v0.5.0

0.5.0 (2026-09-26)

Features

• adapt public JevBench items to the eval harness (A8 preparation) (d4a78a2)
• add advisory calibrate subcommand fitting thresholds on caller-labeled rows (3f3051b)
• add order-sensitivity probe to the eval harness (A2) (e4276ac)
• cache: off-loop IO, a TTL, and oldest-first eviction (ADR-0047 amendment) (8e93953)
• file lists name each reviewed file's action (file_actions) (5a02bf1)
• gate file list verifies claims once, per file review stays per file (25016dd)
• providers: opt-in JEV_MCP_MAX_INFLIGHT cap on concurrent requests (fd26573)
• telemetry: log the metrics snapshot at INFO every 100 tool spans (0c2972b)

Bug Fixes

• address every critique finding on fm/jev-eval-calibrate (P2-1..4, P3-1..6, E1..3) (0f2cdbe)
• cli: code isError envelopes with the one parity mapping (422c17f)
• critique follow-ups H1-H4 (hook gates, cache leftovers, serve wiring test, off-path hops) (14c2569)
• delta-critique-a follow-ups — sweep before no-evict, scaffold guard, worst-last pin (C1, C2, F1) (0c35884)
• delta-critique-b nits — one source_commit call, required control (F3, F4) (831fab4)
• example collects every confidence across the real payload shapes (f2d3c86)
• gate file list keeps isError, sums usage, shares file plumbing (0cf8928)
• hook: match completion commands on tokens, not a string prefix (c6641f7)
• qualify file_actions' worst-entry claim and pin the clamp shapes (D1, G-A, G-B) (38b7b34)
• responses: code every frozen budget refusal input_too_large (dc06e0e)
• review-critique follow-ups on the gate file list (F1-F3, G1-G3) (d33c98e)
• server: configure redacting logging before the judge/gate/hook subcommands (6630fc8)
• server: hand the HTTP listen sockets to uvicorn, closing the probe gap (6862536)

Documentation

• caller guide for writing states and questions (86bdcfb)
• claims are positional strings, not id records (87b65b6)
• fix every critique finding on fm/jev-readme-bench (P1-A..P3-8) (fc70ce6)
• fold caller-docs' EVIDENCE.md pointer into the reworked Measured results; depth now points at guidance.md (89c5369)
• lead with measured speed, cost, and decision quality; add agent setup rules (e386694)
• limits page for caps, defaults, and error codes (66cecf3)
• name the pin behind every budget refusal's error code (47d5f4e)
• per-tool cards with measured evidence and weak spots (0c04ad8)
• pin or source every restated frozen number (1215d47)
• pin the install command's -y exactly; state the public_tasks id join (F1, F2) (a95c097)
• release evidence page and a release step that fills it (ee28b62)
• state the per-file diff review rule on the limits page and tool cards (5d579f0)
• track the JevBench run's machine evidence; point the adapter at the recorded run (F1, F2) (e807384)