Bug Description

The amazon-bedrock-openai provider with api: openai-responses produces multiple incremental phase:"final_answer" text blocks within a single assistant response. Each incremental block is delivered as a separate visible message to downstream channels (Telegram, Discord), resulting in 4–5 progressively-longer duplicate messages for the same user turn.

Repro Evidence

Session history shows a single assistant response containing 4 final_answer text blocks for one user turn (image + calendar request):

A second turn (calendar event creation) showed 5 incremental final_answer blocks for the same response.

Trajectory event confirms the route:

• Provider: amazon-bedrock-openai
• Model: openai.gpt-5.5
• API: openai-responses

Affected Channels

• Telegram DM (direct) — 4–5 duplicate messages for a single user turn
• Discord channel session — 4 duplicate repo-review messages for a single assistant response

Affected Route

• amazon-bedrock-openai/openai.gpt-5.5
• API: openai-responses
• Bedrock Mantle endpoint shape: https://bedrock-mantle.*.api.aws/openai/v1 (redacted)

Hypothesis

The openai-responses streaming adapter/parser for amazon-bedrock-openai treats Bedrock Responses streaming incremental/cumulative snapshots as distinct final-text events, rather than deduplicating or delta-appending to a single final output. This causes the transport layer to emit multiple final_answer phases.

Root-Cause Update (2026-06-03)

Direct curl -N to Bedrock Mantle /openai/v1/responses with stream: true shows only one phase: "final_answer" in the raw SSE stream. The Bedrock Mantle streaming format is correct; the duplicate final_answer phases are produced inside the OpenClaw openai-responses streaming adapter.

See comment below for raw SSE evidence.

Expected Behavior

One final_answer text block per assistant response, containing the complete final text.

Actual Behavior

3–5 incremental final_answer blocks per assistant response, each delivered as a separate channel message.

Not a Duplicate Of

• EmbeddedAttemptSessionTakeoverError fires at ~120s on long Bedrock streams (fence whitelist too narrow?) #89259 (EmbeddedAttemptSessionTakeoverError on long Bedrock streams)
• Bug: Bedrock Converse Streaming silently aborts on long-context agent sessions (~6 min timeout, no retry, no fallback) #87876 (Bedrock Converse Streaming silently aborts)
• [Bug] Regression 2026.5.27→2026.5.28: "No API provider registered for api: bedrock-converse-stream" — pi-ai removal breaks Bedrock provider registration; bearer token auth broken #88707 (provider registration regression)
• openai-responses adapter is unusable against Azure OpenAI: every turn returns a synthetic 0-token refusal (openai-completions works) #79570 (Azure 0-token refusal)
• Tool calls emit empty arguments {} when content parameter is large (write/exec) #70033 (large payload empty args)

Environment / Version

• OpenClaw observed on runtime circa 2026-06-02
• Gateway model: amazon-bedrock-openai/openai.gpt-5.5
• API: openai-responses
• Channel streaming modes already set to off (not channel-level issue)