fix(sse): give extended-thinking targets the reasoning readiness budget

[abhisheksharma2411]

Summary

kiro/claude-sonnet-5-thinking fails with [504] Stream produced no non-ping SSE event within 125000ms when the upstream stays ping-only through a long reasoning turn.

125000 is not a hardcoded Kiro timeout — it is STREAM_READINESS_TIMEOUT_MS (80s) plus the 45s very-large-history/payload bump, capped there because nothing in resolveStreamReadinessTimeout recognised the request as a reasoning target.

The policy already grants a 30s warm-up allowance to codex high-effort models and to Claude-format replicas, but it keys the latter on the provider registry's format field. Kiro serves Anthropic thinking models through its own CodeWhisperer translator, so format: "kiro" excludes it — and the -thinking alias was never a reasoning signal the way -high is. devin, antigravity and chatgpt-web publish -thinking ids from outside claude format too, so they had the same gap.

This detects the -thinking alias suffix directly and applies the same 30s allowance, so kiro/claude-sonnet-5-thinking gets 155s instead of 125s. The three reasoning bumps are mutually exclusive — they model the same one-off warm-up, so a single request cannot stack them into a multi-minute readiness window.

Scoped to the alias suffix (-thinking, -thinking-1m) rather than a substring, so an unrelated id that merely contains the word does not qualify.

Related Issues

• Closes feat(providers): Kiro provider: 125000ms non-ping SSE idle timeout not configurable anywhere #11922

Validation

• Change type: routing (SSE stream readiness)
• Focused tests and category gates from the golden path
• npm run lint — clean on both changed files
• Reconciled with the current active release base (release/v3.8.51); focused checks rerun afterward
• Production-code changes include a new or updated automated test in this PR

Reproduced first: the new #11922 test failed at exactly 125000 !== 155000, matching the reporter's number before any fix was written.

Focused suites, all green after the change:

tests/unit/stream-readiness-policy.test.ts
tests/unit/stream-readiness.test.ts
tests/unit/combo-stream-readiness-fallback.test.ts
tests/unit/runtime-timeouts.test.ts
tests/unit/lmarena-stream-readiness-repro-9306.test.ts
tests/unit/stream-early-eof-breaker.test.ts
→ 76 tests, 76 pass, 0 fail

npx tsc --noEmit reports the same 4782 pre-existing errors on release/v3.8.51 with and without this change — it adds none.

Each new guard was mutation-tested; every mutation killed exactly one test:

mutation	result
drop && !extendedThinking from the claude-format bump	1 fail
drop && !codexHighReasoning from the thinking bump	1 fail
loosen `/-thinking(?:-	$)/to/thinking/`

Tests Added Or Updated

• tests/unit/stream-readiness-policy.test.ts — 5 added:
  • the feat(providers): Kiro provider: 125000ms non-ping SSE idle timeout not configurable anywhere #11922 repro (kiro + large history: 125s → 155s)
  • provider-agnostic small-request case (devin claude-opus-4-6-thinking)
  • no stacking with claude_format_heavy_reasoning
  • no stacking with codex_gpt_5_5_high_reasoning
  • no false positive on an unrelated id containing "thinking"

Coverage Notes

open-sse/utils/streamReadinessPolicy.ts is the only production file changed. Every added branch is covered by the tests above: the bump firing, both exclusivity guards, and the suffix-boundary rejection. Coverage does not move down in any touched file.

Reviewer Notes

• Behavioural change is bounded: a -thinking request gets at most +30s of readiness budget, still clamped by maxTimeoutMs, and nothing else in the request path changes.
• Worth a second opinion on scope: this treats -thinking as a reasoning signal for every provider. That is deliberate — the alias means the client asked for extended thinking regardless of who serves it — but if you would rather gate it to a provider allowlist, say so and I will narrow it.
• I did not touch the Kiro executor or add a per-provider timeout control, which is the other option feat(providers): Kiro provider: 125000ms non-ping SSE idle timeout not configurable anywhere #11922 raises. The env vars (STREAM_READINESS_TIMEOUT_MS, STREAM_READINESS_MAX_TIMEOUT_MS) already exist and do govern this path; the reporter's real problem was the default being too low for thinking targets, not the absence of a knob. Happy to add the dashboard control as a follow-up if you want it.

[abhisheksharma2411]

CI note, so nobody has to dig: the three red checks here are all base-branch state, not this change.

No new ESLint warnings — I ran the job's exact command on both refs. release/v3.8.51 alone:

$ npm run lint:json -- --max-warnings 0
exit=2
There are suppressions left that do not occur anymore. Consider re-running the command with `--prune-suppressions`.

Identical on this branch — same exit code, same message, byte-identical output apart from a [BABEL] size note about a cached dist/ artifact. The exit 2 comes from stale suppression entries, which is exactly what #11924 measured in a clean-room clone.

Build (advisory) and dast-smoke are the hosted-runner OOM in #11946.

All 4 unit-test shards, Vitest, semgrep, docs gates and merge integrity are green, which is the part this PR actually touches.

[diegosouzapw]

Babysit summary (#11959)

Verdict: green for required checks. Not merging (human merge).

Contributor credit preserved: original commit 2ad1f3cc remains authored by @abhisheksharma2411.

Before (head 2ad1f3cc)

• No new ESLint warnings — FAIL (stale suppressions: "There are suppressions left that do not occur anymore")
• dast-smoke — FAIL (hosted-runner OOM; inherited infra ci: Fast Production Build + dast-smoke (hosted 7 GB) OOM on release/v3.8.51 — 19/30 red, paints every PR #11946/ci: docker-publish OOM (ResourceExhausted) on every release/v3.8.51 push — :next stale since 08-22 and the v3.8.51 tag will not build on the hosted runner #11976)
• Build (advisory) — FAIL (hosted-runner shutdown/OOM; advisory)
• Fast Quality Gates, unit shards, Vitest — PASS

What we did

PR was 37 commits behind origin/release/v3.8.51. Merged current base (includes #11940 / #11955 ESLint suppressions re-freeze, #11975) and pushed to the fork (maintainerCanModify: true).

Pushed SHA: 8220a44b334b2a8a226e062e541a0c9055f82172
(2ad1f3cc + merge origin/release/v3.8.51)

No production-file edits beyond the merge. ESLint gate not widened.

After (head 8220a44b)

• No new ESLint warnings — PASS
• Fast Quality Gates — PASS
• Unit Tests fast-path 1/4–4/4 — PASS
• Vitest, Docs Gates, Merge integrity, semgrep — PASS
• dast-smoke — not scheduled on this run (no longer on the PR rail after fix(ci): take the two hosted-runner builds off the PR rail (#11946, option 3) #11962)
• Build (advisory) — FAIL (npm run build cancelled on ubuntu-latest mid-compile; same advisory hosted-runner OOM/shutdown as before — ignored; Mergify already allows this single failure)

Ready for owner merge when the ⭐ gate is applied.