You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Documented daemon API for current native questions and conditional answers? #175
Is there an existing documented API for a client addressing a worker through a Tern daemon to read its currently open native question, including the text, ordered options, and dialog identity/revision, rather than infer it from transcript history?
For ordinary agent prompts only, can an answer be conditional on the exact worker instance and dialog identity/revision, atomically refusing without sending input if either has changed? Is there a receipt/status query that distinguishes response acceptance, confirmed closure of that original dialog, and an unknown outcome after a disconnect, without blindly resubmitting?
Which Tern builds and daemon/client combinations support this? A documentation link and minimal read → conditional-answer → receipt example would be sufficient. We are asking about an existing supported interface, not proposing command names.
Provider safety checks and permission or other protected approval dialogs must retain their required fresh human approval; this inquiry does not request bypassing them.
Adding a concrete plugin use case to this request: ORCH+ needs an opt-in way to accept the answer an agent explicitly recommends in a structured question. We currently cannot enable this through the documented Tern plugin interface.
The route we previously used with Orca
We ran our own orchestrator agent as the user's decision proxy under an explicit, per-session grant. Workers sent recommendation-first message files; the orchestrator accepted appropriate recommendations, logged the decision, and replied through the communication files. For questions already open in a terminal pane, it read the current screen and answered through Orca's terminal CLI, using provider-specific key handling and checking the result.
That workflow worked; it was our automation above Orca, not a built-in Orca recommendation API. It depended on a model reading text or terminal state. Our retained notes also record menu misfires. The original implementation has since been retired, so this is a description of the worked route, not a claim that reusable source is available.
What would remove that workaround in Tern
A supported plugin API that can:
List or subscribe to pending structured questions for the exact agent session/pane, with stable question/dialog ID, revision, ordered choices, and single-/multi-select semantics.
Expose the agent's explicit recommended choice(s) as metadata, rather than infer a recommendation from a label or highlighted row.
Answer that exact question conditionally on the session and dialog ID/revision, with a result distinguishing accepted, answered elsewhere/stale, and unknown. This extends the conditional-answer requirement in the original issue.
The plugin would make this a deliberate user opt-in. Without an explicit recommendation it would leave the question for the user. This is about ordinary agent questions, not bypassing authentication or protected permission approvals. It should not require terminal keystrokes, and it should work for a local agent and for an agent on an attached execution host.
For related upstream context, Orca's PR #31440 describes why terminal-state-dependent answering is unreliable and calls for an active-dialog-ID-bound semantic answer API.
Proposed setting
The screenshot below is our current unavailable-state design. We are hiding this unsupported control until the necessary capability exists; the image is a UI proposal, not evidence that Tern currently supports it.
Environment: Tern 0.9.18, Windows desktop with an attached Linux execution host. If a supported interface already covers this, a minimal pending-question → recommendation metadata → conditional answer → result example and supported versions would be enough.
Activity
Einherjare commentedon Nov 14, 2026
Adding a concrete plugin use case to this request: ORCH+ needs an opt-in way to accept the answer an agent explicitly recommends in a structured question. We currently cannot enable this through the documented Tern plugin interface.
The route we previously used with Orca
We ran our own orchestrator agent as the user's decision proxy under an explicit, per-session grant. Workers sent recommendation-first message files; the orchestrator accepted appropriate recommendations, logged the decision, and replied through the communication files. For questions already open in a terminal pane, it read the current screen and answered through Orca's terminal CLI, using provider-specific key handling and checking the result.
That workflow worked; it was our automation above Orca, not a built-in Orca recommendation API. It depended on a model reading text or terminal state. Our retained notes also record menu misfires. The original implementation has since been retired, so this is a description of the worked route, not a claim that reusable source is available.
What would remove that workaround in Tern
A supported plugin API that can:
The plugin would make this a deliberate user opt-in. Without an explicit recommendation it would leave the question for the user. This is about ordinary agent questions, not bypassing authentication or protected permission approvals. It should not require terminal keystrokes, and it should work for a local agent and for an agent on an attached execution host.
For related upstream context, Orca's PR #31440 describes why terminal-state-dependent answering is unreliable and calls for an active-dialog-ID-bound semantic answer API.
Proposed setting
The screenshot below is our current unavailable-state design. We are hiding this unsupported control until the necessary capability exists; the image is a UI proposal, not evidence that Tern currently supports it.
Environment: Tern 0.9.18, Windows desktop with an attached Linux execution host. If a supported interface already covers this, a minimal pending-question → recommendation metadata → conditional answer → result example and supported versions would be enough.