feat(providers): add chat-completions provider (OpenAI Chat Completions-compatible endpoints)

[Brian Krabach (bkrabach)]

Summary

Adds chat-completions as a first-class provider in amplifier-agent's provider catalog, wiring up the existing official module amplifier-module-provider-chat-completions. It speaks the OpenAI Chat Completions wire format (/v1/chat/completions) and targets any compatible server — llama.cpp, vLLM, LM Studio, LocalAI, SGLang, TGI, and other OpenAI-compatible endpoints.

The module is listed in the ecosystem MODULES.md but was not selectable from amplifier-agent, whose provider set is hardcoded. ISSUES.md already notes this seam ("adding a provider requires editing five literals … nothing enforces agreement across them"); this PR makes them agree for chat-completions.

Note: this is distinct from the existing openai provider, which uses the OpenAI Responses API and therefore cannot be pointed at a llama.cpp/vLLM server. Filling that gap is the point of this provider.

What changed

Code

• config/loader.py: add chat-completions to _VALID_PROVIDER_MODULES so provider.module: chat-completions passes config validation.
• provider_sources.py: add to KNOWN_PROVIDERS and PROVIDER_CATALOG; add a dedicated resolve_credential_detailed branch that reads CHAT_COMPLETIONS_BASE_URL (required) and optional CHAT_COMPLETIONS_API_KEY. A dedicated branch (rather than a PROVIDER_CREDENTIAL_VARS entry) is required because the generic path routes the primary env var into api_key; here the value must land in base_url.
• bundle.md: declare provider-chat-completions in the install-only stub list so it cold-prepares uniformly with the other providers.

Docs: README, docs/spec/providers-and-models.md, docs/spec/host-config.md, docs/CONFIGURATION.md, docs/spec/bundle-and-cache.md, docs/LAYERS_AND_RELEASES.md, docs/INTEGRATION.md, skills/amplifier-agent/SKILL.md, docs/architecture/architecture.dot, CHANGELOG.md.

Verification (Digital Twin, real model — no mocks)

Verified end-to-end in an isolated container running the patched build (amplifier-agent 0.12.0) against a real Ollama server serving qwen2.5:0.5b:

• providers list: chat-completions appears; reports resolvable=false/source=none without CHAT_COMPLETIONS_BASE_URL, and flips to resolvable=true/source=env when it is set — exercises both the catalog entry and the credential-resolution branch.
• Loader negative control: an invalid provider.module is rejected with must be one of [… 'chat-completions' …] — proves the loader gate now includes it (an unpatched build rejects chat-completions outright).
• Real routing: amplifier-agent run drove real POST /v1/chat/completions calls into the endpoint (confirmed in Ollama's own request log, all HTTP 200) and real, non-empty completions flowed back through the provider — not a silent fallback to another provider.
• Cold-prepare: bundle prepare succeeds with the 6-provider bundle.md; the prepared bundle declares all six providers including provider-chat-completions.

Honest caveat: a fully-terminated run envelope was not captured because prompt prefill of a 0.5B model on aarch64 CPU (~12 tok/s) makes each agentic call ~4–5 min; the provider path itself is proven by the routing + real-completion evidence above. On a GPU host or a hosted OpenAI-compatible endpoint the same config completes in seconds.