v0.7.21

Changes

• Enforce distinct served-quality pools for Simple, Medium, and Complex routing tiers.
• Prevent Claude Code wrapper/title side-channel prompts from inflating user prompt complexity.
• Reclassify current model families across economy, balanced, and premium tiers.
• Filter superseded Gemini Pro model IDs when newer successors are available.

Verification

• ruff check uncommon_route tests
• pytest: 714 passed, 1 skipped

v0.7.20

Changes

• Avoid routing to superseded Gemini 2.5 Pro when newer Gemini Pro successors are available in the upstream catalog.
• Tighten contextual follow-up floors so short run/commit follow-ups stay simple instead of inheriting medium or complex tiers.
• Parse Anthropic top-level tool names when estimating agentic step risk.
• Keep Claude Code suggestion-mode prompts capped as simple side-channel traffic.
• Refresh README benchmark/runtime guidance and bump Python/OpenClaw npm package versions to 0.7.20.

Verification

• python3 -m pytest, 707 passed, 1 skipped, 1 warning.
• python3 -m twine check /tmp/ur-build-check/*, passed.
• npm pack --dry-run in openclaw-plugin, passed.
• Live local proxy e2e on 8404 with five Claude Code style weather-cli prompts: SIMPLE/COMPLEX/COMPLEX/SIMPLE/SIMPLE, with no google/gemini-2.5-pro selections.

v0.7.19

Changes

• Remove session sticky previous-model route-pool locking from agent/tool routing.
• Keep session context as routing features without forcing candidates to the last successful model.
• Preserve semantic failure observability in traces with verification_failed and failure_kind fields.
• Bump Python and OpenClaw npm package versions to 0.7.19.

Verification

• PYTHONPATH=. python3 -m pytest -q tests/test_proxy.py tests/test_v2_routing_logic.py tests/test_quality_routing.py
• python3 -m twine check /tmp/uncommon-route-dist-0.7.19/*
• npm pack --dry-run
• Live local proxy check on 8404: same session switched between src/weather/api.py and src/weather/cli.py with no session-sticky/session-retry/previous-model route notes.

v0.7.18

Changes

• Remove the Anthropic tool-safe lane and thinking-context model lock from virtual Anthropic Messages routing.
• Allow Anthropic tool requests to route through OpenAI-compatible upstream transports when selected.
• Strip Anthropic thinking blocks from Anthropic-to-OpenAI request conversion while preserving native Anthropic source-body reuse.
• Bump Python and OpenClaw npm package versions to 0.7.18.

Verification

• python3 -m pytest -q
• python3 -m ruff check uncommon_route tests
• python3 -m twine check dist/uncommon_route-0.7.18*

v0.7.17

Highlights

• Tuned routing signal floors to improve simple vs medium/high routing across follow-up, tool, vision, and structured-output cases without broad keyword whitelists.
• Added shared signal tuning helpers and regression coverage for proxy and router behavior.
• Promoted v2 classifier dependencies into the base Python package so installed builds can load packaged embedding/classifier assets reliably.
• Added benchmark/e2e artifacts and scripts used to validate the routing matrix.

Verification

• python3 -m pytest -q -> 697 passed, 1 skipped
• python3 -m ruff check uncommon_route tests -> passed
• python3 -m build --sdist --wheel --outdir dist/0.7.17 -> passed
• python3 -m twine check dist/0.7.17/* -> passed
• npm pack --dry-run -> passed

v0.7.16

Changes

• Release Python and OpenClaw/npm package version 0.7.16.
• Prevent learned runtime fallback aliases from overriding gateway exact model IDs or known seed aliases, fixing OpenRouter cold-start/gateway model resolution when local learned aliases exist.
• Keep public exports and v2 telemetry lifecycle lint-clean; skip lifecycle embedding telemetry when message length is unavailable to preserve short-message privacy.
• Clean up lint issues across tests and support modules so the full ruff suite passes.

Verification

• python -m ruff check uncommon_route tests
• python -m pytest -q -> 672 passed, 1 skipped
• 3 rounds live Commonstack E2E through /health, /v1/chat/completions (auto and fast), and /v1/messages
• python -m build --outdir /tmp/uncommon-route-dist-0.7.16.POdBzx
• python -m twine check /tmp/uncommon-route-dist-0.7.16.POdBzx/*
• npm pack --dry-run

v0.7.15

Patch release focused on OpenRouter upstream compatibility.\n\nChanges:\n- Detect OpenRouter as a gateway so provider/model IDs are preserved.\n- Update legacy model aliases to current OpenRouter catalog IDs.\n- Treat negative upstream pricing sentinel values as unknown pricing instead of routing on negative costs.\n- Add regression coverage for OpenRouter mapping and pricing parsing.\n- Bump Python and OpenClaw package versions to 0.7.15.

v0.7.14

Patch release focused on routing failure recovery and cost-aware agent step handling.

Changes:

• Distinguish semantic verification failures from environment and test-invocation failures.
• Keep environment/invocation recovery capped to medium instead of escalating as reasoning failure.
• Avoid treating read-only source inspection output that mentions failure words as failed verification.
• Add pressure-rescue routing tests and failure-kind coverage.
• Bump Python and OpenClaw package versions to 0.7.14.

v0.7.13

Patch release for v2 routing runtime readiness and quality/cost safeguards.\n\n- Package v2 routing assets in PyPI wheels/sdists so semantic signals work without manual local asset setup.\n- Add asset-loading tests and expand routing regression coverage.\n- Refine quality priors, benchmark routing context, proxy metadata, and AUTO routing safeguards.\n- Sync Python package and OpenClaw/npm plugin versions to 0.7.13.

v0.7.12

Patch release for Anthropic thinking routing correctness and release version reporting.

• Keep Anthropic native requests with thinking blocks inside a thinking-compatible routing pool instead of hard-locking to the previous model.
• Filter models that cannot accept adaptive thinking payloads from Anthropic Messages thinking continuations.
• Preserve previous-model context in trace reasoning as diagnostic metadata without making the whole step sticky.
• Centralize runtime package version reporting so CLI, proxy health, and support bundles use the pyproject/package metadata source of truth.
• Publish matching OpenClaw plugin version 0.7.12.

Validation:

• Full Python suite: 496 passed, 1 warning.
• PyPI build + twine check passed.
• npm pack dry-run passed.
• npm publish 0.7.12 passed.
• PyPI upload 0.7.12 passed.