Count assistant tool_calls in count_message_tokens

[t-timms]

Why

count_message_tokens drives context compaction. It encoded only string values on a message or one level inside list items, so the nested function dict of assistant tool_calls (tool name and arguments) was never counted.

Observed on a local llama-server with a hard 32K window (KAT-Coder REAP-50, Godspeed agent-in-loop runner, SWE-bench Lite dev): Godspeed logged compaction at tokens=26999 / 27944 / 34473 (its own estimate), yet the server hit n_tokens = 32767, truncated = 1 five times in one run. Each time the reply was cut off in the middle of a tool call, llama-server answered HTTP 500 Failed to parse tool call arguments as JSON ... missing closing quote on every retry, and the session ended with exit_reason=llm_error. Those sessions emitted 80-170 tool calls.

Changes

• count_message_tokens now recurses through lists and dicts (_collect_strings), so tool-call names and arguments count. image_url blocks are still counted flat and not encoded.
• Docstring states what the estimate does not see: the chat template's own markup and the tool schemas sent with the messages.
• tests/test_token_counter.py: tool-call arguments counted, tool-call name counted, plain content unchanged, content blocks counted, image blocks flat-estimated without encoding their URL. The two tool-call tests fail without the change.

Measured effect (real requests)

18 requests sampled from a Qwen3.8-27B agent-in-loop run (captured with a logging proxy, 8 SWE-bench tasks), true tokens = server /apply-template (with tools) then /tokenize:

estimator	true / estimate, all requests	true / estimate, prompts > 8K tokens
main (before)	median 1.29x, max 1.65x	median 1.27x, max 1.29x
this PR	median 1.11x, max 1.65x	median 1.08x, max 1.12x

So on main the estimate is 22-39% too low, and the default compaction threshold of 0.8 on a 32K window corresponds to roughly 0.8 x 1.29 x 32K = 33.8K true tokens, i.e. already past the window: compaction cannot prevent the overflow. With this PR the same threshold corresponds to about 29K true tokens. Tool schemas add a further median of 562 tokens in this runner (more in the full harness) and are still not estimated.

Verification

• 523 passed, 1 skipped across test_token_counter, test_multimodal, test_agent_loop, test_context/. ruff check . and ruff format --check . (CI flags) clean.

Not verified / limits

• The remaining ~8-11% gap is the tokenizer (cl100k versus the model's own) and template markup. The 29K-true-tokens headroom below a 32K window is thin for a turn that returns several large tool outputs at once (KAT emits ~4-5 parallel calls per turn), so a bigger safety margin or calibrating against the server's reported usage.prompt_tokens would be a sensible follow-up; that is not part of this PR.
• The one max-ratio request (1.65x) is a small early-conversation prompt where fixed template overhead dominates.

[github-actions]

🤖 Godspeed Review

LLM review skipped (no API key configured).

Security

No secrets detected in changed files.

[github-actions]

🤖 Godspeed Review

LLM review skipped (no API key configured).

Security

No secrets detected in changed files.