feat: configurable upstream + body-idle timeouts on RecordConfig

[tombeckenham]

Summary

Adds `RecordConfig.upstreamTimeoutMs` and `RecordConfig.bodyTimeoutMs` so callers can lift the hardcoded 30s ceilings used by the proxy. Defaults stay 30s — back-compat preserved.

Motivation

Reasoning models (e.g. Grok 4.3 via OpenRouter with `response_format: json_schema`) routinely leave 30s+ gaps between SSE chunks during the thinking phase under concurrent load. The hardcoded `BODY_TIMEOUT_MS = 30_000` in `makeUpstreamRequest` fires `req.destroy()` mid-stream, and the captured fixture ends up with no `[DONE]` and no `finish_reason` — downstream `JSON.parse(text)` then blows up on `Unterminated string`.

Reproduces deterministically with a 6-way concurrent record run against the same Grok 4.3 + structured-output prompt:

```
BEFORE (hardcoded 30s):
jsonValid: 3/6 elapsed min=23.23s median=30.95s max=31.00s
3 streams cut off at exactly 30.95–31.00s, no [DONE], no finish_reason

AFTER (bodyTimeoutMs: 180_000):
jsonValid: 6/6 elapsed min=31.84s median=38.62s max=42.18s
All 6 finish=stop, all done=true, all JSON valid
```

For comparison, the same 6-way load fired direct to OpenRouter (no aimock in the loop) completes 6/6 up to 52s with clean `[DONE]` — i.e. the upstream is fine, only the proxy's idle timer was clipping us.

Change

• `RecordConfig`: two new optional numeric fields, documented inline in `src/types.ts`.
• `makeUpstreamRequest`: optional 7th `timeouts` arg, falls back to `30_000` for both timers when omitted.
• `proxyAndRecord`: threads `record.upstreamTimeoutMs` / `record.bodyTimeoutMs` into the upstream request.

Test plan

• Existing `recorder.test.ts` "body timeout" test still passes (asserts the default 30_000 is applied when no override is set)
• `pnpm run build` clean
• Manual: 6-way concurrent Grok 4.3 streaming record through aimock with `bodyTimeoutMs: 180_000` — 6/6 finish=stop, vs 3/6 truncated without the override
• (Reviewer:) Worth adding a vitest that verifies a custom `bodyTimeoutMs` is honoured — happy to add if you'd like before merge

Back-compat

No behavior change when the new fields are omitted — defaults stay at the original 30_000.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@197

commit: 41166b9

Posted prematurely — will follow up with a proper review.

[jpr5]

I pushed fixes on top of your commit:

1. CHANGELOG — moved entry from released [1.18.0] to [Unreleased]
2. Input validation — clampTimeout helper rejects 0, NaN, Infinity, and negatives (they silently disable timeouts via Node coercion)
3. Tests — three new: custom bodyTimeoutMs honored, custom upstreamTimeoutMs honored, clamp fallback for zero/negative
4. CLI flags — --upstream-timeout-ms / --body-timeout-ms with validation
5. JSDoc — clarified upstreamTimeoutMs is a socket idle timer, not TTFB
6. Type — inline anonymous type → Pick<RecordConfig, 'upstreamTimeoutMs' | 'bodyTimeoutMs'>
7. Docs — added timeout flags to the record-replay page (CLI table + Upstream Timeouts section with programmatic example) and a README CLI example

Everything passes locally (tsc, 3026 tests, build). CI will need a push from your side to trigger — can you push an empty commit to kick the workflows?

Take a look and let me know what you think.

[jpr5]

I'm sorry for the state thrash — my bot is being a fucken idiot today.

[jpr5]

@tombeckenham OK all yours.

[tombeckenham]

Yes. This works nicely. Thank you!