@matheuskrumenauer/tanya 0.10.0-beta.0 → 0.11.0-beta.0

package/README.md

@@ -267,6 +267,55 @@ schema-invalid tool responses before they reach model history.
 See [docs/mcp.md](./docs/mcp.md) for the full schema, transports, server tools,
 and security model.
 
+## Multi-model routing
+
+Tanya can route each agent step to a different provider/model. Planning and
+simple tool-call turns can use cheap chat models, while synthesis,
+verification, and reasoning turns can use stronger models only when needed.
+
+Default route profile:
+
+| Step | Route | Fallback |
+| --- | --- | --- |
+| `planning` | `deepseek/deepseek-chat` | `qwen/qwen3-coder-plus` |
+| `tool_call` | `deepseek/deepseek-chat` | `groq/llama-3.3-70b-versatile` |
+| `synthesis` | `deepseek/deepseek-reasoner` | `openai/gpt-4.1-mini` |
+| `verification` | `deepseek/deepseek-reasoner` | `openai/gpt-4.1-mini` |
+| `reasoning` | `deepseek/deepseek-reasoner` | `qwen/qwen3-coder-plus` |
+
+Project routes live in `.tania/routes.json`; user-global routes live in
+`~/.tanya/routes.json` with a legacy read fallback from `~/.tania/routes.json`.
+Use `/route` in the REPL to inspect the effective table, `/route show
+<stepType>` to inspect one step, `/route set <stepType> <provider>/<model>` for
+a session-only patch, and `/route reset` to clear session patches.
+
+Escalations are visible: if a cheap route exhausts the malformed tool-call
+repair budget, Tanya emits `escalation_event` and uses the route fallback once,
+up to `TANYA_ESCALATION_CAP` per session.
+
+See [docs/routing.md](./docs/routing.md) for schema, examples, context-window
+guards, per-tool model overrides, and sub-agent model pins.
+
+## Reasoning models
+
+Reasoning routes such as `deepseek-reasoner`, `qwen3-thinking-*`, and
+`grok-3-reasoning` are handled as a separate stream. Tanya archives reasoning to
+`.tania/runs/<runId>/reasoning.jsonl`, emits `reasoning_chunk` events, and keeps
+assistant history reasoning-free so replay and verifier inputs stay stable.
+
+Reasoning tokens appear separately in `/cost` and `/budget`. Route rules can set
+`reasoningCap.maxTokens`; built-in defaults are 2k for planning-like turns and
+8k for synthesis/verification/reasoning turns. If the cap is exceeded, Tanya
+emits `reasoning_truncated` and asks the model to finish.
+
+Use `/memory --reasoning <runId>` to inspect archived reasoning. Use
+`TANYA_HIDE_REASONING=1` to hide reasoning from the human UI while preserving
+JSONL/Cosmo events. Verifier reasoning annotations are off by default; enable
+them with `--verbose-verifier` or `TANYA_VERIFIER_INCLUDE_REASONING=1`.
+
+See [docs/reasoning.md](./docs/reasoning.md) for provider notes, billing math,
+budget defaults, and UX modes.
+
 `--verify` adds required verification commands to the run context. Tanya must run and report each exact command before finishing the coding task.
 
 `tanya benchmark run --all` currently exercises 27 executable low-to-medium regression fixtures: targeted edits, new files, dependency/lockfile updates, framework-style migrations, failing-test repair, frontend smoke checks, artifact/context reuse, streaming long-tool execution, compaction-boundary recovery, run-history logging, dirty worktrees, report repair, and the CosmoHQ mobile/backend smoke profiles.