Independent implementation results

Implemented 2026-10-19 after the operator authorized Sol workers with parent review. The standalone package is staged at package/, named treadle-headless version 0.1.0, import treadle_headless. It has no runtime dependencies and targets Linux with Python 3.11 or newer. No code or tests were imported from the headless branch. Treadle and the Commonplace runtime were not modified.

Delivered

• Immutable requests and returned results, named text inputs and outputs, caller-supplied working directories, and a generic subprocess worker.
• Fresh paths for every physical launch, including retries; explicit delivery outcomes and raw execution evidence.
• Exclusive run ownership, atomically committed result records, request and content checksums, workdir binding, persistent launch allowances, and explicit acknowledgment of uncertain launches before resumption.
• Bounded parallel groups with ordinary Python control flow. No scheduler or generic correction protocol was added. The example owns its content check, fixed comparison snapshot and bounded correction loop.
• Codex, Claude Code and Pi command factories and a standalone JSONL evidence parser. Codex and Pi passed real Sol file-copy/reuse probes; Claude is command-tested and parser-tested only.
• A distributable wheel and source archive, with source-archive documentation and examples included.

Review and corrections

One Sol worker implemented the core; another independently built the ordinary file example and acceptance tests; a third researched and implemented the adapters, then reviewed the core without write authority. The parent reviewed the source, directed fixes, added two regression tests, ran real adapter probes, and verified the built distribution independently.

Review found and fixed:

• Literal braces in generic command arguments being interpreted as template syntax, and recursive substitution when a real path contained token text.
• Mutable mappings inside nominally immutable results.
• Saved results with missing, invented or modified output contents being accepted on reload.
• A failed result commit leaving reusable, uncommitted controller state.
• Process-start record failures bypassing cleanup, and cleanup failures after process start being misclassified as settled launch failures.

The final cleanup-error regression uses a real sleeping subprocess, injects a cleanup error, verifies that its reservation remains unresolved, and explicitly cleans up the test worker. Other interruption cases cover a controller killed while its worker runs, a prelaunch reservation, and exit after result commit.

Validation

Commands run from the package directory unless stated otherwise:

Check	Result
uv run pytest -q	55 passed
uv run ruff check .	Passed
uv build --quiet	Wheel and source archive built
Wheel source comparison	All packaged Python modules exactly matched current source bytes
Source archive contents	Includes examples/ordinary_files.py and docs/design.md
Installed-wheel test suite	55 passed outside the repository, using copied tests and examples
Installed import check	Imported from temporary environment's site-packages; Commonplace was absent
commonplace-validate kb/work/headless-workflow-extraction	Passed before this result record; revalidated after recording it
git diff --check	Passed

The fresh installation used Python 3.13.1 in /tmp/treadle-headless-installed-tqktvy03/.venv. The test command there was uv run --no-project --python /tmp/treadle-headless-installed-tqktvy03/.venv/bin/python python -m pytest -q tests. There was no package source directory in that test location and no Git repository. The isolated test files include their own deterministic subprocess workers; they do not need model credentials.

Wheel: package/dist/treadle_headless-0.1.0-py3-none-any.whl. SHA-256: 3d6c1bfd76f311a3d1f2f77b5f1f6227110ce351eef23cc3f51d5cb68ace0feb.

Source archive: package/dist/treadle_headless-0.1.0.tar.gz. SHA-256: ba840b8a9033491e73dfd5809aaa740972340171de49dad06088f23e3f02a8a2.

Build artifacts are ignored local outputs, reproducible with the package's build command; their hashes identify this tested build, not all future builds.

The harness evidence records versions, commands, raw local record locations and limitations of the real Codex/Pi probes. Both requested gpt-6.1-sol, delivered the exact input copy in one launch and reused that launch on the repeated call. Requested model identity is not substituted for independently observed served identity.

Remaining limits and handoff

The package does not isolate access, pin code, control ambient dependencies, make external effects transactional, or automatically recover orphan workers. Its parallel limit applies per group. Provider quota errors are not universally classified; structured problem-file usage limits are recognized and not automatically retried. The adapter stream parser preserves evidence separately from the generic runtime's result protocol.

The implementation and independent acceptance phases are complete within these support claims. Public naming, final repository placement and external publication remain undecided. Commonplace integration remains a separately commissioned consumer project, not an implicit follow-up in the headless branch. The workshop stays active while those ownership and adoption decisions remain. Nothing has been committed or published by this implementation session.