When the task is build, run, benchmark, test, or debug:

• Proceed without asking for permission if the command is already allowed by settings.
• You may iterate on configure/build/test loops and inspect logs until you have a concrete result. Stop and summarize if the same failure repeats or if progress requires source changes not already requested.
• If the build fails because source files are missing, report the missing files and stop. Do not create source files to fix the build unless I explicitly ask for implementation work.
• Prefer running the smallest useful command first.

When the task is design, API shaping, refactoring, or architecture work:

• Do not edit files immediately.
• First give a short plan and list the exact files you want to inspect or change.
• After giving the plan, stop and wait for my approval before making any edits.
• Show me the relevant file or diff step by step.
• Do not touch files that were not listed without asking first.
• Make changes one file at a time.
• After each file, summarize what changed and why before moving on.

Always:

• Ask before commit, push.
• Announce before creating new source, doc, script, config, or result files.
• Keep generated files and build artifacts out of commits unless I explicitly ask for them.

Local command style

For gate validation and benchmarks, do not use large multi-command Bash scripts.

Prefer:

• one direct command per run
• full paths instead of shell variables
• stdout/stderr redirected to local/*.stdout and local/*.stderr
• separate grep, cat, or tail commands for inspection

Avoid unless explicitly approved:

• shell variables like BIN=..., MODEL=..., RC=$?
• command substitution like COUNT=$(...)
• arithmetic expansion like $((END-START))
• inline conditionals like [ "$A" = "$B" ] && ...
• timing wrappers using START=$(date +%s) / END=$(date +%s)
• grouped benchmark scripts that run many commands in one Bash call

A plan is not permission to edit files. If source or docs will change, stop after the plan and wait for explicit approval.