fix(grok): encode session/cancel as a real notification and interrupt steers

[leoisadev1]

Problem

Grok CLI dropped session/cancel because effect-acp encoded notifications as JSON-RPC requests with id: "" and headers. Stop did not stop. The next send queued behind the old prompt.

Mid-turn sends also waited behind the in-flight ACP prompt instead of interrupting like Claude and Codex.

Changes

Outbound ACP notifications now encode as JSON-RPC with no id or headers. RpcClient Interrupt frames are dropped. AcpSessionRuntime.cancel waits for the cancel write. prompt can signal dispatched once the RPC fiber is registered.

Grok mid-turn sends cancel the in-flight prompt, then send the replacement under a prompt epoch and lifecycle semaphore. A failed steer does not discard the live prompt.

Upstream

Adapted from pingdotgg/t3code#9154 session/cancel and steer work. Depends on #214 for grok-build handling and initialize-only runtime.

Shared ACP encoding also applies to Cursor and OpenCode. Cursor adapter tests still pass. No Mastra Codex/Kimi/OpenCode Go change.

Scope

This PR is cancel encoding and Grok steer only.

Assigned Grok ports:

• fix(grok): health check, model selection, and stop all work against the real CLI pingdotgg/t3code#9154 health probe / initialize / grok-build slug (fix(grok): probe health with initialize and allow in-session model changes #214)
• fix(grok): allow model changes in existing threads pingdotgg/t3code#8392 in-session model change (fix(grok): probe health with initialize and allow in-session model changes #214)
• fix(grok): health check, model selection, and stop all work against the real CLI pingdotgg/t3code#9154 cancel encoding, mid-turn steer, prompt dispatch
• fix(grok): improve skills, plans, usage, and turn reliability pingdotgg/t3code#8358 plans, usage, watchdog, permissions, reasoning
• fix(grok): improve skills, plans, usage, and turn reliability pingdotgg/t3code#8358/fix(providers): discover workspace skills everywhere pingdotgg/t3code#9180 Grok skills (separate stacked PR)

Verification

• effect-acp protocol tests: 21 passed, including cancel frames without id/headers.
• GrokAdapter tests: 23 passed, including mid-turn steer cancel and failed-steer keeps original prompt.
• CursorAdapter tests: 19 passed.
• XAiAcpExtension tests: 10 passed.
• Targeted lint and format passed.

Stacked on #214. Leave both open.

Made by Grok 4.6 High in Grok Build via Orca.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
akeru-bot-landing	Building	Preview	Sep 10, 2026 5:55pm UTC

[greptile-apps]

Greptile Summary

This PR corrects ACP notification framing and makes Grok mid-turn steering deterministic by preserving the replacement prompt’s terminal outcome.

Confidence Score: 5/5

Safe to merge.

No outstanding findings remain.

_{Reviews (2): Last reviewed commit: "fix(grok): keep steered turn outcome on ..." | Re-trigger Greptile}

[leoisadev1]

This is Leo's agent. Independent review found that #228's terminal-emission finding also applies to this exact head, 11ccb78b76cfa08c2f3f25a5e709e850f6c78c62.

I ran the 109 focused protocol and adapter tests successfully, then reproduced the omitted case against the shipped grokTurnCompletionForPromptEpoch:

1. Epoch 2 stores { completedStopReason: "end_turn" } while one older prompt remains.
2. Superseded epoch 1 drains with remainingPrompts: 0 and emitTurnCompletion: false.
3. The helper returns emit: undefined instead of flushing the stored current-epoch completion.

The helper is identical in #224 and #228. This reproduction exercises the real helper; it does not claim a full provider-process interleaving. The existing adapter tests do not cover this final no-emission drain.

I requested a scoped repair and committed lifecycle regression from the existing Grok worker, with the fix propagated to #228. That worker is currently weekly-quota blocked, so no fix has executed. Both PRs remain on hold pending the repair and fresh verification.