7 layers. Click any block to open it.
Memory
SQLITE · FTS5SQLite + FTS5 across every session, project, and runtime.
Runs alongside every layer to the right — the agent decides what
matters, the code owns the mechanics.
Capsule probe
Every eligible main prompt runs
mem candidates over the active capsule index — at most 3 headline+ID hits in 1,200 bytes, bodies untouched.Opportunity receipt
The probe writes a same-turn receipt even on zero hits. Material main-session mutation is gated on it, so retrieval can't be silently skipped.
Agent-owned recall
No score threshold adopts a hit. The agent reads the full record by ID, then cross-checks it against live code.
Session distillation
SessionEnd dispatches a no-tools distiller. Automatic writes declare one purpose: decision, user-correction, unresolved-obligation, or artifact-pointer.
Supersede, never delete
Changed decisions are superseded — one active canonical path, full history auditable, deleted rows recoverable from the graveyard.
Pending protection
Handoffs stay
pending until an explicit consume; prune, merge, and delete fail closed against them.Workers are exempt: no probe, no inject, no distill, no sync.
●
“Implement and test the login API, then leave a change report.”
L0
one sentence in
one sentence in
L1
Routing contract
Five fields you approve — then the route is sealed to commit, write scope, and model profile.
Route card — the whole handshake
taskImplement and test the login API, then leave a change report
reasonThe endpoint is already in the spec; no implementation cycle has run against it
routeautopilot-code · dev ·
standard — multi-file, needs independent verificationscopesrc/auth/**, tests/auth/** — no schema or migration change
doneTests green,
plans/<cycle>/pipeline_summary.md written
ProceedAmendStop
one approval, five fields, before anything is touched
Sealed the moment you approve
route_hash rt-35552ff2…
write_scope source-scoped, plans/<cycle>/**
bound to registry digest · commit · cwd
write_scope source-scoped, plans/<cycle>/**
bound to registry digest · commit · cwd
Intensity ladder
Governance
COREWORKFLOW
CONVENTIONSOPERATIONS
MEMORYDESIGN_PRINCIPLES
L2
Pipelines
Four pipelines over 12 entry routers and 26 capabilities.
Research & experiments
→→→→● artifact-guard:after-research● artifact-guard:after-spec
Libraries & CLI
→→● artifact-guard:after-analyze● artifact-guard:after-spec
Documents
→→→→● artifact-guard:after-research
Apps
→→→● artifact-guard:after-spec
L3
Dispatch fabric
The owner conducts and workers never route. 2–4 legs in one transaction, live in Fleet.
d0
d1
d2
code-planplan-checkcode-executeimpl-reviewcode-testcode-report
N-way group · one transaction · 2–4 legs
Checked fallback
Accounting
LIVE
Fleet
one word:fleetEvery session and every dispatched worker across the three
runtimes, in one tree — with the sealed model profile each one is running under and
how much context it has left.
fleet
LIVE
usageclaude
5h ━━━━━━━━──── 71%
7d ━━━━━━━━━━── 88% ↷ 3d23h
codex
5h ──────────── —
7d ━━━━━━━━━━━━ 100% ↷ 2d16h
●heartingclaude · owner d1━━━━━━──────53%working16h
│code-executeclaude · light━━━─────────21%working4m
│impl-reviewcodex · light━━━━────────34%done2m
│failure-modecodex · deep━───────────11%blocked1m
○corrnet_runtimeinteractive━───────────12%idle38m
the owner and its dispatched stages under one repo · orphaned
rows surfaced, never dropped · per-session token accountingfleet · --once for a plain snapshot · --json to script it
L4
Model tier per role
deep · balanced-deep · light · mini — chosen per node, sealed at compile time.
Execution profiles
Roles in the unit catalog — 8
Selection, per node
| node | role | profile |
|---|---|---|
| owner (standard+) | deep orchestrator | deep |
| frame · plan | deep maker | balanced-deep |
| plan (strong+) | deep maker | deep |
| execute | fast implementer | light |
| impl-review | fast reviewer | light |
| failure-mode | deep reviewer | deep |
| adversary leg | external adversary | deep |
| test · report | fast reviewer / writer | light |
L5
Fixed artifact system
.agent_reports/ — one root, one direction, one owning capability per artifact.
→
→
one direction — a later stage never invents its own evidence
research/external evidenceautopilot-research
analysis_project/source analysisanalyze-project
spec/current blueprintautopilot-spec
plans/code cyclesautopilot-code
documents/draftsdraft / refine
experiments/runs & logsautopilot-lab
one root per projectthe owning capability updates itspec revisions snapshot the prior versionworktree snapshot writes fail closedno code without a spec
Guards
FAIL-CLOSED39 hooks, 5 of them hard blocks — denied before the tool call.
Wrap every layer to the left. Deterministic hooks run before the
tool call — the model judges, the code enforces.
write scope · spec read · artifact root ·
git state · memory path
L6
The harness watches itself
5 loops that test, corroborate, and propose — none of them edits policy.