Select action response-header timeouts from the manifest

[CryptoJones]

Description

Closes #25.

nvpair-engine-manager defined two HTTP clients — a 30s response-header bound for everything and a 10m one reserved for Ollama's cold load — and actions.go chose between them with a literal engine == "ollama" && action == "run_model" comparison. Every other engine got 30 seconds for every action, no manifest field could change it, and the shipped lmstudio.json chat action (and any vLLM manifest) inherited the same ceiling: a long prefill against a large model failed with net/http: timeout awaiting response headers, which reads like a broken engine rather than a client-side limit.

This makes the response-header budget a per-action manifest field:

• New optional timeout_s on http actions (seconds). Validation rejects negative values and placement on non-http actions at manifest load.
• Default unchanged. An action that does not declare timeout_s gets exactly the same 30s budget as before; third-party manifests are unaffected.
• No engine-name special case left in the executor. The ollamaLoadClient field and the name comparison are gone; the client is selected by declared value and cached per distinct value (all default-budget actions share one client).
• Ollama's run_model declares 600s in the bundled manifest — behavior preserved, now data-driven.
• LM Studio's chat declares 600s in the bundled manifest — the action [Bug]: Action HTTP response-header timeout is selected by a hardcoded engine name, so only Ollama gets the long client #25 was reported against now gets a budget that fits a cold model load or long prefill before the first byte. The total call stays bounded by the executor's per-action timeout (default 30 min), so a declared header budget can never unbound a call.

Scope

In: registry.go (field + validation); executor.go (client selection/caching); both bundled manifests; the README engine:action row; tests; versions.json.

Out (deliberately): JSON-RPC surface changes — timeout_s is manifest data, not a wire field, so no broker relay, desktop bridge, or service-contracts changes; per-engine defaults in code (the manifest is the single source of truth); an upper cap on timeout_s (every action call is already wrapped in the 30-min context deadline).

Validation

Go 1.26.0 on Linux (x86_64). Desktop tree untouched, so the npm gates were not run.

Gate	Result
gofmt -l / go vet ./... in nvpair-engine-manager	clean
Bundled manifest JSON validity (lmstudio.json, ollama.json)	valid
go test ./... in services/nvpair-engine-manager (incl. 6 new tests)	ok 46.756s
New behavior tests under -race	ok 1.931s

New tests: TestEngineHTTPClientBoundsResponseHeaders, TestActionTimeoutSFollowsManifestNotEngineName, TestBundledLMStudioChatDeclaresResponseHeaderBudget, TestActionClientCachesPerDeclaredTimeout, TestValidateAcceptsActionTimeoutS, TestBundledOllamaRunModelTimeout. The two TestBundled* tests load the real go:embed manifests and pin that both declaring actions keep their long budgets, so a future manifest edit that silently reverts this fix fails CI.

Independent review. The diff was reviewed by a panel of fourteen independently trained models (Anthropic ×2, OpenAI, Google, Z.AI, NVIDIA Nemotron ×2, Mistral, DeepSeek, xAI, Moonshot AI, Nous Research, Amazon, and a local Qwen), one at a time, each blind to the others. Tally: 2 approve, 9 approve-with-nits, 3 request-changes — all three request-changes lanes on the same finding. Every finding was verified against the code before being acted on: (a) confirmed real and fixed in this branch — lmstudio's chat action never declared timeout_s, so the reported action would still have received the 30s default; it now declares 600s with a regression pin; (b) refuted — no upper cap is needed because every action call is wrapped in the 30-min context deadline; (c) confirmed minor and fixed — the README now documents the field; (d) refuted — the executor already guards non-positive/default values before building a client.

Risk

• Behavior: lmstudio chat's response-header budget rises from 30s to 600s for the first byte only; body streaming is unchanged and the total call remains capped by the per-action timeout. Every other action keeps its existing behavior (30s default). Ollama run_model: same 600s, now declared in data instead of hardcoded.
• Compatibility: additive optional manifest field. No JSON-RPC method or payload changes; desktop tree untouched.
• Conflict note for maintainers: open PR Opt-in, API-key-authenticated LAN access to the compatibility endpoints #38 also bumps product/installer to 0.92.0 in versions.json; if it lands first this branch needs a one-line rebase (the values match).

Versions

nvpair-engine-manager 0.17.4 → 0.18.0 (MINOR: additive optional manifest field on the engine:action surface). Per services/VERSIONING.md, a MINOR component bump implies product/installer 0.91.7 → 0.92.0.

AI disclosure

This change was written with AI assistance and reviewed by a panel of AI models, under the direction of the human author, who tested it on his own hardware and takes responsibility for it.

• Authored: Qwen 3.8 27B (qwen3.8-27b) running as a Hermes agent, operated by Aaron K. Clark (CryptoJones), who set the design decisions (per-action manifest field over an engine-name special case; 600s matching Ollama's existing cold-load budget; MINOR version bump) and ran all validation on his own hardware.
• Quality review: the diff was put to fourteen independently trained models, one at a time, each blind to the others' answers (round 1 is blind by construction — a lane cannot see an answer that does not exist yet); every finding was verified against the code before being acted on, and the adopted ones are listed under Independent review above.

Reviewer	Model
Anthropic	Claude Fable (fable), Claude Opus (opus) — via the claude CLI
OpenAI	gpt-5.6-sol
Google	gemini-3.1-pro-high
Z.AI	z-ai/glm-5.3-flash (via OpenRouter)
NVIDIA	nvidia/nemotron-3-ultra-550b-a55b, nvidia/nemotron-3-super-120b-a12b (via OpenRouter, free tier)
Mistral	mistralai/mistral-large-2512 (via OpenRouter)
DeepSeek	deepseek/deepseek-v4-flash (via OpenRouter)
xAI	x-ai/grok-4.6 (via OpenRouter)
Moonshot AI	moonshotai/kimi-k2.6 (via OpenRouter)
Nous Research	nousresearch/hermes-4-405b (via OpenRouter)
Amazon	amazon/nova-pro-v1 (via OpenRouter)
Local (Ollama)	qwen3.8-27b on the review host

Transcripts of every lane's answer are retained by the author and can be shared with maintainers on request.

Checklist

• I have read the Contributing Guidelines.
• Every commit is signed off (git commit -s), certifying the Developer Certificate of Origin.
• New or existing tests cover the change.
• Relevant documentation is updated.
• I checked the diff, changed filenames, and commit messages for credentials, private data, internal URLs, internal issue identifiers, and generated artifacts.
• I recorded the validation commands and results above.
• I bumped any affected component in services/versions.json, and described user-visible changes above so they reach the release notes.

---

Proudly Made in Nebraska. Go Big Red! 🌽 https://xkcd.com/2347/

🤖 Authored in Hermes (qwen3.8-27b) under the direction of Aaron K. Clark