Everything the ghost does 👻
Sixty-some features. One promise: nothing merges that didn't do what the ticket said.
01verification
Catch the lie
“A pass/fail verdict on intent, derived in code — a lying agent can't fake it.”
- ✓SpecReview verdict — every PR scored against what the ticket actually asked for.
- ✓Per-criterion GitHub checks — each requirement is its own check, pass or fail.
- ✓Merge gate — the verdict posts as a commit status that can block the merge.
- ✓No-spec proposer — no ticket? it infers the spec and asks you to confirm before judging.
- ✓Local verification — run the same verdict on your own machine, before you ever push.
02calibration & the ledger
Prove it gets sharper
“The only merge gate whose precision is measured, published, and improving.”
- ✓The verdict ledger — claim, evidence, reasoning, and cost — kept for every call.
- ✓Calibration — precision tracked per language and model, so you know when to trust a FAIL.
- ✓Human thumbs — your 👍 / 👎 feed straight back into the scoring.
- ✓Fix-escalation ladder — a failed verdict routes to a stronger agent — never a self-fix.
precision — when it says FAIL, a revert follows
TypeScript · Claude86%
Python · Claude81%
Go · GPT74%
Rust · Gemini69%
fix-escalation ladder · never self-fix
fast model→stronger model→strongest + tools
numbers illustrative
03autonomous execution
Your compute, your keys
“Agents run on hardware you control, in parallel, at $0 metered tokens.”
- ✓Dispatcher + parallel slots — many agents at once, each in an isolated git worktree.
- ✓Agent gateway — Claude Code, Codex, Gemini, Aider, or OpenHands — your pick.
- ✓Ollama local pool — route work to models running entirely on your own boxes.
- ✓nohup verify — verification runs on your own subscription, not our meter.
04the glass box
Nothing hidden
“Every verdict, dollar, decision, and reasoning step — visible.”
- ✓The Explorer — 30 live views over runs, verdicts, cost, and decisions.
- ✓Real-time HUD — watch jobs, slots, and spend update as they happen.
- ✓Transparency log — every decision and reasoning step, on the record.
- ✓Code intelligence — a file explainer that grounds answers in your actual code.
- ✓Ask-your-repo — cited Q&A over your PR, verdict, and cost history.
Explorer30 live views
runs
verdicts
cost / issue
decision log
file explainer
ask-your-repo
? ask-your-repo: “why did PR #42 fail?”
→ grounded answer, cited to the verdict + diff.
05governance & audit
Spend on rails
“Hard budget caps that block dispatch, and a provable trail for every AI-authored merge.”
- ✓Budget tiers — GREEN → YELLOW → ORANGE → RED, with RED halting dispatch.
- ✓Cost per verified issue — know what a proven 'done' actually costs.
- ✓CCB sessions — change-control board reviews for the calls that matter.
- ✓Audit export — spec → verdict → model → cost, exportable end to end.
- ✓Secrets rotation — keys rotate on a schedule; nothing sensitive is ever rendered.
budget — hard caps that block dispatch
GREENYELLOWORANGERED
RED — dispatch haltedno new agents spawned
cost / verified issue$0.41
illustrative · every merge leaves an audit trail
06onboarding & flow
Empty repo to verified done
“One-click install to a board where 'Done' is machine-proven.”
- ✓Zero-config provisioning — install on a repo and everything wires itself up.
- ✓Scaffolder — template discovery to go from empty repo to first PR.
- ✓Verified-Done board — a native planner — no external PM tool required.
- ✓Walkthroughs — human-verification checklists for the steps a machine can't judge.
- ✓Skill library + imports — reusable skills, plus Linear / GitHub import.
- ✓Casper — the free, open-source sibling that runs entirely on your machine.
board — “done” is machine-proven
todo
add-webhook-retry
in progress
migrate-users
verified done
add-rate-limit
verdict ✓
fix-session-expiry
verdict ✓
native planner + walkthroughs — no external PM tool required.
Install it and let the ghost start proving your “done”
capabilities available today; the verification loop is new — try it and tell us where the ghost is wrong.