Add `release-kotlin-library` with an adapted Haze release helper. The
workflow reconciles an existing changelog against changes since the
previous release, resolves versions, and follows the repository's checks
and publishing mechanism. Release files are prepared and validated
before publication; completion requires artifact and remote Git
readback.
The helper separates preparation, publication, and recovery, loads
required credentials from `~/.env` without surfacing raw child output,
rejects unrelated staged work and divergent remote branches, and
preserves check-owned edits when preparation fails. README/router
integration and direct, novel, and no-change evaluation cases are
included. Versions are unchanged.
Validation: the full `npm test` suite passed (307 tests, one existing
Linux-only skip on macOS), followed by the final staged-file regression
test. After merging current main, all 40 affected tests, Markdown lint,
and validation of all 94 evaluation cases passed. Python compilation
also passed. Standards and spec reviews have no remaining blocking
findings. Publication tests use isolated repositories and synthetic
credentials; live publishing and the model evaluation benchmark were not
run.
## Summary
- identify `implement` as an external workflow from `mattpocock/skills`
and make its current companion requirements actionable
- document every required, conditional, and optional external provider
used by the bundled workflows
- add missing-provider calibration cases for `implement-with-subagents`
and `run-github-project`
## Testing
- `npm run lint`
- `python3 evals/run.py validate --suite workflows-writing`
- `npm test`
Fixes#63
## Summary
- remove the unsupported paths field from the Kotlin and Compose router
skill
- stop allowing paths in the repository frontmatter schema
- correct the README and add a regression assertion
## Validation
- npm run lint
- python3 -m unittest discover -s evals/tests -v (143 tests)
- git diff --check
Fixes#61
## Summary
- Route raw `Thread` and `Executor` usage through the Kotlin concurrency
skill.
- Add structured migration guidance for suspend APIs, owned scopes and
dispatchers, interruptible blocking work, and executor-backed
dispatchers.
- Add direct, novel, and no-change evaluation coverage, then update the
documented Kotlin suite counts.
## Validation
- Skill quick validation passed.
- `npm run lint` passed.
- Evaluation corpus validation passed with 79 cases, including 22 Kotlin
cases.
- 22 focused CLI and Kotlin Gradle matrix tests passed.
- The direct fixture passed its Gradle red/green check.
- `git diff --check` passed.
Live behavioural model evaluation was not run; the planned forced and
automatic comparison would require 12 model calls.
## Summary
- make the four advanced workflow skills user-invokable only in both UI
metadata and SKILL.md frontmatter
- make automatic evaluation and reporting exclude explicit-only skills
while retaining forced coverage
- document and validate the frontmatter contract
## Validation
- npm test
- npm run evals:validate
- git diff --check
## Not run
- npm run lint (the required remark executable is unavailable in both
sandbox and host environments)
## Summary
- restore evaluation behaviour after the recent skill simplification
changes
- capture subject token, tool-call, turn, elapsed-time, and retry
telemetry
- compare per-skill baseline and automatic efficiency with compact
values and percentage changes
- keep the result tables in the root and evaluation READMEs synchronized
## Validation
- npm test
- npm run lint
- npm run evals:validate
- git diff --check
The evaluation results are advisory and remain non-gating.
## Summary
- switch `shepherd` from narrow CI-driven repair loops to full locally
available CI-equivalent verification after a code-related failure
- batch all discovered repairs into one verified push and report checks
unavailable locally
- update the direct and novel evaluation cases plus the README
## Validation
- `npm run lint`
- `npm test` (250 tests, 1 skipped)
- `npm run evals:validate` (73 cases)
- `python3
/Users/chris/.codex/skills/.system/skill-creator/scripts/quick_validate.py
skills/shepherd`
- `git diff --check`
## Summary
- identify `gpt-5.6-terra` with medium reasoning as the subject behind
the published evaluation scores
- identify `gpt-5.6-sol` with high reasoning as the judge
- link both official model pages and clarify that results are model- and
reasoning-specific
## Testing
- `npm run lint`
- `git diff --check`
## Summary
- publish the current-HEAD 100.0% Automatic result for kotlin-api-design
- publish the current-HEAD 100.0% Restraint result for
kotlin-concurrency-and-flow
- keep the result tables in README.md and evals/README.md synchronized
## Evaluation
- kotlin-api-design Automatic: 12/12 outcomes passed across four cases
and three repetitions
- kotlin-concurrency-and-flow Restraint: 3/3 forced and 3/3 automatic
controls passed
- zero objective failures, violations, forbidden actions, process
failures, or retries
- reported API routing precision was 100.0% and recall was 85.7%; both
routing gates passed
## Validation
- npm test (247 passed, 1 skipped)
- npm run lint
- npm run evals:validate (61 cases)
- git diff --check
## Summary
- Remove per-case forbidden skill and action fields now covered by
global evaluator policy
- Centralize subject skill canonicalization and output validation in the
Codex harness
- Reuse the shared parsing and reporting helpers across experiment
execution
- Update evaluator documentation to describe global safety checks
## Testing
Not run (not requested)
Chris Banes