refactor(global): one vocabulary on the wire, in the ledger and in the code

[rami-hatoum]

The runtime now says what the product says, everywhere a person or a program can read it. A brain has definitions of a type, reasoning, interaction, computation, recall or workflow, and a definition carried out once is a run with a run_id. The HTTP routes, the MCP tools, their fields and values, the events a brain records, its stream names and tables, the settings, the package names and the code all use those words, and nothing serves the old names beside them: no alias, no mapping, no fallback. The product is not live, so this is the moment to change them once.

What a client of the HTTP API changes

Brain routes are relative to /v1/orgs/{org}/brains/{brain}.

Was	Is
POST, GET /specs/{primitive}	POST, GET /definitions/{type}
GET, PUT /specs/{primitive}/{name}	GET, PUT /definitions/{type}/{name}
POST /specs/{primitive}/{name}/retire	POST /definitions/{type}/{name}/retire
POST /specs/{primitive}/{name}/execute	POST /definitions/{type}/{name}/run
GET /executions	GET /runs
GET /executions/{execution_id}	GET /runs/{run_id}
GET /executions/{execution_id}/history	GET /runs/{run_id}/history
POST /executions/{execution_id}/cancel	POST /runs/{run_id}/cancel
POST /executions/{execution_id}/events	POST /runs/{run_id}/events
POST /executions/{execution_id}/answer	POST /runs/{run_id}/answer
path values inference, recollection, orchestration	reasoning, recall, workflow; interaction and computation unchanged
body execution_id of a run	run_id
query primitive of GET /executions and GET /analytics	type, with the new values, as GET /analytics?days=7&type=reasoning
query execution_id of GET /events	run_id
answer specs of a listing of definitions	definitions
answer executions of a listing of runs	runs
primitive of a definition, a listed definition, a run, a listed run and by_function[]	type
spec_version of a run and a listed run	definition_version
execution_id of a run, a listed run, interactions[] and the answer of sending an event	run_id
event type values execution_started, execution_deferred, execution_succeeded, execution_rejected, execution_failed, execution_cancel_requested, spec_created, spec_updated, spec_retired, in list_brain_events' answers and its type filter	run_started, run_deferred, run_succeeded, run_rejected, run_failed, run_cancel_requested, definition_created, definition_updated, definition_retired
an event's data.execution_id, on every event of a run, workflow_input_applied, telling_started, and a call's step_waiting	data.run_id
an event's data.primitive and data.spec_version	data.definition_type and data.definition_version
data.called_by.execution_id, data.emitted_by.execution_id	data.called_by.run_id, data.emitted_by.run_id
sources a published event may not use: /executions/, /specs/, /callers/	/runs/, /definitions/, /callers/
a fact's source /executions/<id>, /specs/<primitive>/<name>, and subject <primitive>/<name>, in a trigger or a recall filter	/runs/<id>, /definitions/<type>/<name>, <type>/<name>, as reasoning/triage
the metadata key a tool server receives with a run's call, com.beonauto/execution_id	com.beonauto/run_id
the settings ORCHESTRATION_MAX_DURATION, ORCHESTRATION_SWEEP_INTERVAL, ORCHESTRATION_NESTED_EXECUTIONS, ORCHESTRATION_MAX_OPEN_CALLS, RECOLLECTION_MAX_FUNCTIONS, RECOLLECTION_MAX_REBUILDS, RECOLLECTION_BRAINS_AT_ONCE	WORKFLOW_MAX_DURATION, WORKFLOW_SWEEP_INTERVAL, WORKFLOW_NESTED_RUNS, WORKFLOW_MAX_OPEN_CALLS, RECALL_MAX_FUNCTIONS, RECALL_MAX_REBUILDS, RECALL_BRAINS_AT_ONCE
the log annotation execution_id	run_id

A type no capability of the server serves is refused with 422 invalid_input at /type, naming the types the server runs, on every route and tool that takes one, where it answered 404 on a definition's routes and an empty page on GET /runs and GET /analytics. A definition's type is type on a resource, and definition_type on a record whose own type already names the record: a ledger event, the data of a brain event, the data of a fact. Unchanged for a client: every other route, the problem types, reasons, kinds and because values, the JSON Schema identifiers, the scopes, /health and the three MCP endpoints. The detail of a problem and the summary of an event are words for people and change only where they named a tool or an old word.

A workflow document calls a definition with call: run_definition and with: { type, name, input }, and its $runtime.metadata is { type: 'workflow' }.

What an agent sees

Eleven tools are renamed and fourteen keep their names; /mcp still serves 25, a key's org endpoint 8 and a brain's endpoint 19.

Was	Is
create_spec, list_specs, get_spec, update_spec, retire_spec	create_definition, list_definitions, get_definition, update_definition, retire_definition
execute_spec	run_definition
get_execution, list_executions, get_execution_history, cancel_execution	get_run, list_runs, get_run_history, cancel_run
send_execution_event	send_run_event

The tools take and answer the fields of the HTTP table. The served instructions no longer carry the sentence that told an agent how to translate the tools' names into the product's words, since there is nothing left to translate; they are 100 characters shorter on /mcp and on a brain's endpoint. The recipes and every tool description name the new tools, each within its bounds.

What the ledger records

A run records on its own stream, runs/<id>, and the workflow engine's run log on run-logs/<id>, a different kind, so the two streams of a run stay apart; a server test checks that no stream of the runs/ kind begins with a run-log event. A brain's definitions of one type are the stream definitions/<type>. The nine event types and the fields above are recorded under their new names. The projections that fold runs take their next versions, run_outcomes_3 (with the column definition_type), open_requests_5 and conversations_2, filled from history by the next server, which drops the versions before. The workflow host keys a run by run_key, the host's org/brain/<id>, beside brain_key, and its reaction backlog keeps the bare run_id. Because every stream name changes, every message id changes with it.

A local database made before this change

Nothing reads the old names, so there is no migration. Delete packages/server/.data/ledger.db for pnpm dev, or the file LEDGER_FILE names, or the PostgreSQL database DATABASE_URL names, before running this build on it. Kept, such a database shows no definition and reads its run logs as runs. The workflow host makes its tables with CREATE TABLE IF NOT EXISTS, so a kept database keeps their old run_id columns: the server starts, and then every read and write of the host's run tables fails, so no workflow run can be recorded, and the due, timer and deferred-start sweeps die on every pass.

The engine's format 7

The workflow engine names its run runId in its state, its inputs, every output an event carries, its call keys and its snapshots, which is state format 7. Format 6 is frozen: its state is read strictly and upcast by renaming executionId to runId, and the event and snapshot schemas of formats 1 to 6 are frozen with it, so each event and snapshot is read with the schemas of the format it names. The corpus gains a format-7 log and keeps formats 1 to 6 loading, and the recorded input logs are recorded again under format 7.

The packages

primitives/ is capabilities/: @beonauto/inference is @beonauto/reasoning, @beonauto/orchestration is @beonauto/coordination, @beonauto/recollection is @beonauto/recall, and @beonauto/specs is @beonauto/definitions, with the same subpaths. The contract a capability package implements is Capability, declared with defineCapability; it states its type and runs a definition with run. A definition's type is its one discriminator: FunctionType (reasoning, interaction, prediction, recall, computation) replaces the verb kinds, and the published taxonomy asset carries type at schema 2. The Dream note is removed; the concepts page already says what it said.

Verification

• pnpm check, without PostgreSQL and against PostgreSQL 18.
• The image's smoke steps of the CI workflow, run against an image built from this change.
• A search test kept in pnpm docs:test finds none of the old words outside docs/decisions, patches/, the lockfile, the frozen engine formats and readers of formats 1 to 6, and the check of internal terms and its test, which must name what they refuse. It allows, each only in the files that use it, the addresses of the Open Workflow Specification's error types and of the CloudEvents specification, the model providers' terms inferenceConfig, inferenceGeo, inference-profile and "application inference profile", toolSpec, code_execution, the model SDK's execution-denied, and the sentence of CLAUDE.md that says what inference names; it fails when an allowed text stops occurring.