[Siddhant-K-code]

Follow-up to #294. Codex fires a subagent's hooks under the parent session_id. Because of that, a subagent's tool calls could not be told apart from the parent's, and the trace did not record what the subagent returned.

Recording (hooks.py)

• For providers that keep subagents in-session (everything outside _SUBAGENT_SESSION_PROVIDERS), tool_call, tool_result and error events get agent_id and agent_type when the payload has an agent_id. Copilot events are unchanged.
• subagent_stop records last_assistant_message as text and also records agent_transcript_path, the same way handle_stop does, and redaction applies. stop_hook_active does not suppress the event.
• Pending-call pairing still matches on tool_use_id first. When there is no tool_use_id, it prefers a pending call from the same agent, then the oldest one.

Rendering

Attribution only appears when agent_id is present, so unattributed traces render exactly as before. I checked this byte-for-byte against main, with and without color.

Command	Change
replay	[subagent <type> <short id>] tag on tool_call/tool_result/error; agent=<id> on subagent_start/stop; returned: "…" line on subagent_stop
replay --expand-subagents (event tree)	same tags; agent=<short id> on start/stop; returned: line
replay --tree, tree (summary)	└─ subagent <type> <short id> N tools ✗ N errors line under the session; tree --json gains subagents only when present
stats	Subagent Tool Calls: section, only when a subagent made tool calls
explain	Subagent: <type> <short id> N tools line per phase

The short ID is the last 8 characters, because Codex agent IDs are UUIDv7 and share a timestamp prefix with the session ID.

Synthetic before/after for replay:

before:
+  1.00s ⤷ subagent_start default
+  2.00s → tool_call Bash
                $ echo synthetic
+  4.00s ⤶ subagent_stop default

after:
+  1.00s ⤷ subagent_start default agent=d0000011-0000-4000-8000-000000000011
+  2.00s → tool_call Bash [subagent default 00000011]
                $ echo synthetic
+  4.00s ⤶ subagent_stop default agent=d0000011-0000-4000-8000-000000000011
                returned: "Synthetic subagent answer."

tree:

<session>  openai-codex  $0.0012  3 tools  4.0s  ✓
└─ subagent default 00000011  2 tools  ✗ 1 errors

Final-answer audit

A subagent_stop now carries text, so I checked every consumer that reads text or a final message to make sure none of them takes this as the session's answer:

• Filter on event type, not affected: explain, review, postmortem, why, timeline, checkpoint/compaction, OTLP and GenAI export, Langfuse, share, compare, freeze, eval scorers, cost, compliance, watch, the web dashboard and the VS Code panel. They read assistant_response or llm_response.
• Fixed: disclosure previously undercounted. It now counts a subagent's returned text as a response and as content-rich.

A test checks that the checkpoint, OTLP, GenAI and timeline final answer stays the same when a subagent_stop with text comes after the session's answer.

Docs

• capture_matrix.json and docs/capture-matrix.md are updated. The remaining Codex gap: no event links a subagent to the spawn_agent call that started it.
• docs/integrations.md and docs/commands.md are also updated.

Tests

New or extended tests cover:

• tool events with and without agent_id
• agent_type without agent_id is ignored
• the SubagentStop fields, stop_hook_active, and both store-level and hook-level redaction
• pairing: interleaved calls, no tool_use_id, and oldest-call fallback
• Copilot is unchanged
• capture-matrix fixture assertions
• replay, tree, summary, JSON, stats and explain output for both attributed and unattributed traces
• disclosure counts, and that the final answer is unchanged

Full suite: 2482 tests OK (2 skipped). I also ran a live check with codex-cli 0.163.0-alpha.11: one subagent ran one shell command in a read-only sandbox, and its tool events carried agent_id/agent_type and rendered as attributed. subagent_stop had text and agent_transcript_path.