release notes · v0.7.0

skylakegrep 0.7.0 — strict evidence and direct/daemon parity

0.7.0 is an accuracy-and-efficiency release for agent workflows. It adds a fail-closed verification mode for high-risk local claims, makes direct and daemon agent-context calls execute the same retrieval contract, and turns a deterministic accuracy/performance comparison into a pull-request CI gate.

The release does not claim that fast retrieval proves external truth. --strict verifies agreement and freshness inside the indexed local corpus; agents must still use authoritative external sources for claims about the outside world.

What changed

• skygrep --strict "<query>" now implies agent context, runs the normal hybrid candidate-recall path plus an independent corpus-wide semantic pass, validates source/index modification-time parity, and emits a strict_verification receipt.
• Strict mode exits 2 when evidence is absent, the indexed source is stale, independent retrieval lanes disagree, semantic rank is insufficient, or the final agent quality remains inconclusive. Only status=passed is verified.
• Direct and HTTP-daemon agent-context calls share run_agent_context_search, including semantic escalation, evidence bundles, confidence summaries, source priors, and strict verification.
• The daemon rejects requests that point at a different project root or index database, and rejects lexical roots outside the configured project.
• Candidate recall coalesces equivalent macOS /var and /private/var path spellings before ranking. Duplicate aliases contribute only the strongest extra score, so lexical evidence is not double-counted.
• Duplicate results retain the union of their candidate_recall_lanes, making the final provenance receipt reflect every contributing lane.
• agent-guidance-v5 teaches Claude Code, Codex, OpenCode, Gemini CLI, and Cursor to require strict receipts for security, release, legal, financial, destructive, and other high-risk local claims.
• Pull-request CI runs benchmarks/ci_agent_contract_benchmark.py through the existing closed-loop regression gate on Python 3.12.

Measured performance and accuracy

The new CI fixture is deterministic and model-free. It exercises the real hybrid candidate-recall implementation over six generic repository tasks, then compares the compact skygrep-first contract with a modeled raw-ripgrep agent workflow. This is a regression fixture, not a universal speed claim.

Live verification also ran the same strict query through the direct and daemon paths. Top-K paths, candidate-recall lanes, agent quality, and strict status were identical; both selected the expected implementation file and returned strict_verification.status=passed.

Accuracy and evidence boundaries

• Strict verification requires independent hybrid and semantic agreement; it does not accept first-pass confidence alone.
• Freshness compares the indexed source timestamp with the file currently on disk. A stale source produces an inconclusive receipt and exit 2.
• The receipt exposes checks, independent lanes, semantic rank, evidence presence, source state, and an error string for machine-readable auditing.
• Direct/daemon parity is asserted by both an HTTP integration test and a live local daemon comparison.
• Daemon project-boundary validation prevents a caller from silently using the server with a different repository database.
• External factual claims remain outside this guarantee. Strict mode validates local evidence, not the current state of a website, service, law, or market.

Compatibility

• Existing 0.6.x indexes are byte-compatible; no rebuild is required.
• The top-level agent JSON value remains a list of result objects. strict_verification and its summary status are optional additive fields.
• --strict is incompatible with answer/agentic synthesis modes because it is an evidence-verification contract, not a generation mode.
• Non-strict agent presets retain their existing bounded-latency defaults.
• Raw regular expressions and exhaustive lexical output should still use rg.

Verification

• Full implementation suite: 390 passed, 20 subtests passed, with two existing Click deprecation warnings.
• Strict-mode tests cover fresh-source success and stale-source rejection.
• Direct/daemon tests cover shared-contract parity, successful and inconclusive strict exit codes, and cross-project request rejection.
• macOS path-alias regression confirms coalescing without lexical score double-counting.
• The six-task CI gate passed with 100% path coverage, evidence coverage, and sufficiency; no configured threshold failed.
• Source compilation and git diff --check passed.
• The local index was rebuilt and independently queried as 190/190 absolute indexed paths before the release candidate was prepared.

Privacy

Source, documentation, tests, benchmark fixtures, and the release candidate use only generic repository-maintenance examples. The source tree and built wheel/sdist are scanned before publication. No real local prompt, private filename or folder, absolute user path, screenshot, document category, name, email, credential, token, or machine-specific identifier is included.

Known follow-ups

• The CI benchmark is deliberately small, deterministic, and model-free. The larger real-repository/Ollama benchmark remains the release-scale empirical check rather than a per-PR dependency.
• PyPI Trusted Publishing still depends on the project-side publisher matching owner danielchen26, repository skylakegrep, workflow release.yml, and environment pypi. If that external configuration rejects the tag job, the authenticated fallback remains required before the release can be called complete.