fix: stabilize reasoning streams and request logs

[rdself]

Summary

• Keep successful streaming requests pending until the chat completion callback finalizes the exact request id, so active request rows do not disappear before durable call-log persistence catches up.
• Re-serialize passthrough OpenAI-compatible stream chunks after splitting mixed reasoning_content plus content deltas, so downstream clients receive a reasoning-only event followed by a content-only event.
• Bind stream chunk debug payloads and completion finalization to the exact pending request id before falling back to the legacy model/provider/connection lookup.
• Stop sending redact-thinking-2026-02-12 by default on Claude Code-compatible relays; expose it as a per-connection CC Compatible toggle for upstreams that require redacted thinking streams.
• Rebase onto release/v3.8.26 and align the quality ratchet baselines to values measured on that release base, keeping this PR neutral against the target branch.

Root cause

There were three independent streaming/logging issues that could combine into the same user-visible failure.

First, createSSEStream() split mixed reasoning/content OpenAI-compatible deltas, but the passthrough branch could still reuse the original serialized SSE payload unless a later mutation forced reserialization. Clients that process reasoning before content, including SillyTavern-style OpenAI-compatible parsers, could then receive a mixed event and treat later reasoning/content boundaries incorrectly.

Second, the successful stream flush() path cleared the pending request before onComplete ran. The upper layer now finalizes by exact pending request id, but the request had already been removed from the pending map, so the in-memory completion fallback could be lost before the durable call_logs row was visible. That matched the observed behavior where the activity row spun for a while and then disappeared even though the upstream provider completed and billed normally.

Third, anthropic-compatible-cc-* relays sent redact-thinking-2026-02-12 unconditionally. With extended thinking enabled, Claude can spend most or all of the output budget on thinking. If the upstream redacts that thinking, OmniRoute receives little or no visible stream content even though the provider reports a successful request.

Changes

• Leave successful stream lifecycle cleanup to onComplete when completion finalization is available; only the legacy no-callback success path clears pending requests directly.
• Keep failure, idle-timeout, and empty-Claude-stream cleanup explicit and idempotent so errored requests still leave the active set.
• Move chatCore stream finalization before usage/cost/persistence side effects, so UI lifecycle state is stabilized even if later accounting work fails.
• Add regression coverage proving successful passthrough streams remain pending until onComplete finalizes the exact request id.
• Preserve provider-delimited reasoning by reserializing split SSE events, without inventing reasoning from unmarked content.
• Make CC Compatible thinking redaction opt-in through connection-level providerSpecificData.requestDefaults.redactThinking and Dashboard add/edit toggles.

Validation

• node --import tsx/esm --test tests/unit/provider-request-failure-pipeline.test.ts
• node --import tsx/esm --test tests/unit/claude-code-compatible-helpers.test.ts tests/unit/request-defaults-store-session.test.ts tests/unit/provider-specific-data-schema.test.ts tests/unit/executor-default-base.test.ts tests/unit/provider-page-helpers-3501.test.ts
• node --import tsx/esm --test tests/unit/request-logger-endpoints.test.ts tests/unit/stream-utils.test.ts tests/unit/active-request-stream-chunks-lifecycle.test.ts
• npx vitest run open-sse/mcp-server/__tests__/glmCodingProviderConfig.test.ts
• npx vitest run 'src/app/(dashboard)/dashboard/providers/[id]/components/modals/__tests__/connModals.test.tsx'
• npm run typecheck:core
• npm run check:file-size
• npm run quality:collect
• npm run quality:ratchet -- --allow-missing
• npm run check:cognitive-complexity
• npm run check:type-coverage
• npm run check:dead-code
• git diff --check

Release-base quality comparison

• origin/release/v3.8.26: npm run quality:collect measured eslintWarnings=3760, eslintErrors=0, openapiCoverage.pct=38, i18nUiCoverage.pct=79.7.
• This PR after rebase: npm run quality:collect measured the same values.
• origin/release/v3.8.26: dedicated gates measured cognitiveComplexity=753 and deadExports=339.
• This PR after rebase: dedicated gates measured the same values.

Live stream checks

• openrouter/deepseek/deepseek-v4-flash: 59.8s stream, 840 SSE events, 379 reasoning_content chunks followed by 456 content chunks, first reasoning at 6.14s, first content at 30.79s, mixedChunks=0, completed with [DONE].
• anthropic-compatible-cc-sp-anthropic/claude-opus-4-6 before disabling default redaction: 46.9s stream, 4 SSE events, 0 reasoning chunks, 0 content chunks, output_tokens=2048, thinking_tokens=2047, stop_reason=max_tokens.
• anthropic-compatible-cc-sp-anthropic/claude-opus-4-6 after disabling default redaction: 38.7s stream, 62 SSE events, 26 reasoning_content chunks, 32 content chunks, first reasoning at 1.73s, first content at 23.86s, mixedChunks=0, completed with [DONE].
• openrouter/deepseek/deepseek-v4-pro@preset/prefer through the local production build: 70.6s stream, 6,434 SSE events, first reasoning at 1.59s, first content at 48.2s, mixedChunks=0, completed with [DONE] and durable call_logs.detail_state=ready.
• anthropic-compatible-cc-sp-anthropic/claude-opus-4-6 through the local production build: 71.8s stream, 107 SSE events, 101 reasoning_content chunks, mixedChunks=0, completed with [DONE] and durable call_logs.detail_state=ready.

[gemini-code-assist]

Code Review

This pull request introduces exact pending request tracking and finalization using a unique request ID, preventing mismatches when multiple requests for the same model overlap. It also updates the call logs API to include completed in-memory fallback rows, extends the completed details TTL to 120 seconds, and ensures proper reserialization when splitting mixed reasoning and content deltas in SSE streams. Feedback on the changes highlights a potential TypeError in open-sse/utils/requestLogger.ts due to inconsistent optional chaining on entry when accessing entry.provider directly.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds stronger in-memory request lifecycle tracking so stream chunks and “just completed” requests reliably show up in usage logs and can be finalized by exact request id.

Changes:

• Introduces finalizePendingRequestById and extends the completed-details cache TTL to support slower UI polling.
• Updates request logging to bind streamChunks to the exact pending request via requestId (with safe fallback matching).
• Extends /api/usage/call-logs to include “completed in-memory” rows and adds test coverage around the new behaviors.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
tests/unit/stream-utils.test.ts	Adds a regression test ensuring passthrough SSE output is consumable by SillyTavern-style reasoning parsing.
tests/unit/request-logger-endpoints.test.ts	Adds tests for completed in-memory rows and requestId-based stream chunk binding.
tests/unit/active-request-stream-chunks-lifecycle.test.ts	Updates lifecycle coverage for overlap finalization and longer completed-details visibility.
src/lib/usage/usageHistory.ts	Adds completed-details TTL, exact-id finalization, and refactors finalization/cleanup logic.
src/app/api/usage/call-logs/route.ts	Prepends pending + completed in-memory rows to persisted call logs with deduping.
open-sse/utils/stream.ts	Ensures mixed reasoning/content split triggers reserialization reliably.
open-sse/utils/requestLogger.ts	Adds requestId option and tightens matching to avoid cross-connection streamChunk attachment.
open-sse/handlers/chatCore.ts	Passes pendingRequestId into logger and finalizes by id first, falling back to legacy behavior.

[diegosouzapw]

Thanks, @rdself! 🙏 Three real, observed fixes in one well-tested PR: (1) pending-request lifecycle — keeping the request pending until onComplete() finalizes by exact id (finalizePendingRequestById) so durable call_logs persist (activity rows stopped vanishing); (2) mixed reasoning/content SSE re-serialization for strict parsers (SillyTavern-style); (3) CC-compatible redact-thinking now opt-in per-connection (it was unconditional and could eliminate output under extended thinking). Guarded fallbacks, behavioral tests, live-verified.

I reconciled the modal conflict with #3921 (just merged): your CcCompatibleRequestDefaultsFields (context1m + redact-thinking) now lives inside the cc-compatible settings block next to the OpenRouter preset input. Validated: typecheck:core clean, 115 behavioral tests green (lifecycle/stream-utils/request-logger/cc-helpers), conn-modals vitest 8/8, lint clean. Merging into release/v3.8.26. 🚀

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.