v3.9.67

Changed

• Restructure or_model_info skill to satisfy CPV strict mode

CPV required:

• SKILL.md under 5000 chars (move detail to references/)

• Error Handling and ## Examples sections present

• Description under 250 chars with "Trigger with ..." phrase
• "Copy this checklist and track your progress" phrase
• Reference files with explicit Table of Contents
• Embedded TOC of each referenced file immediately after its link

New reference files under skills/llm-externalizer-or-model-info/references/:

• errors.md — full error table with 7 error codes and resolutions,
  plus debugging tips (partial-name workaround, :free vs paid id
  distinction, :thinking variants)
• example-output.md — complete sample response for
  nvidia/nemotron-3-super-120b-a12b:free with annotated explanation
  of how to read pricing, latency percentiles, throughput percentiles,
  and uptime
• use-cases.md — six primary scenarios: verify supported params,
  compare provider pricing, debug slow calls, check quantization,
  confirm context length, check reasoning support

SKILL.md now 4062 chars with embedded TOC summaries for each
referenced file so progressive discovery can find the sub-content.

CPV result: CRITICAL=0 MAJOR=0 MINOR=0.

• Add or_model_info tool + llm-externalizer-or-model-info skill

New MCP tool that queries OpenRouter's /v1/models/{exact_id}/endpoints
for any model and returns formatted metadata: architecture, per-endpoint
provider info, context length, pricing (per-M-tokens), supported
request-body parameters, quantization, uptime (30m / 1d), latency
percentiles, and throughput.

Required input: model — the EXACT OpenRouter model id, case-sensitive,
including vendor prefix and any :free / :thinking / :beta suffix. Only
works when the active profile is OpenRouter; returns a clear error
with a suggestion to switch profiles otherwise.

The tool is informational only, not an LLM call — not added to
LLM_TOOLS_SET, does not count toward session usage, no rate limiting.

New skill: skills/llm-externalizer-or-model-info/SKILL.md. Triggers on
phrases like "openrouter model info", "what params does X support",
"show pricing for model", "check model support", etc. Walks the caller
through parsing the exact model id (with fallback to asking for
clarification on partial names) and presents the markdown block.

Primary use cases:

• Verify supported_parameters before integrating a new model —
  Nemotron :free accepts reasoning + temperature + top_p but
  NOT frequency_penalty / presence_penalty / top_k / min_p /
  stop. The paid variant supports all of them. Important distinction.
• Compare pricing across multiple providers hosting the same model.
• Debug slow or failing calls by checking current uptime + latency.
• Look up quantization and max token limits for a specific endpoint.

The results are live — no caching on the MCP side. Every call hits
OpenRouter directly. Safe to call repeatedly.