2026-08-10-pi-coding-agent-extension-tool-call-handlers-can-terminate-a-blocked-batch-without

Extension tool_call handlers can terminate a blocked batch without paying for another model call.

tool_call handlers now return { block: true, reason?: string, terminate?: boolean }. Per docs/extensions.md at v0.84.2: "terminate only applies to a blocked call; the agent stops early only when every finalized result in the batch is terminating." The doc's worked example is exactly the governance case -- intercept bash, and if the command contains rm -rf, return { block: true, reason: "Dangerous command", terminate: true }. The same release contributed by @muyiyr (PR #7715). A parallel terminate: true return from a custom tool's execute() carries the same semantics for structured-output tools that should end the turn.

Channel: tagged-release. Ancestry: Commit 1eb988cfe88fb0ff740ff62583d2f16359f7b6b0 ("feat(agent): allow blocked tool calls to terminate (#7715)", authored 2026-08-06T15:14:48Z). gh api repos/earendil-works/pi/compare/v0.84.1...1eb988cf returned status=behind, behind_by=22 -- ancestor of the stable v0.84.1 tag (prerelease=false).

Operator consequence: Try it if you are building the permission layer Pi deliberately omits. Pi ships no permission popups and no built-in sandbox; the tool_call interception hook is where operators have been implementing their own policy. Until 0.84.1, blocking a call still triggered an automatic follow-up model call -- so a policy that refuses a whole batch burned a round trip explaining itself to the model. Now a wholly-refused batch stops. That makes an extension-layer command policy cheap enough to run on every turn.

Receipt

• https://github.com/earendil-works/pi/blob/v0.84.2/packages/coding-agent/docs/extensions.md