feat(R4): make LLM rate limit configurable via LLM_MAX_CALLS_PER_MIN

[siva23367]

Summary

Replaces the hardcoded >= 3 LLM-calls-per-minute cap in _wait_for_rate_limit with a configurable LLM_MAX_CALLS_PER_MIN setting.

Changes

• repi/core/config.py — Added LLM_MAX_CALLS_PER_MIN field (default 60, ge=1)
• repi/investigation/react_loop.py — __init__ reads from settings, _wait_for_rate_limit uses configurable value
• repi/api/config.py — Improved error handling (400 vs 500 separation)
• README.md — Added rate limiting documentation with provider recommendations

Testing

All 254 tests pass.

Closes #13

[vercel]

@siva23367 is attempting to deploy a commit to the varun7singh's projects Team on Vercel.

A member of the Team first needs to authorize it.

[VarunGitGood]

Review

The goal is sound but there is one fatal wiring bug that makes the feature non-functional, plus a few unrelated changes that should not be in this PR.

[BLOCKING] container.py is never updated — the setting is never actually read

container.py:227 is the only place ReactInvestigationLoop is instantiated, and this PR does not touch it:

return ReactInvestigationLoop(
    llm=llm,
    tools=tools,
    ...
    enable_reflection=settings.ENABLE_REFLECTION,
    reflection_interval=settings.REFLECTION_INTERVAL,
    # llm_max_calls_per_min missing — always None → hardcoded fallback of 60
)

Result: _wait_for_rate_limit always uses 60 regardless of what is in config.json. Fix: add llm_max_calls_per_min=settings.LLM_MAX_CALLS_PER_MIN, to that constructor call.

Once that wire-up exists, the < 1 guard inside __init__ is also dead code — pydantic already enforces ge=1 before the value can reach the constructor via settings.

---

On the positive side: the 400/500 split in api/config.py is a genuine improvement and the field definition + README docs are well written.

[VarunGitGood]

Four-line comment block reads like a changelog entry, not a code comment. Per project style: only comment when the WHY is non-obvious from the code itself. ge=1 already enforces the constraint; the provider recommendations belong in the README (where they already live). Remove this block — the field definition is self-explanatory.

[VarunGitGood]

Remove this comment — it describes what the code below does, which the code already makes clear.

Also: this block defaults to 60 on None, but that is a hardcoded constant, not settings.LLM_MAX_CALLS_PER_MIN. Until container.py is updated to pass the setting at construction time, this fallback silently overrides the user's config with no warning.

[VarunGitGood]

httpx is already a transitive dependency (used directly in repi/llm/adapters.py), so promoting it to an explicit dep is reasonable — but it has nothing to do with this PR. Please move to a separate commit.

[VarunGitGood]

Don't require this patch, the above added variable is good enough

[VarunGitGood]

@siva23367 Hey thanks for the PR, can you please check and resolve the comments and feel free to ask for clarifications

[siva23367]

Hi @VarunGitGood , addressed all review comments:

• Added llm_max_calls_per_min=settings.LLM_MAX_CALLS_PER_MIN to container.py
• Removed 4-line comment block from config.py
• Removed all "what" comments from react_loop.py
• Reverted docker-compose.yml, pyproject.toml, web/package-lock.json

All 254 tests pass.