Skip to main content
A workflow step can bind published skill versions and an optional agent profile. Kiwi keeps those bindings with the workflow revision and run. Publishing another release does not change work already in progress.

​
Publish instructions

Open a project’s Skills area. Create a draft with its instructions, referenced text files, exact skill dependencies, and required tools. To revise a release, start from that exact version and describe the change. Use a higher release number; the selected predecessor remains explicit. For feedback-driven changes, preserve the real failure example, affected consumers, and intended behavior change. Maintenance revisions can leave feedback absent. Create a remote check for the current draft, inspect its findings, and publish a numbered release as a project Admin. Publication links the exact checked content and the Admin’s decision. An incomplete semantic review requires explicit acknowledgement and remains recorded as incomplete. A completed review can contain findings; it does not make a workflow acceptance decision. Published instructions and profile settings are immutable. Archive a version to prevent new workflow bindings and runs from selecting it. Existing runs keep their pinned versions and authorized access.

​
Bind a workflow

Include exact references in the workflow JSON you publish:
These are optional fields on a step; retain its other required fields. The optional workflow-level delegated_skills array declares released skills available for bounded child assignments. The service resolves dependencies and validates the complete manifest. Importing a supplied digest cannot replace that validation. Create a new workflow revision and run to adopt updated skills or profiles. A run keeps its original versions through retries, rework, and handoff. Earlier runs without version records show that tracking was not recorded.

​
Fetch an assignment through MCP

Give a worker explicit run, step, and, when delegated, assignment references. A concise project entry instruction can be:
The entry instruction routes work; the full procedure stays in Kiwi. Agents fetch instructions before starting an eligible step. Delegated workers retrieve their own bounded contract, while parent output requirements are identified separately. After starting work, use report_configuration and report_skill_use to record exposed runtime settings and declared skill use. Unknown values remain unknown. Append a report when settings change. Configuration uses a safe allowlist; credentials, raw prompts, and environment dumps are excluded. get_step_skill returns the exact instruction body, dependency references, required tools, and reference-file manifest. get_skill_file returns a specific file from that release. Reading content records delivery, not proof that the agent followed it. File digests are SHA-256 of the exact content string encoded as canonical JSON, including its quotes and escaped characters. They differ from a raw file-byte hash. Preserve the returned text exactly when verifying its digest.

​
Inspect a run

The run view shows readable skill and profile versions. Select a step to inspect its bindings, parent and child performers, harness versions, configuration history, reported skill use, and instruction deliveries. Version links open the exact release and its related authorized runs. Server-pinned versions and agent-reported settings have distinct provenance. A model identifier can be a requested alias; Kiwi does not infer an unavailable provider revision or hidden setting. Older attempts retain their evidence and participant records after replacement. Compare releases or run attempts to see meaningful changes first. Expand unchanged fields when needed. Skill comparisons cover instructions, files, dependencies, and tools; execution comparisons include versions, participants, profiles, reported settings, and unknowns.

​
Use remote checks

The hosted checker exposes create_skill_check, run_skill_check, and get_skill_check through MCP. Use them for a specific draft or release. Kiwi retains structural observations, semantic findings, exact input and context digests, checker and ruleset versions, and reviewer settings. Jobs can be pending or complete; after an unknown result, read the same job before starting another check. A semantic review that is unavailable or incomplete remains blocked or incomplete in the result, while any completed structural observations remain readable. The page and MCP response show the actual state. A blocked result is not a completed semantic review. A product defect is a failed result. An unusable capture, tool, or criterion produces blocked evidence and a decision request to its owner. The checker helps identify instruction defects and propose repairs; publishing changes and accepting workflow work retain their designated authority.