← graph

design-wg-single-attempt-kernel

Design the single WG attempt-control kernel

Metadata

Statusopen
Agent identity02e879681e52e0a384106169be043416c4d946e850ab26b2269c57681b52a6e7
Created2026-08-06T08:00:17.174487746+00:00

Description

Objective

Design a replacement WG attempt-control kernel from the deletion audit. The target model has exactly one durable Attempt record/reducer, one attempt supervisor, one provider-neutral canonical event cursor, and one resumable completion transaction. Heartbeats and filesystem/stream observations are evidence only; they cannot independently terminalize a verified-live exact owner. Assignment and quality orchestration are visible daemon audit operations, not blocking user tasks requiring Land candidates. Worker graph coordination is distinct from exact-owner runtime/terminal authority.

Specify the minimal states, events, idempotency keys, crash semantics, process ownership, migration, rollback, and CLI projections. Preserve required behavior—worktree/session/route identity, first-terminal-wins, review before publication, retry-in-place, and repeated qualifying Pi compactions—without preserving obsolete implementation paths. Do not introduce a parallel long-lived controller or indefinite dual write. The cutover must replay recorded traces against old and new projections, switch once, migrate once, and delete the replaced paths in the same program.

Deliver docs/design-wg-single-attempt-control-kernel.md with a state table, crash matrix, deletion list, staged cutover, and measurable complexity budget.

Validation

  • One reducer owns attempt state; all other components are adapters, supervisors, or projections with no independent terminal authority.
  • The design collapses completion to one submitted immutable result and one resumable daemon transaction through review/publication/terminalization.
  • It removes blocking synthetic assignment/quality tasks while retaining visible audit history and actionable diagnostics.
  • It normalizes provider events before shared liveness/failure policy and models each distinct qualifying compaction occurrence independently.
  • The final target has fewer authorities, helper processes, journals, transition APIs, and production LOC than the audited baseline; any non-deletion is explicitly justified.
  • Crash/fault matrix covers response loss, daemon restart, PID reuse, heartbeat loss, duplicate events, multiple compactions, terminal races, and observer exit.

Depends on

Required by

Log