ArchitectureA guided teardown of the harness

The mission is the unit of work. Agents are workers inside it.

One durable loop holds the objective, your approved plan, isolated specialists, execution surfaces, and the memory the next mission can use.

harness / cutaway5 layers · exploded view
SF-05 · same agent · same knowledge · bigger surface

Interactive cutaway · hover a layer to inspect

The mission loop

A mission is a state machine, not a chat.

The objective is the durable thing. Its state survives the session, plan mode blocks for approval, and bounded specialist work stays attached to the same local mission.

state
persisted beyond the session
gate
human approval is a state, not a popup
review
separate mutation-disabled profile
Watch one full pass, with timestamps
mission / state machinepass 1 · state persisted
discover · walking imports · pipelines · runtime
delegation packet · w1/4mission rl-311
to
repo agent 1 of 4
model slot
“frontier” · routed by you
repository
api-gateway · branch mission/rl-311
objective slice
enforce per-tenant limits at the edge
pinned contract
rate-limit headers v1 · read-only

tools granted

read repositoryedit isolated worktreerun focused checks

out of bounds: unassigned repositories · unrelated credentials · reviewer mutation

exit: bounded work complete · ready for review

The packet is the whole world that agent sees.

Delegation

Specialists get a packet, not the keys.

The orchestrator splits the plan into work orders: one repository, one objective slice, the pinned contract, an allowed toolset, and a model slot. Everything outside the packet does not exist for that agent.

isolation
one boundary per agent
contract
pinned before any code changes
verifier
independent · writes no code
The context bus

One mission across every available surface.

Code, browser, terminal, monitors, automations, and personal memory remain attached to the same objective instead of becoming unrelated sessions.

Repositories

local git

readsregistered checkouts · files · history

writesisolated specialist worktrees

local

Browser

mission surface

readsvisible and interactive page state

writesclick · type · navigate when allowed

available

Terminal

local process

readsstdout · stderr · exit state

writescommands within role capability

available

Monitors

mission schedule

readsrecurring check results

writesstatus attached to the mission

available

Automations

local harness

readstask and prior execution context

writesscheduled engineering work

available

Memory

local service

readsscoped · cited · relevant records

writesreview · accept · dismiss · forget

reviewed
one mission · multiple local surfaces · role capability stays explicit
The control plane

Autonomy with an observed impact map.

Four standing rules bound every mission. They are not settings buried in a menu — they are how the loop itself is built, and every one of them leaves a trace you can replay.

default
read-only until the plan says otherwise
approvals
consequence-triggered · yours to widen
audit
every action attributable
reviewer proof
agent-profiles.test.ts
control plane / policy4 standing rules

Start read-only

Every mission opens in discovery. Writes are explicit, later states.

r.01

Gate on consequence

Plan mode blocks until you approve the implementation boundary.

r.02

Execute locally

Agents run on your linked Mac with your repositories and local credentials.

r.03

Record everything

Actions, prompts, and approvals land in the mission log, replayable.

r.04

audit / rl-311

t+00:06plan proposed · implementation blocked

t+00:09plan approved · boundary pinned

t+00:33read-only reviewer report attached

knowledge / write-backon mission complete

+outcomerate limits enforced at api-gateway · limits.yaml per plan tier

+constraint429 responses must carry Retry-After · v1 clients surface it

~ownershipplatform-infra owns the redis cluster (updated)

reviewed by dreamer · promoted with evidence links
Knowledge write-back

The machine gets smarter on the way out.

Completion is a write, not a wrap-up. Outcomes, constraints, and ownership facts are committed with evidence links — so the next mission recalls instead of rediscovers.

Questions

The machine, precisely.

Short answers about trust boundaries and moving parts. For the running system, watch a mission replay.

01What persists?+

The mission objective and state survive the session, so work can continue without becoming a new chat-shaped project.

02When does implementation start?+

Plan mode blocks until you approve the plan. The approved boundary then guides specialist work.

03What can specialists use?+

The current harness exposes code, browser, terminal, monitors, and automations according to each role's capabilities.

04What does independent review mean?+

The reviewer receives only file.read, file.search, and agent.report, receives read-only capability operations, and has mutationAllowed set to false.

Housing back on

Now give it an objective.

Skyflo Desktop is free with your own API keys or local models. The whole machine — loop, agents, bus, knowledge — arrives in one install.

Local execution, with no markup on BYOK or provider-direct usage.