[yyjeqhc]

Summary

Reduce tool discovery and response text overhead for the opt-in execution-only MCP CLI surface without changing the canonical Runtime execution, authorization or durable Job lifecycle.

• Keep a focused direct CLI tool inventory (8 tools with a default Project, plus the Project picker without one); discover advanced admitted tools and complete input schemas through read_tool_manifest and call_runtime_tool.
• Compact common CLI text surfaces while preserving exact Project selectors, read revisions, error/recovery evidence and Job continuation calls. Synchronous command head/tail previews adjust only the CLI log-page cursor; server-owned Job state and replay safety remain unchanged.
• Clarify CLI routes and protected edit-path errors; cover normal, pending, truncated, failed and outcome-unknown execution states.
• Extend the local-agent comparison harness with deterministic client-inventory probes, per-call output/usage accounting, and discovery size assertions. Paid live-model E2E was not run during this review.

Review follow-up

• 161ccfe95 fixes discovery size accounting to use actual UTF-8 JSON wire bytes instead of Python's default ASCII-escaped Unicode representation. Added a multilingual regression.

Validation on special

• Rust CLI-related library tests: 216 passed, 0 failed, 1 ignored by configuration.
• MCP tool-description consistency regression: 1 passed.
• Python local-agent metrics fixtures: 5 passed.
• python3 scripts/check_markdown_links.py: 279 files, 1,505 local links, none missing.
• cargo fmt --all -- --check and git diff --check: passed.
• Full workspace production dogfood check: 0 warnings, Cargo exit 0.

No deploy, restart or merge is included.