๐ Overseer Kit
Phased AI work without amnesia, fake DONE, or silent drift โ on plain GitHub.
Optional MuseHub depth when you need provenance, signed identity, and non-invertible history.
The problem
AI-assisted development breaks down at the boundaries between sessions:
- Amnesia โ each new chat re-derives context from scratch.
- Self-approval โ agents mark phases DONE without independent verification.
- Silent drift โ roadmaps and handovers fall out of sync with reality.
- Skipped tests โ expensive remakes when mechanical gates are bypassed.
๐ Overseer Kit makes skipping honesty expensive โ not by running your product, but by governing how phased work proceeds.
L0 in 60 seconds โ roadmap + handover
Every repo gets a living ROADMAP (phase control board) and HANDOVER (session relay with one next step). Fresh AI sessions paste the handover prompt instead of guessing.
./cli/overseer init --regime git-only ./cli/overseer governance-sync --dry-run ./cli/overseer status
L0 alone is useful on any GitHub repo. No Muse install required.
Add L1 โ mechanical checkpoints
When artifacts are measurable (media exports, test suites, data transforms), L1 runs verify scripts before the pipeline advances.
overseer verify-step ยท policy/manifest ยท ARTIFACT_SHA256 binding ยท fail-closed exits
Domain logic stays in your verify scripts โ the kit ships the orchestrator socket, not video QC or accounting rules.
Add L2 โ boss / worker / checker
When handoffs involve money, publish, or audit stakes, L2 adds role-gated verdicts on an append-only ledger.
overseer honesty-status ยท overseer ledger ยท producer/verifier separation ยท hash chain
Producers cannot self-certify. Verifiers run in independent sessions. Tampering breaks the chain.
Optional L3 โ MuseHub substrate
Same CLI commands. When you flip to muse+git-mirror, you gain content-addressed history, safe GitHub mirror export, and drift realign โ without gating L0โL2.
realign ยท isolated mirror PR ยท provenance enrichment
Guardrail: no core governance feature is MuseHub-only. Git-only is a first-class baseline.
Modularity โ lanes, modules, domain packs
Multi-lane docs
Named handover/roadmap pairs per concern (governance-sync --lane).
L1 + L2 modules
Typed checkpoint and honesty sockets โ enable in .overseer/config.yaml.
Domain packs
Consumer-owned verify scripts and hooks โ VideoFactory, research, classroom, etc.
VCS adapters
git-only ยท muse+git-mirror ยท muse-only โ same CLI surface.
Who it is for
AI video studios
Stop remakes and fake approvals before publish.
Research labs
Bind claims to evidence hashes; independent verifier sessions.
Classrooms
Portable integrity curriculum โ L0 default; Muse identity optional.
Treasuries & regulated ops
Dual-control hooks + append-only ledger for audit trails.
Quickstart โ git-only in 5 minutes
- Clone overseer-kit.
- From your consumer repo:
./cli/overseer -C . init --regime git-only --non-interactive - Open
docs/OVERSEER-HANDOVER.mdโ copy the paste-ready prompt. - Run
./cli/overseer governance-sync --dry-runbefore ending each session.
Full guide: Git-only quickstart ยท Consumer pattern: CONSUMER-ADAPTER-PATTERN.md
MuseHub upgrade path
When provenance matters, flip regime โ same commands, deeper history:
- Install Muse and run
muse initin your repo. - Set
regime: muse+git-mirrorin.overseer/config.yaml. ./cli/overseer syncโ seeds bridge workflow + safe deploy script.- Day-to-day:
muse commiton feature branches. Publish viascripts/muse-bridge-deploy.shonly.
Operator runbook: K7 dogfood runbook
Public roadmap
| Phase | Deliverable | Status |
|---|---|---|
| K9b | L1 checkpoint orchestrator | DONE |
| K10 | L2 honesty ledger | DONE |
| K11 | API/CI freeze provider | DONE |
| K12 | Track N landing + scenario gallery | DONE |
GitHub โ Kit โ MuseHub
Start where you are. Add layers when stakes rise. Upgrade substrate only when provenance pays for itself.
Not another agent framework. Governance + honesty for people who already use Cursor and GitHub.