Harness → AgentPlaybookRun alignment

[savitharaghunathan]

Harness → AgentPlaybookRun alignment

Tracking issue for aligning the harness (merged in PR #33) with the AgentPlaybookRun execution model defined in ADR-0001.

The current harness is a full migration CLI with a hardcoded 5-step pipeline. The target state is a thin git-lifecycle wrapper that runs a single stage — pipeline orchestration moves to the AgentPlaybookRun controller, skills move to standalone SkillCard files, and each stage runs in its own container image with exactly one skill.

Subtasks

• 5.1: Slim harness to single-stage git-lifecycle runner #38 — 5.1: Slim harness to single-stage git-lifecycle runner
• 5.2: Extract pipeline steps into standalone SkillCard files #39 — 5.2: Extract pipeline steps into standalone SkillCard files
• 5.3: Per-stage container images (plan, execute, verify) #40 — 5.3: Per-stage container images (plan, execute, verify)
• 5.4: Filesystem watcher for incremental commit+push #41 — 5.4: Filesystem watcher for incremental commit+push
• 5.5: AgentPlaybookRun controller #42 — 5.5: AgentPlaybookRun controller
• 5.6: .konveyor/ file contract (result.json, handoff.md) #43 — 5.6: .konveyor/ file contract (result.json, handoff.md)
• 5.7: OCI packaging via skillimage #44 — 5.7: OCI packaging via skillimage
• 5.8: End-to-end playbook test with Kind #45 — 5.8: End-to-end playbook test with Kind
• 5.9: Hub API integration in harness #61 — 5.9: Hub API integration in harness

Dependencies

• 5.1: Slim harness to single-stage git-lifecycle runner #38, 5.2: Extract pipeline steps into standalone SkillCard files #39, 5.3: Per-stage container images (plan, execute, verify) #40 can proceed in parallel (harness slim-down, skill extraction, images)
• 5.4: Filesystem watcher for incremental commit+push #41 (watcher) depends on 5.1: Slim harness to single-stage git-lifecycle runner #38 (slim harness to integrate into)
• 5.5: AgentPlaybookRun controller #42 (controller) depends on 5.6: .konveyor/ file contract (result.json, handoff.md) #43 (file contract must be defined first)
• 5.7: OCI packaging via skillimage #44 (OCI packaging) depends on 5.2: Extract pipeline steps into standalone SkillCard files #39 (skills must exist first)
• 5.9: Hub API integration in harness #61 (hub integration) depends on 5.1: Slim harness to single-stage git-lifecycle runner #38 (slim harness)
• 5.8: End-to-end playbook test with Kind #45 (e2e test) depends on all of the above

References

• ADR-0001: Agentic Platform CRD Architecture
• ADR-0002: ACP Transport and Agent Observability
• PR ✨ Harness init #33: Harness init (baseline being merged)

[ibolton336]

Empirical data point for the alignment work: I ran the merged #36 controllers E2E today on minikube (controller @ fff23af) with a three-stage playbook (assess → remediate → validate) of goose-based agents against Bedrock, sharing a git branch per the #36 param convention. Once the stages ran to completion, orchestration worked exactly as designed — sequential execution, per-stage status, 9m08s wall clock, all stages Succeeded, each stage's output committed to the shared branch.

Two observations relevant to this issue's direction:

1. Run-to-completion is the load-bearing requirement. An ACP-serving harness (goose serve, per the current interactive convention) never exits, so any playbook composed of such agents hangs on stage 1 indefinitely — stage success is only observable via pod exit 0 → PodSucceeded. The 5.1: Slim harness to single-stage git-lifecycle runner #38 single-stage-runner direction resolves this structurally. As an interim shim I added a mode=batch param to my goose entrypoint (goose run with the stage instructions from KONVEYOR_INSTRUCTIONS, then exit with goose's rc), which was sufficient for the full E2E — happy to share if useful as a stopgap pattern before 5.1: Slim harness to single-stage git-lifecycle runner #38 lands.

2. KONVEYOR_PLAYBOOK_INSTRUCTIONS consumption works as intended: folding the Guide into the agent's context (goosehints) alongside the Agent prompt and stage instructions gave each stage correct ambient awareness of the pipeline ("stage N of 3, shared branch, push before finishing") with the controller staying out of prompt composition — nice separation as designed in ✨ Add AgentPlaybook and AgentPlaybookRun controllers #36.

Related findings filed separately: #51 (failing stage crashloops under restartPolicy: OnFailure instead of failing the run), #52 (param forwarding forces cross-stage param declarations).

[savitharaghunathan]

Will create a separate issue for the migration skills