[Bug]: channel draft streaming is post-hoc chunking; run_tool_call_loop does not consume provider streaming

[Lemonawa]

Summary

Channel draft streaming (Telegram stream_mode=partial/on) appears progressive, but run_tool_call_loop does not consume upstream provider streaming APIs.

Affected component

channel

Severity

S2 - degraded behavior

Current behavior

• run_tool_call_loop sends requests via provider.chat(...) instead of streaming APIs.
• Final answer text is chunked and forwarded through on_delta only after completion.
• This produces post-hoc chunking UX rather than true upstream token streaming.

Evidence on latest main (82b71530):

• src/agent/loop_.rs:1509 uses provider.chat(...).
• src/agent/loop_.rs:1978 starts chunking final text to on_delta after final response is ready.
• src/channels/mod.rs:3672 streaming decision is based on channel draft support, not provider streaming path.
• src/gateway/mod.rs:1478 already uses provider.stream_chat_with_history(...) (control path).

Expected behavior

When channel draft streaming is enabled and provider supports streaming, channel responses should consume upstream provider deltas progressively during generation, while preserving existing tool-call loop semantics and fallback behavior.

Steps to reproduce

# On latest main
rg -n "provider\.chat\(" src/agent/loop_.rs
rg -n "split_inclusive\(char::is_whitespace\)" src/agent/loop_.rs
rg -n "supports_draft_updates\(\)" src/channels/mod.rs
rg -n "stream_chat_with_history\(" src/gateway/mod.rs

Impact

Affected users: Telegram and other channels using draft updates.
Frequency: always on this path.
Consequence: degraded perceived responsiveness; "streaming" is not upstream-real-time.

Logs / stack traces

N/A (source-level deterministic reproduction)

ZeroClaw version

82b7153

Rust version

rustc 1.93.1 (01f6ddf75 2026-02-11)

Operating system

Darwin 25.3.0 arm64

Regression?

Unknown

Pre-flight checks

• I reproduced this on the latest main branch or latest release.
• I redacted secrets/tokens from logs.
• I removed personal identifiers and replaced identity-specific data with neutral placeholders.

[theonlyhennygod]

Implemented and opened PR: #2868

Root cause confirmed:

• run_tool_call_loop used provider.chat(...) and only chunked final text post-hoc, so channel draft streaming was not consuming upstream provider deltas.

What changed:

• Added guarded upstream streaming consumption in run_tool_call_loop when draft streaming is active, provider supports streaming, and no request tools are attached.
• Added suppression/rollback guard for streamed tool-call payload markers (<tool_call, <toolcall, "tool_calls") so internal payloads are not leaked in user-visible draft updates.
• Added fallback to non-streaming provider.chat(...) with trace event llm_stream_fallback on stream failure.
• Skipped post-hoc whitespace chunking when live streamed deltas already delivered the final visible response.
• Added regression tests:
  • run_tool_call_loop_consumes_provider_stream_for_final_response
  • run_tool_call_loop_streaming_path_preserves_tool_loop_semantics

Validation run:

• cargo fmt --all
• cargo test run_tool_call_loop_consumes_provider_stream_for_final_response -- --nocapture
• cargo test run_tool_call_loop_streaming_path_preserves_tool_loop_semantics -- --nocapture
• cargo test process_channel_message_streaming_ -- --nocapture

This preserves existing tool-loop semantics and fallback behavior while making channel draft streaming consume real upstream streaming deltas on the safe path.

[theonlyhennygod]

Merged via rebase in PR #2868 (commit 52b9e6a). Follow-up parity work is tracked in #2869.