101 lines
4.7 KiB
JSON
101 lines
4.7 KiB
JSON
{
|
|
"incident": "multi_tool_use.parallel subagent completion did not wake idle orchestrator",
|
|
"observedAt": "2026-09-03T15:54:57-04:00",
|
|
"environment": {
|
|
"piVersion": "0.84.4",
|
|
"repositorySubagentPackageVersion": "0.1.0",
|
|
"note": "The failing tool surface was the harness-provided functions.subagent tool wrapped by multi_tool_use.parallel. It has not been proven to be implemented by the repository @jbcode/pi-subagents package.",
|
|
"parentSessionId": "01a06885-7f00-7cff-bc2f-334984151941",
|
|
"parentSessionFile": "/home/jb/.pi/agent/sessions/--srv-data-code-jb-pi-harness-tools--/2026-09-03T18-26-14-400Z_01a06885-7f00-7cff-bc2f-334984151941.jsonl",
|
|
"provider": "openai-codex",
|
|
"model": "gpt-5.6-sol",
|
|
"reasoning": "high"
|
|
},
|
|
"failingCall": {
|
|
"tool": "multi_tool_use.parallel",
|
|
"payload": {
|
|
"tool_uses": [
|
|
{
|
|
"recipient_name": "functions.subagent",
|
|
"parameters": {
|
|
"agent": "codebase-analyzer",
|
|
"name": "workflow-architecture",
|
|
"cwd": "/srv/data/code/jb-pi/harness-tools",
|
|
"task": "Read-only research. Analyze the current extensions/workflow and extensions/_workflows architecture end-to-end. Focus on what behavior is rigidly enforced, where failures/error-proneness arise, state/recovery mechanics, human gates, review gates, and extension complexity. Read relevant tests and docs. Return concrete findings with file paths and a proposed split of responsibilities between skills, minimal extensions, and ordinary repo instructions. Do not edit files."
|
|
}
|
|
},
|
|
{
|
|
"recipient_name": "functions.subagent",
|
|
"parameters": {
|
|
"agent": "precedent-locator",
|
|
"name": "workflow-history",
|
|
"cwd": "/srv/data/code/jb-pi/harness-tools",
|
|
"task": "Read-only research. Investigate git history and docs/ADR/workflow artifacts related to the workflow extension. Identify original goals, decisions, known usability issues, failure modes, later fixes, and lessons that should shape a replacement workflow. Return commits/docs/files and concise lessons. Do not edit files."
|
|
}
|
|
},
|
|
{
|
|
"recipient_name": "functions.subagent",
|
|
"parameters": {
|
|
"agent": "codebase-pattern-finder",
|
|
"name": "workflow-patterns",
|
|
"cwd": "/srv/data/code/jb-pi/harness-tools",
|
|
"task": "Read-only research. Find existing skills, extensions, tools, and agent instructions in this repo that already support planning, task tracking, subagent delegation, independent review, human gates, verification, worktrees, continuation, and recovery. Identify reusable pieces and missing capabilities. Return concrete paths and recommended composable patterns. Do not edit files."
|
|
}
|
|
}
|
|
]
|
|
}
|
|
},
|
|
"launchResults": [
|
|
{
|
|
"name": "workflow-architecture",
|
|
"result": "launched",
|
|
"childId": null,
|
|
"childSessionId": null
|
|
},
|
|
{
|
|
"name": "workflow-history",
|
|
"result": "launched",
|
|
"childId": null,
|
|
"childSessionId": null
|
|
},
|
|
{
|
|
"name": "workflow-patterns",
|
|
"result": "launched",
|
|
"childId": null,
|
|
"childSessionId": null
|
|
}
|
|
],
|
|
"expectedLifecycle": [
|
|
"wrapper returns one prompt acknowledgement per child without waiting",
|
|
"each child reaches one terminal state",
|
|
"each terminal result is queued exactly once for the parent",
|
|
"an idle parent is woken for each queued result",
|
|
"UI and tool-visible child state agree",
|
|
"late subagent_message resumes a completed child rather than steering a running child"
|
|
],
|
|
"observedLifecycle": [
|
|
"all three wrapped launches returned successful acknowledgements",
|
|
"the parent continued independent work and then became idle",
|
|
"the UI later appeared to show all three children completed",
|
|
"no terminal result woke the idle parent",
|
|
"the user manually woke the parent and reported completion",
|
|
"subagent_message for each child responded that the message was delivered to a running subagent"
|
|
],
|
|
"controlCase": {
|
|
"tool": "functions.subagent",
|
|
"wrapped": false,
|
|
"name": "skill-workflow-plan-review",
|
|
"cwd": "/srv/data/code/jb-pi/harness-tools/.worktrees/skill-led-development-workflow-plan",
|
|
"result": "automatic completion delivered",
|
|
"elapsed": "14m 9s"
|
|
},
|
|
"missingEvidence": [
|
|
"exact launch acknowledgement timestamps",
|
|
"internal child task IDs",
|
|
"internal child session IDs",
|
|
"terminal-state timestamps",
|
|
"callback queue entries"
|
|
],
|
|
"reproductionRequirement": "A focused fixture must capture the missing internal evidence from instrumented execution rather than infer it from UI state."
|
|
}
|