Summary

Operator R3 Q8a: "I email Kora something I found that I want to save as an idea on a Sea_Ticket." First real product-value use of Kora's inbound-email surface — flips the email path from receive-and-discard (post #173) to receive-recognize-write-confirm.

Joshua emails Kora → identity filter (already gates to KORA_EMAIL_JOSHUA_ADDRESS per the existing handler) → regex intent recognition → sea__create_ticket via in-process IsoKron MCP client → Slack confirmation DM to KORA_SLACK_JOSHUA_USER_ID → audit row on every evaluated email.

Bucket spec: 17_cc_bucket_prompts/KR-INTENT-EMAIL-TO-SEA-TICKET_save_idea_from_inbox.md.

K-DG verdict: no STOP-ASK

The Sea_Ticket write surface IS accessible in-process. mcp_tools.py:_execute_create_sea_ticket (Tool 8 kora__create_sea_ticket) uses get_last_active_provider()._connection.get_mcp_client().invoke("sea__create_ticket", payload). The intent module calls that same path directly (sidesteps the MCP-Tool Caller auth shape — this is operator-driven from the handler, not external-caller-gated).

Regex patterns shipped

All case-insensitive. Highest-confidence match wins; first non-matching tier falls through.

Sample DM confirmation text

Success — sent to KORA_SLACK_JOSHUA_USER_ID after a successful write:

:seedling: Got it — saved *Idea: try the new Haiku model* as Sea_Ticket `#STK-42` (matched `subject_idea_prefix`).

Cap-exceeded / failure variant:

:warning: Could not save *Idea: one too many* as Sea_Ticket (matched `subject_idea_prefix`): hourly cap of 10 Sea_Tickets/hour reached; raise KORA_EMAIL_INTENT_HOURLY_CAP or wait

Sample audit entry

{
  "emitted_at": "2026-05-23T19:42:11.512000+00:00",
  "seam": "intent.email_to_sea_ticket",
  "details": {
    "action": "created",
    "pattern_matched": "subject_idea_prefix",
    "confidence": "high",
    "subject": "Idea: try the new Haiku model",
    "ticket_id": "STK-42",
    "tags": ["email", "idea"]
  },
  "caller_session_id": "email:<m-1@example.com>",
  "source": "email"
}

action ∈ {created, dry_run, logged_only, cap_exceeded, failed}. The caller_session_id is wire-compatible with the engine-side email:{message_id} derivation so a future xref bucket can join inbound JSONL ↔ audit ↔ Sea_Ticket history by one key.

Env vars added

Files

• NEW kora_cli/intent/__init__.py + kora_cli/intent/email_to_sea_ticket.py (~600 lines)
• NEW tests/kora_cli/intent/__init__.py + tests/kora_cli/intent/test_email_to_sea_ticket.py (34 tests)
• MOD kora_cli/audit/jsonl_sink.py — added intent.email_to_sea_ticket to SeamName Literal
• MOD kora_cli/handlers/email_inbound_handler.py — call process_email_intent after _emit_received_event (wrapped in try/except so the intent path can never disturb HANDLED_RECEIVED)

Test plan

• 34 intent-module tests pass: regex pattern matrix + orchestrator branches (created / dry_run / logged_only / cap_exceeded / failed) + hourly-cap behavior + fail-soft (DM failure doesn't block create; orchestrator swallows unexpected exceptions)
• Regression: 667 passed (intent + handlers + audit + clients + test_listeners)
• ruff check clean on all changed files

CC#2 follow-on recommendation: KR-FE-EMAIL-INTENT-LOG-PANEL

The audit-JSONL seam is forensically rich enough today to drive a small cockpit panel without further backend work. Recommended cut:

1. Read endpoint: /api/audit-events?seam=intent.email_to_sea_ticket — already serviceable by the existing read_audit_entries(seam=...) reader (PR feat(kora): KR-PROBE-AUDIT-AND-CONVERT — cheap-cron + wake-event + fix-envelope per probe #163 era). No new backend code needed.
2. Panel rows: per audit row, render the subject, pattern_matched, action, and a deep-link to the substrate Sea_Ticket when ticket_id is present.
3. Filtering: action chip filter (created / dry_run / logged_only / cap_exceeded / failed) is the operator's primary lens — logged_only is the promotion-loop training-data view; failed is the triage view.
4. Optional sparkline: per-day count of created actions, hosted next to the recent_error_count_5min band in HealthHero post-KR-SNAPSHOT-DAEMON-HEALTH — daemon's own listeners/uptime/errors (schema v4) #170.

The panel can ship before any cockpit-side substrate read — the audit JSONL already carries the operator's full mental model (subject + pattern + outcome).

🤖 Generated with Claude Code