2.3 KiB
| created | path | project | tags | type | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 2026-06-04 | Sources/Dev | petal-dispatch-v3 |
|
session-notes |
Phase −1 of the v3 build ran the OAuth viability spike on 2026-06-04 and surfaced findings that materially patch the v3 spec. v3 is immutable; this note is the consumption layer for future readers.
Full results live in trellis thread petal-dispatch-v3-build (id 588), checkpoint event 1084. Raw artifacts in repo petal-power/petal-dispatch (currently petal-dispatch-demo) under phase-minus-1/.
Corrections to v3
D-11 (OS exit code mapping): drop the SIGINT/130 row. Measurement: claude under stream-json treats SIGINT-when-busy as a clean exit; rc=130 is unreachable. The orchestrator's cancel path is SIGTERM (rc=143 → cancelled), the crash path is SIGKILL (rc=137 → crashed). SIGINT plays no role in v1. The rest of the v3 mapping (0 → success, nonzero → failed, 143 → cancelled, 137 → crashed) holds; the SIGINT row was conjectural.
Additions to internalize when reading v3
Init event exposes mcp_servers[*].status (per-spawn). D-21 plumbing: the harness can detect needs-auth MCPs at spawn time and route around them instead of crashing on an interactive prompt. The Phase 0 event taxonomy should treat MCP-status as a first-class signal in init events.
system init fires per user_message in long-lived stream-json mode. Not a one-time event. The orchestrator state machine and the journal's event-type reference table must accept >1 init event per process lifetime. Relevant to Phase 0 and any forward-compat planning for long-lived workers (v4 lifecycle).
rate_limit_event is a status-only signal. Payload reports status (allowed / not) with no remaining-quota gauge. Phase 2's not-clean-return policy for rate_limited (FR-009 / D-18 / Phase 2) must back off on status != allowed rather than on a quota threshold. Forecasting headroom from this signal is not possible; the orchestrator only sees the binary circuit-breaker state.
What did NOT change
D-21 (extractor as Claude Code spawn), D-13 (two-service split), Phase −1's OAuth decision, and the harness-as-sole-substrate-touching-layer containment argument all hold under measurement. Six of seven Phase −1 tests passed cleanly; the one PARTIAL was the SIGINT row above.