fix(anthropic): use model-native output limits instead of hardcoded 16K

[teknium1]

Summary

The Anthropic adapter hardcoded max_tokens=16384 as the default when the user didn't configure one explicitly. Since Anthropic's API requires max_tokens (unlike OpenAI where it's optional), this meant every direct-Anthropic session was capped at 16K output tokens — regardless of the model's actual capability.

This is the root cause of two frequently reported errors:

• "Response truncated (finish_reason='length')" — thinking consumed most of the 16K budget, leaving little for the response
• "Response only contains think block with no content after it" — thinking consumed ALL 16K, zero tokens left for any response

The problem in numbers

Model	Actual max output	What we sent	Wasted capacity
Claude Opus 4.6	128,000	16,384	87%
Claude Sonnet 4.6	64,000	16,384	74%
Claude Sonnet 4.5	64,000	16,384	74%
Claude 3.5 Sonnet	8,192	16,384	(overcounted but harmless)

With adaptive thinking on Claude 4.6, the model self-manages its thinking budget within max_tokens. At 16K, it might think for 14K and have only 2K left for the response — or use all 16K on thinking and produce nothing.

What other agents do

Cross-referenced with Cline, Clawdbot, and OpenCode:

• Cline: Uses the model's full output limit (64K Sonnet, 128K Opus) — sourced from a per-model catalog
• OpenCode: Caps at min(model_limit, 32K) — still 2x our old default
• Clawdbot: Defaults to 8192 via external library

None of them hardcode 16K for models that support 64-128K.

Changes

agent/anthropic_adapter.py:

• Added _ANTHROPIC_OUTPUT_LIMITS — static lookup table with output limits for all Claude model families (sourced from Anthropic docs + Cline's model catalog)
• Added _get_anthropic_max_output(model) — substring matching that handles date-stamped IDs and variant suffixes (:1m, :fast)
• Unknown/future models default to 128K (future models won't have less capacity)
• build_anthropic_kwargs now accepts optional context_length parameter — clamps max_tokens so it never exceeds the context window (protects custom endpoints with limited capacity)

run_agent.py:

• Passes context_compressor.context_length to the adapter for clamping

Why this only affects direct Anthropic

OpenRouter users aren't affected because we don't send max_tokens through the chat_completions path — OpenRouter handles it. The Codex/OpenAI paths also don't send it when unconfigured. Only the Anthropic adapter required it (mandatory API field) and hardcoded 16K.

Test plan

• 19 new tests: model lookup accuracy, date-stamped IDs, variant suffixes, unknown models, explicit override, context_length clamping
• Updated existing test_default_max_tokens to expect model-native limits
• Full suite: 6477 passed, 0 failed

Closes #2706