A task's real work happens in an ephemeral archie-agent container over
NATS. Four things currently destroy the evidence of what it did:
cmd/archie-agent/main.go logs to stderr only, and internal/container/pool.go creates containers with AutoRemove: true —
the log dies with the container.
StageBaselineGate (internal/domain/workflow/implement.go) captures
full gate output via CombinedOutput(), clips it to 200 chars into a Warn log, and discards the rest — the park reason carries only a status
string.
agentexec.Result has no step-level transcript (tool calls, durations,
errors).
agentexec.SubjectForSystem is documented as the return channel for
exactly this ("log dumps, health... the daemon reads these for
observability") but has zero publishers and zero consumers anywhere in
the codebase.
Consequence: a task parked today with stage baseline: baseline red -- go build ./... fails and builder could not auto-fix (status: parked) and there was no way, from the dashboard or from
chat, to see the actual go build error — the container that had it was
already gone.
This matters beyond convenience: archied is meant to run unattended for
weeks (CLAUDE.md deployment model) and potentially for users other than
its original operator. A system only debuggable by someone watching docker ps live doesn't meet that bar.
Decision
NATS is the only transport this targets. [agent] mode = "nats" is the
supported deployment and the agent must already reach NATS to report any
stage result — a container that can't reach NATS is a distinct, narrower
failure (connectivity) than this epic addresses (a connected container's own
output being unrecoverable). No new functionality goes into the in-process
runner, which is being migrated away.
Subtasks
See linked sub-issues.
Imported from beads archie-core-1786637501090-427-c465db57 · GitHub #450 · type epic · priority P2 · created 2026-08-09 · closed 2026-08-09
Design doc: `docs/architecture/observability.md` (this branch/PR).
## Problem
A task's real work happens in an ephemeral `archie-agent` container over
NATS. Four things currently destroy the evidence of what it did:
1. `cmd/archie-agent/main.go` logs to stderr only, and
`internal/container/pool.go` creates containers with `AutoRemove: true` —
the log dies with the container.
2. `StageBaselineGate` (`internal/domain/workflow/implement.go`) captures
full gate output via `CombinedOutput()`, clips it to 200 chars into a
`Warn` log, and discards the rest — the park reason carries only a status
string.
3. `agentexec.Result` has no step-level transcript (tool calls, durations,
errors).
4. `agentexec.SubjectForSystem` is documented as the return channel for
exactly this ("log dumps, health... the daemon reads these for
observability") but has zero publishers and zero consumers anywhere in
the codebase.
Consequence: a task parked today with
`stage baseline: baseline red -- go build ./... fails and builder could not
auto-fix (status: parked)` and there was no way, from the dashboard or from
chat, to see the actual `go build` error — the container that had it was
already gone.
This matters beyond convenience: archied is meant to run unattended for
weeks (`CLAUDE.md` deployment model) and potentially for users other than
its original operator. A system only debuggable by someone watching
`docker ps` live doesn't meet that bar.
## Decision
NATS is the only transport this targets. `[agent] mode = "nats"` is the
supported deployment and the agent must already reach NATS to report any
stage result — a container that can't reach NATS is a distinct, narrower
failure (connectivity) than this epic addresses (a connected container's own
output being unrecoverable). No new functionality goes into the in-process
runner, which is being migrated away.
## Subtasks
See linked sub-issues.
---
_Imported from beads `archie-core-1786637501090-427-c465db57` · GitHub #450 · type epic · priority P2 · created 2026-08-09 · closed 2026-08-09_
sam
added the mediumepic labels 2026-10-27 13:38:26 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Design doc:
docs/architecture/observability.md(this branch/PR).Problem
A task's real work happens in an ephemeral
archie-agentcontainer overNATS. Four things currently destroy the evidence of what it did:
cmd/archie-agent/main.gologs to stderr only, andinternal/container/pool.gocreates containers withAutoRemove: true—the log dies with the container.
StageBaselineGate(internal/domain/workflow/implement.go) capturesfull gate output via
CombinedOutput(), clips it to 200 chars into aWarnlog, and discards the rest — the park reason carries only a statusstring.
agentexec.Resulthas no step-level transcript (tool calls, durations,errors).
agentexec.SubjectForSystemis documented as the return channel forexactly this ("log dumps, health... the daemon reads these for
observability") but has zero publishers and zero consumers anywhere in
the codebase.
Consequence: a task parked today with
stage baseline: baseline red -- go build ./... fails and builder could not auto-fix (status: parked)and there was no way, from the dashboard or fromchat, to see the actual
go builderror — the container that had it wasalready gone.
This matters beyond convenience: archied is meant to run unattended for
weeks (
CLAUDE.mddeployment model) and potentially for users other thanits original operator. A system only debuggable by someone watching
docker pslive doesn't meet that bar.Decision
NATS is the only transport this targets.
[agent] mode = "nats"is thesupported deployment and the agent must already reach NATS to report any
stage result — a container that can't reach NATS is a distinct, narrower
failure (connectivity) than this epic addresses (a connected container's own
output being unrecoverable). No new functionality goes into the in-process
runner, which is being migrated away.
Subtasks
See linked sub-issues.
Imported from beads
archie-core-1786637501090-427-c465db57· GitHub #450 · type epic · priority P2 · created 2026-08-09 · closed 2026-08-09