{
  "models": {
    "providers": {
      "my-ollama": {
        "baseUrl": "http://localhost:11434/v1",
        "models": [{
          "id": "qwen3.5:9b",
          "name": "Qwen3.5 9B",
          "api": "ollama",
          "options": { "num_ctx": 32768 }
        }]
      }
    }
  }
}

## Changed files

- `src/agents/ollama-stream.test.ts` (modified, +154/-0)
- `src/agents/ollama-stream.ts` (modified, +87/-5)
- `src/agents/pi-embedded-runner/run/attempt.ts` (modified, +20/-2)
- `src/config/types.models.ts` (modified, +1/-0)
- `src/config/zod-schema.core.ts` (modified, +1/-0)


---

# PR #51356: fix(agents): accept model `options` in config and safely recover Ollama tool calls (#50894)

- Repository: openclaw/openclaw
- Author: hhhhao28
- State: open | merged: False
- Link: https://github.com/openclaw/openclaw/pull/51356

## Description (problem / solution / changelog)

## Summary

- Problem: Users configuring local Ollama models (e.g. Qwen3.5-9B) in `openclaw.json` cannot pass provider-specific options like `num_ctx` — the config validator rejects the key, and even if the schema were relaxed, `applyConfiguredProviderOverrides` drops the field so it never reaches the runtime. Separately, smaller models that understand tools semantically but lack native `tool_calls` output silently fail to dispatch tool calls.
- Why it matters: Without `num_ctx` passthrough, Ollama defaults to a 2048–4096 token context window, silently truncating conversation history and appearing as complete memory loss. Without tool-call recovery, local models cannot execute tools at all.
- What changed:
  - `ModelDefinitionSchema` now accepts an optional `options` record for provider-specific passthrough settings.
  - `applyConfiguredProviderOverrides` in `model.ts` now propagates `options` so user config reaches the runtime.
  - `options.num_ctx` is used as a fallback for `contextWindow` in both the native Ollama stream path and the OpenAI-compat `num_ctx` injection path. The extraction is done once and reused in both places.
  - New `recoverToolCallsFromText()` parser in `ollama-stream.ts` extracts structured tool calls from `<tool_call>` XML tags only (fenced JSON code blocks are intentionally excluded per security review). Recovered markup is stripped from text content to prevent history pollution.
- What did NOT change (scope boundary): Cloud provider paths, existing `contextWindow` behavior when explicitly set, tool schema generation, the OpenAI-completions/responses tool-call parsing, and fenced JSON code blocks are never interpreted as tool calls.

## Change Type (select all)

- [x] Bug fix
- [ ] Feature
- [ ] Refactor
- [ ] Docs
- [ ] Security hardening
- [ ] Chore/infra

## Scope (select all touched areas)

- [x] Gateway / orchestration
- [x] Skills / tool execution
- [ ] Auth / tokens
- [ ] Memory / storage
- [x] Integrations
- [x] API / contracts
- [ ] UI / DX
- [ ] CI/CD / infra

## Linked Issue/PR

- Closes #50894
- Supersedes #51094 (closed — addressed all bot review feedback from that PR)

## User-visible / Behavior Changes

1. Users can now add `"options": {"num_ctx": 8192, ...}` to model definitions in `openclaw.json` without config validation errors.
2. The `options` field is properly propagated through model resolution (previously dropped by `applyConfiguredProviderOverrides`).
3. When `contextWindow` is not set but `options.num_ctx` is, the value is used as context window size — preventing silent conversation history truncation.
4. Smaller local models that emit `<tool_call>{"name":"write_file",...}</tool_call>` in text output now have those calls recovered and dispatched. The raw markup is stripped from the visible text to keep conversation history clean.

## Security Impact (required)

- New permissions/capabilities? `No`
- Secrets/tokens handling changed? `No`
- New/changed network calls? `No`
- Command/tool execution surface changed? `No` — recovered tool calls go through the same dispatch and policy pipeline as native tool calls. Only explicit `<tool_call>` XML tags are matched; fenced JSON code blocks are intentionally excluded to prevent accidental execution of explanatory examples.
- Data access scope changed? `No`

## Repro + Verification

### Environment

- OS: Windows 11 + WSL 2 + Ubuntu 24.04
- Runtime/container: Ollama (local)
- Model/provider: qwen3.5:9b via Ollama
- Relevant config (redacted):

```json
{
  "models": {
    "providers": {
      "my-ollama": {
        "baseUrl": "http://localhost:11434",
        "api": "ollama",
        "models": [{
          "id": "qwen3.5:9b",
          "name": "Qwen3.5 9B",
          "options": { "num_ctx": 32768 }
        }]
      }
    }
  }
}

## Changed files

- `docs/.generated/config-baseline.json` (modified, +19/-0)
- `docs/.generated/config-baseline.jsonl` (modified, +3/-1)
- `docs/.generated/plugin-sdk-api-baseline.json` (modified, +4/-4)
- `docs/.generated/plugin-sdk-api-baseline.jsonl` (modified, +4/-4)
- `src/agents/context-window-guard.test.ts` (modified, +38/-0)
- `src/agents/context-window-guard.ts` (modified, +11/-4)
- `src/agents/ollama-stream.test.ts` (modified, +217/-0)
- `src/agents/ollama-stream.ts` (modified, +88/-2)
- `src/agents/pi-embedded-runner.e2e.test.ts` (modified, +1/-0)
- `src/agents/pi-embedded-runner.run-embedded-pi-agent.auth-profile-rotation.e2e.test.ts` (modified, +1/-0)
- `src/agents/pi-embedded-runner/compact.ts` (modified, +14/-3)
- `src/agents/pi-embedded-runner/extensions.ts` (modified, +8/-9)
- `src/agents/pi-embedded-runner/model.test.ts` (modified, +48/-0)
- `src/agents/pi-embedded-runner/model.ts` (modified, +6/-1)
- `src/agents/pi-embedded-runner/run.overflow-compaction.harness.ts` (modified, +1/-0)
- `src/agents/pi-embedded-runner/run.ts` (modified, +21/-2)
- `src/agents/pi-embedded-runner/run/attempt.ts` (modified, +17/-12)
- `src/config/defaults.ts` (modified, +69/-2)
- `src/config/model-alias-defaults.test.ts` (modified, +99/-0)
- `src/config/schema.base.generated.ts` (modified, +7/-0)
- `src/config/types.models.ts` (modified, +1/-0)
- `src/config/zod-schema.core.ts` (modified, +1/-0)