A useful session can still leave no durable operation.
Chat history is a fragile place to keep project state. Decisions, discoveries, verification, and the next action need a home another session can inspect.
← The Lab
Case 001Released system / agent operations
Repository-local operating layer
Citadel
Long-running coding-agent work needs a place to route, resume, verify, and hand off.
Agents can write code. The harder problem is operating them across work that has to survive a single chat.
ArtifactReleased open-source system
ProblemOperating agents across real work
RoleRepository-local operating layer
The operating idea
Make the loop visible.
Citadel is the harness around Claude Code and OpenAI Codex. It keeps the work legible between the request and the next useful action.
Citadel / operating layerPublic system readout · 01
Citadel
Operating infrastructure for coding agents.
01Intentname the outcome
02Routechoose the lightest mode
03Worktreebound the change
04Verifykeep evidence
05Handoffleave a way back in
durable state/do routingsafety hookscost telemetryparallel fleets
Current handoffResume from evidence
campaign / phase-03 / verifiedTry it
Choose the work. See the operating choice.
This public-safe reconstruction shows how a request becomes a bounded mode, a traceable operation, and an explicit handoff without exposing a private repository.
Public-safe reconstructionMission Control / operating loop
01 Intent router
Describe the work. Inspect the operating choice.
Selected operating modeCampaign
DurableThe work needs phases, decisions, handoffs, and a concrete place to resume after interruption.
- Campaign state
- Phase evidence
- Session handoff
02 Mission Control
One operating loop. Visible state.
Choose a stage to inspect what Citadel keeps legible.
Current stageRoute
Start with the outcome. Citadel selects the lightest operating mode that fits the work.
/do + routing recordOperationservice-refactorEvidenceInspectableNext actionResume phase 03
The software-shaped problem
The model is only one part of the system.
More agents create an operating problem.
Parallelism is valuable only when scopes are independent, branches are isolated, discoveries travel, and a human can review what should actually land.
“Done” should resolve to evidence.
Citadel records checks and handoffs with passed, failed, blocked, or unknown states. It improves operating discipline; it does not replace code review or guarantee correctness.
The governing model
Engine, operating layer, source of truth.
Citadel does not try to become a new coding-agent runtime. It supplies the workflows and durable state that help existing runtimes operate inside a real repository.
03RepositoryCode, rules, project state, verification evidence
02CitadelRoute, coordinate, safeguard, verify, hand off
01Coding agentClaude Code or OpenAI Codex executes the work
Evidence and limits
Show the claim. Keep the boundary visible.
One natural-language entry point
/do routes a request toward a focused workflow, campaign, fleet, or other bounded operation.
State outside the chat
Campaigns, decisions, discoveries, reports, and handoffs live with the project instead of disappearing with a context reset.
Evidence with named states
Checks remain passed, failed, blocked, or unknown. A missing result does not become success through presentation.
Discipline is not correctness
The harness runs with the agent’s permissions and cannot eliminate human review, repository-specific tests, or consequential approvals.
repository / publicSethGammon / Citadel
open source · documented · reversibleInstall pathdocumentedVerificationnamed statesRemoval pathincluded
Inspect the system
Open source, with the proof boundary attached.
The repository contains the install path, demo workflow, Mission Control specification, operating contracts, verification gates, threat model, and removal path.