openclaw - 💡(How to fix) Fix infer model run should expose a flag to skip agent-context wrapping for raw provider testing [2 comments, 3 participants]

Problem

openclaw infer model run --local --model <provider>/<model> --prompt "..." is documented as "Run a one-shot model turn" and the source path in src/cli/capability-cli.ts:580-680 looks like a thin builder that would dispatch only [{ role: "user", content: <prompt> }]. In practice the embedded runtime it hands off to re-wraps the request with the default agent's full context (bootstrap files, memory chunks, skills, identity profile) before dispatching to the provider.

This makes the command unusable for testing local provider backends with constrained context windows — the very use case --local looks designed for. On a Hailo-NPU host with a HEF-compiled model (effective context ~2-4K tokens), a 5-word user prompt becomes a ~13K-token request after agent wrapping, overflowing the HEF and surfacing as terminated | other side closed.

Reproduction

Host with:

• ~/.openclaw/openclaw.json agent default has any non-trivial bootstrap (e.g., AGENTS.md ≈ 10K chars per openclaw doctor).
• models.providers.ollama.models[*] includes a small-effective-context model. We use qwen2:1.5b served by hailo-ollama on 127.0.0.1:8000.
• models.providers.ollama.models[<qwen2>].params.num_ctx = 2048 set (working as of 2dba9e6a76 in v2026.4.26).

Run: ```bash openclaw infer model run --local --model ollama/qwen2:1.5b --prompt "Reply: pong" --json ```

Expected: thin model-only call, ~5-10 input tokens, fast response.

Actual: session metadata records inputTokens: 13074, status: running, embedded runtime retries 4× with same runId, every attempt ends in embedded_run_agent_end / error: 'terminated | other side closed', command never returns. Direct POST to the same hailo-ollama endpoint with the same model + same num_ctx returns coherent text in ~1s, confirming the provider itself is healthy — the problem is openclaw's pre-dispatch context inflation.

Same behavior with --gateway (where promptMode: "none", modelRun: true is sent in the dispatch payload but doesn't actually skip context loading).

Suggested fix

Add --skip-agent-context (or --raw) to infer model run:

• When set, dispatch with literally [{ role: "user", content: <prompt> }] — no bootstrap, no memory, no skills, no identity injection.
• Useful for: provider conformance testing, debugging provider-specific quirks (context size, tool support, encoding), confirming a provider is reachable on a host without agent-runtime noise interfering with the diagnosis.

--gateway should honor the same flag end-to-end (the promptMode: "none" path already implies this intent; just isn't actually wired through to skip the embedded runtime's context loader).

Environment

• openclaw 2026.4.26 (be8c246)
• node 22.22.2
• Linux 6.12.75-rpi (Raspberry Pi 5, arm64)
• Local provider: hailo-ollama on 127.0.0.1:8000

Related

2dba9e6a76 (v2026.4.26) added params.num_ctx per-model override which addresses runtime-side context allocation — verified working on this host. The agent-prompt-inflation issue is the remaining half. Conceptually related to closed #70782 ("slim mode to pass --bare").