ROADMAP
MEASURED,
NOT PROMISED.
This page is derived from the repository's own state documents at v0.3.8.67: PLAN.md records where the colony measurably is, and AUTONOMY-10.md orders what is left. An item is done only when it has a production call site and a test keyed to what the producer actually emits.
AVAILABLE NOW
- Mission planning and dispatch across twelve contracted roles, with per-role kill switches
- Durable worker runtime: leases, heartbeats, and crash reclamation
- Isolated mission workspaces — detached git worktrees pinned to a base revision
- Patch proposals verified in a sandbox that actually contains the patch, with base-hash staleness refusal
- Policy-inserted Tester and Soldier on every state-changing patch set; deterministic blocks demote outcomes
- Verification from stored evidence — model prose is recorded, never decisive
- One approval queue and lifecycle for proposed changes, with execution-time re-checks and audit events
- Artifact and evidence stores with provenance and content hashing
- Pheromone memory with verified-only reinforcement and decay toward neutral
- Colony recall: what worked, what fails, who solved this, what knowledge exists
- Chat console with streaming answers, projects, attachments, and mission transparency
- Windows and Linux releases, desktop app, LXC/systemd installer, Docker
IN PROGRESS
- Typed collaboration between roles
Typed artifacts travel alongside prose with IDs for replay, but prose is still the primary channel a model reads first.
- Policy-inserted verification
Policy binds the Verifier to Tester and Soldier evidence and inserts one when a plan omits it, but the role itself is still planner-selectable.
- Tester on the patched tree
The tester names the tree it judged, but the materialized patch does not yet outlive the verification sandbox — the named gap of the current line.
- A Queen-driven twelve-role acceptance run
All twelve roles execute together in tests, but no deterministic acceptance mission reaches all twelve through production triggers, and no live twelve-role mission has run against a real model.
- Reputation-aware routing
Role standing is computed from trails; nothing routes on it yet.
PLANNED
- Real qualification
A composed-runtime scenario matrix that qualifies each role before it counts as ready.
AUTONOMY-10 phase 5 - Autonomous coding lifecycle
Inspect a repository, propose an atomic patch set, test, repair, review, verify, and hand back a reviewable branch or pull request.
AUTONOMY-10 phase 7 - Colonies and connectors
Safe operation across external systems with scoped credentials, dry runs, and rollback.
AUTONOMY-10 phase 8 - Safe self-improvement
The colony improving its own procedures without widening its own authority.
AUTONOMY-10 phase 9 - Production qualification
Long-running operation that is observable, budgeted, recoverable, and proven through a soak.
AUTONOMY-10 phase 10
ANTHILL is pre-1.0 software under active development. Aspirational phases are not shipped claims — when in doubt, trust PLAN.md and CHANGELOG.md over any summary, including this one.