feat: add native Mistral AI provider support

[edouard-claude]

Summary

• Adds Mistral AI as a first-class provider (18th provider) alongside existing ones
• Uses OpenAI-compatible API at https://api.mistral.ai/v1 with provider-specific model prefix stripping
• Supports both new model_list format and legacy providers config migration
• Fixes tool calling compatibility: removes extra name/arguments fields from tool_calls JSON serialization that Mistral's strict API rejects (OpenAI silently ignores them)
• References: closes the gap identified in feat: Add Mistral AI provider support #246, builds on the approach from feat: add Mistral AI provider, Firecrawl and SerpAPI tools #193

Changes

• pkg/config/config.go: Add Mistral ProviderConfig to ProvidersConfig, IsEmpty(), HasProvidersConfig()
• pkg/config/defaults.go: Add mistral-small entry in default model_list
• pkg/config/migration.go: Add mistral migration entry for ConvertProvidersToModelList()
• pkg/config/migration_test.go: Update AllProviders test (17 → 18)
• pkg/providers/factory_provider.go: Add "mistral" to OpenAI-compat protocol list + getDefaultAPIBase()
• pkg/providers/factory.go: Add "mistral" case in legacy resolveProviderSelection()
• pkg/providers/openai_compat/provider.go: Add "mistral" to prefix strip list
• pkg/providers/protocoltypes/types.go: Fix ToolCall.Name and ToolCall.Arguments serialization (json:"-") — these internal fields were incorrectly included in API payloads, causing 422 errors on strict APIs like Mistral
• pkg/migrate/config.go: Add "mistral" to supportedProviders
• config/config.example.json: Add mistral config section

Configuration

{
  "model_list": [
    {
      "model_name": "mistral-small",
      "model": "mistral/mistral-small-latest",
      "api_key": "YOUR_API_KEY",
      "api_base": "https://api.mistral.ai/v1"
    }
  ]
}

E2E Test Logs

Simple chat (model=mistral-small-latest):

[INFO] agent: Registered agent {agent_id=main, name=Main Agent, model=mistral-small-latest}
[INFO] agent: LLM response without tool calls (direct answer) {iteration=1, content_chars=40}
[INFO] agent: Response: I'm picoclaw, your helpful AI assistant.

Tool call — read_file:

[INFO] agent: LLM requested tool calls {tools=[read_file], count=1, iteration=1}
[INFO] agent: Tool call: read_file({"path":"/Users/edouard/.picoclaw/workspace/test.txt"})
[INFO] tool: Tool execution completed {tool=read_file, duration_ms=0, result_length=23}
[INFO] agent: LLM response without tool calls (direct answer) {iteration=2, content_chars=61}
[INFO] agent: Response: The file test.txt contains the text: "test file for picoclaw"

Tool call — list_dir:

[INFO] agent: LLM requested tool calls {tools=[list_dir], count=1, iteration=1}
[INFO] agent: Tool call: list_dir({"path":"/Users/edouard/.picoclaw/workspace"})
[INFO] tool: Tool execution completed {tool=list_dir, duration_ms=0, result_length=146}
[INFO] agent: Response: Here are the files and directories in my workspace: AGENT.md, IDENTITY.md, SOUL.md, USER.md, aieos.json, test.txt, memory/, sessions/, skills/, state/

Tool call — exec (shell):

[INFO] agent: LLM requested tool calls {tools=[exec], count=1, iteration=1}
[INFO] agent: Tool call: exec({"command":"echo hello from mistral"})
[INFO] tool: Tool execution completed {tool=exec, duration_ms=6, result_length=19}
[INFO] agent: Response: The output of the command is: hello from mistral

Test plan

• All existing tests pass (go test ./... — 0 failures)
• New migration test passes (18 providers)
• gofmt, go vet clean
• E2E simple chat with mistral-small-latest
• E2E tool calling: read_file, list_dir, exec — all working
• No breaking changes to existing providers (extra fields were silently ignored by OpenAI)

[harshbansal7]

End-to-end tested with mistral-small-latest model via picoclaw agent

@edouard-claude Can you add logs that were generated during that chat in the description just to be sure?

[lukemilby]

@edouard-claude does tool calling work with the HTTP provider? Following @harshbansal7 comment. Would like to see that or create a test ensuring tool calling work.

[edouard-claude]

Thanks @harshbansal7 @lukemilby for the great feedback! Your comments actually led me to discover a tool calling compatibility issue — Mistral's API is stricter than OpenAI's and rejects extra fields (name, arguments) at the top level of tool_calls in assistant messages. I pushed a fix that cleans up the JSON serialization in protocoltypes/types.go (fields are still available internally, just not serialized to the API payload).

Here are the E2E logs with tool calling working:

Simple chat:

[INFO] agent: LLM response without tool calls (direct answer) {agent_id=main, iteration=1, content_chars=40}
[INFO] agent: Response: I'm picoclaw, your helpful AI assistant.

Tool call — read_file:

[INFO] agent: LLM requested tool calls {agent_id=main, tools=[read_file], count=1, iteration=1}
[INFO] agent: Tool call: read_file({"path":"/Users/edouard/.picoclaw/workspace/test.txt"})
[INFO] tool: Tool execution completed {tool=read_file, duration_ms=0, result_length=23}
[INFO] agent: Response: The file test.txt contains the text: "test file for picoclaw"

Tool call — list_dir:

[INFO] agent: LLM requested tool calls {agent_id=main, tools=[list_dir], count=1, iteration=1}
[INFO] agent: Tool call: list_dir({"path":"/Users/edouard/.picoclaw/workspace"})
[INFO] tool: Tool execution completed {tool=list_dir, duration_ms=0, result_length=146}
[INFO] agent: Response: Here are the files and directories in my workspace: AGENT.md, IDENTITY.md, SOUL.md...

Tool call — exec (shell):

[INFO] agent: LLM requested tool calls {agent_id=main, tools=[exec], count=1, iteration=1}
[INFO] agent: Tool call: exec({"command":"echo hello from mistral"})
[INFO] tool: Tool execution completed {tool=exec, duration_ms=6, result_length=19}
[INFO] agent: Response: The output of the command is: hello from mistral

All 3 tool types (read_file, list_dir, exec) work correctly with mistral-small-latest. I'll update the PR description with these logs too.

[PixelTux]

This change hides Name and Arguments from JSON serialization, which could be a breaking change for downstream consumers. Could you share the reasoning behind this decision and whether it’s been tested with other setups outside Mistral?

[edouard-claude]

Good question! These two fields (Name and Arguments) are duplicates of Function.Name and Function.Arguments — they were populated internally after JSON parsing (see the deserialization logic in the provider code) but were never meant to be serialized back to the API.

The OpenAI spec only expects tool_calls[].function.name and tool_calls[].function.arguments — the top-level name/arguments fields don't exist in the spec. OpenAI silently ignores them, but Mistral's stricter validation returns a 422 error.

Since the data already lives in Function.Name/Function.Arguments (which remain serialized normally), removing the duplicates from JSON output doesn't change any behavior — it just stops sending fields that shouldn't be there.

Tested with: Mistral (mistral-small-latest) including tool calling (read_file, list_dir, exec). Other OpenAI-compatible providers are unaffected since they already ignore these extra fields.

[PixelTux]

Thanks for the clarification—everything looks good now. LGTM!

[vietvudanh]

Looks great, I was testing with new model_list and have 422 too on tools as Mistral is not following strictly with OpenAI specs. I was about to code then meet this PR, looks good.

[lukemilby]

Thanks! Looks good

[alexhoshina]

LGTM

[alexhoshina]

This fix is already in main, perhaps we can keep the commits clean by rebasing?

[edouard-claude]

Done! Rebased on main and dropped the shell_test.go commit since it's already in main. Clean 2-commit history now.

[nikolasdehor]

This addresses the feature request in #246 (Add Mistral AI provider support). Linking for tracking.