# Version 1.6.0 — portable autopilot

This record covers the reviewed source and local verification. The
[GitHub release](https://github.com/nobrainer-tech/nobrainer-tech-skills/releases/tag/v1.6.0)
records publication and its tag commit. A local version or passing test is not
itself publication evidence.

## What changed

Clear coding and everyday tasks can execute directly, including chat-only and
non-Git work. Larger tasks retain one clarification round when needed, concise
TODO progress, existing authorization and a bounded corrective loop. Markdown
state suffices for resume without native goals or telemetry. Optional subagents
do not require persistent sessions. Existing project specs, docs and diagrams
are reused; new scaffolding requires an actual acceptance need.

The review records five baseline contract defects,
their fixes and dated ecosystem research. The fifteen-skill inventory is unchanged.

## Evidence and limits

- Baseline `d94dd05`: 121 deterministic tests passed despite the source contradictions.
- Final source and release documentation: 125 deterministic tests passed.
- Codex CLI `0.153.1`: one four-case batch each for baseline/candidate with
  requested `gpt-6-astra` and `gpt-5.6-luna`, low effort. Behavior was acceptable;
  this establishes no improvement percentage or universal model claim.
- A separate frozen four-case holdout passed all 16 prewritten criteria after
  good/bad evaluator controls. One Astra batch, no retries, explicit source
  injection, no tool calls; this is simulation evidence.
  [Sealed cases, calibration, output and verdict](../evals/artifacts/v1.6.0/holdout/receipt.json).
- Claude Code `2.1.241`: blocked before model execution by organization subscription
  access. No API credential substitution; current Claude behavior remains unverified.
- Actual non-Git artifact execution on Codex/Astra passed: a 43-word invitation,
  three-item organiser checklist and completed existing TODO. Independent readback
  checked every brief fact, unchanged skill hashes and absence of new Git/workflow
  scaffolding. Artifacts and receipt.
- Final deterministic, syntax and public-clean release checks are recorded in the
  GitHub release and CI run for the reviewed commit.

The CLI accepted the requested models but did not separately expose backend
model identity. Global skill metadata affected input usage; the holdout emitted
a description-truncation warning. These probes do not prove clean automatic
discovery, alias selection, all client integrations or token-cost savings.

Reproducible public artifacts:

- Frozen four-case prompt
- [Source SHA-256 manifest](../evals/artifacts/v1.6.0/candidate-source-hashes.json)
- [Model receipts, usage and read commands](../evals/artifacts/v1.6.0/smoke-receipts.json)
- Astra candidate output
- Luna candidate output

The source manifest is checked against the [frozen source snapshot](../evals/artifacts/v1.6.0/source-snapshot.zip)
by the test suite. Its 31 files were recovered from release commit `d51d1e10db881182601afef80df9e3196d4779ec`
and verified against the original manifest. Later source changes do not inherit
these runtime results. Raw local CLI logs are not published because host metadata is outside the public package.
Outputs and selected read receipts preserve the relevant bounded evidence.

## Recovery

Pin the reviewed commit or release. Version `v1.5.0` remains available as a source
rollback; a focused revert restores prior behavior without changing private user
skills or unrelated projects. Website publication uses its own scoped source,
preimage backup and public readback.