Final Tool Response Budget Implementation Plan

1. Establish the shared contract

Add structured persisted-output metadata to tool results and response info.
Correct the generic fallback writer so it returns the path and committed byte count.
Separate Shell and MCP persistence triggers from their model previews without changing rich display output.
Add focused tests for primary/fallback persistence, preview shape, Unicode boundaries, and metadata propagation.

Add a shared utility that measures model-facing tool-response text, reuses persisted paths, persists unpersisted large output once, and enforces a final no-I/O hard cap while preserving the enter_plan_mode lifecycle policy outside the output budget.
Preserve response and part order, media parts, error fields, and function response identity.
Recompute content length and cover single-large, many-large, mixed output/error/text, persistence failure, disabled budget, and already-persisted cases.

Replace the scheduler's marker-based batch offload with the shared finalizer and run it on both sides of PostToolBatch.
Finalize interactive executable and synthetic responses together before recording and submission.
Finalize the whole headless turn outside per-call schedulers before recording and submission.
Finalize ACP calls before transcript recording while keeping immediate display events unchanged.
Finalize agent and speculative aggregates before model-facing emission/history mutation.
Add the tool-response-only no-I/O send guard.

Run package-local unit tests for every changed runtime.
Run the E2E baseline and fixed scenarios for parallel Shell output in headless and interactive modes; exercise ACP where the local harness permits it.
Run build, typecheck, and lint.
Inspect every metadata producer and consumer to prove the field is populated and cannot leak into external schemas.
Perform open-ended diff audits until two consecutive passes find no actionable issue.