Problem
WP Codebox agent sandbox runs currently bury the useful agent-runtime result inside nested stdout and artifact review files. In the lab fanout run, Homeboy could see wp-codebox/recipe-run/v1 success and an artifact directory, but not a normalized transcript/actionability summary.
Evidence from lab run:
- Fanout run:
/tmp/homeboy-wp-codebox-audit-bwz45ims/fanout-run.json
- Artifact root:
/tmp/homeboy-wp-codebox-artifacts-pr44ovsz
- Example artifact:
/tmp/homeboy-wp-codebox-artifacts-pr44ovsz/runtime-mpohjxp3-c85xcu
files/patch.diff existed but was 0 bytes.
files/changed-files.json existed but represented no useful edits.
- The nested agent runtime result was inside
result.executions[0].stdout -> output -> agent_runtime.result.
- There was no top-level
transcript.json or normalized agent_result.json artifact for downstream tools to consume directly.
Why this matters
Homeboy Extensions and review tools need a stable WP Codebox artifact contract that says whether an agent run was actionable, what changed, and where to find the transcript. Today they have to parse nested stdout strings and inspect artifact internals.
Acceptance criteria
- Agent sandbox runs write a normalized machine-readable agent result artifact, e.g.
files/agent-result.json or files/transcript.json.
- The recipe-run JSON includes normalized summary fields for actionability: changed files count, patch bytes, transcript path, agent completion status, no-op reason, and workspace-tool availability diagnostics.
- Empty patch/no changed files is represented explicitly as a non-actionable result, even if the runtime command itself succeeded.
- Artifact review metadata includes enough detail for downstream orchestrators to distinguish process success from useful edit success.
- Smoke coverage verifies the output shape for a sandbox agent run that returns prose but no edits.
Related
AI assistance
- AI assistance: Yes
- Tool(s): OpenCode (GPT-5.5)
- Used for: Inspected nested WP Codebox lab artifacts and drafted this issue from observed output shapes.
Problem
WP Codebox agent sandbox runs currently bury the useful agent-runtime result inside nested stdout and artifact review files. In the lab fanout run, Homeboy could see
wp-codebox/recipe-run/v1success and an artifact directory, but not a normalized transcript/actionability summary.Evidence from lab run:
/tmp/homeboy-wp-codebox-audit-bwz45ims/fanout-run.json/tmp/homeboy-wp-codebox-artifacts-pr44ovsz/tmp/homeboy-wp-codebox-artifacts-pr44ovsz/runtime-mpohjxp3-c85xcufiles/patch.diffexisted but was 0 bytes.files/changed-files.jsonexisted but represented no useful edits.result.executions[0].stdout -> output -> agent_runtime.result.transcript.jsonor normalizedagent_result.jsonartifact for downstream tools to consume directly.Why this matters
Homeboy Extensions and review tools need a stable WP Codebox artifact contract that says whether an agent run was actionable, what changed, and where to find the transcript. Today they have to parse nested stdout strings and inspect artifact internals.
Acceptance criteria
files/agent-result.jsonorfiles/transcript.json.Related
AI assistance