dependencies and interrupted streams accurately.
Refs: archie-core-qi91, archie-core-252tj
29 of 59 directives did nothing. Removing each one individually and
re-running the gate revealed no issue from any of them: 20 named gosec,
which is not enabled in .golangci.yml and never has been, plus 4 nilerr
in functions that return no error, 2 dupl where dupl reports nothing
repo-wide, and one each of contextcheck, nestif and unused. The funlen
name on settings.go:147 is gone for the same reason; its gocyclo and
cyclop names are both live.
The unused directive in webui/authorize.go sat on `var _ = slog.Default`,
a statement that only existed to keep the log/slog import alive. Both go.
The 30 remaining directives all suppress a real issue.
Bump github.com/samcharles93/ai-sdk v0.1.40 -> v0.1.41. ai-sdk now sends
max_completion_tokens for reasoning models and omits temperature when the
model rejects it, so the max_tokens workarounds are no longer needed: the
chat-turn bound is set directly and MCP sampling loses its reasoning
branch. Classic models keep their existing bounds.
agentexec's wire contract, harness adapters and log forwarding no longer
import the AI SDK; the in-process model loop, runtime, tool set and sampling
helpers that do now live in agentexec/modelloop, used by the Gateway and
archie-agent only.
internal/domain/workflow imported internal/agentexec for its protocol types,
pulling the AI runtime into every binary that touches workflow types. The
request, result and runner contract now live in internal/domain/agentrun;
agentexec keeps the runtime. The model catalog cache moves from servicekit
into modelcatalog and WaitForPeer into staterpc, so archie-state-store and
archie-messaging no longer link the AI runtime.
Remove PRD/doc paths, bead IDs, phase notes and deleted-test references
from comments; drop test-only hooks left by the test deletion. Includes
in-progress archied bootstrap, gateway and query changes.
A task hosts the same configured MCP servers as a chat turn, but the
worker's provider builder passed no sampling handler, so a task-scoped
server's sampling/createMessage request got a JSON-RPC method-not-found
while the identical request succeeded on the chat path. The worker runs
its own model runtime, so it can answer.
Thread a sampling handler built from the task's builder model into the
worker's MCP provider builder, beside the chat path's handler. Build the
task's model runtime once and share it with the stage runners instead of
creating a second client for sampling.
Move the sampling message mapping and max-token bound to agentexec, next
to the runtime both paths call models through, so a server gets the same
bound on either path. Carry the catalog's reasoning flag in
ModelLimits: a reasoning-class builder model must omit the max-token
bound its provider rejects, exactly as the chat path does.
A remotely-shipped task log entry was re-emitted through the sink with
slog.Logger.LogAttrs, which stamps time.Now(): the entry's own event time
was discarded and replaced with the daemon's receipt time. Because a
container's records arrive in bursts, a multi-minute run collapsed onto a
few milliseconds of receipt timestamps and its timing became
unrecoverable -- which stage ran when, how long an iteration took, whether
a stage stalled.
Write now builds the record from entry.Time (falling back to receipt time
only for an entry that carries none) and dispatches it through the same
handler, preserving the sink's level gate. Receipt time stays available
separately in the record's received_at field, so event time and arrival
time cannot be confused. The agent-side producer already carried the
record's own time; a test now pins that half too.
The temp-repo fixtures in agentexec, domain/workflow and agentworker seeded
a repository and committed without disabling gpg signing, so a host with
commit.gpgsign=true and an agent-only SSH signing key failed those tests at
the fixture's own git call ("Couldn't get agent socket?").
Each producer now writes commit.gpgsign=false into the temporary repository's
own config, which also covers commits made inside the seeded repo by a
subprocess (the fake harness). A regression test per package pins the
fixture's hermeticity under a hostile global git config.
Clears the five open CodeQL alerts. harnessCredential parsed uid and gid
as 32-bit and os.Chown converted them to int, which wraps on a 32-bit
build; they are now parsed to 31 bits, so an out-of-range id is refused
where it is read. mergeMapping sized its map by len(base)+len(override);
it now hints len(base) and lets the map grow.
Claude's input_tokens excludes cache reads and writes; the built-in
loop's Anthropic provider folds them into PromptTokens so budgets count
total prompt size. The harness adapter now does the same, so one budget
means the same on either runner.
Pairs tool_use and tool_result blocks by ID into tool-call reports, keeps
the latest session_id, takes the run's usage from the result event and
falls back to per-message usage, counted once per message ID, for a run
killed before its result. Tested on streams recorded from Claude Code
2.1.282 and scrubbed of host detail.
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 runs worker and harness in one container, separated by user:
the worker keeps its credentials as root, and every harness invocation
runs as the Kit's user with only its declared environment. A nil env
would inherit the worker's, so the runner always passes an explicit one;
root is refused as a harness user, and a root worker refuses to start a
harness with no user. The PRD's container layout now says one container
per run, with the relay forwarding the proxy, NATS and the State Store.
A stage whose request carries a HarnessSpec runs the Kit's CLI as a
subprocess in its own process group, with the mission and archie's rules
as the prompt. After every invocation the runner compares HEAD, every ref
and the git configuration, and the worktree against a pre-run baseline:
any ref or config change, a protected-path edit, or a write in a
read-only stage parks the step whatever the CLI claimed. The runner runs
the gate itself and resumes the same session with the gate's output, up
to the gate's failure cap. The wall clock and cancellation kill the whole
process group, so the CLI's children cannot outlive the step.
ClearTerminalTasks now deletes completed tasks, with a test that holds its
SQL list to taskstate.Terminal. The AllowTools comment says agentloop's
built-in file tools are outside the allowlist and why. The PRD model
diagram lists only what a profile holds (image, tools), and the profile
field and AgentProfile say a workflow's profile name is resolved at
dispatch, not when the workflow is saved.
Workflow definitions declare inputs, repository (none, optional or
required; default required) and profile. A binding assigns each input
from a mapped parameter or a constant and may take its repository from a
parameter, which must name a configured repo; both are checked when the
binding is saved and again at dispatch. Inputs travel as structured task
data (tasks.inputs, the task brief, and a JSON block in the agent's
mission), never as body text.
A task with no repository runs in a per-task scratch workspace with no
clone and no publication grant. The agent.run step runs one mission
without a repository and finishes the task in the new completed status.
Agent profiles live under [containers.profiles.<name>]: an image and an
allowlist over the tools archie adds (MCP, scripts, plugins).
[containers].image is the default profile; an unconfigured profile parks
the task. The workflow is now pinned before the container starts so the
profile can choose its image.
Also stop two archied tests writing embedded NATS state into the package
directory, refuse an empty state_dir for embedded NATS, and drop the NATS
store files an earlier lane commit tracked.
The selection criteria alone left workflow optional and defaulted an omitted
or unrecognised value to implement, so a decide call could still skip the
decision and send an unsettled capability straight to the builder.
Omission is now a retryable rejection: CaptureTool gains RequiredWhenTrue, a
conditional requirement a flat RequiredFields entry cannot express, because
the close-without-change branch has no workflow to name and must not be made
to invent one. A rejected call is not recorded, so it does not consume
MaxCalls and the model retries with the field.
A value outside the enum now fails the stage rather than degrading to
implement: reaching OnResult with one means the model ignored its own schema.
Splits validateCaptureArgs into per-constraint checks (cyclop) and carries
the regenerated contracts.json.
The events table had no attempt column, so two attempts of the same task were
indistinguishable: a retried task's stage events interleaved into one timeline and
nothing could say which run produced a given result.
- events gains an `attempt` column through the existing presence-gated
migrateTasks loop rather than a version bump, so user_version stays 3 and an
existing archie.db is migrated in place. A legacy-database test builds the
pre-change schema and proves both that existing rows survive unattributed
(attempt = 0) and that a new value round-trips.
- Event.attempt = 11 and ReadTaskLogRequest.stage = 9 are field additions to
existing messages: no new RPC, no new table, and no change to any canonical error
message that crosses the state-store wire.
- Every task event the workflow engine emits now carries the attempt. Producers
that have no attempt yet (task_queued, binding_dispatch_failure) stay
unattributed rather than guessing one.
- Changed files are captured as a durable changes_captured event at commit and push
time, reusing the existing three-dot diff primitive: paths from FilePatches(),
counts from Stats(), base_sha recorded as the merge base actually diffed rather
than the base tip, a 200-file cap with totals kept complete. Consumption goes
through an unexported optional interface so workflow.Trees -- the yaegi capability
surface for interpreted stage plugins -- is not widened. A capture is also taken
once the pull request exists, so the capture can link it.
- The per-attempt configuration is captured as a durable config_captured event
carrying the non-secret config.TaskConfig, instead of a new table plus two new
state-store RPCs: events are already this system's per-attempt provenance stream,
so a table would have given one feature two provenance mechanisms. Redaction is
proven against canary secrets. Measured payload: 539 bytes typical, 2333 bytes at
20 models.
- Captures carry a schema marker and are refused unless their key set matches the
declared payload shape, so drift surfaces as "recorded but could not be read"
rather than as a capture of zeroes.
- Three read endpoints serve the attempt rail, the changed-file captures and a raw
debug view, mapping to UI-owned response types rather than dumping store types.
- The task-log read gains a stage filter, bound to the stage logger for that
stage's own lines; agent and tool output records no stage.
A task-scoped grant may already insert events on its own task, so this provenance
is exactly as trustworthy as the rest of that task's event stream. That limit is
documented rather than papered over.
- toolset.go: extract requireApproval out of toolExecute (nestif)
- gateway.go: extract handleDeny/handleVersion/handleUpdateCheck/
handleRestartAdapter out of dispatchLocal (gocyclo 26 -> under 20)
- cronstore/store.go: extract applyPatch/applyScheduleChange out of
Update (nestif, gocognit)
Behavior-preserving.
Replaces hand-written JSON Schema + manual unmarshal with
core.NewTypedTool[In], now available at the pinned ai-sdk v0.1.30. This
removes the class of bug where a hand-rolled schema drifts from the
struct a manual decode unmarshals into.
Task runs now preserve real model context limits, provider usage, baseline-repair spend, commit identity, durable agent-finish telemetry, and useful timeline details without imposing a token cap. Dashboard forge links are canonical per owning GitHub or Gitea identity. Chat turn claims acquire SQLite writer intent before reading, eliminating the concurrent SQLITE_BUSY snapshot race.
Autonomous workflows now have exactly one production boundary: a task-scoped archie-agent receives a full-task core-NATS request and runs the worker-local stage loop. The legacy [agent] section remains decode-only for migration; it no longer selects runtime behavior.