Release v0.5.0 · bilalimamoglu/sift

0.5.0

Added a conservative hook beta with explicit sift hook match and sift hook run entry points for a tiny known-preset command set.
Added a shared adoption contract layer so doctor, install, agent previews, help, and docs all agree that sift exec is the default path and hook mode is only an optional shortcut.
Added a tiny Codex-native sift skill and a tiny Claude-native .claude/commands/sift/ command pack.
Added a shared runtime-payload rendering layer so managed blocks and native payloads reuse the same product truth instead of drifting by copy-paste.
Added lifecycle ownership checks for native payloads so generated Codex and Claude files are only updated or removed when sift can clearly prove ownership.
Added contract, smoke, dist, and packaging coverage for the adoption path and runtime-native packaging path.
Added local gain and discover commands so Sift can show small, local-only usage/value signals without turning into shell telemetry.
Added a narrow contract-drift preset for explicit snapshot, golden, OpenAPI, manifest, and generated-artifact drift output.
Added canonical scenario manifests for public test-status examples and a canonical live-session benchmark manifest bootstrap so docs and benchmark fixtures stop drifting independently.

Improved install honesty: runtime installs now explain the difference between the CLI as the real runtime and the generated native guidance files as small helper surfaces.
Improved runtime parity: Codex installs now produce both managed instructions and the generated skill; Claude installs now produce both managed instructions and the generated command pack.
Improved safety for low-level lifecycle commands by refusing to overwrite or delete custom SKILL.md files or custom Claude command files.
Improved first-contact docs and CLI help so users can understand what gets installed for Codex versus Claude without overestimating runtime parity or mistaking the native surfaces for a second execution system.
Improved gain confidence language so low-signal preset runs are no longer treated as decision-grade product truth.
Improved insufficient handling for repo-evidence shapes like prose docs, grep hits, path lists, diff stats, and structured config output.
Improved diagnose JSON trust by making read_targets more explicit about anchor kind, confidence, and whether a target is exact versus representative.
Improved benchmark truth surfaces by moving the existing live-session scorecard fixture behind a canonical manifest contract instead of a hardcoded fixture blob.

Completed Phase 2 adoption-friction work and Phase 3 runtime-native packaging work.
Completed the first trust-first hardening wave for evidence quality, diagnosis confidence, and scenario/benchmark source-of-truth cleanup.
Updated package version to 0.5.0.

gain and discover are intentionally local-only and conservative; they are product truth helpers, not full analytics.
contract-drift is intentionally narrow and should not be treated as generic repo analysis or broad mismatch detection.