TypeAgent Program Macros
Use the typeagent-macros MCP server for macro catalog and lifecycle work.
Use the configured owner; never switch owners or learning preferences for the
user. Routing mode, owner and learning preference are separate settings.
Run A Macro
- If the macro reference is not already known, use
search_macrosorlist_macrosto find it. - If required inputs are not known, use
get_macro_requirements. Collect required inputs without exposing secret values in chat. - Call
run_macrowith the macro reference and current inputs. The server resolves and validates the current approved immutable program revision; do not callinspect_macroas an execution prerequisite. - For
completed, deliver the actual output without inventing or rewriting it. For any non-completed outcome, report its actual structured status without claiming success. Inspectget_macro_runfor durable dispatch/uncertainty evidence.
Never launch a planning agent to replay steps, reconstruct a program, or retry an accepted failed, cancelled, denied, or uncertain run through another provider. Symbolic programs make no model calls; hybrid programs call only declared, bounded semantic functions. Backend authorization remains authoritative.
Lifecycle
- Use
inspect_macrofor explicit catalog inspection and draft review, not as a prerequisite for execution. - For inspection-only requests, inspect an existing candidate; do not create one when the catalog is empty. Report that no candidate is available and obtain recording/learning status before proposing another operation.
- A recording token from
armedorclaimedis not a trace ID. Selected successful learning recordings enter program learning automatically; do not bypass a pending or failed job by making a step-only draft or rerunning the task. - Native selected learning requires the actual token-bound root success receipt
and subsequent final answer. Attest success only after the task succeeds;
the receipt is not independent correctness verification.
awaitingAuthoris not ready. New user work may preempt a child with a recoverable lease; Off/Cancel rejects late work rather than authorizing another provider. - Learning derives independent fixtures, validates source and executable, and publishes under the workspace preference: Off, Prepare, Read-only, or All. Prepare requires review; Read-only and All can auto-publish validated programs. Never change that preference for the user.
- Validate staged programs before explicit approval. Approval never runs a program and does not grant backend permission.
Owner-specific capabilities
typeagent(default): the initial broker supports read-only MCP tools available to TypeAgent; native-only tools fail explicitly.plugin(opt-in): supports nativeview,rg,globand exact explicitly classified read-only MCP identities. Optional TypeAgent action support requires the selected action's read-only contract, not blanket classification of the generic execute tool.off: recipe routing/learning is disabled; do not select another owner.
Both production brokers reject writes and unsupported schemas/provenance. Read-effect classification and publication do not grant permission. Normal native hooks, content exclusions and backend authorization remain authoritative. All does not enable writes. A clean pre-acceptance miss may stay native; an accepted/ambiguous failure must never be repeated through a native/planner route.