fix: streaming timeout, Anthropic-native fallback, and routing on the provider-architecture branch

[hungcuong9125]

Summary

This PR integrates the streaming/fallback/timeout stabilization work from the original PR #80 onto the new provider-architecture base (Provider Abstraction + Unified Request Model + Routing Policy Engine, PR #82), and adds two fixes surfaced during the post-merge review.

The original PR #80 (#80) was opened against the pre-#82 architecture and is now obsolete; it has been closed. This PR is its replacement, re-applied cleanly on top of the current main (4fe96a7).

Why a new PR instead of rebasing #80

PR #80 was authored against the pre-architecture-rewrite base. After the provider architecture (#82) was merged, the streaming/fallback code paths moved from the handler/transformer layer into the new provider/router layer (scenarios.go, fallback.go refactor, request transformer split, Anthropic/Zen/Responses/Gemini routing policy engine). A clean re-integration on the current base is more readable than a rebase, and lets the new architecture own the work from the start.

What's included

From PR #80 (re-applied on the new architecture)

• streaming_timeout_ms for both opencode_go and opencode_zen, with helpers RequestTimeout(model) and StreamingTimeout(model) on the client.
• HTTP client no longer relies on a shorter global http.Client.Timeout; per-request contexts carry the timeout. The proxy never kills a stream that is actively producing bytes, and the server-level WriteTimeout is set to 0. Each upstream read uses a per-Read deadline via http.ResponseController.SetReadDeadline that is renewed on every successful byte.
• handleStreaming derives each attempt from the client request context and binds body reads via the new ctxio.NewCtxReadCloser so streaming_timeout_ms aborts mid-stream (returns ErrStreamReadCanceled).
• ExecuteWithFallback short-circuits on parent ctx.Err() and on context.DeadlineExceeded, and does not record client cancellation as a circuit-breaker failure.
• Raw Anthropic streaming pauses SSE keepalive writes during passthrough to prevent :keepalive injection into event frames (the client disconnected during anthropic stream error="context canceled" / Unexpected EOF failure mode seen on Claude Code).
• responseWriter serializes concurrent writes (heartbeat + stream body copy) with a mutex, and exposes a flushWriter wrapper for raw passthrough so events are not buffered in net/http's bufio.Writer.
• transformTools hardening: skip empty/whitespace names, normalize null/{}/missing schemas, validate type == "object", validate properties is an object, guard schemaObj nil for whitespace null (panic fix from PR fix: stabilize Anthropic-native streaming, timeout handling, and fallback cancellation #80 review).
• Streaming Scenario Routing (enable_streaming_scenario_routing) documented in CONFIGURATION.md and README.md.
• Reload messaging in atomic.go reports timeout changes as effective immediately for both timeout_ms and streaming_timeout_ms on both Go and Zen.
• walkthrough.md documents the integration and timeout fix for future maintainers.

New on this branch (post-#80 review findings)

• Timeout config overlap fix in loader.go::applyDefaults: when stream_timeout_ms is unset, the loader now inherits streaming_timeout_ms before falling back to timeout_ms. Without this, a user-configured 600s streaming_timeout_ms was silently downgraded to 300s by the idle watchdog (StreamIdleTimeout reads StreamTimeoutMs only). New regression test: TestDefaults_StreamingTimeoutFallback.
• Anthropic-native routing fix for Go provider: IsAnthropicModel and isAnthropicNativeGo now route minimax-m2.5, minimax-m2.7, minimax-m3, and qwen-plus through /v1/messages directly. These models reject OpenAI-format streaming with tools (400 invalid params, function name or parameters is empty); the previous code only routed qwen3.7-max through the Anthropic-native path on the Go side, so all other Minimax/Qwen variants fell back to the broken Chat Completions branch.

Commits (5)

dfa4c35 fix: integrate PR #80 streaming/fallback fixes with timeout config fix
328646c fix: integrate streaming context reader into new provider architecture and update streaming error handlers
70ebb18 Merge branch 'fix/streaming-fallback-anthropic-raw' into fix/streaming-fallback-new
7dd3ce4 fix: restore schemaObj nil check to prevent panic on whitespace null schema
feb71f0 fix: address PR review — bind streaming read to attempt ctx, guard nil schema

Tests

• TestRequestTimeout_* and TestStreamingTimeout_* for both Go and Zen providers, including default fallback, configured override, and small-value edge cases.
• TestExecuteWithFallback_* for cancelled parent context, parent deadline exceeded, per-model timeout, circuit-breaker accounting under cancellation.
• TestHandleStreaming_* for configurable timeout, client-context cancellation, mid-stream disconnect, per-model timeout fallback, Anthropic raw no-keepalive injection, concurrent response-writer behavior.
• TestHandleNonStreaming_ParentContextCanceled_No502 and TestHandleNonStreaming_ParentDeadlineExceeded_No502 to prevent the false-502 all models failed regression.
• TestTransformTools_* covering empty names, whitespace names, null/{} schemas, missing type, missing properties, malformed JSON, valid schema preservation, and the whitespace-null panic guard.
• TestDefaults_StreamingTimeoutFallback for the loader fix.
• TestNewCtxReader_* and TestNewCtxReadCloser_* for the new ctxio helpers.

All packages pass under go test -count=1 ./....

Validation

go build ./...
go test -count=1 ./...

Manual checks recommended:

• long Claude Code streaming session with tools=29+, long_context → minimax-m3
• confirm no 2-minute cutoff
• confirm no Unexpected EOF / no client disconnected during anthropic stream error="context canceled" for Anthropic-native raw streams
• confirm normal client disconnect does not produce false all models failed
• confirm streaming_timeout_ms: 600000 in config.json actually results in a 600s idle gap (verify via log: stream idle timeout reached fires after ~600s of upstream silence, not 300s)

Files

22 files changed, +2369 / -112:

• CONFIGURATION.md (+46)
• README.md (+1)
• walkthrough.md (new, +46)
• configs/config.example.json (+6/-2)
• internal/client/opencode.go (refactored timeout helpers)
• internal/client/opencode_test.go (+157)
• internal/config/atomic.go (timeout reload messaging)
• internal/config/config.go (StreamingTimeoutMs field)
• internal/config/loader.go (timeout fallback to streaming_timeout_ms)
• internal/config/loader_test.go (+42)
• internal/handlers/messages.go (responseWriter mutex, flushWriter, ctxio binding, parent-ctx short-circuit)
• internal/handlers/messages_test.go (+1106)
• internal/handlers/streaming.go (+2)
• internal/provider/opencode_go.go (isAnthropicNativeGo expanded)
• internal/router/fallback.go (parent-ctx/parent-deadline short-circuit, no circuit-breaker on cancel)
• internal/router/fallback_test.go (new, +267)
• internal/transformer/ctxio.go (new, +76)
• internal/transformer/ctxio_test.go (new, +101)
• internal/transformer/request.go (transformTools hardening)
• internal/transformer/request_test.go (+205)
• internal/transformer/stream.go (heartbeat-paused flag for raw Anthropic)
• internal/transformer/stream_test.go (+49)

Related

• Original PR fix: stabilize Anthropic-native streaming, timeout handling, and fallback cancellation #80: closed in favour of this PR.
• Upstream PR Provider Abstraction + Unified Request Model + Routing Policy Engine #82: provider architecture (the new base this PR is built on).

[samueltuyizere]

LGTM!