# Execution integration checklist

## Implemented and verified

- [x] Connect Ptah's `CoderAgent` to `AuthorizedExecutionService`.
- [x] Use shared identity, permissions, policy, approvals, and audit through
  `PtahBridge`.
- [x] Default to Bubblewrap; fail closed if the sandbox is unavailable.
- [x] Require explicit Python or shell permission scopes.
- [x] Record task-correlated permission, approval, tool-result, and agent-result
  audit events.
- [x] Return sandbox output and failure status through the Ptah step result.
- [x] Cover allow, approval, deny, timeout, missing scope, sandbox failure, and
  audit correlation in tests.
- [x] Run pytest, Ruff, mypy, and whitespace checks; see the integration report
  for recorded results.

## Before production use

- [x] Review Bubblewrap host prerequisites, namespaces, mounted runtime paths,
  writable workspace handling, and resource limits. The engineering review is
  in `EXECUTION_SECURITY_REVIEW.md`; it does **not** approve hostile-code use.
- [ ] Decide and configure default CoderAgent scopes per deployment; keep shell
  and Python scopes separate.
- [x] Add opt-in SQLite storage for Ptah task steps, approvals, and audit
  records, with schema migrations, restart behavior, backup support, and
  explicit retention. See `docs/data/sqlite-storage.md`.
- [ ] Expose execution configuration and review decisions through the API/UI;
  no production approval interface is implemented.
- [x] Configure CI to install Bubblewrap and run the full tests, Ruff, and
  package-wide mypy. The mocked missing-Bubblewrap test runs with the same
  suite; hosted CI execution remains to be observed.
- [x] Review staged, unstaged, and untracked repository changes and classify
  them in `BASELINE_REVIEW.md`. The working tree remains uncommitted and is not
  yet a clean release baseline.

## Implemented local Data Plane slice

- [x] Add versioned SQLite schema for task jobs, approvals, audit events,
  project memories, and provenance.
- [x] Persist named JSON tasks and Ptah roadmap steps; interrupted jobs require
  explicit retry to avoid automatic repetition after uncertain side effects.
- [x] Add trust-labeled keyword memory, local provenance hashes, and heuristic
  redaction before project-memory writes.
- [ ] Add ChromaDB, Notebook, vector retrieval, and a deployment retention
  scheduler.

## Later roadmap work

- [ ] Add separately authorized execution support for other agent roles only
  when their tool workflows are defined.
- [ ] Add container or microVM providers behind the same execution protocol.
- [ ] Continue distributed registry/dispatch and Governance Plane
  implementation.