fix(bedrock): raise default max_tokens from 4096 to 16384

[liorfranko]

Problem

The Bedrock Converse API path hardcodes max_tokens=4096 when no explicit value is configured. This is far below Claude Opus 4.6's 32768 output limit, causing silent output truncation (finish_reason='length') on any response exceeding ~3K tokens of tool calls or long-form output.

This affects any Hermes user running Claude via AWS Bedrock inference profiles. Claude Code uses the same Bedrock profiles without this issue because it sets a higher max_tokens in its API calls.

Root Cause

AIAgent.__init__ accepts max_tokens: int = None, and there is no config-level key to set it (#15037 is open but unimplemented). When self.max_tokens is None, the Bedrock path falls through to hardcoded 4096:

# run_agent.py:8462
max_tokens=self.max_tokens or 4096,

Fix

Raise the hardcoded default from 4096 to 16384 (half of Opus 4.6's 32768 max) — a safe middle ground that eliminates truncation for typical agent workloads while leaving headroom for the model's own budget management.

Files changed:

• run_agent.py — fallback in _call_bedrock_converse()
• agent/transports/bedrock.py — BedrockTransport.build_kwargs() default
• agent/bedrock_adapter.py — build_converse_kwargs(), call_converse(), call_converse_stream() function signatures

Impact

• ✅ Fixes truncation for all Bedrock users without any config change
• ✅ No behavioral change for users who explicitly set max_tokens
• ✅ 16384 is well within Claude Opus/Sonnet 4.x limits (32768 max output)
• ✅ Minimal diff — 6 lines changed across 3 files

Related

• feat(config): support per-model max_tokens in custom_providers config #15037 — feat(config): support per-model max_tokens in custom_providers config (long-term fix)
• tracking: provider transport refactor (agent/transports/) #13473 — provider transport refactor (BedrockTransport extracted in PR feat: add BedrockTransport + wire build_kwargs #13467)

[teknium1]

Thanks for isolating the Bedrock output-cap issue. The unset-cap behavior is still present on current main, but this branch needs a small transplant before it can address it.

Problems

• build_api_kwargs was extracted from run_agent.py in 4b25619bc; the live native Bedrock request path is now agent/chat_completion_helpers.py:692-705, and it still sends agent.max_tokens or 4096 at line 702. The PR edits the old location and is currently conflicting.
• agent/auxiliary_client.py:1403-1412 independently calls Converse and explicitly falls back to 4096 at line 1408. Changing only the adapter default leaves that sibling fallback unchanged.

Suggested changes

• Move the agent-loop edit to agent/chat_completion_helpers.py:702 and retain the explicit model.max_tokens override behavior resolved in agent/agent_init.py:1602-1628.
• Cover the unset default in regression tests; tests/agent/transports/test_bedrock_transport.py:44-51 currently covers only an explicit cap.

Automated hermes-sweeper review.

[teknium1]

This request builder was extracted in 4b25619bc; current main executes agent/chat_completion_helpers.py:702 for native Bedrock requests. Please transplant this fallback change there, otherwise the live path will continue to pass 4096.