[bborbe]

Summary

session-self-exit.py resolved the anchor task by claude_session_id alone. When a planning or design session hands its task to an implementation session, that field moves to the new session and the original matches no task — so it can never exit, even after the task goes terminal.

This adds a second resolution path: a task where the session appears in metrics_sessions while claude_session_id names a successor. That branch skips the terminal check, because the session's own work is already on the record and the successor's progress is not its proof. The pane-ownership and open-card guards still refuse.

Surfaces

• scripts/session-self-exit.py — read_frontmatter now also returns the raw block; new metrics_session_ids reads the nested list the flat reader cannot see; anchor_task returns a fourth handed_off flag; main branches on it and prints SELF-EXIT: exit — handed off to <sid>.
• scripts/session-self-exit-test.sh — fixtures and cases for a handed-off task while in progress and after completion, the open-card refusal, an unrelated owner, and an id present only in the task body.
• CHANGELOG.md — ## Unreleased bullet.

Verification

make test — green. session-self-exit: 18 passed.

[ben-s-pull-request-reviewer]

{
  "verdict": "approve",
  "summary": "PR adds a correct 'handed-off' exit path to session-self-exit.py: when a session's claude_session_id has moved to a successor, the original session can exit if it appears in the task's metrics_sessions. Implementation is correct, tests are comprehensive, and all four pre-identified concerns are either intentional design or minor test gaps that don't affect correctness.",
  "comments": [],
  "concerns_addressed": [
    {
      "concern": "metrics_session_ids() indentation-based state machine could misinterpret edge cases like empty lines",
      "disposition": "not-an-issue",
      "detail": "metrics_session_ids (py:132-151) correctly handles blank lines: the 'not line.strip()' skip preserves in_block state, so a blank line inside metrics_sessions does not end the section. The state machine is only reset by a non-indented line, which is the correct boundary for a YAML list-of-maps structure."
    },
    {
      "concern": "handed_off branch skips terminal_verdict check entirely — session handed off will exit even if task is in abnormal state",
      "disposition": "not-an-issue",
      "detail": "Intentional design per the docstring and comments: 'A handed-off session skips this guard: its own work is on the record.' Tests confirm the behavior: handed_off_task with status=completed phase=done exits, and open-card guard still applies."
    },
    {
      "concern": "anchor_task returns handed_off=False path when claude_session_id matches, even if session also in metrics_sessions",
      "disposition": "not-an-issue",
      "detail": "Correct behavior: when claude_session_id == sid, the session owns the task as its anchor and should use the normal (non-handed-off) exit path. The handed-off path is specifically for when claude_session_id names a different (successor) session. The test 'an id present only in the task body refuses' confirms body-only appearance is rejected."
    },
    {
      "concern": "handed_off_task fixture only tests execution and done phases, not hold or planning",
      "disposition": "not-an-issue",
      "detail": "The handed_off path skips terminal_verdict entirely (phase is irrelevant), so status=hold with handed_off is not a regression risk. The existing test 'hold is parked, not finished' covers hold for the normal path. The handed_off test gaps are minor and don't affect correctness."
    }
  ]
}