CONSOLE / MISSIONV0.6.5 / ON MAIN
Mission control for Claude Code.
Evidence before done.
A Claude Code plugin that shows the work as it really is: the milestones Claude reported, the agents the host observed, the checks that actually ran. A task reads DONE only when the evidence is in.
claude plugin marketplace add echelong/cobalt-cockpitThen claude plugin install cobalt-cockpit@cobalt-cockpit, restart Claude Code and type /cockpit.
01/13 STANDBY
Waiting for a prompt
STANDBY No task, no progress. The HUD shows nothing it has not observed.
Scripted from the repository’s own demo scenario and computed by a port of Cockpit’s completion, flow and readiness rules. How the demos work
- VERSION
- 0.6.5
- from .claude-plugin/plugin.json
- DEFAULT MODE
- OBSERVE
- orchestration: false
- OPTIONS
- 43
- declared in the manifest
- COMMANDS
- 26
- documented in the README
- NETWORK
- NONE OF ITS OWN
- see Safety and trust
- LAST COMMIT
- 2026-10-16
- main @ 2dbeaf1
01 / DOCTRINE
Five rules the plugin holds itself to.
Each one is a property of the code, stated in its documentation, with the page that says so.
Observe by default.
Installed as it ships, Cockpit watches the work Claude Code already does. It enforces no model routing, no subagent budget, no Fable block and no authentication rule until you switch one on. Two command guards are on from the start: it asks before genuinely destructive commands and keeps AI attribution out of commits.
Where it says soEnforce by opt-in.
orchestration: true enforces the Opus, Sonnet and Haiku policy, pool budgets and task ownership. blockFable, subscriptionOnly, cobaltStrict, the Stop guard and owner alerts are separate switches, each off until you choose it.
Where it says soEvidence before DONE.
A completed turn does not complete a task. DONE needs every milestone complete and every required gate passed or marked not applicable; for a coding task above the LIGHT level, a goal check with evidence for each acceptance criterion.
Where it says soNever an approval.
Cockpit takes no part in Claude Code’s permission check. Its guards run before it and can only refuse a call or ask you about it. Your permission rules still decide everything it lets through.
Where it says soLocal, with one exception you choose.
No fetch, no model call, no MCP call of its own. The one program it starts that can reach a network is the decision router in JEV mode, and only in a session where you selected it.
Where it says so
02 / INSTRUMENTS
Six modules. One band of truth above the prompt.
- MODULE 01Agent orchestrationBounded assignments recorded before anything spawns; admission by dependencies, capacity and resource ownership.swarm toolorchestration: truedefault: observe
- MODULE 02Neural Flow and progressINPUT to DONE in one HUD row, every state derived from real milestones, gates and observed agents./cockpit flow/cockpit statuson by default
- MODULE 03Agent InspectorEvery swarm task, why it waits, and what was reported versus what the commander verified./cockpit agents/cockpit agents <task>read-only
- MODULE 04Verification ReadinessThe completion rules applied to the recorded task: what still stands between it and DONE./cockpit readyread-onlyapproves nothing
- MODULE 05Cobalt GuardOwner alerts when a run needs a person, and an optional Stop hold until your own check passes./cockpit guard/cockpit alertsopt-in
- MODULE 06Run Ledger and replayA bounded local record of runs, agents, tools, usage and verification, with replay and checkpoints./ledger/replay/parklocal
03 / BOUNDARIES
What it will not do.
Cockpit supplements Claude Code. It does not replace its permissions, its model controls or your judgement, and its guards are heuristics, not a sandbox.
Read the trust modelApprove a tool call
There is no hook on the permission check. A guard refuses or asks; it never allows.
SourceCall a model from the dashboard
The HUD, Mission Control and the Ledger are drawn from what was observed. The dashboard makes no model calls.
SourceStore hidden reasoning
Tool calls, paths, agent events, usage, Git state and verification are observed. Hidden reasoning is never stored.
SourceCall a claim verified
Gates record observed exit status and reported evidence, and say which is which. That is a hold on progress, not a proof of correctness.
Source
04 / START
Two commands, a restart, /cockpit.
claude plugin marketplace add echelong/cobalt-cockpitclaude plugin install cobalt-cockpit@cobalt-cockpit/cockpitIn Claude Code, after a restart. It observes: models, budgets and ownership are not enforced until you configure them.