∞skylakegrep

release notes · v0.5.15

skylakegrep 0.5.15 — first-class agent presets and release gates

0.5.15 turns the 0.5.14 agent playbook into explicit CLI and release contracts. The search engine is the same generic router; this release removes flag-bundle ambiguity for LLM callers, makes setup freshness auditable, and replaces the broken token-based PyPI workflow with Trusted Publishing.

What changed

  • --agent-fast is now the official path-anchor preset for LLM callers. It expands to JSON output, no snippets, top-10 candidates, and no rerank.
  • --agent-context is now the official first evidence preset. It expands to JSON output, standard snippets, and no rerank.
  • --agent-daemon makes daemon-first repeated calls explicit. It uses SKYGREP_DAEMON_URL when set, otherwise http://127.0.0.1:7878, and falls back in-process if no daemon is running.
  • skygrep setup --check reports whether managed Claude Code, Codex, OpenCode, Gemini CLI, and Cursor snippets are current, stale, missing, or broken without modifying user files.
  • Managed setup snippets now carry an internal guidance version and teach the shorter agent presets first, while keeping the explicit long-form commands documented for auditability.
  • benchmarks/closed_loop_regression_gate.py turns saved closed-loop benchmark JSON into a machine-checkable release gate.
  • .github/workflows/release.yml now uses PyPI Trusted Publishing (OIDC) instead of PYPI_API_TOKEN.

Agent command contract

Use these commands when another LLM will consume the result:

skygrep --agent-fast "where is token refresh implemented?"
skygrep --agent-context --include "src/**" "what does token refresh do?"
skygrep serve --port 7878
skygrep --agent-daemon --agent-context "what does token refresh do?"

The explicit long-form equivalents still work:

skygrep --json --no-content --top 10 --no-rerank "<query>"
skygrep --json --content --detail standard --top 8 --no-rerank "<query>"

Setup freshness

Existing managed setup blocks are still refreshed automatically on normal interactive use, and skygrep setup still updates only the managed BEGIN/END block while preserving user-authored instructions. The new check mode gives users and CI a non-mutating audit command:

skygrep setup --check

Benchmark gate

The universal benchmark remains the real measurement path:

python benchmarks/universal_closed_loop_benchmark.py \
  --repo self --repo django --repo react --repo tokio --summary-only \
  > /tmp/skygrep-closed-loop.json

The new release gate checks the saved report:

python benchmarks/closed_loop_regression_gate.py /tmp/skygrep-closed-loop.json

Default gate thresholds require skygrep-first to preserve at least 90 % path coverage, 90 % evidence coverage, 85 % sufficiency, 2× context reduction, no estimated agent elapsed regression, and no work-quality per-minute regression. Release runs can tighten those thresholds when the benchmark fixture is fully prepared.

Publishing

The GitHub release workflow now expects PyPI Trusted Publishing with:

  • owner: danielchen26
  • repository: skylakegrep
  • workflow: release.yml
  • environment: pypi

That removes the long-lived PYPI_API_TOKEN failure mode.

Verification

  • Targeted CLI/setup/benchmark tests cover the new agent presets, daemon expansion, setup status checks, and regression gate.
  • Full test suite and privacy release scans were run before publishing.
  • Public examples in this release use only generic placeholder questions. No local private path, filename, screenshot, or personal benchmark content is included.