- The quality gate for
AI-assisted development
- Evaluate before you build. Score before you ship. Catch regressions before they ship for you — from one binary, backed by one SQLite database.
- -+ A safe, disciplined, + time‑saving workspace + for AI-assisted dev. +
+Janus captures every coding session automatically, proves every change through 9 quality gates, and reports honest adoption metrics — from one Rust binary, backed by one SQLite database. No commands to remember.
+ +Why Janus
-One binary. Every guardrail.
-A harness that governs how a change reaches merge, and an eval harness that proves your skills still behave — sharing a single queryable database.
-9 quality gates
Pre-flight, acceptance criteria, recall, matrix, evidence, review, push permission, Jira tier, reconciliation — scaled by risk lane.
14 eval commands
run · baseline · diff · status · promote · trend · accept · apply · ab · rebaseline · analyze · suggest · goal · quality.
9 check kinds
Shell, JSON-path, file, transcript grep (±), LLM judge, prompt & response quality, and context relevance.
5 attribution classes
On a regression, Janus says why: skill, fixture, model, cross-skill change, or unknown drift.
Native & tiny
A 5.4 MB Rust binary with zero runtime dependencies. No bash, no jq, no python. Cold start ≈ 10 ms.
Cost-aware
Per-case token & time caps, a daily spend budget, and model pricing baked in for honest accounting.
Watch Janus work
+Pick a command. This is the real CLI shape — the same output you get in your own terminal.
+Illustrative session — numbers reflect a real Janus database after a week of use.
One binary. Every guardrail.
How it works
-Intent → Gate → Merge
-Every change enters one intake and is routed to a risk lane. The lane decides which gates apply.
-Classify the risk
Record the work; Janus picks a lane — tiny, normal, or high-risk — from its risk checklist.
--summary "Add OAuth" --lane normal
Prove each criterion
Acceptance criteria are extracted, mapped to tests, and proven with evidence — reviewer ≠ implementer.
harness-cli query matrix
Guard against drift
Behavior-regression evals run on push. A regression is attributed, not just flagged red.
--skill=code-reviewer --mode=2tier
Intake → Gate → Persist
+Every change enters one intake, is routed to a risk lane, proven through the lane's gates, then captured for good.
+Classify the risk
+Record the work; Janus picks a lane — tiny, normal, or high-risk — from a 10-item risk checklist.
+--summary "Add OAuth" --lane normal
Prove each criterion
+Acceptance criteria are extracted, mapped to tests, and proven with evidence — reviewer ≠ implementer.
+harness-cli query matrix
Capture, automatically
+On session end the hook records the trace; nothing relies on you remembering to log it.
+# adoption green · 98% instrumented
The 9 gates, in order
-Eval harness
-Does this skill still do what it did?
-Pick a mode, run the checks, and on a regression get the likely cause — not just a red X.
-smoke
One sample per case. Fast feedback for every commit and pre-push hook.
full
Three samples per case. Thorough scoring for release-grade confidence.
2-tier
Smoke first, then escalate only the failures to full. Best of both.
The dashboard that keeps you honest
+Instrumentation is the whole point — a trace with no tokens can't tell you what a session cost. Janus tracks that too.
+Promote a skill from warn-only to blocking once it's earned trust — a blocking regression exits 12 and stops the push.
Compare
-More surface, fewer moving parts
-Counts below are verifiable from the source and harness-cli --help.
+More guardrails, fewer moving parts
+Everything below is verifiable from the source and harness-cli --help.
+| Capability | Bash harnesses | Janus |
|---|---|---|
| Capability | Bash / script harnesses | Janus |
| Automatic session capture | hook · zero-config | |
| Adoption & instrumentation metrics | query coverage | |
| Honest RAG (grey on thin data) | built in | |
| Quality gates | 0 | 9 |
| Eval commands | ~6 | 14 |
| Check kinds | ~6 | 9 |
| Attribution classes | ~4 | 5 |
| Eval modes | 1 | 3 |
| LLM judge (your model) | ||
| Single binary, zero deps | ||
| SQLite-native state | ||
| Windows support | ||
| Single binary, zero deps | ||
| SQLite-native state | ||
| Opt-in eval harness (--features eval) |
Team-level ROI numbers (e.g. “bugs caught”, “hours saved”) are scenario projections, not measured benchmarks of this repo. The counts above are real.
+The eval harness (behaviour-regression testing) ships as an opt-in cargo feature — kept out of the default build to stay lean.
More surface, fewer moving parts
The name
-Why “Janus”?
-Why “Janus”?
In Roman myth, Janus is the god of gates, doorways and transitions — depicted with two faces, one turned to the past and one to the future.
-The harness does the same: it weighs what was (your baseline) against what is (your current run), and guards the threshold between them. The mark's checked seam is the gate, passed.
+In Roman myth, Janus is the god of gates, doorways and transitions — shown with two faces, one turned to the past and one to the future.
+The tool does the same: it weighs what was (your baseline, your last session) against what is (your current run), and guards the threshold between them. The mark's checked seam is the gate, passed.
Why “Janus”?
Install
-Up and running in seconds
-One command installs the binary and git hooks into your project.
-Up and running in seconds
+One command installs the binary and wires the git + session hooks into your project.
+Up and running in seconds
The Unified Eval Engine for AI-Assisted Development.
-MIT License © 2025–2026 Hoài Nhớ. All rights reserved.
+A safe, disciplined, time-saving workspace for AI-assisted development.
+MIT License © 2025–2026 Hoài Nhớ.