Summary

Closes the gap CC#2 caught via STOP-ASK on KR-REASONING-PANEL-EMAIL-XREF: `email_outbound_log.jsonl` was missing the model_used / tokens / duration / reasoning_error / caller_session_id fields that slack_dm's outbound has had since PR #131. Mirrors that exact pattern for the email path.

Bucket spec: `17_cc_bucket_prompts/KR-EMAIL-OUTBOUND-REASONING-META_extension.md`

Reference PRs:

• feat(kora): KR-FEAT-AI-RESPONSE-LOOP ST2 — handler swap + cred priority flip + anthropic runtime dep (re-opened post #129) #131 — slack_dm pre-/post-pattern (the model being mirrored)
• feat(kora): KR-FEAT-EMAIL ST1 — Purelymail outbound via SMTP #124 — original PurelymailClient + outbound JSONL writer
• feat(kora): KR-FEAT-EMAIL-INBOUND-IMAP ST2 — EmailInboundHandler + AUTO_REPLY #138 — EmailInboundHandler + AUTO_REPLY plumbing

Surface (additive + backwards-compatible across 3 files)

caller_session_id format — documented + enforced consistent

Per CC#2's spec proposal: `f"email:{message_id}"`. This matches the reasoning engine's own derivation at `kora_cli/reasoning/anthropic_engine.py:869-871` (in `_derive_caller_session_id` for `source="email"`). Both sides on the same literal string lets KR-REASONING-PANEL-EMAIL-XREF join audit ↔ outbound JSONL rows by a single field.

Test `test_caller_session_id_matches_engine_derivation_shape` proves this: it imports the engine's private helper, constructs an `IncomingMessage` exactly the way the handler does, and asserts the handler's derived `caller_session_id` equals the engine's output for the same `ParsedIncomingEmail`.

§1 K-DG verification at HEAD `b959737b`

All 3 line-refs from the bucket spec still accurate:

• `purelymail_client.py:521` `_append_outbound_log()` ✓
• `email_inbound_handler.py:611` `client.send_email()` call site ✓
• `email_inbound_handler.py:648` `_call_reasoning_engine()` ✓

JSONL contract (post-bucket)

Reasoning-driven send entry shape:
```json
{
"sent_at": "...", "from": "...", "to": [...],
"subject": "...", "in_reply_to": "...",
"send_status": "ok", "message_id": "...",
"smtp_code": 250, "error": null, "retry_count": 0,
"caller_actor_kind": null,
"model_used": "claude-opus-4-7",
"input_tokens": 100, "output_tokens": 50,
"reasoning_duration_ms": 1500,
"caller_session_id": "email:msg-1@example.com"
}
```

Non-reasoning send (MCP `kora__send_email` tool / notification) entry: identical to pre-bucket — the 6 reasoning-meta keys are simply absent (opt-in inclusion).

Test coverage

Handler side (7 new tests):

• Happy path: full meta threaded (model + tokens + duration + caller_session_id; reasoning_error=None)
• Engine unavailable: reasoning_error="engine_unavailable", SDK fields None, caller_session_id still derived
• Engine raises: reasoning_error="engine_exception:RuntimeError"
• Engine result.error set (e.g., cost_ladder_halted): SDK meta preserved + error code passed through
• Engine empty text: handler overrides reasoning_error to "empty_response_text" while preserving SDK meta
• caller_session_id format matches engine's `_derive_caller_session_id`
• AUTO_REPLY off: meta threading is gated (send_email never called)

Client side (5 new tests):

• Non-reasoning send: NONE of the 6 new keys appear in JSONL
• Full reasoning send: all 6 keys land (with opt-in for reasoning_error=None)
• Partial meta (engine-unavailable): only reasoning_error + caller_session_id appear
• SMTP failure with reasoning meta: meta still recorded — useful for the panel to show "we tried this model + spent these tokens, then SMTP rejected"
• `send_email_internal` forwards kwargs

Test plan

• 12 new tests pass
• 80 existing tests in touched modules still pass (backwards-compat)
• 505/505 cross-bucket regression (clients + handlers + test_listeners + reasoning)
• Ruff clean

Cascade

Unblocks CC#2's KR-REASONING-PANEL-EMAIL-XREF — the panel can now project the new fields into the same xref shape it uses for slack_dm.

🤖 Generated with Claude Code