fix: guard oversized JSON responses in code-generation prompt

[tkatta-stack]

What

generate_code() (in integuru/util/print.py) builds a single LLM prompt per DAG node to write that node's request-parsing function. For a text/html/application/javascript response it already guards against oversized bodies (>100000 chars) by substituting short context snippets instead of the full response. The application/json branch had no such guard — it always embedded the complete raw response text verbatim, on top of the key_paths already computed that pinpoint exactly where each extracted variable lives in the parsed JSON.

Why this causes #30

A single large JSON API response captured in the HAR file (a bulk data dump, a paginated listing, etc.) gets embedded in full. By itself this can push one prompt well past the model's context window — matching the reported "messages exceeded available capacity by roughly 3x" after a normal run, and the second reporter's identical experience.

Fix

Mirrors the existing 100000-character threshold from the HTML/JS branch onto the JSON branch. Past that threshold, the prompt drops the raw response text and relies solely on the already-computed key_paths, which are sufficient for the model to navigate the parsed JSON and write correct extraction code without seeing the raw payload. Small/typical JSON responses are unaffected — same prompt as before.

Testing

Added tests/test_generate_code.py:

• a response with a 200,000-character value confirms the raw text is excluded from the prompt while the resolved key path is still present
• a small response confirms the original (full-text) prompt is unchanged

Both pass with the fix; I also confirmed the large-response test fails against the pre-fix code (i.e. it genuinely catches the regression).

Note: tests/test_integration_agent.py currently fails locally/in CI independent of this change — it needs a test.har fixture file that isn't committed to the repo. Not touched here since it's unrelated to this fix; flagging for visibility.

Fixes #30

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}