OMP parent shows idle/stopped while internal task subagents are still active; nativeHandle points at missing .jsonl

[thatdaveguy1]

Summary

When an OMP-backed Paseo agent fans out work with OMP-internal task subagents, the parent agent flips to idle / UI “stopped” even though the OMP process is still actively writing child session logs.

Paseo’s stored persistence.nativeHandle points at a missing .jsonl file, while the real session is a directory with the same stem containing live child *.jsonl files.

This makes long-running fan-out audits look finished/stopped in the agent list when they are not.

Environment

• Paseo: 0.1.110 (/Applications/Paseo.app)
• OMP: omp/17.0.5
• Provider: omp / model cursor/cursor-grok-4.5-high-fast
• OS: macOS 26.5.2 (Darwin 25.5.0, arm64)
• Daemon: 127.0.0.1:6767
• OMP provider config uses params.sessionDir: ~/.omp/agent/sessions

Repro

1. Create/run an OMP agent that spawns several internal task subagents (fan-out audit / research style).
2. Wait until the parent turn parks on hub waits / child completion.
3. Observe:
   • paseo inspect <parent-id> --json → Status: idle
   • UI shows parent as stopped/idle
4. Meanwhile:
   • omp process for that session is still alive
   • child *.jsonl files under the session directory continue updating
   • lsof -p <omp-pid> shows open writes to those child files

Observed (exact case)

Parent agent:

• ID: 0d43efa4-2b73-4645-bf79-940b8f82be48
• Title: critically audit the entire repo, data feed, api usage, pipe
• Status: idle
• Session ID: 019f7c27-d1fd-7000-918f-fd49e1e66a6a

Stored handle:

~/.omp/agent/sessions/-Antigravity-unusualwhales/2026-07-19T20-53-32-285Z_019f7c27-d1fd-7000-918f-fd49e1e66a6a.jsonl

• exists = false

Actual session:

~/.omp/agent/sessions/-Antigravity-unusualwhales/2026-07-19T20-53-32-285Z_019f7c27-d1fd-7000-918f-fd49e1e66a6a/

• exists as a directory
• contains live child logs such as:
  • ApiBudgetAudit.jsonl
  • PipelineFeedAudit.jsonl
  • Btc5mStratHealth.jsonl
  • LegacyStratHealth.jsonl
  • HarnessPnlAudit.jsonl
  • etc.

At observation time, parent OMP pid 27988 still had open write FDs on multiple child jsonls in that directory.

Contrast: currently healthy running agents still have existing .jsonl nativeHandle files.

Expected

While OMP-internal child task sessions are still active for a parent:

1. Parent should remain running (or some explicit waiting / delegated status), or
2. Child work should be visible in Paseo’s agent/subagent tree, and
3. nativeHandle should resolve to a real session artifact (file or directory), not a missing path.

Actual

• Parent status becomes idle / UI stopped
• Child work continues invisibly under ~/.omp/agent/sessions/.../<session-stem>/
• Paseo list no longer reflects that the job is still in progress
• Attach/logs UX is degraded because the parent looks finished

Impact

• False “done/stopped” signal for multi-agent OMP runs
• Hard to babysit long audits from the agent list
• Easy to interrupt/restart/abandon work that is actually still running
• Confusing freeze triage: parent looks idle while child hub waits continue

Workarounds (no Paseo patch)

1. Watch session-dir mtimes / open FDs:
   • ~/.omp/agent/sessions/.../<session-stem>/*.jsonl
   • lsof -p <omp-pid>
2. Use paseo attach <id> / external watcher for activity
3. Prefer Paseo-level agents (paseo run / handoff / advisor) over OMP-internal task when list visibility matters
4. paseo send can flip the parent back to running briefly, but that does not restore correct child tracking

Suspected gap

Paseo appears to treat OMP parent turn completion / missing parent .jsonl handle as terminal idle, without accounting for OMP’s session-directory + internal task-subagent layout.

Ask

Please consider:

1. Keep parent non-idle while OMP child session files under the session stem are still active
2. Make nativeHandle support OMP session directories (or rewrite handle to the real artifact)
3. Surface OMP-internal task children in the agent tree / status model
4. Or document clearly that OMP-internal task fan-out is intentionally invisible to Paseo status

Happy to provide more logs / a minimal repro script if useful.

[paseo-bot]

Keeping this open as a current OMP bug. On current main, OMP turn completion still depends only on isStreaming and isCompacting, so active internal task subagents are not part of the idle decision (source).

The directly scoped fix in PR #2245 remains open and unmerged, with unresolved event-order and reconciliation cases in its review thread. The merged idle-runtime change in PR #2590 prevents idle garbage collection from terminating background work, but it does not correct the false idle status reported here.

Marking this p2: it is valid and actionable, but limited to OMP internal fan-out workflows with available workarounds.

[tiezbro]

Paseo now integrates seamlessly with Antigravity, with additional optimizations. Repository:https://github.com/tiezbro/paseo-agy-acp

[leadingproblemsolver]

The unresolved event-order cases suggest the parent status probably should not be driven by a single isStreaming || isCompacting predicate plus ad-hoc child checks. I would make "idle" a derived projection over explicit activity evidence.

Something like:

type ActivityEvidence = {
  parentTurn: "active" | "terminal";
  internalChildren: Map<ChildId, "active" | "terminal" | "unknown">;
  providerProcess: "alive" | "dead" | "unknown";
  observedAt: number;
  generation: number;
};

Then define the transition conservatively:

idle is allowed only when:
  parentTurn == terminal
  AND every known child == terminal
  AND there is no newer provider activity contradicting that snapshot

The unknown state matters. A missing parent .jsonl should not become evidence that work is finished when the provider process and child artifacts are still demonstrably live. It should mean "this particular handle is stale/unresolved" until reconciliation establishes the replacement artifact.

For event ordering, attach a monotonically increasing provider/session generation (or equivalent sequence) to observations. Then a late parent-terminal event from generation N cannot overwrite child activity observed in generation N+1.

A useful regression matrix would be:

parent terminal + no children                    => idle
parent terminal + child active                   => waiting/delegated
parent terminal + child unknown + provider alive => reconciling, not idle
late parent-terminal event after newer child activity => ignore for projection
all children terminal + provider quiescent       => idle

I would also separate nativeHandle from the logical session identity. The handle is an observed provider artifact that can change shape (file -> directory); the logical session ID should remain stable while a reconciliation step resolves the current backing artifact.

That keeps the state model from making filesystem representation authoritative. The filesystem is evidence about provider activity; current lifecycle is a projection over that evidence.