[rng1995]

Summary

Prepare SkillEvaluator v0.6.0, following the v0.5.0 release PR (#185).

Bump package metadata, the source version fallback, citation metadata, the developer guide example, and the root lockfile entry to 0.6.0. Cut the dated 0.6.0 - 2026-10-20 changelog section and leave a fresh Unreleased section. Dependency versions are unchanged.

A second commit moves the four workflow-hardening ### Security entries from #126 into 0.6.0. They had landed under 0.3.0, but #126 merged after v0.5.0. The text is unchanged; only its position moves.

Release Impact

This release includes everything merged since v0.5.0. Every merge that touches src/ has a changelog entry. #184 changes only tests and has none.

• Models: GPT-6.1 Sol and Claude Opus 5.5 become the defaults across Tiers 1–3, with pinned Tier 3 agent CLIs, per-family token limits, and reasoning effort (feat: default to GPT-6.1 Sol and Claude Opus 5.5 across Tiers 1-3 #188). SkillSpector keeps its own default model, and the Tier 1 card shows its version (feat(report): show the SkillSpector version in the Tier 1 HTML card #187).
• Tier 3 backends: Kata Containers (feat(tier3): run hardened Docker trials under Kata Containers #182), Harbor Runta and Mosaic (feat(tier3): expose Harbor's Runta and Mosaic backends #183), GKE mode, --ek forwarding, and Claude Code on Vertex AI with skill MCP config validation (feat(tier3): support Claude Code on Agent Platform and add active preflight probe #140).
• Scoring: one overall-score policy (fix(tier3): use one overall-score formula #124), accuracy derived from criterion verdicts (fix(tier3): derive accuracy scores from criterion verdicts #190), no credit for failed skill reads (fix(tier3): exclude failed skill reads from execution credit #201), inconclusive results for never-activated skills (fix(tier3): mark never-activated skill evaluations inconclusive #209), and per-step token and ATIF content-part evidence (fix(tier3): honor per-step tokens when detecting runtime failures #199, fix(tier3): include ATIF content parts in judge answer evidence #203). Also target dependency coverage and result integrity (fix: target dependency coverage and Tier 3 result integrity #196).
• Reporting: the skillevaluator dashboard comparison tables (feat(reporting): add quantitative skill performance dashboard #179). BENCHMARK.md publication now requires complete tier evidence (fix(reporting): require complete tier evidence for publication #77).
• Docker agents: no more redundant Node/npm install during agent setup (fix: avoid redundant Node/npm during Docker agent setup #193).
• Other: multiline Tier 2 catalog descriptions (49b2b43), docs refresh (docs: bring the published docs in line with v0.5.0 #186), and CI workflow hardening (ci: cover every workflow with the hardening guards, and fix the gaps they surface #126).

Upgrade notes (details in CHANGELOG.md):

• Default models change to gpt-6.1-sol and claude-opus-5-5, so scores and cost may shift against 0.5.0 baselines.
• The overall-score formula changes. Lift directions recorded by 0.5.0 can differ.
• Docker Claude Code and Codex containers no longer get distro nodejs/npm, which on Debian images also provided npx, gcc, and make. Skills that need them must install them through scripts/apt-packages.txt or a custom Dockerfile.
• --output-dir or --results-dir inside the evaluated skill is now rejected.
• validate --block-on-agent-eval passes only on a completed Tier 3 PASS.
• PASS benchmark cards must record Source digest and Tier 3 run ID. Cards rendered by earlier releases must be regenerated.

This PR prepares the release metadata. The v0.6.0 tag and GitHub Release can be published against its merge commit after validation.

Verification

• Reviewed the Contributing Guidelines; commits include DCO sign-off.
• Updated version documentation and CHANGELOG.md; no credentials, private datasets, or proprietary benchmark content added.
• uv lock --check and uv sync --all-extras --locked --python 3.13.
• ruff check .
• Fern 5.121.0 check (lockfile install): 0 errors. The authenticated redirect check was skipped because FERN_TOKEN was not set.
• uv build --python 3.13 --no-sources: wheel and source distribution built successfully.
• OSS boundary scan passed for source, wheel, and source distribution.
• Public benchmark fixtures passed with required source provenance.
• twine==6.2.0 check --strict passed for both distributions.
• Installed-wheel smoke test on Python 3.12 and 3.13: metadata and runtime both report 0.6.0 with Harbor 0.24.0. Imports, the bundled verifier template, --help, models --help, doctor, and the progress reporter pass.
• Full local pytest -n 4 on Python 3.13: 13,735 passed, 23 skipped, 2 failed.
  • Both failures are in test_harbor_secure_docker_environment.py::test_sidecar_retains_*structurally_required_nonsecret_*, and the cause is the host environment.
  • With a minimal environment, the whole file passes (305 passed). env -i ... MY_AUTH_FLAG=1 alone reproduces the failure.
  • Any host variable with a sensitive-looking name and a short value, such as 1, that also appears in the test image python:3.13-slim is treated as a protected value.
  • The same tests pass in main's Linux CI. This PR makes no test or runtime changes.

No new behavioral tests are needed for this metadata-only change; existing tests validate the release.

🤖 Generated with Claude Code