The question that stalls AI adoption on a serious team isn’t whether the agent writes good code. It’s who approved it, what was agreed, and how you prove it later. Spaccy answers all three — with human approval before every write and a full trail of what was done.
The process shows itself, on its own
The project’s history, with nobody typing it in
Every work session logs its own start, end, scope and outcome, and authorship comes from git — not from a field someone filled in. Reading the history means reading what actually happened, in the order it happened.
in your repodocs/history.md
Status is read, not reported
Each feature’s progress is derived from the real execution and written to a versioned file. No board to drag a card across, no standup to figure out where things stalled: the panel shows the state of the repository.
in your repo.harness-state/
Every feature with its own complete trail
Proposal, specification, review, use cases and validation reports live in the folder of the feature that produced them. What was agreed, what was delivered and what was verified sit side by side — ready for review, for audit, or for whoever arrives later.
in your repodocs/features/
Nothing left over disappears
Whatever fell outside a cycle becomes a tracked item, with severity, owner and status — instead of a lost comment on some thread. Technical debt stops being a feeling and becomes a list.
in your repodocs/followups/
The follow-ups of this very project — no sample item.
And the limits of AI stay contained
Stops when context is missing, instead of making it up
When a precondition isn’t met or essential information is missing, the flow halts and states exactly what’s missing — without writing anything. It’s the opposite of the agent that fills the gap with a plausible guess you discover three days later.
Critical checks aren’t the model’s opinion
Compliance, layering rules, security verification and traceability run on deterministic engines: they give the same result every time, for any model. What the agent “thinks” is never the last word on whether something passed.
Nobody writes to your code without your approval
Before any risky action, a human approves it — with configurable levels by operation type, four-eyes approval whenever you want it, and a record of who cleared what. You choose how much autonomy to grant, per operation.
Interrupted work never becomes lost work
If a run stalls midway, Spaccy reconciles it and resumes the queue from the right point. And when a write drifts from the expected pattern, it’s reverted, and the revert is proven by checksum — nothing ends up duplicated or accidentally erased.
Explicit permission, real identity
Every operation requires a declared permission, and every team role gets only what it needs. The identity behind a write comes from git — the trail points to a person, never to “the agent.”
Everything on this page runs today and governs Spaccy's own development. Request early access and watch the first cycle from the inside.