Add LLM cost change analysis to CI

[Jwrede]

Summary

Add tokentoll GitHub Action to analyze LLM API cost changes on pull requests.

When a PR modifies code that makes LLM API calls (model swaps, max_tokens changes, new call sites), tokentoll posts a comment showing the projected cost impact. It stays silent on PRs that don't touch LLM code.

What it detects

tokentoll uses Python AST analysis (zero runtime dependencies) to find calls to:

• OpenAI (chat.completions.create, responses.create, embeddings.create)
• Anthropic (messages.create)
• Google GenAI (generate_content)
• LiteLLM (completion, embedding)
• LangChain (ChatOpenAI, ChatAnthropic, init_chat_model)
• Zhipu AI / GLM (ZhipuAiClient, ZhipuAI)

Current scan of this project

tokentoll Scan Report

File	Line	SDK	Model	Est. Cost/Call	Monthly
engine/utils/api_tier_detector.py	145	google_genai	gemini-2.5-flash	$0.000013	$0.01
engine/utils/gemini_client.py	122	google_genai	gemini-2.0-flash (default)	$0.000869	$0.87
engine/utils/revise.py	136	google_genai	gemini-3-flash-preview	$0.05	$49.40

Total estimated monthly cost: $50.28

Assumptions

• 1000 calls/month per call site

Generated by tokentoll

How it works

• Runs only on PRs, compares base vs head
• Posts a sticky comment (updates in place, no duplicates)
• Only comments when LLM calls are added, removed, or modified
• First run posts an initial scan (this PR) so you can see the output
• Zero runtime dependencies, ~1s scan time for most projects
• PyPI | GitHub

Changes

• Adds .github/workflows/llm-costs.yml (17 lines)
• No code changes, no new dependencies

[federicodeponte]

Maintainer triage after the Quality Gates fix and #43 packaging merge: leaving this unmerged.

This adds a new PR workflow with pull-requests: write permission and a third-party action that comments on PRs. Even though the action is SHA-pinned, it changes CI behavior for every PR and has no checks reported here. Needs maintainer approval for the bot/commenting behavior before merge.