design-simplified-authoritative — WorksGood live mirror

Metadata

Status	done
Assigned	`agent-875`
Agent identity	`603daf069cd2ae4a26357aa66e8c6124154b648711a0816ffd8b660ee3b23815`
Created	2026-07-26T10:23:22.725165512+00:00
Started	2026-07-26T11:09:32.371732769+00:00
Completed	2026-07-26T11:39:51.394620626+00:00
Tokens	2563382 in / 28058 out

Description

Objective

Turn the completed lifecycle/coordinator deep survey into a ratified simplification design. Reduce the number of transition authorities and eliminate implicit task reopening/resume/failure behavior while preserving durable recovery, attempts, worktrees, messages, and optional evaluation.

Required decisions

Define one authoritative transition function/event ledger and which actors may request each transition.
Messages are data by default: they must not change task status/liveness/eligibility unless the task explicitly waits on a message condition.
Separate task state, attempt state, worker process state, resource admission, provider health, worktree merge state, and evaluation evidence; never encode one as another.
Make resource backpressure a deferral; process launch/runtime failure a failed attempt; evaluator failure evaluation evidence—not source execution failure.
Specify idempotent daemon restart/replay/reconciliation and exact ownership/fencing rules.
Decide which auto-repair/rescue/scaffold mechanisms to remove, retain, or make explicit operator commands.
Define lazy evaluation records/satellites and advisory vs hard-gate policy.
Provide backward-compatible migration from existing graph events and ambiguous historical states.
Define operator-visible diagnostics and a model-based transition conformance suite.

Produce docs/design-simplified-task-lifecycle.md; no production code changes.

Validation

Every current failure class has a deterministic state/transition in the proposed model.
No ordinary message can reopen or keep alive a task.
Number of status writers/transition authorities is explicitly reduced.
Crash, restart, retry, worktree, evaluation, and admission semantics are unambiguous.
Migration and staged implementation can land without a flag-day graph rewrite.
Acceptance brief identifies permanent smoke/model-based scenarios.

## Objective
Turn the completed lifecycle/coordinator deep survey into a ratified simplification design. Reduce the number of transition authorities and eliminate implicit task reopening/resume/failure behavior while preserving durable recovery, attempts, worktrees, messages, and optional evaluation.

## Required decisions
- Define one authoritative transition function/event ledger and which actors may request each transition.
- Messages are data by default: they must not change task status/liveness/eligibility unless the task explicitly waits on a message condition.
- Separate task state, attempt state, worker process state, resource admission, provider health, worktree merge state, and evaluation evidence; never encode one as another.
- Make resource backpressure a deferral; process launch/runtime failure a failed attempt; evaluator failure evaluation evidence—not source execution failure.
- Specify idempotent daemon restart/replay/reconciliation and exact ownership/fencing rules.
- Decide which auto-repair/rescue/scaffold mechanisms to remove, retain, or make explicit operator commands.
- Define lazy evaluation records/satellites and advisory vs hard-gate policy.
- Provide backward-compatible migration from existing graph events and ambiguous historical states.
- Define operator-visible diagnostics and a model-based transition conformance suite.

Produce `docs/design-simplified-task-lifecycle.md`; no production code changes.

## Validation
- [ ] Every current failure class has a deterministic state/transition in the proposed model.
- [ ] No ordinary message can reopen or keep alive a task.
- [ ] Number of status writers/transition authorities is explicitly reduced.
- [ ] Crash, restart, retry, worktree, evaluation, and admission semantics are unambiguous.
- [ ] Migration and staged implementation can land without a flag-day graph rewrite.
- [ ] Acceptance brief identifies permanent smoke/model-based scenarios.

Depends on

done deep-survey-of

Required by

✉ Messages 4 messages (4 unread)

#1user2026-07-26T11:06:11.543419695+00:00read

Ratification must explicitly eliminate the reproduced Done -> Open on pending-message path; make completion/failure first-terminal-wins per attempt; make evaluation append-only evidence; make worktree ownership a fenced lease released/transferred atomically; and make ownership conflicts/reset inconsistencies breaker-neutral, self-diagnosing states rather than repeated spawn failures.

#2design-simplified-authoritative2026-07-26T11:09:41.534369047+00:00read

Acknowledged — the ratified design will explicitly prohibit Done→Open on pending-message paths; define first-terminal-wins per attempt; make evaluation append-only; fence and atomically release/transfer worktree leases; and make ownership/reset inconsistencies breaker-neutral diagnostic states rather than spawn failures.

#3user2026-07-26T11:47:30.983932167+00:00read

Add explicit attempt-finalization transaction: Active -> Suspect/NeedsFinalization -> durable candidate checkpoint -> deterministic validation -> evaluation against candidate -> merge/repair -> terminal. No-main-file heuristic; no terminalization while old PID can still write; no evaluator against main when candidate is isolated. Process exit without wg_done/wg_fail must trigger same-session completion probe, not inferred Done/Failed.

#4.respond-to-design-simplified-authoritative2026-07-26T12:15:45.796828654+00:00read

Done — commit 14fbcdd1 revises the ratified design with an explicit resumable attempt-finalization transaction: unreported exit → Suspect → same-session typed completion probe; all candidate writers quiesced; durable leased-worktree checkpoint; deterministic validation; exact-digest evaluation; merge/fenced repair; then one terminal commit. Main-file/diff heuristics and evaluation fallback to main are explicit hard violations. Added permanent scenario lifecycle_attempt_finalization_transaction.

Log

2026-07-26T10:23:22.710013646+00:00 Task paused
2026-07-26T10:25:30.969036025+00:00 Task published
2026-07-26T11:09:32.455924663+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
2026-07-26T11:09:50.750007552+00:00 Starting ratification design from the completed lifecycle survey; working inline because the deliverable is one tightly integrated design document whose decisions must remain internally consistent.
2026-07-26T11:14:15.081157877+00:00 Completed dependency review. Ratified target: one lifecycle kernel and append-only event ledger; separate task generations, attempts, processes, admission, provider, worktree leases, and evaluation records; explicit generation creation replaces all implicit reopening.
2026-07-26T11:24:30.191052621+00:00 Drafted docs/design-simplified-task-lifecycle.md: one transition kernel/ledger; generation and attempt state model; message-only correlated waits; fenced worktree leases; lazy append-only evaluation; deterministic failure map; replay/migration/diagnostics; and permanent acceptance suite.
2026-07-26T11:39:12.549812370+00:00 Validated task criteria: the design deterministically classifies every surveyed gate/failure, prohibits ordinary-message lifecycle effects, reduces 44 writer files to one kernel, specifies crash/retry/worktree/eval/admission semantics, gives lazy migration, diagnostics, and 18 permanent scenarios. Local link and code-fence checks pass; cargo fmt --check and cargo build --locked pass. Full cargo test is baseline-red in unrelated stale route tests (28 bin tests under add/edit/recover); the initial parallel run also exposed 7 profile test isolation failures, whose primary test passes alone. No production code changed.
2026-07-26T11:39:35.472653423+00:00 Committed: b12ef533 — pushed to remote
2026-07-26T11:39:51.394627326+00:00 Task marked as done
2026-07-26T11:51:07.989113116+00:00 Resurrection: created child task '.respond-to-design-simplified-authoritative' (1 pending message(s), downstream active)