fix(llma): two $ai_generation fidelity bugs in the Claude Code session ingest

[josephruocco]

Fixes #92. Both bugs are still on main as of 257a7d5. One commit each.

Cache tokens missing the $ai_ prefix

build_ai_generation emits these two unprefixed while every other key in the dict is namespaced:

"$ai_project_name": project_name,
"$ai_agent_name": agent_name,
"cache_read_input_tokens": cache_read_tokens,
"cache_creation_input_tokens": cache_creation_tokens,

They land as ordinary custom properties, so cost calculation ignores them.

I checked the expected names against posthog-python rather than taking them from the issue. Both are in _TOKEN_PROPERTY_KEYS in posthog/ai/utils.py, and around line 754 the anthropic provider always emits them even at zero, unlike the other providers. Kept that behaviour here since this builder only ever runs for anthropic.

Thinking blocks arriving as assistant text

_finalize_generation appended thinking and text into the same text_parts list, and build_events wrapped the joined result in a single {"type": "text"} block. Reasoning ends up indistinguishable from the actual answer in $ai_output_choices.

Added output_blocks alongside output_text, keeping each block's type in the order they first appear:

if gen.get("output_blocks"):
    content_blocks.extend(gen["output_blocks"])
elif gen["output_text"]:
    content_blocks.append({"type": "text", "text": gen["output_text"]})

output_text still holds the flattened join, so nothing reading it changes, and the elif keeps older parsed payloads working. Block shape matches anthropic_converter.py.

Note on the test diff

Two passing tests asserted the old behaviour, so the diff touches tests that were previously green:

• test_cache_tokens checked the unprefixed key
• test_thinking_preserved_when_split_across_chunks expected thinking to come back as a text block

Both updated. Probably part of why these lasted as long as they did.

Testing

89 passing, up from 86. Ran both commands from tests.yml locally:

python3 -m pytest tests/test_session_parser.py tests/test_posthog_llma.py
./tests/test_gate_exec_write.sh

The two commits are independent, so happy to split this into separate PRs or drop the second one if you'd rather take the prefix fix on its own.