v0.7.11

Patch release for routing correctness and provider checks.

• Filter image/audio/video/embedding catalog entries out of chat routing pools.
• Add current Gemini/xAI model aliases and avoid fuzzy-mapping chat requests to image-preview models.
• Add AUTO/FAST cost sanity guard so exploration cannot prefer materially more expensive same-quality candidates.
• Treat OpenAI nano models as economy in reasoning lanes.
• Prevent short implementation prompts from being flattened to SIMPLE while avoiding premium over-escalation.
• Add provider-check User-Agent headers to avoid false 403s from Cloudflare/WAF rules.

Validation:

• Full Python suite: 493 passed, 1 warning.
• Routing/provider targeted suite: 97 passed.
• PyPI build + twine check passed.
• npm pack dry-run passed.

v0.7.10

Routing bugfix release.

• Enforce served-quality floors so high-risk routing cannot leak economy candidates.
• Normalize Moonshot/Kimi provider aliases to balanced quality.
• Route explicit reasoning/thinking requests through reasoning-aware features and medium floors where appropriate.
• Treat malformed or incomplete upstream pricing as conservative unknown pricing, not free.
• Disable bandit randomness for tool/agent steps even when clients do not provide a session id.
• Keep Dashboard route preview aligned with the live selector and avoid lifecycle pollution.

Validation:

• python -m pytest -q: 488 passed, 1 warning
• python3 -m twine check release artifacts: passed
• npm pack --dry-run: passed

v0.7.9

Fixes\n- make the feedback prior strength configurable in routing bandit config\n- reduce the default local feedback prior from 20 virtual samples to 5 so single-user feedback influences routing much sooner\n- add regression coverage proving a few weak signals can overturn a stale benchmark prior in local mode\n\n## Validation\n- PYTHONPATH=. python -m pytest tests -q\n- python3 -m build\n- python3 -m twine check dist/*\n- cd openclaw-plugin && npm pack --dry-run

v0.7.8

Patch release focused on remaining open bug issues.

Included in 0.7.8:

• fix OpenAI-to-Anthropic stream conversion when tool call deltas arrive before an id
• resolve legacy internal model ids against current live successors so doctor no longer reports false unresolved mappings
• bypass Anthropic-style cache hints for Gemini on OpenAI transport to avoid cached-content request shape conflicts
• add NO_PROXY guidance in Claude Code setup and doctor for local proxy environments

Validation:

• 441 passed, 1 warning

v0.7.7

Patch release for issue #20.

• Lock Anthropic thinking follow-ups to the previous session model
• Fail closed when the previous thinking model is unavailable
• Speed up feedback and Anthropic compatibility tests

v0.7.6

Patch release for Anthropic native thinking compatibility.

Highlights:

• Preserve Anthropic thinking and redacted_thinking blocks during internal compatibility conversion.
• Force native source-body reuse for Anthropic requests that contain thinking blocks so signatures are not regenerated or dropped.
• Add regression coverage for thinking block preservation and native Anthropic reuse.

v0.7.5

Patch release for Anthropic native caching reliability.

Highlights:

• Fix Anthropic cache breakpoint TTL ordering so injected cache_control blocks do not produce invalid 1h-after-5m request bodies.
• Preserve existing client-provided Anthropic cache_control metadata instead of overwriting it.
• Add regression tests for Anthropic cache breakpoint behavior.

v0.7.4

Highlights

• add absolute served-quality buckets with capability lanes
• keep strong continuity on Anthropic tool follow-up turns
• expose served quality and lane details in traces, stats, and dashboard

v0.7.3

Transport-safe Anthropic tool routing patch release.

Highlights:

• filter Anthropic tool/agentic requests to models that can preserve native Anthropic tool semantics
• fail closed with a clear routing constraint error when no transport-safe model exists
• add a reusable UncommonRoute release skill plus version-sync helper for future releases

v0.7.2

Highlights

• Add transport-aware routing so protocol selection is routed alongside model selection instead of defaulting everything to OpenAI-compatible transport
• Surface transport diagnostics in the Dashboard with a request-level EXPLAIN view and aggregate TRANSPORT MIX activity view
• Preserve Anthropic-style tool follow-up payloads for supported transports and record transport reasoning in traces for support/debug workflows

Packages

• Python: uncommon-route==0.7.2
• npm: @anjieyang/uncommon-route@0.7.2