feat(core): operator-owned task ledger and outcome-bound plans #462

Merged
karsten merged 10 commits from fix/repeated-validator-context into main 2026-11-14 12:41:31 +00:00
Owner

WIP — review candidate, no live rollout

Implements the agreed first Core operator-order boundary, independent of Dispatcher and session history. Pi-extension work remains separate.

Root cause

Odo repeatedly corrects one retained read-only step; Peter repeatedly re-assesses old parent mandates and emits Resume turns. Reproductions demonstrate tool-free validator admission bypassing stagnation, scheduler metadata counted as progress, and late worker results reopening technically completed parents.

Core candidate

  • Agent-wide operator disposition separate from technical result, immutable original/clarifications, revisions and generations, within existing execution-store commit/ownership.
  • Done/unresolved/stale orders are fenced at Main, Dispatcher, subagent and tool boundaries; historical unknown tasks held rather than reactivated.
  • Authenticated /orders, checked natural-language close/reopen/clarify proposals, optional research/skill-aware plans; simple work requires no plan.
  • Plan/outcome/checklist outbox commit atomically. Failed work cannot adopt plans. Scoped validation context and mandatory preservation of operator constraints under byte budget.
  • Durable archival before task-receipt capacity; old event idempotence preserved.

Verification

  • Typecheck/build PASS; full suite 2534 tests: 2533 pass, 0 fail, 1 skip.
  • Separate model review found five concrete ingress/identity/persistence/budget defects, all fixed with regression tests and targeted rereview. This is model review, not human/cross-provider approval.
  • Details: docs/task-lifecycle.md and development-evidence/core-task-lifecycle.md.

Remains WIP

  • Direct validator-originated plan-patch integration (currently plan-aware validation plus Main-submitted checked patches).
  • Registered recurring-series reopen/generation contract.
  • Explicit operator inactivity policy (no invented default), migration-inventory decisions and installed acceptance for Odo/Peter.
  • No bulk reopening of historical work and no deployment changes. Do not merge/activate as if all intended integrations were complete.

Update: review-blocked, do not merge

Validator plan protocol merged as Plugins #107 (8a34ff0) and is pinned; targeted tests and full 2537-test run passed. Real scheduler recurrence tested as distinct occurrence, not reopening old authority. Independent Core review FAIL identifies cumulative task-book capacity and request-size compatibility. No production rollout. Resolve durable capacity/lookup/fencing and verify actual size boundaries before re-review.

Capacity fix and rebase

Capacity finding addressed: immutable hash-verified pages referenced by the partitioned execution ledger, >10000 historical orders/bindings regression, long and worst-case escaped 16MiB requests, offline reclamation requiring exclusive home lock. Rebasing onto Core #460 required integration with partitioned roots; full test suite now 2567: 2566 passed, 0 failed, 1 skipped. Review PASS with truncated-diff limitation. No rollout performed.

## WIP — review candidate, no live rollout Implements the agreed first Core operator-order boundary, independent of Dispatcher and session history. Pi-extension work remains separate. ### Root cause Odo repeatedly corrects one retained read-only step; Peter repeatedly re-assesses old parent mandates and emits Resume turns. Reproductions demonstrate tool-free validator admission bypassing stagnation, scheduler metadata counted as progress, and late worker results reopening technically completed parents. ### Core candidate - Agent-wide operator disposition separate from technical result, immutable original/clarifications, revisions and generations, within existing execution-store commit/ownership. - Done/unresolved/stale orders are fenced at Main, Dispatcher, subagent and tool boundaries; historical unknown tasks held rather than reactivated. - Authenticated `/orders`, checked natural-language close/reopen/clarify proposals, optional research/skill-aware plans; simple work requires no plan. - Plan/outcome/checklist outbox commit atomically. Failed work cannot adopt plans. Scoped validation context and mandatory preservation of operator constraints under byte budget. - Durable archival before task-receipt capacity; old event idempotence preserved. ### Verification - Typecheck/build PASS; full suite **2534 tests: 2533 pass, 0 fail, 1 skip**. - Separate model review found five concrete ingress/identity/persistence/budget defects, all fixed with regression tests and targeted rereview. This is model review, not human/cross-provider approval. - Details: docs/task-lifecycle.md and development-evidence/core-task-lifecycle.md. ### Remains WIP - Direct validator-originated plan-patch integration (currently plan-aware validation plus Main-submitted checked patches). - Registered recurring-series reopen/generation contract. - Explicit operator inactivity policy (no invented default), migration-inventory decisions and installed acceptance for Odo/Peter. - No bulk reopening of historical work and no deployment changes. Do not merge/activate as if all intended integrations were complete. ### Update: review-blocked, do not merge Validator plan protocol merged as Plugins #107 (8a34ff0) and is pinned; targeted tests and full 2537-test run passed. Real scheduler recurrence tested as distinct occurrence, not reopening old authority. Independent Core review FAIL identifies cumulative task-book capacity and request-size compatibility. No production rollout. Resolve durable capacity/lookup/fencing and verify actual size boundaries before re-review. ### Capacity fix and rebase Capacity finding addressed: immutable hash-verified pages referenced by the partitioned execution ledger, >10000 historical orders/bindings regression, long and worst-case escaped 16MiB requests, offline reclamation requiring exclusive home lock. Rebasing onto Core #460 required integration with partitioned roots; full test suite now 2567: 2566 passed, 0 failed, 1 skipped. Review PASS with truncated-diff limitation. No rollout performed.
karsten force-pushed fix/repeated-validator-context from d7119173bb
All checks were successful
decision-interface-ci / full-validation (pull_request) Successful in 2m58s
to 6a9aff8a81
All checks were successful
decision-interface-ci / full-validation (pull_request) Successful in 3m27s
2026-11-14 12:10:56 +00:00
Compare
Author
Owner

PI-REVIEW-RECEIPT

State: FAIL
Head: 6a9aff8a8136
Summary: FAIL: The new mandatory task ledger introduces hard limits that can prevent normal requests from running.
Findings: 2

Top findings:

  • P1 Ledger capacity eventually blocks all new orders (src/task-lifecycle.ts:validateTaskBook)
  • P1 Requests longer than 8,000 characters cannot be enrolled (src/task-lifecycle.ts:validateTaskBook)
<!-- PI-REVIEW-RECEIPT --> ## PI-REVIEW-RECEIPT State: **FAIL** Head: `6a9aff8a8136` Summary: FAIL: The new mandatory task ledger introduces hard limits that can prevent normal requests from running. Findings: 2 Top findings: - **P1** Ledger capacity eventually blocks all new orders (src/task-lifecycle.ts:validateTaskBook) - **P1** Requests longer than 8,000 characters cannot be enrolled (src/task-lifecycle.ts:validateTaskBook)
Author
Owner

PI-REVIEW-RECEIPT

State: PASS
Head: 08bac469b680
Summary: PASS — no actionable findings identified in the supplied diff excerpt. The diff was truncated, so the review could not cover the complete implementation.
Findings: 0

<!-- PI-REVIEW-RECEIPT --> ## PI-REVIEW-RECEIPT State: **PASS** Head: `08bac469b680` Summary: PASS — no actionable findings identified in the supplied diff excerpt. The diff was truncated, so the review could not cover the complete implementation. Findings: 0
karsten force-pushed fix/repeated-validator-context from 08bac469b6
All checks were successful
decision-interface-ci / full-validation (pull_request) Successful in 3m3s
to 1ea2592f9a
All checks were successful
decision-interface-ci / full-validation (pull_request) Successful in 3m36s
2026-11-14 12:36:33 +00:00
Compare
Author
Owner

PI-REVIEW-RECEIPT

State: PASS
Head: 1ea2592f9ace
Summary: PASS — No blocking findings identified in the supplied diff. The review is limited by truncation of the diff.
Findings: 1

Top findings:

  • P2 Task lifecycle orchestration is coupled to the central Harness (src/harness.ts:715)
<!-- PI-REVIEW-RECEIPT --> ## PI-REVIEW-RECEIPT State: **PASS** Head: `1ea2592f9ace` Summary: PASS — No blocking findings identified in the supplied diff. The review is limited by truncation of the diff. Findings: 1 Top findings: - **P2** Task lifecycle orchestration is coupled to the central Harness (src/harness.ts:715)
karsten changed title from WIP: feat(core): operator-owned task ledger, terminal intent and outcome-bound plans to feat(core): operator-owned task ledger and outcome-bound plans 2026-11-14 12:41:25 +00:00
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
harness/core!462
No description provided.