local-model workers: prompted-tools text mode breaks tool calls + dominates prompt size

[webdevtodayjason]

Found during the 2026-06-10 overnight conductor-demo rehearsal (isolated gateway, LM Studio).

Evidence

• The lmstudio/openai-compat provider path injects ALL tool schemas as prompt text ('You have access to the following functions: {…}') instead of the native tools API field. A 119-tool turn ships a ~43k-token prompt (LM Studio rejected with n_keep 43019 vs n_ctx 4096 until the model was reloaded at 64k).
• Under the execution-worker job prompt, both gemma-4-12b and qwen3.6-35b-a3b narrate intent and emit malformed pseudo-calls ('<|tool_call>call:tasks.search{…}', '[ tasks: list, status: open ]') that never execute → no board change → runs auto-block after maxNoProgressAttempts. A plain foreground chat turn on the same model DID parse a correct call, so the format-following is prompt-sensitive.
• Worker mechanics verified fine end-to-end with the same template (cycle picks up task, simulate-mode external-tool policy enforced, governance metadata recorded) — the failure is purely the tool-call channel on local providers.

Asks

1. Use native tool-calling (OpenAI 'tools' field) for lmstudio/ollama/openai-compatible providers where the server supports it (LM Studio does).
2. This is also ~the entire perf: 113-tool JSON-schema block (~17k tokens) sent on every turn regardless of intent #407 prompt-size tax made measurable: tool schemas as text dominate the 43k. Selective injection + native tools together make small local models viable workers — which matters for client-site Mac minis with no cloud keys.

Related: #407, #423 (silent tool-skip class). Demo kit: scripts/demo/conductor-demo.mjs (PR #441).

[webdevtodayjason]

Diagnosis correction (verified against pi upstream source, 2026-06-11):

pi-ai's openai-completions provider ALWAYS sends native tools — buildParams sets params.tools = convertTools(context.tools, compat) (openai-completions.ts:538-539), and convertTools (:993) produces standard ChatCompletionTool[]. There is no text-injection mode in the provider layer.

The 'You have access to the following functions' text observed in lms log stream is LM Studio's own chat-template rendering of those natively-passed tools — that's the normal mechanism by which local models consume tool definitions, not a bug. Proof the native round trip works: a plain foreground chat turn on qwen3.6-35b-a3b through the full Argent stack executed a real tasks tool call.

Revised problem statement: under the execution-worker job prompt specifically, gemma-4-12b and qwen3.6-35b emit off-format pseudo-calls that LM Studio's parser doesn't recognize (they emit CORRECT calls under plain chat). So the fix surface is worker-prompt format adherence for small models (prompt shape/sampling/template-parser config), not the provider layer. The prompt-size tax (119 tool defs rendered into the template ≈ most of the 43k tokens) remains real but its fix is #407 selective injection.

Retitling implications: this issue = 'local models drop tool-call format under the worker job prompt'.

[webdevtodayjason]

Contract v2 results (2026-06-11 overnight, isolated rig, 6-ticket conductor):

• Worker guard shipped (PR fix(worker): textual pseudo-tool-call detection + format correction (#442) #444): native-call-only directive in the execution prompt + detection of the observed pseudo-call dialects + FORMAT CORRECTION re-prompt on the next attempt + diagnostic log line, before honest auto-block. Detection fired correctly on both models.
• Verdict under the current ~40k-token worker prompt: gemma-4-12b-QAT and qwen3.6-35b-a3b both still fail to hold native tool-call format even WITH correction → local-model conductors are gated on the perf: 113-tool JSON-schema block (~17k tokens) sent on every turn regardless of intent #407 prompt shrink, not on provider plumbing (native tools confirmed) and not fully on model size.
• Quant matters a lot: plain-turn A/B on the same machine — gemma-4-12b-QAT (Q4_0) fires a correct native tool call and answers (114s round trip incl. 64k-ctx prefill); gemma-4-12b Q4_K_M emits the textual pseudo-call dialect even on a plain turn (115s). Use the QAT for any tool work.
• gemma-QAT's plain-turn success + worker-prompt failure brackets the problem precisely: it's prompt-mass-induced format drift. Strong signal that post-perf: 113-tool JSON-schema block (~17k tokens) sent on every turn regardless of intent #407 a 12B QAT local conductor is viable.

Also filed: runNow no-ops while a blocked task exists (separate issue).