Delegated editing and trusted lifecycle¶

A host can give the editor a limited connection while retaining initialization, SAVE, replay and export on its own connection. Supply the editor with the target ID and enough task context to choose and verify an edit. Treat profile selection as a tool boundary. The server checks every call against the selected profile and applies the same service validation.

Observed editor¶

Use observed_editor to delegate structural edits with the existing observation and observed submission tools. The editor can read or search text, obtain a scope receipt and submit guarded operations. The trusted host consumes the accepted receipt and handles lifecycle calls on a separate connection. See the observed editor workflow for the complete editor and host sequence. For an explicit snapshot envelope and compact commit receipt, choose observed editor v2.

Unobserved editor¶

Use unobserved_editor when the delegate should read, search and submit exact text edits. For example, give it a target ID and a request to change a heading. It reads the current version, disambiguates the heading if needed and submits a guarded replacement. It returns the acceptance or rejection to the host for the next lifecycle action.

The delegate has no structural observation, SAVE or replay. This is the read/search/submit part of the existing-target text workflow. Strict observation blocks its unobserved submissions.

Hybrid editor¶

Use hybrid_editor when the host already supplies the exact source context and current version or hash. The delegate can call only patchdiff_submit_edits. For example, the host reads a target, then gives the delegate the target ID, the relevant exact text and version, and the requested change. If the guard fails or the match is ambiguous, the host obtains fresh context before the delegate prepares another submission.

Use unobserved editor if the delegate should read or search for itself. Hybrid editing also depends on admission of unobserved writes.

Full¶

Use full for a trusted operator who needs the complete 17-tool catalog. It can initialize targets, read and edit them, SAVE, replay, export, inspect audit history and support host-owned finalization. A typical trusted sequence initializes a target and hands its ID to a restricted editor, then checks the accepted version, checkpoints it, replays the target or exports it as needed. See the host lifecycle recipe.

The full catalog includes the underlying structural tools. The observed v2 tool aliases are specific to that profile. Strict observation still applies at the service boundary even on a full connection. Consult the shared tool contract for the complete catalog and argument rules.