Factory schema

Explicit agent-loop tool policy for AGENT_WORKER definitions. Omit or set policy DISABLED to run agent loops without advertising or executing tools.

Additional command arguments passed to the configured command.

Hosted-worker authentication contract. V1 hosted workers accept only auth.secretRef.

Inline worker instructions or script body when the worker is authored directly in factory config.

Command to execute when this worker runs through a command or script provider.

Optional localized customer-facing explanation of this worker.

Execution mechanism. Use `ACP` for ACP-backed workers and put the configured integration identity (for example `cursor-acp`) in modelProvider. `SCRIPT_WRAP` remains the command-wrapper compatibility value; legacy named executor identities remain accepted during migration.

Optional durable public identifier for this worker. When present, graph and layout references should use this id instead of the mutable name.

Provider-specific configuration for the built-in hosted LINEAR worker.

Model identifier to request from the configured model provider when this worker uses model execution.

Provider locality for this model capability declaration. Use `LOCAL` for embedded or host-managed inference and `CLOUD` for remote provider execution.

Canonical provider identity used for model routing and provider diagnostics, or an exact invocation-parameter placeholder such as `${modelProvider}`. For `executorProvider: ACP`, this names the configured ACP integration, such as `cursor-acp`. Extension identities use lowercase standardized syntax; built-in values such as `CLAUDE` and `CODEX` remain compatibility conveniences.

Worker name referenced by Workstation.worker.

Provider-agnostic model operations that this worker can execute, including named input and output slots.

Built-in hosted provider identity when this worker uses repository-owned hosted execution.

Resource capacity this worker requires before it can be dispatched.

When true, bypasses permission checks for providers that support permission gating.

Marker that tells model-oriented workers where to stop generated output when the provider supports it.

Optional Go duration that caps one worker execution attempt.

Worker implementation family to instantiate for this definition.

Required executor policy for agent-loop tool use on this worker.

Linear issue field name to use when deriving optional assignee claim metadata.

Optional claim-related configuration that v1 hosted Linear polling allows.

Deterministic mapping fields for canonical work submission generation.

Optional Go duration that controls how often the hosted Linear worker polls for updates.

Optional Linear issue-state identifiers that bound the poll source.

Optional Linear team identifiers that bound the poll source.

Canonical submitted work state emitted for matched Linear issues.

Canonical submitted work type emitted for matched Linear issues.

Referenced secret name that resolves the hosted provider API key at runtime.

Named operation input slots this worker can consume.

Named operation output slots this worker can produce.

Uppercase content types accepted or produced by this slot.

• minItems1

Stable slot name used by workstation-side bindings and diagnostics.

Whether this input slot must be resolved before invocation starts. Output slots omit this field when not needed.

Scheduling behavior for this workstation, such as STANDARD, REPEATER, or CRON execution.

Inline workstation instructions or script body when authored directly in factory config.

Explicit label-to-destination routing used only by CLASSIFIER_WORKSTATION definitions. Each route must declare a unique non-empty label and one or more outputs.

Copy supported referenced script files into the expanded workstation layout when config expand runs.

Cron trigger configuration for workstations whose behavior is CRON.

Optional localized customer-facing explanation of this workstation.

Environment variables added to the workstation execution context.

Guarded loop breakers should use `VISIT_COUNT` guards here with a `LOGICAL_MOVE` workstation instead of top-level exhaustion rules.

Optional durable public identifier for this workstation. Graph and layout references should use this id instead of the mutable name.

Work states this workstation can consume before it dispatches.

Retry and execution ceilings applied to this workstation.

Customer-authored workstation name used by guards, diagnostics, and authored references.

Optional destination emitted when the workstation makes partial progress and should continue iterating. Classifier workstations must not declare onContinue.

Optional destination emitted when the workstation fails permanently.

Optional destination emitted when the worker rejects the current work without a hard failure. Classifier workstations must not declare onRejection.

Uppercase provider-agnostic operation requested by `MODEL_INVOKE` workstations, such as `TTS`.

Optional workstation-authored slot bindings that resolve operation inputs from runtime content or static config content.

Optional worker-output parsing mode for model workstations. When set to `decision-envelope`, agent output is parsed as a reviewer/checker JSON envelope that maps directly onto WorkResult outcome, feedback, output, and optional recorded output work instead of stop-token routing.

JSON schema string used to validate or parse structured model output when configured.

Work states emitted after a non-classifier workstation succeeds. Classifier workstations must use classificationRoutes instead of normal success outputs.

Path to a prompt template file loaded for model-oriented workstation execution.

Resource capacity this workstation consumes while one dispatch is in flight.

Optional workstation-specific runner override. When omitted, dispatch falls back to the factory runner, then worker modelProvider compatibility when no explicit runner is configured, then the default codex runner.

Stop words authored on the topology entry for model-oriented dispatches.

Runtime workstation implementation type, equivalent to the workstation AGENTS.md frontmatter type.

Optional policy for whether downstream work uses the workstation output payload or preserves the consumed input payload.

Name of a worker declared in the workers list.

Go template resolved from token tags at dispatch time.

Go template resolved and passed as the worktree path to CLI dispatchers.

Case-sensitive classifier label that must match the trimmed classifier output exactly.

One or more authored destinations emitted when this classifier label is selected.

Propagation mode for downstream work payload selection after this workstation succeeds.

Positive Go duration interval, such as 30s, 5m, 1h, or 1h30m, used instead of schedule.

Positive Go duration after due_at before a stale cron time token expires and can be consumed by the system expiry transition. Defaults to the duration until the next scheduled cron fire when omitted.

Non-negative Go duration used as the maximum deterministic delay added to scheduled time tokens. Defaults to "0s".

Standard five-field cron expression used to produce internal time work while the factory service is running.

When true, service startup submits one immediate internal time work item before waiting for the next scheduled cron fire.

For `MATCHES_FIELDS` guards, the field-selector configuration used to compare candidate inputs.

For `SAME_NAME` and `SAME_TRACE_ID` input guards, the peer input workType name from another input in the same workstation.

For `VISIT_COUNT` guards, the fixed visit ceiling.

• minimum1

Optional invocation argument whose positive integer value tightens the fixed visit ceiling.

For parent-aware input guards, the parent workType name from another input in the same workstation.

For dynamic fanout input guards, the workstation that spawns the children for count tracking.

Guard condition to evaluate for this workstation-level attachment.

For `VISIT_COUNT` guards, the workstation whose visits are counted.

Per-input guards that must pass before this specific input can be used.

Name of the work state consumed or emitted for the referenced work type.

Name of the work type consumed or emitted at this edge of the workstation.

Go duration limit for one dispatch attempt before it times out.

Fixed maximum number of Work items one accepted worker-emitted FACTORY_REQUEST_BATCH may contain.

• minimum1

Optional invocation argument whose positive integer value tightens the fixed generated-Work ceiling.

Offset added to the invocation argument before applying the generated-Work ceiling.

• minimum0

Maximum number of retry attempts after a failed dispatch before the workstation gives up.

Static authored content bound directly or used as the first fallback when runtime input does not match.

Optional final fallback content when neither runtime input nor config content resolves the slot.

Ordered runtime-input selector used before falling back to config or default content.

Stable input slot name declared by the worker operation.

Match a content part by its label field.

Match a content part by its role field.

Match a content part by its authored slot field.

Match a content part by its uppercase public type.

Optional localized customer-facing explanation of this work type.

Optional CLI routing markers for this work type. Factories used with you run --factory must declare handlingBehavior DEFAULT on exactly one work type.

Optional durable public identifier for this work type. When present, graph and layout references should use this id instead of the mutable name.

Customer-authored work type name referenced by workstation inputs, outputs, and submitted work.

Lifecycle states available for work items of this type.

• enum"AUDIO"

• enum"BINARY"

Optional artifact identifier for externally materialized content.

Optional MIME content type for file-backed or structured parts.

Optional caller-defined label for slot binding or diagnostics.

Optional semantic role for model-operation authoring.

Optional slot name used by model-operation binding selectors and diagnostics.

• enum"image" | "IMAGE"

Arbitrary JSON value preserved in canonical part order.

• enum"JSON"

Optional durable public identifier for this state within its work type. When present, graph and layout references should use this id instead of the mutable name.

Customer-authored state name referenced by workstation inputs and outputs.

Lifecycle category for this state, such as initial, processing, terminal, or failed.

Inline text content preserved in canonical part order.

• enum"text" | "TEXT"

JavaScript-specific orchestrator configuration. Required when kind = JAVASCRIPT.

Petri-specific orchestrator configuration. Required only when kind = PETRI and additional Petri options are authored.

Operator worker preset inherited by child calls using this agent id.

• minLength1

Named child-agent roles and their operator worker preset defaults.

• additionalProperties/$defs/FactoryOrchestratorJavaScriptAgent

JSON Schema object describing workflow invocation arguments.

• additionalPropertiestrue (open)

Default JavaScript workflow policy object applied when no runtime override exists.

• additionalPropertiestrue (open)

Optional JavaScript dialect label for the authored workflow source.

Optional exported entrypoint or phase name used to start the workflow.

Inline workflow source when the factory carries source text directly.

Free-form JavaScript orchestrator metadata for authoring and diagnostics.

Optional content hash for the resolved workflow source.

Factory-relative or authored reference to the workflow source file.

Declared content encoding for the inline workflow source.

• enum"utf-8"

Inline JavaScript workflow source text.

Optional customer-facing hint for the factory's primary output shape.

Declared invocation parameters keyed by canonical parameter name.

Policy for named inputs that do not match any declared parameter binding.

Structured invocation arguments; values are never parsed or executed while loading the Factory.

Localized customer-facing explanation of what the example does.

Stable example name.

• minLength1

Output media type hint for docs, API consumers, and dashboard affordances.

Human-readable summary of the primary output contract.

Suggested file extension when the output mode writes a file.

High-level output contract mode exposed to callers.

Parameter name that controls the destination path when the factory writes output to disk.

Additional accepted named-argument keys that normalize to this parameter.

Accepted invocation bindings for this parameter across positional, named, and stdin sources.

Optional allowed string values for this parameter.

Default string value used when an omitted parameter resolves to one effective value.

Default string values used when an omitted parameter resolves to multiple effective values.

Customer-facing description rendered in help, docs, and form controls.

Preferred named-argument key shown to callers, such as `output`.

Internal canonical parameter name used for normalized argument maps and interpolation.

When true, invocation normalization must reject requests that omit this parameter.

When true, diagnostics must preserve names and source metadata but redact concrete values.

String-first hint that guides parsing, docs, and dashboard form selection.

Declares whether the parameter consumes one value, repeated values, variadic values, or file contents.

Binding kind used to route invocation input into the parameter.

1-based positional slot used when kind is POSITIONAL.

• minimum1

Return selection policy for this factory.

Authored terminal state name used by EXPLICIT policy selection.

Optional authored work name filter used by EXPLICIT policy selection.

Work type name used by EXPLICIT policy selection.

Optional model name to scope throttling more narrowly than the provider-level window.

Provider whose inference-throttle history controls this factory-level guard.

Duration string that controls how long the factory should keep re-checking throttle history before allowing the lane again.

Factory-level guard condition to evaluate before dispatch-ready transitions can proceed.

Field selector resolved against each candidate input, such as `.Name` or `.Tags["_last_output"]`.

• minLength1

For `MATCHES_FIELDS` guards, the field-selector configuration used to compare candidate inputs.

For `SAME_NAME` and `SAME_TRACE_ID` input guards, the peer input workType name from another input in the same workstation.

For `VISIT_COUNT` guards, the visit threshold.

• minimum1

For parent-aware input guards, the parent workType name from another input in the same workstation.

For dynamic fanout input guards, the workstation that spawns the children for count tracking.

Guard condition to evaluate for this input-level attachment.

For `VISIT_COUNT` guards, the workstation whose visits are counted.

Input type name. The reserved name "default" is implicit.

Managed runtime backend identifier for `MODEL` resources, such as `LLAMACPP`. Backend selection stays provider-agnostic in customer-facing factory config.

Total units of this resource available to the factory at one time.

• minimum1

Optional durable public identifier for this resource. When present, graph and layout references should use this id instead of the mutable name.

Managed runtime load policy for `MODEL` resources, such as `ON_DEMAND` or `EAGER`.

Stable managed runtime identity for `MODEL` resources, such as `OMNIVOICE_Q4_K_M`. Packaged and authored factories declare the same managed-runtime dependency through this field plus matching `MODEL_WORKER.model` values.

Resource name referenced from worker requirements and workstation resourceUsage entries.

Provider identity associated with this resource, especially for `PROVIDER_QUOTA` resources.

Optional uppercase resource family, such as `MODEL`, `PROVIDER_QUOTA`, or `INVOCATION_SLOT`.

Durable bundled-file identifier used by portable layout and graph editor references. When omitted on input, the canonical targetPath is materialized as the stable identifier.

Canonical factory-relative restoration target for the bundled file. Absolute paths, backslash-separated paths, and paths that require dot-segment normalization are rejected.

Portable file class. SCRIPT entries target factory/scripts/..., DOC entries target factory/docs/..., INPUT entries target factory/inputs/<work-type>/<channel>/..., and ROOT_HELPER entries target supported project-root helper files such as Makefile only when explicitly declared in bundledFiles. Shared-factory INPUT entries snapshot current source inputs at share time instead of creating a live link.

• enum"SCRIPT" | "DOC" | "INPUT" | "ROOT_HELPER"

Declared content encoding for the inline payload. V1 bundled files use UTF-8 text content.

• enum"utf-8"

Inline bundled file content carried in the manifest. SCRIPT and DOC files under factory/scripts/ and factory/docs/ may be discovered during flatten, but supported root helper paths such as Makefile are bundled only when they appear as explicit ROOT_HELPER entries in bundledFiles.

Executable lookup token that must resolve on PATH.

Human-readable tool name used in manifests and validation output.

Optional explanation of why the portable factory requires this tool.

Optional argument vector used by future validation flows to probe the tool version without changing the executable lookup token.

Portable bundled files that belong inside the factory boundary. Entries are explicit only, use factory-relative target paths, and must stay under the canonical script, docs, or inputs roots for SCRIPT, DOC, or INPUT entries, or match the supported root-helper allowlist for ROOT_HELPER entries. Export, share, flatten, and materialize flows auto-discover SCRIPT and DOC files under the documented factory subtrees, but ROOT_HELPER entries such as Makefile are opt-in manifest entries that travel only when explicitly declared here. In v1 shared-factory flows, INPUT entries capture the source factory's current starter work at share time and are restored as independent recipient copies.

Declarative external tools that must already resolve on PATH. These entries are validated but not embedded or installed.

• minimum1

Optional inert positioned notes and embedded-raster images that decorate the canvas without becoming graph topology.

Optional authored graph edge geometry keyed by canonical graph edge id.

Optional flat background groups keyed independently from topology.

Optional authored graph node geometry keyed by canonical graph node id.

Portable layout contract schema version. Version 1 is the initial public layout contract.

• minimum1

Stable annotation identifier unique within this layout.

• pattern\S
• minLength1

Horizontal canvas coordinate between -100,000 and 100,000 inclusive.

• minimum-100000
• maximum100000

Vertical canvas coordinate between -100,000 and 100,000 inclusive.

• minimum-100000
• maximum100000

Positive authored height no greater than 10,000 canvas units.

• maximum10000
• exclusiveMinimum0

Positive authored width no greater than 10,000 canvas units.

• maximum10000
• exclusiveMinimum0

Authored group height.

Authored group width.

Left graph layout coordinate.

Top graph layout coordinate.

Canonical graph edge id such as workstation-output:workstation:review->work-state:task:done.

Optional authored intermediate edge points in graph canvas space.

Literal empty-state text. It is not rendered as HTML or Markdown.

• pattern\S
• minLength1
• maxLength500

Optional authored group accent or fill color.

Stable authored group id for future layout editing.

Optional visible group label.

Optional authored group lock flag for future editor affordances.

Canonical graph node ids visually contained by this group.

Reserved for future nested groups. Omit or set null for flat groups.

Literal alternative text for the embedded image.

• pattern\S
• minLength1
• maxLength500

Strict padded base64 payload for the embedded raster source, limited to 2 MiB after decoding.

• minLength4
• maxLength2796204

Source variant discriminator. EMBEDDED carries portable base64 raster data.

• enum"EMBEDDED"

Declared media type for the embedded raster.

• enum"image/png" | "image/jpeg" | "image/webp"

Canonical graph node id such as workstation:<workstationId>.

• pattern\S
• minLength1

Optional authored node lock flag for future editor affordances.

Required literal plain-text note body.

• pattern\S
• minLength1
• maxLength4000

Optional literal plain-text note title.

• maxLength160

Horizontal graph layout coordinate in authored canvas space.

Vertical graph layout coordinate in authored canvas space.

Preferred authored graph direction for portable layout rendering.

• enum"UP" | "DOWN" | "LEFT" | "RIGHT"

Authored node height.

Authored node width.

Authored viewport horizontal offset.

Authored viewport vertical offset.

Authored viewport zoom factor.

Monotonic Lamport-style logical component derived from the persisted factory definition version. Serialized as a decimal string so JavaScript clients can round-trip the 64-bit value without precision loss.

• pattern^[0-9]+$

UTC physical timestamp component for the persisted factory definition version.

Optional stable metadata identifier; consumers must not render it as display copy.

Canonical BCP 47 locales for which the base value was authored.

• uniqueItemstrue

Discriminator for localized customer-facing metadata.

• enum"LOCALIZABLE_ASSET"

Required base value returned when no exact locale override exists.

• minLength1

Exact canonical BCP 47 locale tags mapped to localized overrides.

• additionalPropertiestrue (open)