List the outputs a given step can read — i.e. what upstream steps make available to it. Use it to bind a step's inputs to real upstream outputs instead of guessing field names. Stateless, and safe for read-only service tokens.

Apply structural mutation commands to a workflow draft (add/remove steps, connect edges, set conditions). Read the workflow first for its version. Commands are applied IN ORDER, one backend call each, with the version threaded automatically — pass the version you read, not one per command.

RESPONSE IS A SUMMARY, not the document: version, appliedCount, launchable, validationIssues, stepCount, edgeCount. A real workflow document can exceed the response limit, which would turn a call that SUCCEEDED into an error you must not retry. Pass returnDocument: true, or call read_workflow, when you need the document itself.

Not transactional: if command N fails, commands before it stay applied and the error reports how many landed and which index failed — resume from there, do NOT replay the list. On a 409 conflict, re-read and retry.

Each command is { "type": "add_step", "payload": { … } }; the flat form { "type": "add_step", … } is accepted too. Valid types: update_workflow, add_trigger, update_trigger, remove_trigger, update_trigger_operator, group_triggers, ungroup_triggers, move_trigger, add_step, add_loop, update_step, remove_step, move_step, add_edge, update_edge, remove_edge, add_decision_condition, convert_to_loop, setup_loop_decision, advanced_edit, attach_output_schema, detach_output_schema.

PAYLOAD CONTRACT:

• add_step builds a COMPLETE step in ONE call: { name, position: {after|before: , …} } are required, and id, type, typeId, config, contract, inputs, outputs and edgeType are all accepted alongside them. No follow-up update_step is needed merely to fill the step in.

• add_edge accepts the document's own from/to as well as fromStepId/toStepId, and HONOURS an id you supply rather than replacing it, so ids stay diffable against an environment you are reproducing.

• Payloads are STRICT: an unknown or misspelled key is a 422 that names it, never a silent drop. Read the error instead of assuming a value landed.

TWO INVARIANTS RUN AFTER EVERY EDIT, and the response reports what they did in systemAdjustments (addedEdges / removedEdges / rederivedInputs):

• Every step is guaranteed to reach the END point; the platform wires that edge itself. An edge you did not ask for is this, and it is listed.

• A bound input's contract, its description included, is DERIVED from the output it reads and cannot be authored on its own. Set the text on the PRODUCING step's output; a downstream description changing by itself is this rule at work, and the inputs it re-derived are listed.

Archive a run — hides it from the default run list without deleting it. Reversible with unarchive_run.

Scope a policy to a agent so the agent uses it. Verify it landed with list_agent_policies before request_publish_agent — the link lives outside the entity body, so re-reading the agent will not show it.

Place a library prompt snippet into a flow step's prompt stack. Choose the channel (target: system/user) and its order in that channel. This is what makes a snippet actually shape the step's prompt — creating the snippet alone does nothing. Returns the created link (with its linkId) so you can reorder or detach it later.

Scope a reference to a agent so the agent uses it. Verify it landed with list_agent_reference_docs before request_publish_agent — the link lives outside the entity body, so re-reading the agent will not show it.

Scope a skill to a agent so the agent uses it. Verify it landed with list_agent_skills before request_publish_agent — the link lives outside the entity body, so re-reading the agent will not show it.

Scope a skill to a workflow so the workflow uses it.

Read the authoring conventions BEFORE creating or updating anything: the draft→publish lifecycle, optimistic locking, per-entity create contracts (required fields, valid enum values, what 'complete' means), how to WIRE tools→agents, agents/tools→workflows and delegation, and how to reproduce a setup by creating new entities and remapping ids (never copy an id across tenants).

Archive up to 500 runs at once. Reversible — archived runs can be unarchived individually. Returns counts of succeeded and failed.

Delete up to 500 runs at once. IRREVERSIBLE — unlike archiving, deleted runs do not come back, and their traces and cost history go with them. Prefer bulk_archive_runs unless a human has explicitly asked for deletion. Returns counts of succeeded and failed; a partial failure is possible.

Soft-delete several workflows at once (each recoverable with restore_workflow, like single delete_workflow). Each target needs its own current version for optimistic locking — read each first. Returns a per-workflow success/error result; a partial failure is possible.

Cancel an in-flight run. Stops further steps; already-completed steps and their side effects are not undone.

Fork a flow into a new, independent, tenant-owned copy — including a framework-provided one. Prefer this over editing a framework flow directly: it is visible and NOT actually protected from update_flow/ delete_flow (a backend gap, despite what its docstrings claim), so editing one in place risks changing something other tenants share.

Duplicate a prompt snippet into a new, independent copy.

Create a new agent draft. Pass the entity's fields (camelCase) in fields; the backend validates them. Returns the created agent with its id and version.

Create an agent's persona. Pass name, characterText and optional status.

Create a conditional trigger on a workflow: an agent checks conditionText every repeatIntervalMinutes and fires the workflow when it is met.

Create a connector draft (a tool of type 'connector'). Do NOT put real credentials in authConfig — use placeholders; a human fills secrets in Axonity.

Schedule a workflow to run on a cron expression.

Create a new flow draft. Pass the entity's fields (camelCase) in fields; the backend validates them. Returns the created flow with its id and version.

Create a new output_schema draft. Pass the entity's fields (camelCase) in fields; the backend validates them. Returns the created output_schema with its id and version.

Create a new policy draft. Pass the entity's fields (camelCase) in fields; the backend validates them. Returns the created policy with its id and version.

Create a new prompt_snippet draft. Pass the entity's fields (camelCase) in fields; the backend validates them. Returns the created prompt_snippet with its id and version.

Create a new reference_doc draft. Pass the entity's fields (camelCase) in fields; the backend validates them. Returns the created reference_doc with its id and version.

Create a new skill draft. Pass the entity's fields (camelCase) in fields; the backend validates them. Returns the created skill with its id and version.

Create a new tool draft. Pass the entity's fields (camelCase) in fields; the backend validates them. Returns the created tool with its id and version.

Create a webhook trigger on a workflow. Returns the trigger AND its plaintext token — shown ONCE and never retrievable again. Hand the token to your human immediately; do not store it in an entity field.

Create a new workflow draft. Pass the entity's fields (camelCase) in fields; the backend validates them. Returns the created workflow with its id and version.

Delete a agent. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_agent (it will not appear in list_agents until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a agent's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_agent_version until it is purged for good.

Delete a conditional trigger. HARD delete — the row is removed and cannot be restored.

Delete a cron schedule. HARD delete — the row is removed and cannot be restored. The workflow stops running on this schedule immediately.

Delete a flow. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_flow (it will not appear in list_flows until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a flow's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_flow_version until it is purged for good.

Delete a output_schema. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_output_schema (it will not appear in list_output_schemas until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a output_schema's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_output_schema_version until it is purged for good.

Delete a persona. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_persona (it will not appear in list_personas until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a persona's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_persona_version until it is purged for good.

Delete a policy. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_policy (it will not appear in list_policies until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a policy's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_policy_version until it is purged for good.

Delete a prompt_snippet. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_prompt_snippet (it will not appear in list_prompt_snippets until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a prompt_snippet's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_prompt_snippet_version until it is purged for good.

Delete a reference_doc. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_reference_doc (it will not appear in list_reference_docs until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a reference_doc's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_reference_doc_version until it is purged for good.

Delete ONE run. IRREVERSIBLE — the run and its trace and cost history go with it. Prefer archive_run unless a human has asked for deletion. For many at once use bulk_delete_runs.

Delete a skill. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_skill (it will not appear in list_skills until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a skill's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_skill_version until it is purged for good.

Delete a tool. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_tool (it will not appear in list_tools until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a tool's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_tool_version until it is purged for good.

Delete a webhook trigger. HARD delete — the row is removed and cannot be restored, and any caller using its token starts failing immediately.

Delete a workflow. Destructive, but RECOVERABLE — it is soft-deleted and can be brought back with restore_workflow (it will not appear in list_workflows until then). Read it first for its version. A 409 here may be a stale version (re-read and retry) or a reference conflict (something published still uses it — re-reading will not help; the error tells you which).

Permanently remove one entry from a workflow's version history. The working draft and the currently published version are protected and cannot be targeted this way. Recoverable with restore_deleted_workflow_version until it is purged for good.

Unscope a policy from a agent. This only removes the link — the policy itself is not deleted and stays available to attach elsewhere. The response is the SERVER'S answer, not a success claim by this connector: a detach that had nothing to remove reports that instead of failing, so read the body rather than assuming the link is gone. Confirm what is left with list_agent_policies.

Remove a prompt snippet from a flow step's stack. This only unlinks it — the snippet itself is not deleted and stays available to place elsewhere. Identify the link by its linkId (from list_flow_step_prompts). This route answers 204 with no body, so the result confirms the call was accepted and nothing more — read the stack back with list_flow_step_prompts to see what is actually attached.

Unscope a reference from a agent. This only removes the link — the reference itself is not deleted and stays available to attach elsewhere. The response is the SERVER'S answer, not a success claim by this connector: a detach that had nothing to remove reports that instead of failing, so read the body rather than assuming the link is gone. Confirm what is left with list_agent_reference_docs.

Unscope a skill from a agent. This only removes the link — the skill itself is not deleted and stays available to attach elsewhere. The response is the SERVER'S answer, not a success claim by this connector: a detach that had nothing to remove reports that instead of failing, so read the body rather than assuming the link is gone. Confirm what is left with list_agent_skills.

Unscope a skill from a workflow. This only removes the link — the skill itself is not deleted and stays available to attach elsewhere. The response is the SERVER'S answer, not a success claim by this connector: a detach that had nothing to remove reports that instead of failing, so read the body rather than assuming the link is gone.

Discard uncommitted draft edits on a agent, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Discard uncommitted draft edits on a flow, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Discard uncommitted draft edits on a output_schema, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Discard uncommitted draft edits on a persona, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Discard uncommitted draft edits on a policy, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Discard uncommitted draft edits on a prompt_snippet, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Discard uncommitted draft edits on a reference_doc, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Discard uncommitted draft edits on a skill, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Discard uncommitted draft edits on a tool, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Discard uncommitted draft edits on a workflow, resetting it to its last published state. Fails if it has never been published — there is no published state to fall back to.

Prove a SAVED tool by running its OWN stored code in the sandbox. This is the run that satisfies the publish gate: a clean pass stamps the tool's current version and clears the dry_run_required blocker, so request_publish_tool can succeed. Nothing about the code is supplied by you — only sample input — which is exactly why it counts and why execute_tool does not. Editing the tool afterwards bumps its version and invalidates the proof, so dry-run again after your last edit, not before it. Expect a refusal for tools that have no code to prove: 403 for a platform-shipped (locked) tool, 422 for a connector, a builtin-backed validator, or a tool with an empty implementation.

Test-run an ALREADY-SAVED connector by its tool id — the backend loads its stored authConfig and decrypts the real secret server-side; nothing credential-shaped ever passes through this tool or through you. Takes only input parameters, nothing else — there is no way to point this at a URL or auth config of your choosing (that admin-only, body-supplied form exists on the backend precisely so a client like this one can never reach it, since it would otherwise let a caller exfiltrate a decrypted secret).

Run tool code directly, WITHOUT saving it — the way to test a tool you are authoring, before create_tool/update_tool. Returns stdout/stderr, the result, and a typed errorType (timeout/memory/import/runtime/validation) on failure. Pass toolId to run against an already-saved tool's context.

This does NOT satisfy the publish gate, however cleanly it runs. It executes the functions YOU supply, which need not be the tool's stored code, so a pass here proves nothing about what would actually ship. To clear the dry_run_required blocker on a SAVED tool, use dry_run_tool — that runs the tool's own stored implementation and is the only run that counts.

Format Python tool code with Black and return the formatted source. Stateless and safe for read-only service tokens — it does not save anything.

Read one publish approval by id — lets you poll a specific request rather than re-fetching the whole list. Its readiness is recomputed on this read, so it reflects the entity as it stands NOW; readinessAtRequest keeps what the requester saw and readinessChanged flags a difference. This is how you check whether a blocker (for a tool, typically dry_run_required) has since been cleared.

List the policy currently scoped to an agent — the read-back for attach_policy_to_agent. Use it to confirm a link took, or to read a source agent's wiring before recreating it. Read-only.

List the reference_doc currently scoped to an agent — the read-back for attach_reference_doc_to_agent. Use it to confirm a link took, or to read a source agent's wiring before recreating it. Read-only.

List all agents in your Axonity tenant (id, name, status, version).

List the skill currently scoped to an agent — the read-back for attach_skill_to_agent. Use it to confirm a link took, or to read a source agent's wiring before recreating it. Read-only.

List a agent's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_agent_version.

List the company document's version history — checkpoints and major versions.

List a workflow's conditional triggers — an agent evaluates a condition on an interval and starts the workflow when it holds.

List a workflow's cron schedules, with their next and last fire times.

List soft-deleted agents — the restore candidates for restore_agent.

List a agent's deleted version rows — restore candidates for restore_deleted_agent_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List soft-deleted flows (reusable workflow fragments) — the restore candidates for restore_flow.

List a flow's deleted version rows — restore candidates for restore_deleted_flow_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List soft-deleted output schemas (reusable step/agent output contracts) — the restore candidates for restore_output_schema.

List a output_schema's deleted version rows — restore candidates for restore_deleted_output_schema_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List soft-deleted personas (an agent's character, 1:1 with the agent) — the restore candidates for restore_persona.

List a persona's deleted version rows — restore candidates for restore_deleted_persona_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List soft-deleted policies (rules and guardrails for agents) — the restore candidates for restore_policy.

List a policy's deleted version rows — restore candidates for restore_deleted_policy_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List soft-deleted prompt snippets (reusable prompt fragments) — the restore candidates for restore_prompt_snippet.

List a prompt_snippet's deleted version rows — restore candidates for restore_deleted_prompt_snippet_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List soft-deleted reference docs (background knowledge for agents) — the restore candidates for restore_reference_doc.

List a reference_doc's deleted version rows — restore candidates for restore_deleted_reference_doc_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List soft-deleted skills (reusable know-how for agents) — the restore candidates for restore_skill.

List a skill's deleted version rows — restore candidates for restore_deleted_skill_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List soft-deleted tools (functions, connectors, validators, evaluators) — the restore candidates for restore_tool.

List a tool's deleted version rows — restore candidates for restore_deleted_tool_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List soft-deleted workflows (business processes) — the restore candidates for restore_workflow.

List a workflow's deleted version rows — restore candidates for restore_deleted_workflow_version. Deleted version numbers stay reserved (not reused), so gaps in the numbering are expected, not corruption.

List all flows (reusable workflow fragments) in your Axonity tenant (id, name, status, version).

List the prompt snippets attached to a flow step — its system and user prompt stacks, each entry with its link id and display order. This is how you see a step's current prompt composition before changing it. Excludes wildcard snippets (see list_wildcard_prompts). Read-only.

List a flow's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_flow_version.

List all output schemas (reusable step/agent output contracts) in your Axonity tenant (id, name, status, version).

List a output_schema's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_output_schema_version.

List all personas (an agent's character, 1:1 with the agent) in your Axonity tenant (id, name, status, version).

List a persona's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_persona_version.

List all policies (rules and guardrails for agents) in your Axonity tenant (id, name, status, version).

List a policy's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_policy_version.

List all prompt snippets (reusable prompt fragments) in your Axonity tenant (id, name, status, version).

List a prompt_snippet's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_prompt_snippet_version.

List this tenant's publish approvals and their status — how you find out whether a request_publish_* was approved or rejected. Optionally filter by status, and page through with limit/offset. Approving and rejecting are human-only actions in Axonity; no tool can do them.

Readiness is recomputed WHEN YOU READ, not frozen at request time: readiness is the verdict right now, readinessAtRequest is what the requester saw, and readinessChanged says whether it moved. So a blocker you saw an hour ago may already be resolved — and 'ready' means ready now.

List all reference docs (background knowledge for agents) in your Axonity tenant (id, name, status, version).

List a reference_doc's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_reference_doc_version.

List workflow runs across the tenant, newest first. Archived runs are excluded unless includeArchived is true. To list runs of ONE workflow, use list_workflow_runs instead — this route has no workflow filter.

List all skills (reusable know-how for agents) in your Axonity tenant (id, name, status, version).

List a skill's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_skill_version.

List the tenant's system-tool catalog (id, label, description, category). There is no grant/revoke tool — enable one for an agent by adding its id to systemToolIds via update_agent.

List all tools (functions, connectors, validators, evaluators) in your Axonity tenant (id, name, status, version).

List a tool's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_tool_version.

List a workflow's webhook triggers. Secrets are never returned — a trigger's token is shown only once, when it is created or rotated.

List the tenant's wildcard prompt snippets — those that apply to EVERY flow step (they sit at the top of each step's prompt stack). Use this for cross-cutting instructions that should hold everywhere, rather than attaching the same snippet to every step by hand. Read-only.

List the runs of one workflow. Note this route takes archivedOnly (show ONLY archived runs), which is not the same as list_runs' includeArchived.

List all workflows (business processes) in your Axonity tenant (id, name, status, version).

List a workflow's version history — checkpoints (every save) and major versions (named releases). Use this to find the versionId you need for restore_workflow_version.

Give a agent's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_agent_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Give one of the company document's major versions a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_company_versions), not a checkpoint or a versionId. Like everything else on this singleton it takes no id. This only renames; it does not create, publish, or roll anything back.

Give a flow's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_flow_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Give a output_schema's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_output_schema_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Give a persona's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_persona_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Give a policy's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_policy_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Give a prompt_snippet's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_prompt_snippet_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Give a reference_doc's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_reference_doc_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Give a skill's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_skill_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Give a tool's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_tool_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Give a workflow's major version a name — label a release so it is easy to find in history. Targets a MAJOR version by its integer number (from list_workflow_versions), not a checkpoint or a versionId. This only renames; it does not create, publish, or roll anything back.

Read one agent by id, including its current draft and version. Read before you update — you need the current version.

Read an agent's persona (voice/character). A persona is 1:1 with an agent — there is no standalone persona list.

Read the LIVE published version of a agent — what the runtime actually uses, as opposed to the draft that read_agent returns. Diff the two to see what a pending publish would change.

Read one historical version of a agent by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_agent_versions) — not by its versionId. Use it to diff two versions before rolling back.

Read this tenant's company document (its current draft and version). There is one company per tenant — no id. Read it before you update, to get the version for optimistic locking.

Read the LIVE published company document — what the runtime uses, as opposed to the draft read_company returns. Diff the two to see what a publish would change.

Read one historical company version by its CHECKPOINT NUMBER (the integer version from list_company_versions), including its document.

Read one flow by id, including its current draft and version. Read before you update — you need the current version.

The flow-scoped counterpart of read_workflow_prompt_stacks: resolve a flow to its steps and their system/user prompt stacks (flowStepId, link ids, order). Read-only.

Read the LIVE published version of a flow — what the runtime actually uses, as opposed to the draft that read_flow returns. Diff the two to see what a pending publish would change.

Read one historical version of a flow by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_flow_versions) — not by its versionId. Use it to diff two versions before rolling back.

Read one output_schema by id, including its current draft and version. Read before you update — you need the current version.

Read the LIVE published version of a output_schema — what the runtime actually uses, as opposed to the draft that read_output_schema returns. Diff the two to see what a pending publish would change.

Read one historical version of a output_schema by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_output_schema_versions) — not by its versionId. Use it to diff two versions before rolling back.

Read one persona by id, including its current draft and version. Read before you update — you need the current version.

Read the LIVE published version of a persona — what the runtime actually uses, as opposed to the draft that read_persona returns. Diff the two to see what a pending publish would change.

Read one historical version of a persona by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_persona_versions) — not by its versionId. Use it to diff two versions before rolling back.

Read one policy by id, including its current draft and version. Read before you update — you need the current version.

Read the LIVE published version of a policy — what the runtime actually uses, as opposed to the draft that read_policy returns. Diff the two to see what a pending publish would change.

Read one historical version of a policy by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_policy_versions) — not by its versionId. Use it to diff two versions before rolling back.

Read one prompt_snippet by id, including its current draft and version. Read before you update — you need the current version.

Read the LIVE published version of a prompt_snippet — what the runtime actually uses, as opposed to the draft that read_prompt_snippet returns. Diff the two to see what a pending publish would change.

Read one historical version of a prompt_snippet by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_prompt_snippet_versions) — not by its versionId. Use it to diff two versions before rolling back.

Read one reference_doc by id, including its current draft and version. Read before you update — you need the current version.

Read the LIVE published version of a reference_doc — what the runtime actually uses, as opposed to the draft that read_reference_doc returns. Diff the two to see what a pending publish would change.

Read one historical version of a reference_doc by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_reference_doc_versions) — not by its versionId. Use it to diff two versions before rolling back.

Read one run in full: status, per-step state, the workflow snapshot it ran against, trigger input, agent invocations, and validator verdicts. The verdicts are the closest thing to evaluator findings — there is no separate findings endpoint.

Read a run's token and cost breakdown, per agent and model.

Counts of runs by state across the tenant (running, waiting, completed, failed, expired, total).

Read a run's chronological tool-call trace — each entry has the step, the tool, its arguments, its output, and any error. This is what you read to work out WHY a run behaved as it did.

Read one skill by id, including its current draft and version. Read before you update — you need the current version.

Read the LIVE published version of a skill — what the runtime actually uses, as opposed to the draft that read_skill returns. Diff the two to see what a pending publish would change.

Read one historical version of a skill by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_skill_versions) — not by its versionId. Use it to diff two versions before rolling back.

Read one tool by id, including its current draft and version. Read before you update — you need the current version.

Read the LIVE published version of a tool — what the runtime actually uses, as opposed to the draft that read_tool returns. Diff the two to see what a pending publish would change.

Read one historical version of a tool by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_tool_versions) — not by its versionId. Use it to diff two versions before rolling back.

Read one workflow by id, including its current draft and version. Read before you update — you need the current version.

Resolve a workflow to its flow steps and each step's prompt stack in one read: per step you get its flowStepId, name, and its system/user prompt snippets (with link ids and order). This is how you discover the flowStepIds you need before placing prompts, and how you see a workflow's whole prompt composition. Read-only.

Read the LIVE published version of a workflow — what the runtime actually uses, as opposed to the draft that read_workflow returns. Diff the two to see what a pending publish would change.

Read the input parameters a workflow's triggers expect — what a caller must supply to start it. Use this before authoring a webhook/cron/conditional trigger, or before start_workflow_run, so the trigger input matches. Read-only.

Read one historical version of a workflow by its CHECKPOINT NUMBER (an integer like 3, from the version field in list_workflow_versions) — not by its versionId. Use it to diff two versions before rolling back.

Set the full ordering of a flow step's prompt snippets in one call. Pass the snippet ids in the order you want them; the backend replaces the step's ordering with this list. Read the step first (list_flow_step_prompts) so you include every snippet — omitting one is a reorder, not a safe no-op.

Replace the ENTIRE workflow document in one atomic call. Read the workflow first to get the current version, then call this again only after rereading on a stale 409. Unlike apply_workflow_mutations, this is a single PUT and treats the payload as the complete editable document — prefer the mutation commands, which the backend validates and normalises per edit.

The response is the same SUMMARY (version, launchable, validationIssues, stepCount, edgeCount), not the stored document; pass returnDocument: true or call read_workflow for that.

Request that a agent be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Request publication of MANY entities in one call — the same human approval queue as request_publish_*, one entry per entity, up to 200. Taking a whole tenant live means hundreds of individual requests otherwise, which turns the human review it exists for into clicking through. This does NOT publish and does NOT approve: a human still decides, in Axonity.

NOT TRANSACTIONAL: each entry carries its own outcome and one failure does not undo the entries before it. The response is { results: [{ id, success, status, error }], succeededCount, failedCount }. Read the per-entry results — retrying the whole list because failedCount is non-zero double-requests everything that worked.

Request that the company document be published. Like request_publish_* for every other entity, this does NOT publish — it creates a pending approval a human approves in Axonity. Company is a singleton, so unlike the others this takes NO id: the server resolves the one company from your tenant. (Direct company publish is closed to service tokens — this is the only path.)

Request that a flow be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Request that a output_schema be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Request that a persona be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Request that a policy be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Request that a prompt_snippet be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Request that a reference_doc be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Request that a skill be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Request that a tool be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Request that a workflow be published. This does NOT publish it — it creates a pending approval a human approves in Axonity. Returns the pending approval (with its readiness).

Undo delete_agent — bring a soft-deleted agent back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live agent now holds a name/slot this one also claims.

Roll a agent's DRAFT back to an earlier version. Takes the versionId (a UUID from list_agent_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Roll the company draft back to an earlier version. Takes the versionId (a UUID from list_company_versions), not the checkpoint number. Changes only the draft; what is live stays live until a human publishes.

Undo delete_agent_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_agent_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_flow_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_flow_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_output_schema_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_output_schema_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_persona_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_persona_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_policy_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_policy_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_prompt_snippet_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_prompt_snippet_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_reference_doc_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_reference_doc_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_skill_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_skill_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_tool_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_tool_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_workflow_version — bring a removed version row back into history. This does NOT change the current draft (unlike restore_workflow_version); it only makes the row visible/restorable again. Takes no version check.

Undo delete_flow — bring a soft-deleted flow back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live flow now holds a name/slot this one also claims.

Roll a flow's DRAFT back to an earlier version. Takes the versionId (a UUID from list_flow_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Undo delete_output_schema — bring a soft-deleted output_schema back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live output_schema now holds a name/slot this one also claims.

Roll a output_schema's DRAFT back to an earlier version. Takes the versionId (a UUID from list_output_schema_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Undo delete_persona — bring a soft-deleted persona back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live persona now holds a name/slot this one also claims.

Roll a persona's DRAFT back to an earlier version. Takes the versionId (a UUID from list_persona_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Undo delete_policy — bring a soft-deleted policy back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live policy now holds a name/slot this one also claims.

Roll a policy's DRAFT back to an earlier version. Takes the versionId (a UUID from list_policy_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Undo delete_prompt_snippet — bring a soft-deleted prompt_snippet back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live prompt_snippet now holds a name/slot this one also claims.

Roll a prompt_snippet's DRAFT back to an earlier version. Takes the versionId (a UUID from list_prompt_snippet_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Undo delete_reference_doc — bring a soft-deleted reference_doc back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live reference_doc now holds a name/slot this one also claims.

Roll a reference_doc's DRAFT back to an earlier version. Takes the versionId (a UUID from list_reference_doc_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Undo delete_skill — bring a soft-deleted skill back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live skill now holds a name/slot this one also claims.

Roll a skill's DRAFT back to an earlier version. Takes the versionId (a UUID from list_skill_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Undo delete_tool — bring a soft-deleted tool back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live tool now holds a name/slot this one also claims.

Roll a tool's DRAFT back to an earlier version. Takes the versionId (a UUID from list_tool_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Undo delete_workflow — bring a soft-deleted workflow back. Takes no version check (a restore must never lose to a stale-version guard). Can 409 if a live workflow now holds a name/slot this one also claims.

Roll a workflow's DRAFT back to an earlier version. Takes the versionId (a UUID from list_workflow_versions), NOT the checkpoint number. The current draft is snapshotted first, so this is itself undoable. It changes only the draft — what is live stays live until a human approves a new publish.

Rotate a webhook trigger's token. Returns the new plaintext token ONCE. The old token keeps working for a short grace period so callers can be migrated.

Start a real run of a workflow — the way to TEST a workflow you just authored. This EXECUTES: it runs the PUBLISHED workflow against live infrastructure (models, connectors), so it can incur cost and cause real side effects. It does not run your unpublished draft. After starting, follow it with read_run / read_run_trace / read_run_cost. Optionally pass triggerInput to feed the workflow's trigger.

Restore an archived run to the default run list.

Update a agent draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Save the company draft as a WHOLE document (full-document PUT, not a field merge). Read it first for its version; on a 409 conflict re-read and retry. This writes the draft only — a human publishes it in Axonity (company is not part of the request_publish_* approval flow).

Change a conditional trigger's condition, interval, agent, or enabled flag. Send only what you are changing.

Update a connector draft (a tool of type 'connector'). Read it first for its version; keep authConfig as placeholders only.

Update a flow draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Change an existing flow-step prompt attachment: move it between the system and user channel, and/or change its display order. Identify it by its linkId (from list_flow_step_prompts or the attach response), NOT the snippet id. Send only what you are changing.

Update a output_schema draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Update a persona draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Update a policy draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Update a prompt_snippet draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Update a reference_doc draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Update a skill draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Update a tool draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Update a workflow draft. Read it first to get expectedVersion; on a 409 conflict, read again and retry. Pass only the fields you are changing (camelCase) in fields.

Check Python tool code for syntax errors and banned patterns before saving it on a tool. Returns valid plus errors (line, column, message, severity, functionName). Stateless and safe for read-only service tokens.

Check a workflow for problems. Returns launchable, catalogChecked, and an issues list (each with code, message, severity, stepIds, edgeIds).

Pass EXACTLY ONE of workflowId or document — they answer different questions:

• workflowId → validates the STORED workflow against this tenant's catalog. This is the real 'can this run?': it confirms the agents, tools and flows the steps reference actually exist. Answers catalogChecked: true.

• document → validates the document you pass, STATELESSLY. No database, so read-only service tokens can call it, but with no catalog the cross-entity checks cannot run: launchable here means 'structurally sound', NOT 'this will run'. Answers catalogChecked: false. Use it for a draft you have not saved, or for fast feedback while editing.

A green verdict is what a reader remembers, so check catalogChecked before telling anyone a workflow is ready. Call this after apply_workflow_mutations and before request_publish_workflow.