[BUG] Strands SDK + Anthropic Provider Incorrect Token Estimates

[tylersalminen]

Description

AnthropicTokenCounter._count_message_estimated does not properly count message content blocks generated by Strands SDK.

To Reproduce

1. Wrap bedrock model with HeadroomStrandsModel configured with some transformer that is dependent on exceeding a model_limit threshold and an Anthropic model
2. Generate Strands agent from model
3. invoke agent with block content
4. any transformer that depends on model_limit threshold being exceeded never applies it's transforms due to Anthropic tokenizer never including block contents in token estimate

Expected Behavior

Properly estimates tokens

Actual Behavior

Tokenizer does not account for messages not containing a "type" property leading to not counting message contents in token total.

Code Sample

from os import environ

from strands.models import BedrockModel
from strands.types.content import ContentBlock
from strands.agent import Agent
from headroom import HeadroomConfig
from headroom.config import IntelligentContextConfig
from headroom.integrations.strands import HeadroomStrandsModel
from json import dumps


environ["HEADROOM_MODEL_LIMITS"] = dumps(
	{
		"anthropic": {
			"context_limits": {
				"us.anthropic.claude-sonnet-4-6": 200000
			}
		}
	}
)

async def _test():

	model = HeadroomStrandsModel(
		config=HeadroomConfig(
			intelligent_context=IntelligentContextConfig(
				enabled=True
			)
		),
		wrapped_model=BedrockModel(
			model_id="us.anthropic.claude-sonnet-4-6"
		)
	)

	agent = Agent(
		model=model
	)

	await agent.invoke_async(
		[
			ContentBlock(
				text="Some question" * 1000
			)
		]
	)

	# IntelligentContextConfig is never utilized b/c
	# token count never includes content / system blocks

Environment

• Headroom version: 0.5.2
• Python version: v3.12.10
• OS: Windows 11
• LLM Provider: Anthropic

Additional Context

Example Strands SDK message block:

{'role': 'user', 'content': [{'text': 'test'}]}

See in debugger image below where will not possibly ever include message content blocks in token count due to type property not existing.

[chopratejas]

Fixed in 8f18468 on main.

Root cause

_count_content_parts() in the base tokenizer only matched content blocks with a "type" field. Strands SDK sends blocks without "type":

# Anthropic format (worked)
{"type": "text", "text": "Hello"}

# Strands format (fell through to json.dumps fallback)
{"text": "Hello"}

For text blocks, the json.dumps fallback mostly worked (slight overcounting from key overhead), but for toolUse and toolResult blocks with nested content, the counting was wrong — nested text inside toolResult.content wasn't recursed into.

Fix

Added explicit handling for Strands-format blocks:

• {"text": "..."} → count text directly (same as {"type": "text"})
• {"toolUse": {"name": ..., "input": ...}} → count name + input
• {"toolResult": {"content": [...]}} → recurse into nested content blocks

8 new tests covering text blocks, tool blocks, mixed formats, and multi-block messages.

Try it

pip install git+https://github.com/chopratejas/headroom.git

This is on main and will be in the next PyPI release. Let us know if this resolves the threshold issue for IntelligentContextConfig!

[tylersalminen]

Perfect! Also, is there a reason that other content like reasoningContent.reasoningText.text and document.source.bytes are not included (among other possible content)?

[chopratejas]

@tylersalminen — good follow-up. Fixed in e7008f6.

Added counting for all Strands content block types:

Content type	Counting method
reasoningContent	Exact — count_text() on the reasoning text
image	Accurate — decodes with Pillow for Anthropic's (w*h)/750 formula; falls back to byte-size estimate
document	Estimated — ~1500 tokens/page heuristic (provider extracts text internally, can't replicate without PDF parser)
video	Estimated — ~1000 tokens/frame heuristic (provider samples frames internally)

The key fix for your use case is reasoningContent — it's pure text that was being missed entirely. Now counted exactly with count_text(), which means IntelligentContextManager thresholds will trigger correctly.

14 tests covering all content block types. On main now.

[tylersalminen]

Perfect! thank you