feat: add Claude/Gemini CLI and Anthropic API LLM providers

[23f3001304]

Summary

Closes #204

Adds support for driving the pipeline through the Claude Code and Gemini command-line tools (and the Anthropic Claude HTTP API) as LLM backends, alongside the existing Ollama and Gemini API providers. The CLI backends reuse your existing claude / gemini login, so you can run the agent without managing API keys.

This change is purely additive. Ollama and the Gemini API continue to work exactly as before.

New providers

LLM_PROVIDER	Backend	Auth
ollama	local models (existing)	none
gemini	Google Gemini HTTP API (existing)	GEMINI_API_KEY
claude	Anthropic Claude HTTP API (new)	ANTHROPIC_API_KEY
claude_cli	Claude Code CLI claude -p (new)	your claude login
gemini_cli	Gemini CLI (new)	your gemini login

Select a backend in .env:

LLM_PROVIDER=claude_cli
DEFAULT_MODEL=opus        # or claude-sonnet-4-6, gemini-2.5-pro, gemma3:4b, ...

Provider selection

• When LLM_PROVIDER is set, it is authoritative.
• When unset, the provider is inferred from the model name (existing behavior, unchanged).
• If a provider's prerequisite is missing (API key unset, or CLI not installed/logged in), it logs a warning and falls back to Ollama.

Implementation notes

All providers conform to the existing chat() contract and return the same {"message": {"content": ...}} shape, so pdf.py / evaluator.py / github.py did not need changes to call them.

The CLI providers are built to be robust:

• Lean, non-agentic one-shot. claude -p runs with --system-prompt (replacing the large built-in coding-agent prompt), --tools "" (no tool use), and --output-format json. This greatly reduces per-call token overhead and avoids agentic behavior in headless mode.
• Use the CLI's own login. API-key env vars are stripped from the CLI subprocess so a stray/placeholder key in .env can't cause 401 Invalid API key.
• Windows-friendly. npm .cmd shims are resolved via PATH and run through cmd /c; the full prompt is piped over stdin to dodge command-line length/quoting limits.
• Resilient. Transient failures retry with backoff, and the real error (including the CLI's JSON error envelope) is surfaced.

Other improvements

• Extraction progress logging (pdf.py): each section logs [i/n] progress with timing at INFO level, so runs aren't silent.
• GitHub rate-limit handling (github.py): logs whether GITHUB_TOKEN is present (and the 60 vs 5000 req/hour implication), and replaces the previous up-to-1-hour blocking sleep with a short, configurable cap (GITHUB_MAX_RATE_LIMIT_WAIT, default 60s) that skips the wait instead of hanging.

Config / docs

• requirements.txt: adds anthropic.
• .env.example: documents all providers, CLI options, and GITHUB_TOKEN; API-key placeholders are commented out so they don't break CLI/API auth by accident.
• README.md: updated prerequisites, configuration table, and provider details.

Backward compatibility

Existing ollama and gemini setups behave identically. No changes to the pipeline flow, prompts, or output schema.

Testing

• Ran claude_cli and gemini_cli end-to-end on Windows: section extraction and the evaluation step both return valid, schema-conforming JSON.
• Verified provider selection across configurations (env-set vs model-name inference, and missing-prerequisite fallback to Ollama).
• Confirmed the existing Ollama/Gemini selection and import paths are unaffected.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds Anthropic Claude support (API + CLI) and CLI-based routing, plus improved logging for GitHub rate limiting and PDF section extraction.

Changes:

• Add Anthropic (Claude) API provider and Claude/Gemini CLI providers with shared CLI execution utilities.
• Add provider inference/override logic via LLM_PROVIDER, new env vars, and model/provider mappings.
• Improve operational logging for GitHub API rate limits and per-section PDF extraction timing.

Reviewed changes

Copilot reviewed 7 out of 8 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
requirements.txt	Adds Anthropic SDK dependency for Claude API support.
prompt.py	Adds Claude model entries, provider explicitness handling, and new env vars for API keys/CLI commands.
pdf.py	Adds progress/timing logs while extracting resume sections.
models.py	Introduces Claude API + CLI providers, Gemini CLI provider, and shared CLI execution helpers.
llm_utils.py	Adds provider resolution (explicit vs inferred) and initializes new providers with fallback behavior.
github.py	Improves GitHub rate-limit handling/logging and reduces blocking waits.
README.md	Documents new supported providers, config, and CLI setup.
.env.example	Updates example configuration for new provider modes and CLI options.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[sp2hari]

Thanks for adding Claude support! Since this was opened, #298 made the provider layer fully config-driven — Anthropic ships an OpenAI-compatible endpoint, so Claude now works with zero new code via a providers.json block (see #365). That supersedes the bespoke provider class here. Closing as superseded — genuinely appreciate the contribution.