Changelog

Release notes that build product trust.

Follow what shipped, what stabilized, and what changed in provider routing, verification, docs, and membership workflows.

8.22.52026-06-03

Planner role + cache hit observability for DeepSeek/Grok

Planner role now active for cache-isolated planning sessions (Reasonix-style): Plan agent defaults to "planner", fork sessions deliberately skip parent transcript prefix to protect main DeepSeek/Grok high-model cache.
Cache hit % surfaced in UX: post-compact messages now show "Context compacted → XXk · 缓存命中 87%" (or cache read count); /provider roles output mentions the observability.
Full Reasonix adoption: 65% proactive compact for large windows, 10min summary timeout + emergency truncate guarantee, separate planner/executor sessions, high-discipline prompts for native chat path.
Updated model-roles docs with deepseek recs (planModel/summarize to cheap chat variants); all tests + npm run check green (1845 pass, 0 fail).

8.17.02026-05-29

Terminal title now shows live token usage (e.g. 323K / 512K 63% with high-usage markers).
New compact footer GrokTokenBar with ◆◇ symbol progress, exact token breakdown, remaining to auto-compact, and session stats (Turns / Tool calls / Compactions).
Elegant bottom "Waiting for model... Xs" indicator with live elapsed time and color feedback on long thinking.
Reasoning blocks upgraded to ◇ Thought for Xs (live updating) + diamond-bullet hierarchy for important steps.
Clearer visual distinction: bash commands are bold, important content gets stronger hierarchy vs routine text.

2.2.02026-05-29

Published [email protected] to npm with the new synthesis risk-tier routing layer.
Added public SDK APIs: inferSynthesisRiskTier, inferSkillRiskTier, routeSynthesisCandidates, SynthesisRiskTier, and SynthesisRoutingBuckets.
Surfaced risk tiers in ToolSearch, skill prompts, doctor registry counts, and orchestration buckets without changing engine permissions or schema versions.

8.13.102026-05-22

Skill directories are user content, not config: bypass-immune safetyCheck narrowed to <root>/skills/<name>/SKILL.md only. Business code under .claude/skills/<name>/scripts/ no longer triggers an ask on every Edit.
File permission dialog adds "Yes, always allow edits in <dir>/ for this project" — writes a directory-scoped Edit allow rule to .clavue/settings.local.json so the choice survives restart.
Reactive compact gains CLAUDE_REACTIVE_COMPACT_PRESERVE_GROUPS=<1-10> trailing-window control (default 1). Warning buffer raised 20k → 40k so warning/error/autocompact thresholds are now strictly monotonic, giving real two-tool-round early warning.
Critical-heap cooldown no longer silently skips: when autocompact is suppressed by the 60s window or 10% heap-relief gate, Clavue actively force-GCs once and surfaces a failureReason with a /clear hint to the query loop.
4 contract test files / 15 subtests locked in; npm run check: 1634 tests / 1628 pass / 6 skipped / 0 fail.

8.8.1182026-05-04

Added GPT-5.5 as the default OpenAI-native max-capability preset.
Added adaptive effort and verbosity gears so high-risk development work gets more model capability while lightweight routes downshift for speed.
Kept Responses function calling, reasoning effort, text verbosity, package docs, and release surfaces aligned for GPT/Codex gateways.

8.8.1172026-05-03

Updated the website, package README, and pinned install commands to the current Clavue release line.
Kept public npm, npx, curl, docs, and changelog surfaces aligned around one version.
Reinforced /mao, provider routing, verification, and release-readiness guidance across public docs.

8.8.932026-04-27

Expanded the homepage into a complete product and onboarding story.
Added richer usage examples, provider-routing docs, best practices, and team rollout guidance.
Updated public install commands and product copy to Clavue 8.8.93.

8.8.582026-04-16

8.8.562026-04-09