feat: model Codex subagent workflows instead of showing raw subagent_notification JSON

[tpn]

Summary

Codex multi-agent sessions should be modeled as real subagent relationships instead of surfacing raw <subagent_notification> JSON blobs as ordinary user messages in the transcript.

Problem

Codex sessions currently record subagent activity as a sequence like:

1. spawn_agent function call from the parent session
2. function_call_output containing the spawned agent_id
3. wait function call for that agent_id
4. function_call_output containing status keyed by agent_id
5. a synthetic user response_item whose content is wrapped in <subagent_notification>...</subagent_notification> and contains JSON like:
   • agent_id
   • status.completed / status.errored

Today, agentsview does not interpret that Codex-specific sequence as a subagent relationship. The visible symptom is that long raw JSON/status payloads show up in the User lane of the transcript.

That is semantically wrong in two ways:

• the notification is not a real user-authored message
• the spawned child agent is not linked into the existing subagent UI path

Current State

A stopgap local fix can hide <subagent_notification> rows as system/intermediate messages, but that is only a presentation band-aid.

The proper product behavior should reuse the existing subagent/session-linking model already used for other agent types:

• tool_calls.subagent_session_id
• relationship_type = "subagent"
• inline child transcript rendering from the originating tool call
• session tree grouping under the parent

Desired Behavior

For Codex sessions:

1. Detect spawn_agent as a subagent-spawning tool call.
2. Extract the returned agent_id from the matching function_call_output.
3. Resolve that agent_id to the spawned Codex child session.
4. Annotate the originating tool call with subagent_session_id.
5. Reuse the existing transcript/sidebar subagent UI so the child session can be expanded inline and grouped correctly.
6. Treat <subagent_notification> messages as system/intermediate bookkeeping:
   • either suppress them from the normal transcript
   • or fold their final status into tool result/status presentation instead of showing raw JSON

Implementation Notes

Likely work items:

• extend Codex parsing to understand the spawn_agent / function_call_output / wait / function_call_output sequence
• add Codex-specific mapping from spawned agent_id to child Codex session ID
• annotate parsed tool calls with subagent_session_id
• avoid storing or rendering raw <subagent_notification> payloads as normal user transcript content
• ensure incremental sync does not lose Codex subagent mapping when the linking events arrive across sync boundaries

Acceptance Criteria

• raw <subagent_notification> JSON no longer appears as a user message in Codex transcripts
• a parent Codex spawn_agent call links to the actual child Codex session when the child session exists locally
• Codex child sessions participate in the existing subagent UI and sidebar grouping
• behavior works for both full parse and incremental sync
• regression tests cover:
  • raw notification suppression/normalization
  • spawn_agent to child-session linkage
  • incremental append cases where the linking events land after an earlier sync

Related Issues

• feat: Sub-agent & team tree view in sidebar #136 introduced the generic subagent/team tree UI
• Incremental sync: DAG boundary validation, ordinal tracking, and atomicity #162 already tracks incremental-sync edge cases around subagent linkage for Claude; Codex likely needs parallel treatment for its own event model

[tpn]

Planning a PR follow-up for this next. I fetched the latest upstream/main and set up an isolated branch/worktree for the implementation work so the Codex subagent/session-linking changes can land cleanly.

[tpn]

PR is up now: #242