Repository navigation
Home
Flagship QA audit for Devin sessions: it checks that what a session claims it delivered is backed by what the tool calls actually did. Verdicts: PASS, PARTIAL, UNVERIFIED.
Agents say "tests pass" and "pushed the fix" whether or not the evidence exists. devin-qa-pack reads sessions.db, tool_call_state and message_nodes, extracts claims (tests, commits, files, pushes, HTTP) and verifies each against recorded evidence and git state.
pipx install devin-qa-pack
# or: pip install devin-qa-pack (PyPI)
# source: pipx install "git+https://github.com/Icaro0310/devin-qa-pack"Requires Python 3.10+ unless noted. Linux, Personal Windows and Corporate Windows are supported; see the repo's platform READMEs for differences.
| Command | What it does |
|---|---|
devin-qa-pack audit --session <id> |
audit one session, print PASS/PARTIAL/UNVERIFIED |
devin-qa-pack audit --latest |
audit the most recent session |
| `devin-qa-pack audit --transcript --format aider | claude-code` |
devin-qa-pack report |
render the audit as HTML/JSON report |
devin-qa-pack session-end |
hook entry point: audit at session end |
Reads: sessions.db, tool_call_state, message_nodes, git history, transcript files.
Writes: report output only. Read-only on stores. Offline by default; URL checks need explicit --online --allow-domain.
-
UNVERIFIEDmeans evidence is missing, not that work is bad. - Transcript adapters (Aider, Claude Code) are experimental; the Devin-native path is the stable one.
- It verifies claims, not code quality.
Unofficial community project. Not affiliated with Cognition AI. Part of the devin-* ecosystem.