feat(config): expose full pi-ai model compat fields in config schema

[ikari-pl]

Summary

• Expose all model.compat fields that pi-ai already supports at runtime but the config schema rejects
• Adds 7 fields to ModelCompatConfig type and ModelCompatSchema zod validator

Problem

The pi-ai openai-completions provider reads 11 compat fields from model.compat via getCompat(), but OpenClaw's config schema only exposes 4 of them (types.models.ts:9-14, zod-schema.core.ts:13-23). Because the zod schema uses .strict(), any user who sets one of the missing 7 fields in openclaw.json gets a validation error on gateway startup.

What changed

Added fields:

Field	Type	Used at	Purpose
thinkingFormat	"openai" | "zai" | "qwen"	L360-L368	Controls how thinking/reasoning is requested. "qwen" sends enable_thinking: true (Qwen3/DashScope), "zai" sends thinking: { type: "enabled" }, "openai" sends reasoning_effort (default).
supportsStrictMode	boolean	L513	Whether to include strict: false in tool definitions. Some endpoints reject unknown fields.
supportsUsageInStreaming	boolean	L291	Whether to request stream_options: { include_usage: true }.
requiresToolResultName	boolean	L462	Whether tool result messages need a name field (Mistral).
requiresAssistantAfterToolResult	boolean	L413	Whether to insert synthetic assistant messages after tool results.
requiresThinkingAsText	boolean	L387	Whether to convert thinking blocks to plain text (Mistral).
requiresMistralToolIds	boolean	L383	Whether to normalize tool call IDs to 9-char alphanumeric (Mistral).

Relation to open PRs

This is a superset of two open PRs that each add a subset of these fields:

• #6683 adds supportsStrictMode only (1 of 7 missing fields)
• #9451 adds thinkingFormat plus auto-detection for DashScope URLs (1 of 7 missing fields + auto-detect logic)

This PR exposes the complete set of 7 missing fields so the config schema stays in sync with what pi-ai already supports at runtime. Auto-detection (as in #9451) is orthogonal and can be layered on top.

Example config

{
  "models": {
    "providers": {
      "lmstudio": {
        "baseUrl": "http://127.0.0.1:1234/v1",
        "api": "openai-completions",
        "models": [{
          "id": "qwen3-32b",
          "name": "Qwen3 32B",
          "reasoning": true,
          "compat": {
            "thinkingFormat": "qwen",
            "supportsStrictMode": false,
            "supportsDeveloperRole": false
          }
        }]
      }
    }
  }
}

Test plan

• pnpm build passes
• pnpm test — no existing tests break
• Verified locally: Qwen3-32B on LM Studio with thinkingFormat: "qwen" sends enable_thinking: true and thinking blocks stream correctly
• Setting supportsStrictMode: false omits strict from tool definitions

Greptile Overview

Greptile Summary

• Expands ModelCompatConfig and its Zod validator to accept additional compat flags already supported by the pi-ai openai-completions provider at runtime.
• Adds 7 new optional fields (stream usage/strict mode/thinking format + Mistral/tooling behavior toggles) to keep the config schema in sync with provider behavior.
• No other config sections are modified; schema remains .strict() so only the allowlisted compat keys are accepted.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk.
• Changes are limited to widening the config schema/types with optional fields that were previously rejected due to .strict() validation; no runtime logic changes were introduced, and the additions keep types and Zod schema in sync.
• No files require special attention

_{(2/5) Greptile learns from your feedback when you react with thumbs up/down!}

[Takhoffman]

Merged via squash.

• Merge commit: d134c85
• Verified: pnpm build, pnpm check, pnpm test
• Notes: full test suite passed except transient timeout in extensions/lobster/src/lobster-tool.test.ts during one run; isolated rerun passed.
• Changelog: CHANGELOG.md updated=true required=true opt_out=false

Thanks @ikari-pl!