feat(wiring): OpenAI-compatible local-model adapter (Ollama / LM Studio / vLLM / …)

[telivity-otaip]

Summary

One adapter, many backends. createOpenAICompatibleCaller targets any HTTP server speaking the OpenAI-compatible /v1/chat/completions API — Ollama, LM Studio, vLLM, llama.cpp server, OpenRouter, Azure OpenAI. This unlocks ASIL for regulated industries (healthcare, gov, finance, defense) that can't send code to cloud APIs.

What ships

Adapter (in `packages/asil-runners/src/wiring.ts`):

```ts
createOpenAICompatibleCaller({
baseUrl: 'http://localhost:11434/v1',
apiKey: 'optional',
modelId: 'llama3.1:8b-instruct-q4_K_M',
estimateTokens: t => Math.ceil(t.length / 4), // optional override
}): LLMCaller

createOpenAICompatibleCodexCaller({ ... }): CodexCaller // for the adversarial gate
```

Shares a small internal POST helper with the existing `createCodexCaller` — same fetch pattern, same error-shape.

Env switching (run-a + run-b):

Var	Effect
`ASIL_LLM_BASE_URL`	Use local adapter for primary LLM
`ASIL_LLM_MODEL`	Model id passed verbatim to the server
`ASIL_LLM_API_KEY`	Optional Bearer
`ASIL_CODEX_BASE_URL` / `ASIL_CODEX_API_KEY`	Same for adversarial gate

When `ASIL_LLM_BASE_URL` is unset, behaviour is unchanged from cloud-only deployments.

loadEnv relaxation: `ANTHROPIC_API_KEY` is no longer required when `ASIL_LLM_BASE_URL` is set. Air-gapped users don't have to set a dummy value.

Cost-controller graceful: `calculateCallCost` returns `Decimal(0)` when the model id isn't in the pricing table. Token caps still bite — wire cost is $0 in local mode, but token counts (from the server's `usage` field, or chars/4 fallback) still flow through.

What this changes for cloud users

Nothing. Without the env vars set, run-a and run-b construct the same Anthropic + OpenAI callers they always did. The adapter additions are pure-add.

Walkthrough

New `examples/local-llm.md` covers:

• Ollama setup (full env block + `ollama pull` + verification)
• LM Studio (port 1234 default)
• vLLM (GPU server flag)
• Mixed deployments — cloud-for-execution + local-for-adversarial-gate (and reverse)
• Cost accounting in local mode (token caps still enforce, dollars report as $0)
• When NOT to use local mode (production grinds against high-value codebases, same-family adversarial gate)

Tests

• Before: 363 (47 + 45 + 165 + 54 + 52)
• After: 371 (47 + 45 + 165 + 54 + 60). New tests in `wiring.test.ts`:
  • `createOpenAICompatibleCaller` — request shape (POST to `/chat/completions` with system+user messages), Authorization header presence/absence, token-usage parsing from server, chars/4 estimation fallback when usage absent, custom estimator override, trailing-slash baseUrl normalization, non-OK error handling
  • `createOpenAICompatibleCodexCaller` — single user message + `{content}` return shape (no token surface)
  • `loadEnv` — local-mode skips the ANTHROPIC_API_KEY throw
• Build + typecheck clean across all 5 packages
• No skips, no `.only`

Files

Modified:

• `packages/asil-runners/src/wiring.ts` — `createOpenAICompatibleCaller` + `createOpenAICompatibleCodexCaller` + shared `postOpenAICompatible` helper; `loadEnv` local-mode skip
• `packages/asil-runners/src/run-a.ts` — env-driven adapter selection
• `packages/asil-runners/src/run-b.ts` — env-driven adapter selection
• `packages/asil-cost-controller/src/cost-estimator.ts` — graceful $0 for non-tier model ids
• `packages/asil-runners/src/tests/wiring.test.ts` — 8 new tests

New:

• `examples/local-llm.md`

Updated docs:

• Root `README.md` — new subsections for analyzer, local models, Python profile

Closes the four-PR sequence

This is the fourth of four merged PRs from the planning session:

• feat: canary gates — pre-flight safety verification #4 — canary gates (pre-flight safety verification)
• feat(asil-analyzer): productize 5-mode reasoning-quality analyzer #5 — asil-analyzer (5-mode deterministic reasoning audit)
• feat(scanner): LanguageProfile interface + TypeScript reference + Python profile #6 — LanguageProfile + Python profile
• feat(wiring): OpenAI-compatible local-model adapter (Ollama / LM Studio / vLLM / …) #7 — local-model adapter (this PR)

🤖 Generated with Claude Code