Local-first OSS for human-agent work
Governed continuity for human-agent work.
Useful context shouldn't reset every session. Govinuity puts a governance layer between agent work and reuse — so what persists has earned it.
- Surface
- candidate signal
- Ratify
- human authority
- Measure
- measured reuse
The pipeline
From surface to measured reuse
Surface
Candidate decisions are extracted from agent session files. The Harvest page lets you scan a session directory, paste conversation text from any tool, or run the harvest script from the CLI.
Each candidate carries rationale, scope, reversibility, confidence, and possible conflicts. Nothing enters the review queue without a reason for surfacing.
python3 scripts/harvest_proposals.py --submit # Any agent tool via --input: cat session.txt | python3 scripts/harvest_proposals.py \ --input - --source cursor
Review
A human assesses each proposal in the review queue. The review card surfaces rationale, reversibility, scope, confidence, and why it was surfaced.
Nothing becomes governing context without a ratification decision. The queue is the only gate.
All DB migrations must be human-reviewed before production.
Ratify
Approve, defer, reject, or supersede. Only approved decisions become eligible for injection. Deferred candidates wait. Rejected decisions remain in the log.
Supersession atomically replaces an existing decision — the old record is marked superseded and permanently excluded from injection.
Inject
Eligible decisions reach agent context via GET /api/memory or a generated GOVERNED_CONTINUITY.md file. Every session is logged as a continuity run.
Transfer tiers control when a decision is included: always, by project, or only when explicitly requested.
GET /api/memory?project=my-project # Or generate a file: --- generated_at: 2026-04-15T10:00:00Z decisions_injected: 2 decisions_excluded: 1 --- ### DB migration review gate ratified_by: review · scope: always active All DB migrations must be human-reviewed.
Measure
Every injection is logged — what was injected, what was excluded, and why. The harvest script detects outcome signals and posts them as run annotations.
Signals: decision followed, correction required, context restatement, stale leakage. Each run is a data point for whether continuity is actually working.
Product console
A local governance console — no cloud, no accounts.
The product is intentionally operational: proposal intake from work, a review queue for incoming candidates, a decision log for ratified context, and run history for measuring reuse.
git clone https://github.com/Taprooted/Govinuity.git
cd Govinuity
npm install
cp .env.example .env.local
npm run devQuickstart
Run it locally.
Start the Next.js dashboard, submit a proposed decision, ratify it, then generate or request governed context for future agent sessions.
Current public release
Early, local-first, and built for technical users.
- Local-first software with no accounts and no telemetry.
- The full pipeline is in the public repo: proposal intake, review, ratified decisions, memory injection, run logging, and annotations.
- Technically literate users can run it themselves, inspect the SQLite-backed data, and integrate it with agent workflows.
- Proposal surfacing is core to the system. The Harvest page lets you scan session files, paste conversation text from any agent tool, or run the harvest script directly — including automatic correction signal detection and run annotation.