MAAS is a Codex-first control plane for supervising autonomous work.
It exists for the gap between “an agent can do work” and “an operator can safely trust an always-on system.” MAAS turns goals into issue inventories, runs Codex inside a governed execution loop, keeps review and recovery explicit, promotes reusable memory, and makes the entire control flow observable through runs, logs, traces, and incidents.
MAAS is not a generic dashboard and not a thin Codex wrapper. The current product direction is:
Codexas the MVP execution runtimeGoal,Issue,Run,Agent,Event, andIncidentas the core control objects- one human operator supervising execution instead of manually driving every task
- explicit review, recovery, memory, and delivery loops around autonomous work
Autonomous execution only becomes useful when the surrounding control loop is trustworthy. MAAS is designed to answer:
- what is the system doing right now?
- what needs operator judgment?
- why is progress blocked?
- what output did Codex actually produce?
- what memory, checks, and recovery rules changed the result?
GoalIssueRunAgentEventIncident
Command: what needs judgment nowTheater: live issue ownership, run posture, and branch lineage in one execution mapWork: sharedList | Boardview of issues plus detail and execution historyIssues: approvals, blocked work, failed runs, grouped review, and recovery actionsAgents: active ownership, execution threads, spawned work, and healthRuns: live and historical execution truthSystem: logs, metrics, queue posture, and machine healthProjects: create, import, clone, archive, delete, and posture management
- project-level autopilot instead of manual “run cycle” babysitting
- first-class run history, traces, and live execution truth
- explicit review policy and grouped review packets
- recovery playbooks instead of raw failure queues
- retrieval-backed project memory with promotion, attribution, and usefulness tracking
- delivery prep, GitHub draft PR sync, and verification-gated handoff
Today MAAS is a substantial local prototype built around:
- a Python/FastAPI backend with SQLite-backed state
- a React control plane for Command, Theater, Work, Issues, Agents, Runs, System, and Projects
- goal intake, issue synthesis, autopilot, review, recovery, retrieval, delivery-prep, and GitHub sync flows
- reconciliation-backed truth inspection and repair for stale run, task, agent, and delivery linkage
- retry-safe external effects for notifications, provider jobs, git workspaces, and GitHub PR sync
- persisted trust runs with deterministic fault injection and replayable incident evidence
- a GitHub Project-driven execution workflow for tracked issues, PRs, and review state
- live and simulated Codex execution paths with operator-visible logs, traces, and artifacts
The implementation history is long because the product has pivoted several times. The public direction above is the one that matters now; the detailed historical roadmap is still preserved below for implementation tracking.
Install the backend and frontend dependencies, then launch the API and web app locally:
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
cd web
npm install
cd ..
./scripts/maas-dev up
./scripts/maas-dev statusThe managed local preview workflow:
- creates or reuses a demo workspace under
/tmp/maas-dev/workspace - imports this repo into that workspace as a brownfield project
- starts the API on
0.0.0.0:8000and the web app on0.0.0.0:5173 - keeps pid files and logs under
/tmp/maas-dev - remembers the last selected workspace and port settings under
/tmp/maas-dev/dev-config.json
Stop or restart the stack with:
./scripts/maas-dev down
./scripts/maas-dev restartRun a persisted local trust soak with deterministic failure injection:
PYTHONPATH=src ./scripts/maas-trust-run --project-root /tmp/maas-dev/workspace --cycles 12 --sleep-seconds 60If :8000 or :5173 is already occupied on your machine, pass explicit overrides such as
./scripts/maas-dev --web-port 5183 up or ./scripts/maas-dev --api-port 8001 --web-port 5183 up.
status, restart, and down reuse the last selected settings from /tmp/maas-dev/dev-config.json.
If you want localhost-only binding instead of LAN exposure, pass --api-host 127.0.0.1 --web-host 127.0.0.1.
You can still bootstrap and run MAAS manually from the CLI when needed:
PYTHONPATH=src python3 -m maas init --project-root .
PYTHONPATH=src python3 -m maas db migrate --project-root .
PYTHONPATH=src python3 -m maas api --project-root .
cd web && npm run devRelevant design and roadmap documents:
- Autonomous Organization Pivot
- UI Reset
- Codex MVP Shape
- Codex MVP Integration Plan
- Codex MVP Hardening Plan
- Codex MVP Next Batch Plan
- Codex MVP Autonomy-Scale Plan
- Codex MVP Autopilot and Memory Plan
- Codex MVP Control-Loop Hardening Plan
- Codex MVP Doctor, Planning, and Delivery Loop Plan
- Codex MVP Delivery Execution and Verification Plan
- Codex MVP Explainability, Review, and Memory Plan
- Codex MVP Control-Loop Governance and Observability Plan
- Codex MVP Brownfield Depth Pass
- Codex MVP Brownfield Execution Leverage
- Codex MVP Brownfield Drift, Refresh, and Trust
- Execution Theater Foundation
- Execution Theater Field and Agent Motion
- Execution Theater Branch, Worktree, and PR Lineage
- Execution Theater Internal-Production Readiness
- Local Dev Lifecycle Scripts
- Unattended Local Trust v1: Invariants and Reconciliation
- Unattended Local Trust v1: Idempotent Side Effects and Retry Safety
- Unattended Local Trust v1: Stop States and Operator Truth
- Unattended Local Trust v1: Overnight Soak and Fault Injection
- Unattended Local Trust v1: Final Closure and Trust Gate
There is also a standalone product mockup for the current direction in mockups/maas-codex-mvp/README.md.
Active planning and execution now live in GitHub, not in the numbered implementation docs.
- execution layer: MAAS Delivery & Execution
- current-truth docs: README.md, docs/implementation/STATUS.md, and docs/implementation/WORKFLOW.md
- history/reference docs: the numbered implementation docs under docs/implementation
Use one GitHub issue per tracked task or roadmap item, keep project fields truthful, and link the PR once code work starts. Reconciliation now also repairs stale merged-state cards for this repo's own GitHub Project when a tracked issue is already closed by a merged PR.
For unattended local use, rely on the System surface trust gate rather than README prose. MAAS now exposes an explicit unattended-mode gate that only arms after a fresh passing trust soak, clean truth reconciliation, and healthy launch posture.
The rest of this README tracks shipped implementation and numbered delivery history as archive/reference.
Legend:
[x]completed in the current numbered delivery sequence[ ]not yet completed in the current numbered delivery sequence
Use the stacked-branch references below to see whether a completed item is already on main or only exists on stacked branches.
Historical stacked development chain above main:
#82exists oncodex/project-aware-supervisor-orchestration#83exists oncodex/brownfield-file-backed-planning#84exists oncodex/recovery-circuit-breakers#85exists oncodex/project-isolated-provider-runtime#86exists oncodex/provider-job-queue#87exists oncodex/provider-job-queue#88exists oncodex/file-linked-task-scopes#89exists oncodex/brownfield-runbook-command-catalog#90exists oncodex/brownfield-runbook-command-catalog#91exists oncodex/brownfield-runbook-command-catalog#92exists oncodex/queue-capacity-controls#93exists oncodex/session-runner-envelopes#94exists oncodex/policy-driven-self-healing-v2#95exists oncodex/brownfield-onboarding-review-v2#96exists oncodex/remote-executor-worker-pool#97exists oncodex/cross-project-scheduler-fairness#98exists oncodex/repo-grounded-plan-synthesis#99exists oncodex/verification-runners-evidence-capture#100exists oncodex/git-aware-task-workspaces#101exists oncodex/cross-project-command-center#102exists oncodex/queue-worker-capacity-governance#103exists oncodex/queue-worker-capacity-governance#104exists oncodex/queue-worker-capacity-governance#105exists oncodex/queue-worker-capacity-governance#106exists oncodex/queue-worker-capacity-governance#107exists oncodex/ux-product-redesign#108exists oncodex/ux-product-redesign#109exists oncodex/ux-product-redesign#110exists oncodex/ux-product-redesign#111exists oncodex/ux-product-redesign#112exists oncodex/ux-product-redesign#113exists oncodex/ux-product-redesign#114exists oncodex/ux-product-redesign#115exists oncodex/ux-product-redesign#116exists oncodex/ux-product-redesign#117exists oncodex/linear-vibekanban-cockpit#118exists oncodex/linear-vibekanban-cockpit#119exists oncodex/linear-vibekanban-cockpit#120exists oncodex/linear-vibekanban-cockpit#121exists oncodex/linear-vibekanban-cockpit#122exists oncodex/linear-vibekanban-cockpit#123exists oncodex/linear-vibekanban-cockpit#124exists oncodex/linear-vibekanban-cockpit#125exists oncodex/linear-vibekanban-cockpit#126exists oncodex/linear-vibekanban-cockpit#127exists oncodex/linear-vibekanban-cockpit#128exists oncodex/linear-vibekanban-cockpit#129exists oncodex/linear-vibekanban-cockpit#130exists oncodex/linear-vibekanban-cockpit#131exists oncodex/linear-vibekanban-cockpit#132exists oncodex/linear-vibekanban-cockpit#133exists oncodex/linear-vibekanban-cockpit#134exists oncodex/linear-vibekanban-cockpit#135exists oncodex/linear-vibekanban-cockpit#136exists oncodex/linear-vibekanban-cockpit#137exists oncodex/linear-vibekanban-cockpit#138exists oncodex/linear-vibekanban-cockpit#139exists oncodex/linear-vibekanban-cockpit#140exists oncodex/linear-vibekanban-cockpit#141exists oncodex/linear-vibekanban-cockpit#142exists oncodex/linear-vibekanban-cockpit#143exists oncodex/linear-vibekanban-cockpit#144exists oncodex/linear-vibekanban-cockpit#145exists oncodex/linear-vibekanban-cockpit#146exists oncodex/linear-vibekanban-cockpit#147exists oncodex/linear-vibekanban-cockpit#148exists oncodex/linear-vibekanban-cockpit#149exists oncodex/linear-vibekanban-cockpit#150exists oncodex/linear-vibekanban-cockpit#151exists oncodex/linear-vibekanban-cockpit#161exists oncodex/codex-mvp-shell-integration#162exists oncodex/codex-mvp-shell-integration#163exists oncodex/codex-mvp-shell-integration#164exists oncodex/codex-mvp-shell-integration#165exists oncodex/codex-mvp-shell-integration#166exists oncodex/codex-mvp-shell-integration#167exists oncodex/codex-mvp-shell-integration#168exists oncodex/codex-mvp-shell-integration#169exists oncodex/codex-mvp-shell-integration#170exists oncodex/codex-mvp-shell-integration#171exists oncodex/codex-mvp-hardening#172exists oncodex/codex-mvp-hardening#173exists oncodex/codex-mvp-hardening#174exists oncodex/codex-mvp-hardening#175exists oncodex/codex-mvp-hardening#176exists oncodex/codex-mvp-hardening#177exists oncodex/codex-mvp-hardening#178exists oncodex/codex-mvp-hardening
The current operator-value and capability sequence in docs/implementation/14-codex-mvp-next-batch-plan.md is now implemented on codex/codex-mvp-operator-scale. It adds truthful queue posture, readiness-aware launch strategy, shared Work/Issues scopes, first-class run detail reads, stronger agent/system execution diagnostics, verification-driven auto-approval, and cleaner fresh-project lifecycle controls.
The current autonomy-scale sequence in docs/implementation/15-codex-mvp-autonomy-scale-plan.md is now implemented on codex/codex-mvp-autonomy-scale. It adds first-class Runs, backend-owned exception grouping, retrieval across issues/runs/artifacts/events, clone-for-fresh-run project lifecycle, stronger stale-run diagnostics, cross-project supervision in Projects, and an async attention loop with optional desktop notifications.
The current autopilot-and-memory sequence in docs/implementation/16-codex-mvp-autopilot-memory-plan.md is now implemented on codex/codex-mvp-autopilot-memory. It adds project templates, project-level autopilot, memory promotion and retrieval-backed Codex prompts, backend-owned batch review, and stronger execution-state truth across Command, Issues, System, and issue detail.
The current control-loop hardening sequence in docs/implementation/17-codex-mvp-control-loop-hardening-plan.md is now implemented on codex/codex-mvp-control-loop-hardening. It adds durable autopilot lease state, a backend-owned operator inbox, lifecycle-safe clone posture resets, grouped review packet truth, notification failure integration, and fresher execution-memory attribution.
The current doctor, planning, and delivery-loop sequence in docs/implementation/18-codex-mvp-doctor-delivery-loop-plan.md is now implemented on codex/codex-mvp-doctor-delivery-loop. It adds an environment doctor, first-class goal creation and synthesis, delivery candidate reads plus PR-draft preparation, stronger autopilot governance gates, goal-scoped review packets, and usefulness-aware execution memory.
The current product-modeling sequence on codex/linear-vibekanban-cockpit now covers the cockpit pivot (#127-#136), the Linear/Vibekanban-inspired workflow cleanup (#137-#146), and the clarified Cockpit/Board role split (#147-#151).
The current Codex-MVP integration sequence on codex/codex-mvp-shell-integration carries that product reset into the real app with a new shell, canonical issue identity, issue detail and agent detail read models, and integrated Command / Work / Issues / Agents / System surfaces.
The current hardening sequence on codex/codex-mvp-hardening makes the integrated MVP truthful in live use: explicit run-cycle vs launch posture controls, review-first issue detail, simulation/live warnings, project delete plus fresh workspace creation, and tighter lifecycle regressions.
- MAAS is now a substantial single-project, greenfield, operator-supervised prototype.
- The core loop exists end to end: bootstrap, board, supervisor, provider execution, failure handling, quarantine, recovery, and artifact inspection.
- For that current prototype shape, the repo is roughly
85-90%complete. - For the broader roadmap vision, the repo is still materially incomplete.
- Greenfield bootstrap with seeded goals, tasks, agents, alerts, and sessions
- Board-first API and React control room
- Ready-queue refresh, acceptance evaluation, and first-pass idle-agent allocation
- Steering controls for review, reprioritize, reassign, pause/resume, and halt
- Escalation queue for risky steering approvals
- Failure-memory logging, quarantine visibility, repeated-failure alerts, incident-specific alert actions, and task recovery for failure-blocked work
- Manual recover-and-requeue for failure-blocked tasks
- Timed-out and failed-session auto-retry with tracked retry state
- Explicit scheduler scoring, board-visible scheduler rationale, and adaptive replan guidance
- Manual replanning queue plus dead-letter routing for retry-exhausted work
- Quarantine queue workflow with restore, dismiss, reopen, and restore+requeue actions
- Recovery for agents left in
error - Real local Claude Code CLI integration behind explicit provider config
- Real local OpenAI Codex CLI integration behind explicit provider config
- Provider status visibility with effective mode, runtime controls, config warnings, preflight readiness, recent run history, manual run controls, mode switching, and editable settings
- Artifact browser and artifact-state visibility in the control room
- Artifact browser supports preview, guarded download, compare, lineage/provenance pivots, and task/session export bundles
- Artifact browser operator actions for restore, restore-and-requeue, dismiss, and reopen on quarantined artifacts
- Shared live transport with websocket, SSE, and polling fallback status in the control room shell
- Product UX simplification, clearer mental model, and stronger first-run guidance
- Visually strong dual light/dark theme and a real design system instead of the current admin-tool feel
- Broader automated restart, retry, backoff, and self-healing workflows beyond the current DLQ path
- Broader external provider coverage beyond the current local CLI paths
- Higher-level artifact retention policy automation beyond the current browser, provenance, and export flows
- Deeper brownfield onboarding and multi-project execution support
- Stronger sandboxing and isolation guarantees
- Project-aware background orchestration beyond the current multi-project read scope
-
#80Provider runtime preflight and readiness checks -
#81Multi-project write path and project lifecycle -
#82Project-aware supervisor and background orchestration -
#83Brownfield file-backed planning and repo navigation -
#84Policy-driven self-healing and circuit breakers -
#85Sandboxed provider runners per project -
#86Remote or queued provider execution beyond local CLI paths -
#87Brownfield rescan and drift detection -
#88File-linked task scopes and acceptance criteria -
#89Brownfield runbook and command catalog -
#90Portfolio view across projects -
#91Background orchestration daemon -
#92Queue and worker capacity management on top of the provider job queue -
#93Stronger runner sandbox envelopes beyond the current per-project runtime isolation -
#94Policy-driven self-healing v2 -
#95Brownfield onboarding review v2 -
#96Remote executor or worker pool -
#97Cross-project scheduler fairness and capacity policy -
#98Repo-grounded plan synthesis and refresh -
#99Verification runners and evidence capture -
#100Git-aware task workspaces and diff review -
#101Cross-project command center -
#102Queue and worker capacity controls -
#103Policy-driven approval and risk routing -
#104Cost, runtime, and quota controls -
#105Notifications and outbound integrations -
#106Incident timeline and replay -
#107Information architecture reset and navigation collapse -
#108Design system and dual light/dark theme foundation -
#109Home command center with recommended actions -
#110Guided onboarding and first-run experience -
#111Unified Work surface for board, plan, and task detail -
#112Unified Runs surface for agents, providers, verification, and outputs -
#113Unified Incidents surface for failures, alerts, recovery, and timeline -
#114Portfolio and project-management UX redesign -
#115Command palette, contextual actions, empty states, and inline guidance -
#116Accessibility, responsiveness, and visual-polish pass -
#117Shell density reset -
#118Default control room layout -
#119Compact kanban redesign -
#120Agent roster and interaction view -
#121Curated live ticker -
#122Goal/subgoal/task relationship explorer -
#123Incident rail and playbooks -
#124Evidence and verification drawer -
#125Project status and portfolio command bar -
#126Remove legacy hero UX and final dense visual pass -
#127Seraph-style cockpit shell pivot -
#128Panel workspace and windowed composition -
#129Cockpit command bar and telemetry strip -
#130Dense agent rail -
#131Center kanban workspace rewrite -
#132Right ops rail for incidents and ticker -
#133Inspector and evidence workspace -
#134Utility drawers for projects, providers, and settings -
#135Cockpit typography and theme system -
#136Remove legacy page surfaces and finish the cockpit interaction model -
#137Guided onboarding takeover -
#138Compact issue cards and inspector-first steering -
#139Board-first workflow defaults -
#140Import and create drawers -
#141Project next-step workflow -
#142Curated live-ops priorities -
#143Linear/Vibekanban-inspired board polish -
#144Utility and settings demotion -
#145Dense typography and dark-theme cleanup -
#146Remove remaining inline control clutter -
#147Distinct Cockpit and Board roles -
#148Guided brownfield review/start flow -
#149Unified Run action and advanced control demotion -
#150Intent-first project setup flow -
#151Inspector-visible board interaction model
-
#81is already shipped onmain -
#82is implemented oncodex/project-aware-supervisor-orchestration -
#83is implemented oncodex/brownfield-file-backed-planning -
#84is implemented oncodex/recovery-circuit-breakers -
#85is implemented oncodex/project-isolated-provider-runtime -
#86is implemented oncodex/provider-job-queue -
#87is implemented oncodex/provider-job-queue -
#88is implemented oncodex/file-linked-task-scopes -
#89is implemented oncodex/brownfield-runbook-command-catalog -
#90is implemented oncodex/brownfield-runbook-command-catalog -
#91is implemented oncodex/brownfield-runbook-command-catalog -
#92is implemented oncodex/queue-capacity-controls -
#93is implemented oncodex/session-runner-envelopes -
#94is implemented oncodex/policy-driven-self-healing-v2 -
#95is implemented oncodex/brownfield-onboarding-review-v2 -
#96is implemented oncodex/remote-executor-worker-pool -
#97is implemented oncodex/cross-project-scheduler-fairness -
#98is implemented oncodex/repo-grounded-plan-synthesis -
#99is implemented oncodex/verification-runners-evidence-capture -
#100is implemented oncodex/git-aware-task-workspaces -
#101is implemented oncodex/cross-project-command-center -
#102is implemented oncodex/queue-worker-capacity-governance -
#103is implemented oncodex/queue-worker-capacity-governance -
#104is implemented oncodex/queue-worker-capacity-governance -
#105is implemented oncodex/queue-worker-capacity-governance -
#106is implemented oncodex/queue-worker-capacity-governance -
#107is implemented oncodex/ux-product-redesign -
#108is implemented oncodex/ux-product-redesign -
#109is implemented oncodex/ux-product-redesign -
#110is implemented oncodex/ux-product-redesign -
#111is implemented oncodex/ux-product-redesign -
#112is implemented oncodex/ux-product-redesign -
#113is implemented oncodex/ux-product-redesign -
#114is implemented oncodex/ux-product-redesign -
#115is implemented oncodex/ux-product-redesign -
#116is implemented oncodex/ux-product-redesign -
#117is implemented oncodex/linear-vibekanban-cockpit -
#118is implemented oncodex/linear-vibekanban-cockpit -
#119is implemented oncodex/linear-vibekanban-cockpit -
#120is implemented oncodex/linear-vibekanban-cockpit -
#121is implemented oncodex/linear-vibekanban-cockpit -
#122is implemented oncodex/linear-vibekanban-cockpit -
#123is implemented oncodex/linear-vibekanban-cockpit -
#124is implemented oncodex/linear-vibekanban-cockpit -
#125is implemented oncodex/linear-vibekanban-cockpit -
#126is implemented oncodex/linear-vibekanban-cockpit
The current numbered #81-#126 sequence is fully implemented on the stacked branch chain above main.
-
#107Information architecture reset and navigation collapse: reduce the current top-level control-room sprawl to a smaller set of primary surfaces with clearer user-language labels and a stronger product mental model. -
#108Design system and dual light/dark theme foundation: introduce semantic color, spacing, typography, and elevation tokens plus persistent light/dark modes so the UI no longer feels like an internal admin tool. -
#109Home command center with recommended actions: replace the current overloaded overview posture with a “what needs attention / what should I do next” landing experience. -
#110Guided onboarding and first-run experience: add create/import/setup flows that explain what MAAS is doing, what the project state means, and what the next safe step is. -
#111Unified Work surface for board, plan, and task detail: merge board, goal tree, and repo-grounded planning into a single execution workspace with richer task detail and evidence drawers. -
#112Unified Runs surface for agents, providers, verification, and outputs: stop splitting execution state across separate silos and present one coherent operator view of active work and produced evidence. -
#113Unified Incidents surface for failures, alerts, recovery, and timeline: turn incident response into one guided workbench with clear playbooks instead of four separate mechanism-heavy pages. -
#114Portfolio and project-management UX redesign: improve project switching, lifecycle actions, cross-project health, and multi-project supervision without burying the user in policy forms. -
#115Command palette, contextual actions, empty states, and inline guidance: make advanced functionality discoverable without overwhelming the default UI. -
#116Accessibility, responsiveness, and visual-polish pass: finish the redesign with mobile/tablet behavior, keyboard-first interactions, stronger hierarchy, and usability QA.
-
#117Shell density reset: remove the oversized landing-page shell and replace it with a compact top strip, tighter navigation, and smaller controls. -
#118Default control room layout: make the landing screen a dense three-pane operator cockpit with agents on the left, kanban in the center, and ops context on the right. -
#119Compact kanban redesign: replace oversized cards with compact execution cards that expose assignee, goal, evidence signals, and failure pressure at a glance. -
#120Agent roster and interaction view: turn agents into first-class visible actors with status, current work, heartbeat, and quick intervention hooks. -
#121Curated live ticker: add a dense meaningful-event feed so the system feels alive without turning into raw telemetry spam. -
#122Goal/subgoal/task relationship explorer: expose selected-task goal lineage, sibling work, repo-plan matches, and recent task-specific history in one inspector. -
#123Incident rail and playbooks: surface actionable incidents directly in the right rail instead of forcing operators to hunt through separate admin pages. -
#124Evidence and verification drawer: put verification state, git diff evidence, artifacts, and task history next to the selected task. -
#125Project status and portfolio command bar: move project selection, health, alert load, and transport/runtime status into a compact top command strip. -
#126Remove legacy hero UX and final dense visual pass: shrink typography, card heights, and button scale across the control room so the product reads like an operations cockpit instead of a landing page. -
#127Seraph-style cockpit shell pivot: replace the page-and-card shell with a fixed-height cockpit shell, compact top bar, internal scroll regions, and explicit workspace modes such as ops, focus, and review. -
#128Panel workspace and windowed composition: reorganize the default experience into persistent left-rail, center-workspace, and right-rail panels with panel-scoped overflow instead of stacked sections. -
#129Cockpit command bar and telemetry strip: collapse navigation, project switching, live transport state, run controls, and quick commands into one dense operator bar. -
#130Dense agent rail: rebuild agent visibility as a compact live rail with status, current work, recent meaningful events, and fast intervention hooks. -
#131Center kanban workspace rewrite: make the board the primary workspace with denser cards, fewer empty lanes, scoped backlog access, and better ops/focus/review modes. -
#132Right ops rail for incidents and ticker: merge the curated feed, approvals, alerts, and incident shortcuts into one compact ops rail instead of separate dashboard blocks. -
#133Inspector and evidence workspace: turn selected-task context into a true inspector with goal path, repo scope, verification, git diff, artifacts, and history in one persistent panel. -
#134Utility drawers for projects, providers, and settings: move lifecycle forms, provider settings, quotas, and policy editors out of the primary workspace into drawers and advanced utility windows. -
#135Cockpit typography and theme system: replace the current generic card styling with a denser mono-forward cockpit system inspired by Seraph’s shell without copying its game flavor. -
#136Remove legacy page surfaces and finish the cockpit interaction model: delete or demote the old dashboard/page paradigm so MAAS reads as one coherent cockpit instead of multiple competing UI systems.
-
#81Multi-project write path and project lifecycle -
#82Project-aware supervisor and background orchestration -
#83Brownfield file-backed planning and repo navigation -
#84Policy-driven self-healing and circuit breakers -
#85Sandboxed provider runners per project -
#86Remote or queued provider execution beyond local CLI paths -
#87Brownfield rescan and drift detection -
#88File-linked task scopes and acceptance criteria -
#89Brownfield runbook and command catalog -
#90Portfolio view across projects -
#91Background orchestration daemon -
#92Queue and worker capacity management on top of the provider job queue -
#93Stronger runner sandbox envelopes beyond the current per-project runtime isolation -
#94Policy-driven self-healing v2 -
#95Brownfield onboarding review v2 -
#96Remote executor or worker pool -
#97Cross-project scheduler fairness and capacity policy -
#98Repo-grounded plan synthesis and refresh -
#99Verification runners and evidence capture -
#100Git-aware task workspaces and diff review -
#101Cross-project command center -
#102Queue and worker capacity controls -
#103Policy-driven approval and risk routing -
#104Cost, runtime, and quota controls -
#105Notifications and outbound integrations -
#106Incident timeline and replay -
#107Information architecture reset and navigation collapse -
#108Design system and dual light/dark theme foundation -
#109Home command center with recommended actions -
#110Guided onboarding and first-run experience -
#111Unified Work surface for board, plan, and task detail -
#112Unified Runs surface for agents, providers, verification, and outputs -
#113Unified Incidents surface for failures, alerts, recovery, and timeline -
#114Portfolio and project-management UX redesign -
#115Command palette, contextual actions, empty states, and inline guidance -
#116Accessibility, responsiveness, and visual-polish pass -
#106Incident timeline and replay
PYTHONPATH=src python3 -m maas init --project-root .
PYTHONPATH=src python3 -m maas db migrate --project-root .
PYTHONPATH=src python3 -m maas task ready --project-root . --refresh
PYTHONPATH=src python3 -m maas task allocate --project-root .
PYTHONPATH=src python3 -m maas supervisor --project-root . --once
PYTHONPATH=src python3 -m maas api --project-root .The project bootstrap creates:
project.yaml.maas/.maas/state.db.maas/artifacts/.maas/logs/.maas/quarantine/
GET /api/healthGET /api/boardGET /api/goalsGET /api/agentsGET /api/activityGET /api/alertsGET /api/escalationsGET /api/failuresGET /api/artifactsGET /api/artifacts/exportGET /api/quarantineGET /api/liveWS /api/live/wsGET /api/overviewGET /api/goals/treeGET /api/providersGET /api/tasks/readyGET /api/tasks/{task_id}/capabilitiesPOST /api/escalations/requestPOST /api/escalations/{escalation_id}/actions/approvePOST /api/escalations/{escalation_id}/actions/rejectPOST /api/providers/{provider_id}/actions/run-taskPOST /api/tasks/actions/refresh-readyPOST /api/tasks/actions/allocate-readyPOST /api/tasks/{task_id}/actions/evaluatePOST /api/tasks/{task_id}/actions/recoverPOST /api/tasks/{task_id}/actions/recover-and-requeuePOST /api/tasks/{task_id}/actions/resolve-repeated-failuresPOST /api/quarantine/{queue_id}/actions/restorePOST /api/quarantine/{queue_id}/actions/dismissPOST /api/agents/{agent_id}/actions/assign-nextPOST /api/agents/{agent_id}/actions/recoverPOST /api/supervisor/run
The primary operational surface is the Kanban board returned by /api/board.
maas task ready --project-root . --refreshmaas task allocate --project-root .maas task allocate --project-root . --agent-id <agent_id>maas task evaluate --project-root . --task-id <task_id>maas task recover --project-root . --task-id <task_id> --actor-id <agent_id>maas task recover-and-requeue --project-root . --task-id <task_id> --actor-id <agent_id>maas task resolve-repeated-failures --project-root . --task-id <task_id> --actor-id <agent_id>maas agent recover --project-root . --agent-id <agent_id> --actor-id <agent_id>maas supervisor --project-root . --oncemaas failure list --project-root .maas quarantine list --project-root .maas quarantine restore --project-root . --queue-id <queue_id> --actor-id <agent_id>maas quarantine dismiss --project-root . --queue-id <queue_id> --actor-id <agent_id>maas escalation list --project-root .maas escalation request --project-root . --project-id <project_id> --actor-id <agent_id> --action-type halt_task|reassign_task|pause_agent|resume_agent --resource-type task|agent --resource-id <resource_id>maas escalation approve --project-root . --escalation-id <escalation_id> --actor-id <agent_id>maas escalation reject --project-root . --escalation-id <escalation_id> --actor-id <agent_id>maas worker --project-root . --provider-type python_script|claude_code|openai_codex ...
These commands expose the current dependency-aware ready queue, allocator flow, acceptance-gate evaluation, supervisor orchestration pass, and escalation approval flow from the CLI.
- board and alert actions are gated by role-baseline
board_actionspermissions fromproject.yaml - task execution now uses task-scoped capability grants, so lifecycle writes are limited to the assigned agent and task
- board cards and the task capabilities API expose the currently active task grants
- risky task and agent interventions can now be routed through an escalation queue instead of being executed immediately
- failed and timed-out sessions are now recorded in failure memory and can raise repeated-failure alerts
- quarantined failure artifacts are isolated under
.maas/quarantine/and surfaced through the failure-memory reads - first-class quarantine queue reads and actions now track open, restored, and dismissed artifact incidents
- recent failure and overview surfaces expose direct operator actions for recovery, restore, dismiss, reopen, and repeated-failure resolution
- operators can return failure-blocked tasks to the planning queue without resuming the old execution context
- timed-out and failed sessions can auto-retry under project recovery policy with tracked retry state
- operators can recover timeout-stranded agents from
errorback toidleonce no active session remains
python_scriptis the reference local worker adapterclaude_codesupports both the simulated adapter and a real localclaude -ppath when enabled inproject.yamlopenai_codexsupports both the simulated adapter and a real localcodex execpath when enabled inproject.yaml/api/providersand the Providers view expose configured mode, effective mode, config warnings, recent provider runs, safe manual run targets, mode switching, and editable runtime settings/api/artifactsand the Artifacts view expose artifact state, missing-file detection, quarantine metadata, and server-side filtering- artifact detail now includes preview, guarded single-file download, task/session export bundles, same-task compare, same-session lineage, and dependency-linked provenance pivots