A connector executes only the action that was judged #1990

Closed
opened 2026-11-14 14:42:48 +00:00 by sam · 2 comments
Owner

The action a connector runs must be exactly the one guards judged and, where needed, the owner approved; anything else is refused and reported. For an automation template, the judged action is the template's declared effect with its parameters and targets.

Done when: an action altered after judgement or approval is refused, and a template launch is judged on its declared effect.

Spec: section 5 rule 6, decision tree 8.5, R18, R32.

The action a connector runs must be exactly the one guards judged and, where needed, the owner approved; anything else is refused and reported. For an automation template, the judged action is the template's declared effect with its parameters and targets. Done when: an action altered after judgement or approval is refused, and a template launch is judged on its declared effect. Spec: section 5 rule 6, decision tree 8.5, R18, R32.
sam added the featurehigh labels 2026-11-14 14:42:48 +00:00
sam added a new dependency 2026-11-14 14:42:59 +00:00
Author
Owner

From the slice 1 kill-mid-run check. Killing the binary after a tool call is recorded resumes cleanly: the run replays the recorded calls and repeats neither the model call nor the tool. But a kill during a tool call leaves nothing recorded, so on resume that call runs again instead of being marked unknown (R15, the spike's "resumes without repeating the action, marking it unknown").

That's harmless for today's read-only lookups. Before any action with an effect runs, execution needs a durable "started" record written before the call (action, digest, idempotency key), so a resume that finds a started call with no outcome records it as unknown and never re-runs it. The agent loop's journal records only outcomes, so this belongs in the execution path this issue builds, beside the digest check.

From the slice 1 kill-mid-run check. Killing the binary after a tool call is recorded resumes cleanly: the run replays the recorded calls and repeats neither the model call nor the tool. But a kill **during** a tool call leaves nothing recorded, so on resume that call runs again instead of being marked unknown (R15, the spike's "resumes without repeating the action, marking it unknown"). That's harmless for today's read-only lookups. Before any action with an effect runs, execution needs a durable "started" record written before the call (action, digest, idempotency key), so a resume that finds a started call with no outcome records it as unknown and never re-runs it. The agent loop's journal records only outcomes, so this belongs in the execution path this issue builds, beside the digest check.
Author
Owner

Done on staging. An action's digest is taken over its canonical input (RFC 8785, raw integers kept), so an action that comes back re-encoded from a step record or decision body still runs as judged, and one whose input actually changed is refused as altered (3a700cddf). A Semaphore template launch is judged on the effect the operator declares for it at acceptance (archipelago 8df55d3): on staging the deletion template was held at propose and wait for reversibility and precedent, approved, and launched once under its key.

Done on staging. An action's digest is taken over its canonical input (RFC 8785, raw integers kept), so an action that comes back re-encoded from a step record or decision body still runs as judged, and one whose input actually changed is refused as altered (3a700cddf). A Semaphore template launch is judged on the effect the operator declares for it at acceptance (archipelago 8df55d3): on staging the deletion template was held at propose and wait for reversibility and precedent, approved, and launched once under its key.
sam closed this issue 2026-11-15 17:23:22 +00:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Reference: sam/archie#1990