Change skills

These are the operations the Admin Assistant uses to modify your workspace. They only run when the Allow changes toggle is on, and they split into three stages that map to the workflow: propose (stage a change), apply (commit an approved change), and test (verify a saved change).

The key boundary: proposing stages a change in memory and validates it; nothing reaches your workspace until you approve a specific proposal and the Operator applies it.

Propose — staging a change

Each propose-* tool takes a draft, renders it into the exact export format the workspace-definitions zip uses, computes the "before" from the live configuration with the same exporter, and then validates it:

• round-trips the code through the same parser the workspace uses to load it;
• compiles code bodies against the matching execution scope and returns line/column diagnostics on failure;
• runs the secret deny-list over the change;
• flags policy regressions (for example, widening an access mode) as blocking concerns.

A proposal that fails validation is never marked ready to approve. Nothing is written to the workspace at this stage — a proposal is a reviewed before/after diff waiting for your approval.

Code surfaces

Tool	Proposes a change to
propose-endpoint	A code endpoint.
propose-tool	An AI tool.
propose-agent	An agent.
propose-skill	A reusable agent skill.
propose-prompt-template	An AI prompt template.
propose-code-task	A custom-code scheduled task.
propose-data-connector	A data-connector scheduled task.
propose-migration	A migration task.
propose-entity-post-processing	An entity post-processing script.
propose-shell-script	A shell-script task.

Schema

Tool	Proposes a change to
propose-schema-node	A node type.
propose-schema-edge	An edge type.
propose-schema-renderer	A node renderer.
propose-schema-style	A node style.

Indexes, NLP, and search

Tool	Proposes a change to
propose-index	A graph index.
propose-code-index	A custom-code index.
propose-search-code-index	A custom-code search index.
propose-nlp-spotter-from-node	A node-derived NLP spotter.
propose-nlp-pattern-spotter	An NLP pattern spotter.
propose-nlp-pipeline	An NLP pipeline.
propose-nlp-linking	An entity-to-node linking configuration.
propose-search-settings	Search settings.
propose-search-facets	Search facets.
propose-search-synonyms	Search synonyms.
propose-search-replacement	A forced search replacement pattern.

Settings and access

Tool	Proposes a change to
propose-ui-settings	Front-end UI settings.
propose-embedding-settings	Embedding / model preferences.
propose-ai-provider	An AI provider configuration.
propose-access-group	An access group.
propose-access-grant	An access grant.

Managing proposals

The assistant can also reshape its own set of staged proposals before you approve them: list the pending proposals, read one proposal's full before/after diff, and discard a proposal it has decided to abandon. These act only on staged proposals in the current session, not on the workspace.

Apply — committing an approved change

Applying is the only stage that writes to your workspace, and only the Operator agent can do it. It re-runs the Reviewer's static checks as a final guard, consumes a one-time approval token tied to the exact diff you saw, and wraps the write in the workspace's configuration-tracking pipeline so each applied change is a single git commit.

Tool	What it does
apply-proposal	Applies one approved proposal.
apply-batch	Applies a set of approved proposals together (respecting their dependencies).
rollback-last	Reverts the most recent applied change.

Test — verifying a saved change

After a change is applied, the Verifier runs test-* tools against the saved configuration to confirm it behaves as intended. These tools do not modify configuration; they execute the saved artefact (so they may have the same side effects running it normally would) and report the result. On failure, the assistant can escalate to rollback-last.

Tool	What it verifies
test-endpoint	Runs a saved endpoint.
test-tool	Runs a saved AI tool.
test-agent-run	Runs a saved agent end-to-end.
test-code-task	Runs a saved scheduled task.
test-index-materialize	Materializes a saved custom index.
test-search	Runs a search against the current configuration.
test-spotter	Runs a saved NLP spotter over sample text.

Next steps

• Read-only skills — the inspection tools that run before any change.
• Examples — end-to-end tasks that exercise propose, apply, and verify.