Optional transcript locator on agent_session

[zhiyao]

Version: herdr 0.9.3, protocol 22.

Apps that show an agent's conversation, such as mobile companions, have to work out where each agent keeps its transcript and resolve the session ID to a file themselves. Integrations often already have that path:

• Claude's and Codex's SessionStart input both include transcript_path.
• Herdr's hook scripts receive it and then drop it.

Proposal

An optional field on pane.report_agent_session and in agent_session:

"transcript": { "path": "/abs/path/to/session.jsonl", "format": "codex-rollout-jsonl" }

• Herdr would store it, but not read or check the file.
• Consumers would still have to keep the path under the agent's own root and match it to the session ID, so a wrong or hostile value can't widen what they read.
• Format tags would be stable strings owned by each integration, for example claude-jsonl or codex-rollout-jsonl.
• Integrations that don't know the path just omit the field, so nothing changes for them.

Why in Herdr

Herdr already gets the session ID from these hooks. Recording the path the agent itself reported would replace a separate, per-agent path guess in every consumer. It would also help agents whose layout isn't documented.

This is separate from the shared-daemon misattribution in #4649. The locator is only as trustworthy as the session report that carries it.

[sebastianst]

Another concrete case for this: herdr-annotate's "review the agent's last reply" (plannotator-tui herdr last) gets the session ID from herdr. It then looks for the transcript under its own CLAUDE_CONFIG_DIR, or ~/.claude without one. It runs under the herdr server, not in the agent's environment. So for any Claude session started with another CLAUDE_CONFIG_DIR (I use ~/.claude and ~/.claude-max side by side), it fails with no Claude Code session <id> in ~/.claude/projects and falls back to the pane's screen. With the path the hook already receives, it could open the right file directly.