devin-evals — MOVED

This repository was absorbed into the devin-assure monorepo.

The code now lives at packages/evals/ and the CLI is unchanged: pip install devin-evals / uv tool install devin-evals still installs the same package, now released from devin-assure.

# development moved
git clone https://github.com/Icaro0310/devin-assure
cd devin-assure/packages/evals

The repository is archived; open issues and PRs belong to devin-assure. History remains readable here for reference.

---

Original README (pre-archive)

Part of the DEVIN ecosystem
Track: Verify · Nature: product
For: QA engineers, AI engineers
Interface: CLI / Python library
Path: QA engineers · step 2/3 — after devin-assure, before devin-judge

devin-evals

Unofficial community project. Not affiliated with, endorsed by, or sponsored by Cognition AI. "Devin" is a trademark of Cognition AI.

Linux · Personal Windows · Corporate Windows

Part of the awesome-devin ecosystem: the curated hub for the devin-* tools.

An evaluation harness for agent work: define graded cases (session + rubric), replay them against recorded Devin sessions, and score quality over time — so "is the agent getting better?" has a number.

The problem

You tune prompts, rules files and models — and judge the result by vibes, one anecdotal run at a time. There is no regression signal. Meanwhile every Devin session already records, in sessions.db, its full transcript and the tool_call_state table: which tools ran, with what arguments, with what exit codes. That ground truth is sitting unused on your disk.

Prior art

Generic eval frameworks (OpenAI evals, promptfoo, Braintrust) grade output text — usually through an LLM judge. SWE-bench-style harnesses grade public repos, not your agent's real sessions. devin-evals adapts the rubric/graders idea; it does not reinvent it. What it adds is the corpus: deterministic checks over Devin's recorded tool-call ground truth.

What makes it Devin-native

Rubrics like "must call devin_redact before publishing" become checkable facts. Graders read tool_call_state via devin-internals-spec, so a check asserts "tool run_shell was invoked with pytest in its args and exited 0" — a fact, not an LLM judgement call.

• Side-by-side: promptfoo cannot assert "tool X was called with args containing Y" — it never sees Devin's tool-call table.
• No-Devin: no sessions.db, no tool-call graders, no replay mode.

Install

Requires Python ≥ 3.10 and pipx or uv. Per-OS setup lives in the platform guides: Linux · Personal Windows · Corporate Windows.

From PyPI: uv tool install devin-evals or pip install devin-evals

From a checkout: pip install -e .

Runtime deps: devin-internals-spec only — no network, no LLM calls.

Usage

Write cases in evals/*.json:

{
  "id": "redact-before-publish",
  "description": "Report workflow must stay hygienic",
  "session_ref": "Refinery session 2026-09-29",
  "rubric": [
    { "grader": "tool_called", "name": "devin_redact" },
    { "grader": "tool_called", "name": "run_shell", "args_substr": "pytest" },
    { "grader": "exit_code", "value": 0, "mode": "all" },
    { "grader": "contains", "text": "all tests pass" },
    { "grader": "no_secrets" }
  ]
}

session_ref matches a session id or title in sessions.db. Then:

# Windows PowerShell
devin-evals list --evals evals
devin-evals run --evals evals --sessions-db "$env:APPDATA\devin\cli\sessions.db" --out report

# Linux
sessions_db="${XDG_DATA_HOME:-$HOME/.local/share}/devin/cli/sessions.db"
devin-evals list --evals evals
devin-evals run --evals evals --sessions-db "$sessions_db" --out report

run writes report/report.json + report/report.md (per-case PASS/FAIL/SKIP/ERROR + aggregate score; reruns are byte-identical) and exits 0 if everything passed, 1 on failures, 2 on usage/IO errors.

Try it without a real Devin install:

python -m devin_evals.demo demo.db
devin-evals run --evals evals --sessions-db demo.db

The shipped evals/ directory contains a passing case, an intentionally failing case, and a tool-call-ground-truth case.

Graders

grader	what it checks
contains / not_contains	literal substring in the transcript (case-sensitive)
tool_called	tool name called ≥min_calls, optional args_substr on call JSON
file_exists	path on disk — relative resolves under the session's working_directory
exit_code	recorded exit codes match value per mode (all/any/last)
no_secrets	zero secret-shaped strings (vendored devin-redact patterns) in transcript + tool JSON
no_pii	zero PII-shaped strings (email, CPF) in transcript + tool JSON
tool_output	text presence in tool-call output JSON (present inverts)
regex	regex text match in tool-call output JSON (present inverts)
no_split_secrets	no secret-shaped match spanning the seam of two tool payloads

Works with Devin alone (Devin-only mode)

devin-evals scores recorded sessions with deterministic rubrics — no LLM calls, no network access, nothing beyond Devin's own session data and Python.

Platform support

Pure stdlib Python — identical behavior on Windows, Linux and macOS. CI runs the suite on windows-latest + ubuntu-latest; the target file or directory is always an explicit argument, so there are no platform-specific paths.

Rubric packs (EV-4)

Reusable check sets for common session types, so replay works out of the box. Built-ins shipped with the package: bugfix, feature, refactor — each is a small hygiene rubric (clean exit codes, no tracebacks, no secrets) meant to be extended by the case's own rubric.

{
  "session_ref": "my-session-id",
  "packs": ["bugfix"],
  "rubric": [ { "grader": "contains", "text": "test_regression" } ]
}

Pack checks run before the case's own checks. List them with devin-evals packs; override or add your own with --packs-dir <dir> (a <name>.json file there shadows the built-in of the same name).

Synthetic sessions (dream)

The dream subgroup generates synthetic Devin sessions with known verdicts — regression and adversarial fixtures. This was the standalone devin-dream repository, absorbed into this package:

devin-evals dream unit   --out out/                 # one dir per defect + expected.json
devin-evals dream unit   --out out/ --defect D01 D03
devin-evals dream inject --out adv/ --n 5           # adversarial (D07/D08) + scorecard
devin-evals dream fleet  --out big/ --sessions 2000 --seed 1 --manifest

Secrets and PII in generated fixtures are always obviously fake public-documentation values — nothing real is ever generated or read.

Golden corpus (EV-3)

devin-evals corpus materializes and replays a deterministic corpus of labeled synthetic sessions — ten defect classes (D01–D10, generated by devin_evals.dream, the absorbed devin-dream catalogue) with a known verdict each — plus the matching evals/*.json cases whose rubrics encode those verdicts in gradable form. It is the CI gate that proves evals, fixtures and graders agree:

devin-evals corpus generate --out .corpus   # sessions.db + evals/ + corpus.json
devin-evals corpus verify  --corpus .corpus # expected-vs-actual per case

Everything is deterministic given --seed and synthetic only: the corpus must never point at a real sessions.db.

The corpus is also versioned in the repo: corpus/evals/*.json and corpus/corpus.json are committed, while the sessions*.db files are always regenerated in place (never committed — *.db is gitignored). tools/regen-corpus.py rebuilds them deterministically and gates CI on drift:

python tools/regen-corpus.py          # rebuild corpus/ in place
python tools/regen-corpus.py --check  # exit 1 if committed corpus diverges
python tools/regen-corpus.py --verify # rebuild + replay expectations

--check reuses the seed recorded in corpus/corpus.json; the committed corpus is generated by the bundled devin_evals.dream package, so the gate is fully self-contained.

Each case carries expected_status (the verdict the case should reach: pass for the clean control D03, fail where a defect must be caught, error for the D06 schema-drift canary whose v18 db is refused at open time). verify prints MATCH / GAP / MISMATCH per case and exits 1 on any undocumented mismatch; --strict also fails on documented gaps.

The D05/D07/D09 grader gaps are closed: no_pii covers PII, tool_output scans tool-call output for injected instructions, and no_split_secrets catches credentials split across payloads. Remaining known limits: verdict granularity is coarser than the source catalogue (qa-pack's UNVERIFIED/PARTIAL both collapse to fail), and D08's "quarantined" is graded by a transcript-level proxy (not_contains on the unsafe policy).

session_ref accepts selectors (EV-2), each resolving to the most recent match: latest, project:<substr> (matches working_directory), window:<YYYY-MM-DD>:<YYYY-MM-DD> — plus exact id or title as before.

devin-evals judge <case> --question "…" (EV-1, opt-in) asks a live LLM to grade one case. Non-deterministic, off by default, fail-closed: needs DEVIN_BRIDGE_CMD (drives devin-bridge, which gates via policy) and uses the free model unless DEVIN_JUDGE_MODEL overrides. Sessions are labelled judge:<case> so devin-janitor can reap the noise.

ab-run (EV-5/G3, opt-in) is the A/B gate: a suite of tasks run in two arms (variant prefixes A vs B), k attempts per arm, graded by the same deterministic rubrics — the with-skill/without-skill proof loop. Consumes real tokens, fail-closed without DEVIN_BRIDGE_CMD; sessions are labelled g3-ab:<task>:<variant>:<attempt> for janitor.

devin-evals ab-run --evals evals --tasks evals/tasks \
  --attempts 5 --seed 73001 --max-sessions 80 --yes \
  --sessions-db "$sessions_db" --out g3-report.json

The suite lives in a tasks manifest dir (default <evals>/tasks, falling back to ./tasks, or --tasks DIR). Two layouts are accepted: an index manifest.json (a list of {"id", "kind", "type", "prompt", "dir"|"workspace"} entries), or one *.json per task ({"id", "kind": "bugfix|feature|refactor", "type": "trigger|control", "prompt", "workspace"}). The repo ships a hermetic 8-task pack under tasks/ (5 trigger + 3 control — see tasks/README.md). A run requires ≥5 trigger + ≥3 control tasks; trigger tasks measure the effect, control tasks detect collateral regressions. Each attempt runs in its own shutil.copytree of the task workspace under --work-dir (default <tmp>/g3-<ts>) — the originals are never touched and _solution/ reference fixes are never copied into an attempt. Order is deterministic: tasks shuffled by --seed, arms interleaved ABBA/BAAB across tasks to spread temporal drift.

Success = all checks pass. The default grader replays the session through the rubrics (--sessions-db, autodetected): the eval case matching the task id when present, else the pack matching the task kind. For hermetic workspace packs, --workspace-check instead runs the task's own deterministic check inside the attempt copy (pytest tests -q, plus check_structure.py when present) — combined with the session rubric when a sessions.db resolves.

Budget caps (--max-sessions, --max-total-time, --session-timeout) abort the whole run when hit and report what ran; over-cap plans need --confirm (interactive) or --yes (headless). --dry-run prints the full session plan for free. Timed-out/failed attempts count as failures, are recorded separately, and are never retried.

Preregistration: tasks, k, seed, caps and every verdict threshold are frozen into the report's design block before the first session runs (preregistered: true). The report (g3-report/0.1, --out) contains only aggregates and session ids — never session content.

Verdicts (Wilson CIs on arm rates; bootstrap CI over the mean per-task Δ = rate_B − rate_A on trigger tasks, 10k fixed-seed resamples):

verdict	condition
regresses	Δ CI entirely < 0, OR any control task Δ ≤ −0.4, OR safety worsens (denials_b > denials_a when denials_a == 0)
improves	Δ CI over trigger tasks entirely > 0 AND no regress condition AND calibration ok
no-detectable-effect	Δ CI contains 0 AND CI width ≤ 0.30
inconclusive	everything else: CI wider than 0.30, baseline ceiling (≥95%) or floor (≤5%), >20% aborts in an arm, calibration absent/failed where needed

Calibration: improves additionally requires a passing A/A calibration — run ab-run --aa (arm B gets the arm A prefix) to measure the harness's own false-positive rate, then point --calibration prior-report.json at it. Without it, an otherwise-improving run is honestly reported inconclusive.

Honest power note: with the minimum suite (8 tasks) at k=5 the Δ CI is roughly ±0.22 — only effects ≥ ~0.25 are detectable. This gate catches large regressions, not subtle improvements; add tasks, not attempts, to shrink the CI.

Passing --task still runs the deprecated simple mode (exactly two sessions, ab-run:<tag>:<variant> labels) — kept for quick smoke checks, not a gate.

Limitations

• Offline replay only (M1): grades recorded sessions, cannot spawn new ones. prompt_context-only cases report SKIP.
• Deterministic only (M1): no LLM-as-judge; contains is a literal, case-sensitive substring — it cannot tell "all tests pass" from "not all tests pass" semantically.
• Depends on Devin's private, versioned internals — a sessions.db schema bump makes devin-internals refuse loudly rather than misread.
• tool_called infers the tool name from name/tool_name/tool/kind keys in tool_call_json; unknown shapes degrade to "tool never called" details, not crashes.
• file_exists checks the filesystem now — replaying an old session whose workspace was cleaned will fail that check.

Development

pip install -e ".[dev]"
python -m pytest     # 177 tests

When to use this

• You changed a prompt, rules file or model and want a numeric regression signal across recorded sessions.
• You want to assert tool-call behavior — e.g. "must call devin_redact before publishing" — as a checkable fact.
• You need reproducible scoring: deterministic graders make reruns byte-identical, no LLM judge involved.
• You want to try it without a Devin install: python -m devin_evals.demo demo.db builds a sample DB.

When NOT to use this

• You need semantic judgement of free text — contains is a literal, case-sensitive substring and there is no LLM-as-judge (M1).
• You need to spawn new sessions — this is offline replay of recorded sessions only.
• Your agent is not Devin — the ground truth comes from sessions.db/tool_call_state.

FAQ

How do I regression-test changes to my Devin prompts or rules? Define cases in evals/*.json pairing a session_ref (session id or title) with a rubric, then run devin-evals run --evals evals --sessions-db <path> --out report. Each rubric item is a deterministic grader over the recorded tool calls, so reruns produce identical scores — a real before/after comparison.

Does devin-evals use an LLM judge? No. All graders (contains, tool_called, file_exists, exit_code, no_secrets) are deterministic checks over tool_call_state and the transcript. That makes scores reproducible but also literal — it cannot evaluate semantic quality of prose.

Can I use devin-evals without Devin installed? Yes, for a demo: python -m devin_evals.demo demo.db creates a synthetic sessions.db and the shipped evals/ directory contains passing and intentionally failing cases. For real use you need a sessions.db from actual Devin sessions.

License

MIT — see LICENSE.

---

If this saved you debugging time, a ⭐ on the repo helps others find it.