Skip to main content

​
Skills in GreatBook workflows

A skill may be a bounded read, one guarded decision in a longer flow, or a catalogue declaration for work that cannot start yet. The registry keeps those distinctions explicit: current action skills return review artifacts with canonical_write_implied: false, while blocked and queued profiles declare no usable execution path. This page is pinned to skills registry 1.0.0 and capability registry 1.14.1.

​
The five graph contracts

The skills registry declares five versioned LangGraph owners. Their checkpoint contracts describe the stages they own; they do not make every similarly named skill executable through that graph. Only /new-document declares direct graph execution, through greatbook.clerk@1, and it is blocked because its typed intake-review boundary is not released. The current profiles execute MCP tools directly; guarded action skills start a staged workspace rather than bypassing an agent checkpoint or canonical service.

​
Intake and posting are one chain with two human boundaries

The intake family begins with uploaded file references, classification, extraction, validation, and a persisted review state. /new-document and the intake queue, item, masters, audit, approve, reject, and resubmit profiles are all blocked, so none should be presented as a working MCP intake command. The Bookkeeper’s posting flow begins from approved canonical intake, proposes balanced lines, waits for exact confirmation, and posts through its owning writer. The /posting-review, /approve-posting, and /reject-posting profiles are blocked. After a valid post exists, the current /transactions and /transaction reads can inspect canonical accounting events and lineage; they are observations of the result, not posting commands.

​
Document lifecycle

The Typewriter graph constrains and calculates document proposals and ends ready for review. The registry separates that proposal work from later state changes:
  1. /draft-document and /preview-document are blocked, so MCP cannot currently start composition or preview as skills.
  2. /documents is current and reads canonical Typewriter document records.
  3. /approve-document is current but only stages a hash- and version-bound approval for review through greatbook_workspace_create.
  4. /submit-document is current but only stages a hash- and version-bound submission for review through the same workspace boundary.
  5. /document-history and /export-document remain blocked.
Approval and submission therefore belong to a lifecycle; neither is a generic document writer and neither implies a canonical write at skill completion.

​
Close and adjustments

The Accountant graph reads and analyzes Close state, produces workpapers, and stages proposals ready for review. Current /close-checklist, /close-readiness, /close-runs, and /fiscal-periods skills are standalone reads of canonical Close records. Current /allocations, /costing, /adjustments, and /workpaper skills are standalone reads of the Accounting Desk’s allocations, costing records, and adjustment workpapers. The action profiles are later steps, not replacements for analysis:
  • /approve-adjustment and /reject-adjustment stage one exact workpaper decision through the guarded workspace.
  • /close-period and /reopen-period stage an exact version-bound fiscal transition through the guarded workspace.
  • /propose-adjustment, /close-adjustments, /create-fiscal-year, /run-close, /statutory-report, and /statements are blocked or queued and cannot start.
Each current action completes at ready_for_review with canonical_write_implied: false. The owning capability’s canonical writer remains responsible for the eventual commit after authorization and exact confirmation.

​
Bank reconciliation

Bank reconciliation is a multi-step evidence and decision flow rather than one all-purpose call.
  1. /bank-statement-schema describes the accepted evidence shape.
  2. /bank-evidence reads bounded evidence, settlements, chains, and candidates.
  3. /bank-facets reads procure-to-pay, order-to-cash, or float document-chain facets.
  4. /decide-bank-match stages an accept, reject, or undo decision bound to the proposal UUID, candidate SHA-256, exact decimal amount, and currency.
Those four profiles are current, subject to their distinct read or write roles. The broader /reconcile-bank and /bank-workbench profiles are queued, while statement import, handoff, proposal, manual match, unmatch, purpose classification, state, assurance, and matching-graph profiles remain blocked. The current decision skill is one guarded W3 step; it is not evidence that the whole W1–W7 workflow is available as a skill.

​
Audit and assurance

The Auditor graph scopes requests, reads or recomputes under a read-only boundary, and reports. Current /audit-runs and /findings inspect the resulting canonical audit-run and finding records and can return inconclusive_read. The job catalogue, manual run, retry, and finding acknowledgement profiles are blocked. That separation keeps reporting available without turning the Auditor into a mutation channel.

​
Standalone calls versus workflow steps

Read skills with completion state complete return non-committable pages or records and declare no canonical writer. They can still fail with permission_denied, not_found, bound_exceeded, or inconclusive_read as their profile specifies. Current action skills use greatbook_workspace_create, return a committable typed-action artifact, and stop at ready_for_review with no implied canonical write. The workspace adds revision, selection, validation, confirmation, and receipt stages before a closed adapter may call the canonical writer. Blocked and queued skills have no execution target and cannot join either path.