[Sequela02]

Part 7 of the move to Effect became fixing the places in the engine's loop that do not hear a stop (debated with GPT-6 Astra and GPT-6.1 Sol, and measured; the owner approved fixing them directly rather than moving loop.ts to Effect). This is a prompt stopped while it waits to start. Implemented by Sol (gpt-6.1-sol), who found it reading loop.ts; reproduced and checked by Claude, who added the print-mode case.

Defect

prompt() (packages/core/src/loop.ts) waits for the session's startup (session_start handlers) and for a manual compaction before it owns a controller; abort() stopped only what was already running. A prompt stopped while it waited then started with a fresh controller and ran. Reproduced on dev: a session_start handler held, a prompt sent, abort() awaited, the handler released; the run ended completed and its tool wrote its file.

Change

• A prompt is admitted before its first wait (admitting); abort() and dispose() stop waiting admissions, and the prompt rechecks after each wait. A stopped prompt is refused: it rejects with the stop's reason before recording a request, so no run is added and the previous run's outcome stays. When it runs, its admission becomes the run's controller.
• A prompt sent after the stop runs as before.
• Print mode (print.ts) takes that refusal, after the operator's Ctrl+C or a termination, as the stop rather than an error: it ends "without a run" and exits 1, as a stop does, instead of printing the rejection with its stack. The shell already shows a rejected prompt's reason as a note.

Checks

• New packages/core/tests/stop-admission.test.ts: a prompt stopped while waiting for startup, and one stopped while waiting for a manual compaction, run nothing, and a prompt after the stop runs. Both fail on dev (the stopped prompt's tool wrote its file) and pass with the change; checked by Sol and again by Claude.
• With compaction-session, conformance and receipt-command tests: all pass.
• bun run check (Sol, Windows): root 264 passed; engine 333 passed, 3 skipped; types and lint clean. The budget failed on shell memory on this machine (120.4 MB, limit 115 MB); CI decides.

Not covered by a test: the print-mode case (a stop by signal before the request starts).