Tooling + closure services on the shared engine
Goal
- Run tooling/review/commit/status through the shared invocation protocol.
- Direct invocation creates a root; auto-triggered invocation becomes a child.
- Child tooling returns local acceptance + root obligations without terminal closeout.
Implementation sketch
- Adapt improve_tooling_runtime phases to root/child entry and typed returns without a skip-closeout boolean.
- Adapt add-bug, review-plan, and review-section as capability-bound workflows with typed returns; no child-authored resume commands.
- Keep gap/design/implementation/tests/self-test/lint/harness activation safety local; delegate aggregate review, final version publication, commit/push, and terminal status when folded.
- Represent harness compatibility changes as version intents aggregated once by the root.
- Run review-work, tpr-review, impl-hygiene-review, and sync as durable engine invocations over the frozen union with frames, task tokens, typed results, finding/cure children, and symbolic returns.
- Make commit-push and status-report root finalizers with idempotent receipts; publication must succeed before the sole terminal report and seal.
- Compose status/reporting from two labeled sources: authoritative plan/tracker workflow state via its API and root invocation/finalizer protocol facts via the execution store. Neither status-report nor an LLM may synthesize the done verdict; only engine closeout emits TerminalReceipt.
- Permit recursive tooling/bug children while preserving root authority and engine epoch.
Test strategy / acceptance
Test direct-root and folded-child tooling/review/add-bug, recursive chains, parent activation safety, one aggregate version bump, stale review invalidation/reactivation, publication-pending retry, source-labeled root reporting, engine-only TerminalReceipt, self-upgrade bundle resolution, and return after crash.
Work Items
- Adapt improve-tooling to direct-root and folded-child entry with the generic invocation/return protocol
- Split improve-tooling gates into local acceptance/activation safety versus root-delegated aggregate review, version publication, commit, and status obligations
- Aggregate child harness compatibility intents into one root fingerprint/version/changelog/migration decision
- Name the legacy-skill-as-child adapter contract: wrap an un-migrated skill dispatch (e.g. /tpr-review, /tp-help pre-migration) as an engine invocation with task token, typed result contract, fold receipt, and symbolic return; pin review.json scope explicitly as the closure-service wrapper only (typed findings in / cure children out) with reviewer-dispatch internals remaining legacy per the tpr-review-multi-agent-dispatch locked design until the future /tpr-review migration plan
- Run review-work, TPR, implementation-hygiene, and sync as durable root-scoped invocations with task tokens, typed findings/cures, fold receipts, and symbolic returns; keep commit-push/status as idempotent finalizers
- Implement semantic-closed/publication-pending/published/sealed transitions with idempotent commit/push retry and no premature terminal report
- Support recursive improve-tooling, add-bug, and fix-bug children with root engine-epoch inheritance and stale-return fencing
- Render source-labeled root status for domain and virtual roots: plan/tracker API supplies workflow state, root store supplies invocation/obligation/generation facts, and only engine closeout supplies the sealed TerminalReceipt; suppress child, client, status-report, and LLM-authored terminal claims
- Migrate add-bug, review-plan, and review-section to engine-issued child capabilities, typed local acceptance, result ingestion, and symbolic return targets
- Run end-to-end thirty-bug acceptance across canonical batch selection, bug-domain adaptation, root aggregate review, finalization, publication, and terminal status
Intel dossier — pointer and tier disposition
Dossier: tooling-closure-services-on-the-shared-engine--s-ec3fa023.intel.md (this section’s content/intel/ sidecar)
Read that dossier from line 1 through EOF before acting on this section. This block is a pointer and an audit record, never a substitute: not the BLUF, not a summary, not selected tiers.
| Dossier tier | Title | Lines | Disposition | Where / why |
|---|---|---|---|---|
| 0 | 0. Bottom line up front | 41 | pending | |
| 1 | 1. The target, verbatim | 30 | pending | |
| 2 | 2. Terrain — symbols and files | 45 | pending | |
| 3 | 3. Code-graph recon | 103 | pending | |
| 3S | 3S. Assigned defect-class sweep — gates whose inability to observe their subject is indistinguishable from a pass | 122 | pending | |
| 3D | 3D. Diagnostic / observability surface | 27 | pending | |
| 3H | 3H. Hygiene constraints the downstream work inherits | 31 | pending | |
| 4 | 4. Cluster / family — the sibling sections and what they inherit | 64 | pending | |
| 5 | 5. Conformance audit | 25 | pending | |
| 6 | 6. Plan ownership | 18 | pending | |
| 7 | 7. Prior art | 25 | pending | |
| 8 | 8. Sentiment | 7 | pending | |
| 9 | 9. Declared coverage gaps — what is dark, and why | 35 | pending | |
| 10 | 10. Recon entry points — the ordered read path | 21 | pending |