文件最后提交记录最后更新时间
fix(router): preserve session store across config reloads and strengthen judge prompt - Allow external SessionRouterStore injection via RouterRuntimeDeps so the store survives config-reload cycles that destroy and recreate RouterRuntime - Skip clearing externally-owned store on shutdown() - Remove deterministic isShortContinuation bypass; let the judge LLM decide - Strengthen judge prompt with CRITICAL RULE for continuation messages so the model correctly inherits the previous tier instead of downgrading Co-authored-by: Cursor <cursoragent@cursor.com> 14 天前
feat(router): introduce in-process router module with scenario routing, fallback, and zero-usage retry Replace the agent's direct model dependency and AgentRecoveryPolicy with a new src/router/ module that owns scenario decision, sticky session state, fallback chain, zero-usage retry, and (skeleton) tokenSaver / orchestrate / customRouter / stats. Wire it into PolitConfig, createLocalGateway, and the agent loop; update e2e + agent test helpers; add docs/router/ and full tests/router/ unit suites. Co-authored-by: Cursor <cursoragent@cursor.com> 24 天前
fix(router): preserve session store across config reloads and strengthen judge prompt - Allow external SessionRouterStore injection via RouterRuntimeDeps so the store survives config-reload cycles that destroy and recreate RouterRuntime - Skip clearing externally-owned store on shutdown() - Remove deterministic isShortContinuation bypass; let the judge LLM decide - Strengthen judge prompt with CRITICAL RULE for continuation messages so the model correctly inherits the previous tier instead of downgrading Co-authored-by: Cursor <cursoragent@cursor.com> 14 天前
fix(router): prevent duplicate content yield and harden tokenSaver judge - Track lastHasYieldedContent to avoid re-yielding buffered events on fallback when the primary attempt already streamed content to the caller - Add retry loop (max 2 attempts) in classifyAndRoute for empty/unparseable judge responses before falling back to defaultTier - Strip markdown code fences from judge output before parsing tier tag - Add tests for code-fence stripping, empty-response retry, and double-empty fallback Co-authored-by: Cursor <cursoragent@cursor.com> 19 天前