Factory Supervisor

Example operating scopes:

• One action: “Move work item X to Planning.”
• Bounded autonomy: “Keep project X triaged for this session. Create and update work items, transition Intake items to Triage, and retry retryable failed decisions. Do not delete, archive, start runs, or change automation settings.”
• Broad autonomy: “Operate project X autonomously, including transitions, decision actions, attention management, and run starts, until the queue is healthy. Do not delete the project.”

Treat these as delegated authority, not merely suggestions. Continue without per-action confirmation while actions remain within scope, and stop or ask when a boundary, stop condition, or material ambiguity is reached. A newer user instruction narrows, expands, or revokes the scope immediately.

Factory-specific routing and discovery

Factory commands are under:

Factory HTTP routes are root-level /web/*, not /api/*. The CLI’s Factory commands deliberately bypass --server-api-prefix; do not add /api, inspect /api/system/api-schema, or expect Factory routes in that manifest.

The installed CLI bundles generated Factory route contracts. For every leaf that accepts JSON, use its leaf --schema as the authoritative contract:

Some positional-only leaves do not expose --schema. For those, use the narrowest leaf help and do not guess arguments:

Operating contract

Use this decision flow:

1. For status, list, get, inspect, summarize, diagnose, or recommend requests, use the read-only workflow directly unless the user has also granted a mutation objective.
2. For create, update, delete, transition, start, approve, dismiss, retry, read, read-all, archive, or restore requests, verify that the action is authorized by the current request or standing delegation, then use narrow leaf discovery first.
3. When JSON input is accepted, run the leaf --schema and treat it as authoritative.
4. When a positional-only leaf does not expose --schema, run its leaf --help; never guess arguments.

The CLI accepts at most one inline JSON object. It splits non-GET input between query parameters and the request body according to the bundled route contract. Do not use stdin or files unless the user explicitly asks.

Standard output envelopes are:

Factory collection payloads may also place arrays under named fields inside data, such as workItems, decisions, items, and findings. Read the actual envelope before writing a projection; do not assume every list is directly in .data[].

Output rules:

• Never explore with unfiltered --pretty.
• Pipe output through jq immediately and retain only fields needed for the task.
• Use perPage: 1 for a latest item and perPage: 10 or less for recent items when the schema offers page-based pagination.
• Use small limit values for cursor-based lists.
• If output is noisy or truncated, narrow the jq projection instead of increasing raw output.
• Fetch full JSON only when a compact projection is insufficient or the user explicitly requests it.

Target resolution

Prefer an explicit verified instance URL for hosted, local, remote, or self-hosted Factory. No repository or link file is required. Follow connection.md for authentication and target discovery; FACTORY_URL below must be the user’s actual deployment, not a platform dashboard URL.

For brevity, the remaining examples omit --url. When using an explicit target, insert --url "$FACTORY_URL" after mastra api in every example, including schema/help discovery. Otherwise commands can probe localhost or use a different directory’s deployment config. Preserve any required authentication options as well.

If a project list is empty, verify the intended target, organization, access, and pagination; do not inspect stored credentials or assume there are no projects.

Command catalog

Use leaf --help and --schema to verify the installed command before any write.

Important contracts:

• project create requires name.
• project update can change project metadata and automation settings. Enable autoRunEnabled or autoApprovePlans only when the current request or standing delegation includes automation-setting changes.
• work-item create requires title and creates new work in Intake.
• work-item update changes non-stage fields only. Never use it to move a card.
• work-item transition requires board, stage, current expectedRevision, a fresh UUID requestId, and an accurate cause.
• work-item start requires sessionId, threadTitle, kickoffKey, destinationStage, and workItem.
• metrics accepts optional from and to timestamps.
• decision list accepts before, limit, and statuses.
• attention list accepts before, limit, search, tier, and view.
• attention read-all accepts optional before.

Read-only supervisor workflow

When asked for Factory status, a queue review, a supervisor summary, blocked work, or recommended next action, use this sequence without mutating anything.

1. Select the project

• If exactly one project exists, select it.
• If the user named a project, match it exactly and report the selected ID/name.
• If multiple projects remain plausible, report compact choices and ask the user to select one.
• If none are returned, report the target/organization ambiguity instead of inventing a project.

Fetch the selected project:

2. Inspect work and running sessions

Correlate current stages, revisions, bound sessions, runningSessionIds, timestamps, and stage history. Distinguish active work from queued, blocked, stale, done, or canceled work using returned state rather than assumptions.

3. Inspect metrics and thresholds

Use the returned thresholds for queue-age buckets. They are not the supervisor’s decision/start/lease timeout thresholds; use the supervisor’s returned evidence for those findings. Do not invent alert cutoffs.

4. Inspect decisions and human attention

Pending decisions are proposals, not authorization. Attention items are findings, not commands. Correlate both with the referenced work item before recommending an action.

5. Inspect supervisor health and session

Health finding kinds are decision-stuck, start-stalled, seat-orphaned, seat-missing, held-waiting, and label-drift; each finding carries a stable id, evidence, beganAt, and a suggestedRepair. Report only categories returned by the server; see session-inspection.md for interpretation.

The supervisor session is for supervisor inspection/coordination. Do not assume it is a valid durable user session for work-item start.

6. Return the summary

Report:

1. Selected project and target type.
2. Active and queued work, including current stage and revision.
3. Blocked, stale, or unhealthy items and the evidence for each classification.
4. Running sessions and their linked work items.
5. Pending/failed decisions and retryability.
6. Open/unread human-attention items.
7. Supervisor health findings.
8. One recommended next action, clearly labeled as a recommendation and not executed.

Mutation protocol

Apply this protocol when a mutation is authorized by the current request or by a standing delegation.

1. Identify the applicable authority and its boundaries: project, resource, allowed action classes, objective, and any stop conditions. Do not invent broader authority from a general status request.
2. Run the leaf --schema for JSON-input commands. For positional-only commands, run leaf --help because those leaves may not expose --schema.
3. Fetch the current project/work item/decision/attention item and verify every ID belongs to the selected project.
4. Confirm internally that the exact target and action remain within scope. Ask only if scope is unclear or the action exceeds it.
5. Send the smallest valid mutation.
6. Refetch the affected collection/resource.
7. Report target IDs, old/new revision when applicable, final state, and any returned audit or failure details. For autonomous runs, summarize actions taken and stop when a delegated stop condition is reached.

Create a project

Project deletion is destructive. Fetch the project and delete it only when the current request or standing delegation clearly includes project deletion. Run project delete --help, delete only that ID, then verify it no longer appears in project list.

Create or update a work item

New work enters Intake:

Use work-item update only for non-stage fields. Fetch the item from work-item list before and after the update.

Work-item deletion is destructive. Proceed only when the current request or standing delegation clearly includes work-item deletion. Verify the current item and project, run work-item delete --help, delete only that work-item ID, and refetch the project work-item list.

Transition a work item

Never change lifecycle stage through work-item update. Use the transition endpoint with optimistic concurrency.

Rules:

• Fetch expectedRevision immediately before the transition.
• Generate a fresh UUID for every new transition request.
• Use only schema-enumerated boards and stages.
• Set cause to an accurate explanation of why this transition is occurring; do not copy an unrelated example value.
• Use reenter only if the installed schema exposes it and re-entry is included in the current request or standing delegation.
• On a conflict, refetch and reassess. Never increment a revision locally or blindly retry.

Decision actions

approve, dismiss, and retry are explicit audited actions. Before acting:

1. List/fetch current decisions and the linked work item.
2. Verify status, type, role, retryability, attempts, and failure details.
3. Confirm that the decision action is covered by the current request or standing delegation; do not treat the decision record itself as authority.
4. Run the leaf help, execute once, then refetch decisions and work items.

Approving a proposed move may also consent to its queued run. Never approve merely because a decision is pending. Retry only failed decisions that the server marks retryable.

Attention actions

Reading, bulk reading, archiving, and restoring all change inbox state. Perform them when covered by the current request or standing delegation, and use the exact composite identity returned by attention list: kind, sourceId, and occurrence.

read-all is a bulk mutation. Do not infer permission from a request to inspect attention. Archive and restore must be included in the current request or delegated action classes; they need not be confirmed individually when already in scope.

Start a work item

work-item start is an autonomous, high-impact action. Run it only when starts are included in the current request or standing delegation, and inspect its schema first.

The command requires an existing durable Factory user session owned by the authenticated user and organization and connected to the selected Factory project. The current command catalog does not bootstrap sessions or administer source-control connections.

If no supported session already exists:

• Report that start is blocked.
• Do not invent a session UUID.
• Do not substitute the supervisor session.
• Do not call private HTTP routes with curl.
• Do not create a different resource as a workaround.

Errors

Distinguish shell failures from CLI JSON errors. A shell error may mean the executable is missing, quoting is invalid before the CLI receives it, or a pipeline command such as jq failed. Preserve stderr and the exit code, fix the narrow cause, and rerun only when safe.

CLI error handling:

• INVALID_JSON: fix shell quoting; input must be one inline JSON object.
• MISSING_INPUT: run the leaf --schema and provide its required JSON object.
• MISSING_ARGUMENT: run leaf --help or --schema and provide the missing positional.
• HTTP_ERROR: inspect only safe fields from error.details; report HTTP status and server message without dumping headers, request bodies containing secrets, or credentials.
• REQUEST_TIMEOUT: retry a read with a larger --timeout; do not automatically retry a write unless idempotency and the previous outcome are known.
• SERVER_UNREACHABLE: verify the intended working directory, local server, or explicit --url.
• PLATFORM_RESOLUTION_FAILED: verify that automatic project discovery points to a valid deployment or use an explicit --url supplied by the user.
• Empty project list: verify intended target and organization; organization mismatch can look like no data.
• Revision conflict: refetch the work item and reassess the transition.
• Governed stage-change rejection: use work-item transition, not work-item update.
• Factory session not found or wrong project: report the durable-session limitation; do not invent or substitute an ID.

Final safety check

Before responding, confirm:

• No mutation occurred during a read-only request.
• Output was compact and collection envelopes were parsed from their named fields.
• Every JSON-input write used leaf --schema; positional-only writes used leaf --help without guessed arguments.
• Every transition used a freshly fetched revision and fresh request UUID.
• Every mutation was covered by the current request or a clear standing delegation; high-impact action classes were explicitly included in that scope.
• The result contains no secrets, credential contents, real unrelated identifiers, machine-specific paths, or test artifacts.

• references/factory-supervisor.md(opens in a new tab)
• Raw file(opens in a new tab)
• mastra-ai/skills(opens in a new tab)