codebase-receipts-cli 1.0.1

Receipts — resume claim verifier & codebase-grounded interview prep CLI

RECEIPTS

Your resume said it. Your code proves it.

Quick Start • What It Does • How It Works • Providers • Setup Guide • Deep Dive

The anti-flattery resume tool. Most resume reviewers feed your bullets into an LLM and hand you back compliments. Receipts does the opposite — it embeds your actual codebase, searches it for evidence behind every claim you made, and tells you the truth. If the code doesn't back it up, you'll know before the interviewer does.

Quick Start

pip install codebase-receipts-cli

receipts check                                    # verify provider is working
receipts ingest ./my-project                      # index your codebase
receipts verify resume.tex ./my-project           # verify every claim
receipts verify resume.tex ./my-project --rewrite # + honest rewrites
receipts tui resume.tex ./my-project              # full-screen dashboard
receipts ama resume.tex ./my-project              # mock interview mode
receipts ledger                                   # token usage & costs

[!TIP] See SETUP.md for the full walkthrough — provider config, model setup, and first-run instructions. See LEARN.md for how every piece works under the hood.

What It Does

Claim Verification

Every bullet in your .tex resume gets broken into individual claims — percentages, counts, latency numbers, technology names — and each one is checked against the code.

Results: 5 verified, 3 plausible, 2 unsupported

UNSUPPORTED CLAIMS:
  [percentage] 40%
    bullet: Reduced API latency by 40% through query optimization...
    reason: No evidence of latency measurement or optimization in codebase

The rule: if the knowledge base has no relevant code for a claim, it's marked Unsupported without even asking the LLM. The model never gets a chance to talk itself into justifying something it can't find.

Honest Rewrites

Weak bullets get rewritten to be exactly as strong as what the code can actually justify. No invented numbers. No inflated language.

ORIGINAL:  Reduced API latency by 40% through query optimization
REWRITTEN: Optimized database queries in the API layer using indexed lookups

Keyword Gap Analysis

Paste a job description. Get back what's addable, what's ungrounded, and what's missing:

Can add (code supports it):
  + Docker — Dockerfile and compose config found in codebase

Ungrounded (resume claims, no code):
  ? Kubernetes — mentioned in resume but no k8s config in codebase

Gaps (JD wants, you don't have):
  - Terraform

Interactive TUI

A full-screen terminal dashboard with a file tree, sortable claims table, side-by-side rewrite diffs, live LLM activity log, and a running token counter.

Shortcut	Action
c	Claims tab
r	Rewrites tab
l	Log tab
q	Quit

Mock Interview (ama)

An AI interviewer that targets your weakest bullets. Unsupported claims get asked about 3x more often than verified ones. For each bullet:

1. Pointed technical question about the specific claim
2. Your answer
3. Natural follow-up based on what you said
4. Your follow-up answer
5. Assessment: did it hold up?

Ends with a summary of strengths, weaknesses, and overall readiness.

Token Ledger

Every LLM call is metered and logged — even when the provider is free. receipts ledger shows lifetime totals across all sessions.

How It Works

receipts/
  cli.py ........................ Typer CLI: check, ingest, verify, tui, ama, ledger
  config.py ..................... BYOK config (env vars / .env, never committed)
  ama/
    interviewer.py .............. Mock-interview loop, weighted toward weak claims
  ingest/
    scanner.py .................. Tree-style walk respecting .gitignore
    git_source.py ............... Clone remote repos to temp dir
    artifact_extractor.py ....... Tree-sitter extraction within a token budget
    secrets_scanner.py .......... Pre-embedding secrets scan + redaction
  ledger/
    token_ledger.py ............. SQLite-backed token/cost log
    pricing_table.py ............ $/1K-token rates per provider/model
  llm/
    provider.py ................. Abstract LLMProvider (complete, embed)
    factory.py .................. Reads config, returns the active provider
    ollama_provider.py .......... Local, free, no key needed
    gemini_provider.py .......... Free-tier Google AI Studio key
    anthropic_provider.py ....... BYOK adapter (paid)
    openai_provider.py .......... BYOK adapter (paid)
    fake_provider.py ............ Deterministic offline stub for tests
  resume/
    tex_parser.py ............... Parse .tex into structured sections/bullets
    claim_extractor.py .......... Extract numeric + technology claims per bullet
  tui/
    app.py ...................... Textual App with panelled layout
    widgets/ .................... StatusBar, ClaimsTable, DiffView
  verify/
    kb.py ....................... ChromaDB-backed vector store over code artifacts
    verifier.py ................. Classify: Verified / Plausible / Unsupported
    rewriter.py ................. Propose honest rewrites for weak bullets
    keyword_gap.py .............. JD keyword gap analysis grounded in code

[!NOTE] For a narrative walkthrough of every module — what it does, why it exists, and how to explain it in an interview — see LEARN.md.

Provider Support (BYOK)

Provider	Default	Key	Install
Ollama	Yes	None	Built-in
Gemini	No	Free	Built-in
Anthropic	No	Paid	pip install codebase-receipts-cli[anthropic]
OpenAI	No	Paid	pip install codebase-receipts-cli[openai]

Zero bundled keys. Zero silent paid API calls. Ever.

[!IMPORTANT] Anthropic does not offer an embeddings API, so Claude needs a split config — chat via Claude, embeddings via a provider that has them: RECEIPTS_PROVIDER=anthropic + RECEIPTS_EMBED_PROVIDER=gemini (or ollama). The embed provider must match whatever embedded your knowledge base — re-ingest if you switch embedding models.

Development

uv sync --group dev          # install everything
uv run pytest -v             # 233 tests, fully offline
uv run ruff check receipts/  # lint
uv run black --check receipts/ tests/  # format check
uv run pre-commit install    # git hooks

The entire test suite runs with zero network access — a deterministic fake provider handles all LLM calls in tests.

Publishing

git tag v0.1.0
git push origin v0.1.0

GitHub Actions runs the full test suite on Ubuntu + Windows (Python 3.10 + 3.12), then publishes to PyPI via trusted publishing. No tokens in secrets.

License

MIT

---

_{Built by Ishaan Sharma}