Local-first Engineering Intelligence for software repositories.
DevTime helps a repository explain itself from evidence. It scans code, tests, configs, routes, and decisions to identify the concepts inside a codebase, show the evidence behind them, surface uncertainty, and warn about risky changes.
It does not execute your code. It does not send your code anywhere. It does not require AI. It does not pretend to know things without evidence.
No cloud. No telemetry. No code execution. No AI required.
Git remembers code. It does not remember understanding - why a behavior exists, what evidence supports it, or what nobody has decided yet. As AI tools generate code faster than teams can review it, that missing understanding becomes the bottleneck.
DevTime builds evidence-backed repository memory: a local layer that says what a repository can prove, and - just as importantly - what it cannot prove yet.
V0 detects six supported concept families - it does not discover arbitrary domain concepts yet:
- Authentication
- Billing Webhooks
- Background Jobs
- Data Export
- Admin Permissions
- File Uploads
Anything outside these six is out of scope for V0. See LIMITATIONS.md.
- Detects concepts - the six supported families above - from routes, tests,
configs, dependencies, and docs, with word-sense gates so a coincidental keyword
(a job title, an avatar URL, a
session_idtrace) does not invent a concept. - Explains from evidence - every claim links to the files/signals behind it.
- Surfaces uncertainty - when evidence is missing (e.g. no decision record), it says so instead of guessing.
- Scores understanding - an Understanding Score (higher = better) with an Understanding Debt label (low/medium/high) and the causes shown.
- Warns about a narrow set of risky changes -
dtc risk --diffreviews a git diff against local memory and flags advisory findings for the change classes it supports (e.g. JWT algorithm weakening, billing-webhook retry without dedupe tests). It reports explicit states:review_failed,no_findings,unsupported_change_class(a known-concept file changed but no rule covers it), andfinding. "No findings" never means "could not inspect". - Records decisions -
dtc decision addstores rationale locally, which reduces uncertainty and improves understanding.
- It does not execute your code.
- It does not send code or data over the network.
- It does not require or call an AI model.
- It does not guarantee correctness or safe changes.
- It does not replace code review or architecture decisions.
- It is not a documentation generator, a static analyzer, an observability tool, a productivity tracker, or an AI coding agent.
See LIMITATIONS.md for the full, honest list.
- DevTime stores local repository memory in
.devtime/(a local SQLite database). - No network access during a scan.
- No code execution during a scan.
- Ignored directories are pruned before scanning; ignored files and secrets must never become evidence or claims.
- Every claim must link to evidence - no claim without evidence.
- Weak evidence produces uncertainty, not confidence.
- Usage is not decision: that a dependency is used does not mean someone decided why.
- Risk review is advisory by default - it does not block PRs.
DevTime scans the current directory, so the demo runs from inside the demo app.
cd examples/demo-saas
dtc init
dtc scan
dtc concepts
dtc explain "Billing Webhooks"
# ...make a change, then:
dtc risk --diff
# A decision only clears uncertainty when the code backs it up (corroborated):
dtc decision add --concept billing_webhooks \
--title "Use Stripe for billing" \
--body "We use Stripe as the payment provider and verify webhook signatures."
dtc explain "Billing Webhooks"The narrative:
- Before a decision: Billing Webhooks has strong evidence (route, signature
verification, test) and uncertainty - no decision explains its key choices.
Understanding Score is
58/100. - Risk review: changing retry behavior without updating duplicate-delivery tests is flagged high severity.
- After a corroborated decision: the reasoning is now in repository memory (and matches the code), the uncertainty clears, and the Understanding Score improves. A decision that the code does not back up stays flagged as uncorroborated.
A full, copy-pasteable walkthrough is in DEMO_SCRIPT.md.
Watch the 2-minute demo: DevTime scans a repo locally, explains concepts from evidence, surfaces uncertainty, catches a risky diff, and shows how a corroborated decision improves understanding.
Requires Python ≥ 3.11 and git.
git clone https://github.com/Shakargy/devtime.git
cd devtime
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
pytest # the full test suite should pass (77+ tests)This installs the dtc command. See QUICKSTART.md for a
step-by-step first run and troubleshooting.
| Command | Purpose |
|---|---|
dtc init |
Create local .devtime memory. |
dtc scan |
Scan the current repository and extract evidence-backed signals. |
dtc concepts |
List detected concepts with confidence and Understanding Debt. |
dtc explain <concept> |
Explain a concept: claims, evidence, confidence, uncertainty, Understanding Debt. |
dtc context <concept> |
Create a governed Context Pack for agents or humans. |
dtc risk --diff |
Review a git diff for risky changes using local evidence (advisory). |
dtc decision add |
Add a local decision record that can reduce uncertainty. |
(Also available: dtc evidence, dtc debt, dtc status, dtc doctor --privacy,
dtc export, dtc reset.)
$ dtc explain "Billing Webhooks"
Concept: Billing Webhooks
Supported claims:
- Billing Webhooks is present and supported by behavior evidence.
type: concept confidence: 0.86 evidence: src/billing/stripe-webhook.ts, tests/stripe-signature.test.ts
- Billing Webhooks has active route handling.
type: behavior confidence: 0.82 evidence: src/billing/stripe-webhook.ts
- Billing Webhooks verifies webhook signatures.
type: behavior confidence: 0.85 evidence: src/billing/stripe-webhook.ts
Uncertainty:
- No decision was found explaining key choices for Billing Webhooks.
Understanding Score: 58 / 100
Understanding Debt: medium
causes:
- missing or uncorroborated decision evidence
- no confirmed owner
Understanding Score is higher = better understanding; Understanding Debt is a label (low/medium/high), not the same number.
DevTime runs on examples/demo-saas and on real repositories. During Reality
Validation it detected - and then learned from - real failures (Next.js App Router
blindness, a false Billing Webhooks detection on a generic webhook system, a DB
migration mis-counted as Background Jobs evidence, and more). Each failure became a
fixture so it cannot silently regress.
- Tests grew from 13 to 77+ as each real failure became a fixture.
- Scan time on a 355-file real repo dropped from ~27.3s to ~0.48s after ignored- directory pruning.
Full evidence, before/after examples, and the validation reports are in
PROOF.md and reports/reality-validation/.
- Runs entirely locally; nothing leaves your machine during a scan.
- No code execution and no network calls during scanning.
- Secrets and ignored files are excluded from evidence by design (
dtc doctor --privacyreports the boundaries). dtc resetdeletes local memory; your source code is never modified.
DevTime is a heuristic scanner, not a full compiler or semantic analyzer. It is currently strongest on TypeScript / Next.js / Express / FastAPI-style repositories that resemble its fixtures. False positives and false negatives are possible. Understanding Debt is a product signal, not an objective universal truth.
Read the full list - including framework coverage, risk-review scope, and what is intentionally not built yet - in LIMITATIONS.md.
This is an early, local-first V0 focused on being trustworthy before being large. Not yet built (intentionally): git-history signals, wired MCP transport, an AI provider, a UI, and any cloud/team/enterprise features. See LIMITATIONS.md.
The most valuable contribution is a fixture: a small repository pattern plus the
expected concepts, allowed claims, forbidden claims, and required uncertainty. If
DevTime gets something wrong on your code, that wrong output can become a fixture so
it never regresses. See fixtures/ for the format and tests/ for how they run.
Licensed under the Apache License 2.0. See LICENSE.