The journal records no sub-agent events, so a delegating run cannot be replayed

[M3gA-Mind]

Summary

The durable journal at {workspace}/tinyagents_store/journal carries no sub-agent events at all, and no link from a child run to its parent. A delegating turn therefore cannot be reconstructed from the journal: the delegate's turn span, its iterations, and its tool and model spans are all absent on replay.

Split out of #6419, where the symptom surfaces as a [agent-tracing][journal-shadow] parity divergence warning. PR #6520 makes that warning specific and corrects a doc comment that described the gap wrongly; it does not close this. This is the actual defect.

Evidence

Surveyed the real journal on one developer machine — 871 MB, 6,970 run files.

Zero sub-agent events, across a random sample of 900 run journals. Every event kind that does occur:

count	kind
51,985	middleware_started
51,985	middleware_completed
14,834	model_delta
1,006	budget_reserved
895	run_started
755	tools_advertised
748	tools_filtered
592	run_failed
568	middleware_failed
439	model_started
418	model_completed
406	usage_recorded
282	run_completed
203	budget_reconciled
154	tool_started
154	tool_completed
	plus a tail of retry_scheduled, handoff_transform_applied, steered, control_applied, unknown_tool_call, limit_reached, invalid_tool_args

Sub-agent events: 0. Journals containing any: 0 of 900.

Meanwhile crates/openhuman-core/src/agent/progress_tracing/journal_projection.rs carries live match arms for AgentEvent::SubAgentStarted, SubAgentCompleted, SubAgentFailed and SubAgentReused. They are dead in production — a consumer waiting on a producer that does not exist.

And the subtree is not reachable indirectly. Every journalled run is its own root:

• root_run_id == run_id in all 588 run_started records sampled, 0 exceptions.
• No root_run_id spans more than one run file, across 1,200 files.
• No run file contains more than one distinct run_id.

So a child run is neither recorded in its parent's journal nor linked from it, and read_run_events(run_id, offset) on the parent cannot reach it by any route.

Why this matters more than the warning

The journal is the only durable record that captures runs when nothing is observing. Langfuse is fail-closed outside staging (LANGFUSE_PUSH_ENVIRONMENTS = &["staging", "development"], and the backend answers 403 FEATURE_DISABLED outside staging), and only the web progress bridge installs a live collector at all — so orchestration, subconscious, cron and meeting turns are untraced live.

That makes the journal the thing you would read to answer any question about a past turn. It is not ground truth for any delegating turn, and delegation is most of what the orchestrator does. Any eval, replay or investigation built on it silently under-reports delegated work — silently being the problem, since the record looks complete.

What would close this

Either is sufficient; the first is better.

1. Emit sub-agent lifecycle events into the journal — the projection already handles all four kinds, so the spans appear the moment the events do, with no host change.
2. Populate root_run_id on a child run so a delegating run's tree is discoverable, and have the replay read the tree rather than one run id.

Provider/harness run instrumentation is owned by vendor/tinyagents, so this is an upstream change plus a gitlink bump — filed here per the contribution workflow.

Acceptance criteria

• A delegating run's journal yields sub-agent spans on replay, or the child run is discoverable from the parent's root_run_id.
• Asserted against a journal produced by a real delegating turn, not a hand-built fixture — a fixture encoding the event shape we assume is exactly what would go green while production stayed empty.
• The "Known parity gaps" list in journal_projection.rs is updated when the gap closes (PR fix(agent-tracing): say what the journal-shadow parity check found, and correct the comment that denied it #6520 documents it as open).

Related

• agent-tracing journal-shadow parity divergence on most turns (journal records ~1/3 of live spans) #6419 — the parity-divergence symptom, and PR fix(agent-tracing): say what the journal-shadow parity check found, and correct the comment that denied it #6520 which makes the warning specific without closing this.
• Two attribute-level gaps remain documented separately in journal_projection.rs and survive this fix: cost is a journal-only estimate, and SubagentAwaitingUser has no journal source.

[tinysweeper]

The journal records no sub-agent events, so a delegating run cannot be replayed.

Labelled priority: p1.