Use repository skills for detailed procedures. Keep project-wide policy here and task-specific commands in the skill that owns them.
Use inet-agent-orchestration for nontrivial work with independent evidence lanes, production C++/NED/MSG changes, unexplained runtime behavior, standards-to-implementation analysis, statistical analysis, or independent regression/review. Keep simple lookups and obvious one-file edits in the root thread.
- Use the canonical path
/home/user/omnetpp_ws/inet-skills/.agents/.... - For the first skill read, invoke the read with the required path access/escalation immediately; do not probe the symlink first with
ls,wc, orrealpath. - If the sandbox rejects the command before execution, retry the identical read using required path access.
- Use the canonical path
/home/user/omnetpp_ws/inet-skills/.codex/.... - For the agent definition read, invoke the read with the required path access/escalation immediately; do not probe the symlink first with
ls,wc, orrealpath. - If the sandbox rejects the command before execution, retry the identical read using required path access.
- Before changing anything under
src/inet/, useinet-architectural-requirementsto check sealing, architecture, naming, exception-ledger, audit, and review requirements. - Use Cmdenv for automated and reproducible runs. Use Qtenv only for interactive inspection or when requested.
- Use command-line overrides for temporary logging, tracing, capture, and result recording; do not edit
omnetpp.inionly to enable diagnostics. - Start investigations with one configuration and one run/seed. Expand only when the task requires a campaign or the narrow case is understood.
- Base claims about delivery, loss, retransmission, scheduling, or protocol behavior on logs, captures, event logs, results, source inspection, or debugger evidence.
- Use debug mode for every INET/OMNeT++ build, simulation, and test run, with matching debug runners and libraries (
MODE=debug,inet --debug,-m debug,opp_run_dbg, andlibINET_dbg.soas applicable). Never build or execute release-mode artifacts. Use-j$(nproc)for parallel builds unless the user requests otherwise. - For code changes, run only unit, module, and fingerprint tests that are directly related to the changed paths, symbols, or behavioral contracts. Record that mapping and invoke an explicit filter; never run an unfiltered or broader suite. If no directly related test can be identified, report the coverage gap instead of broadening the test selection.
- Never update fingerprint CSV files without explicit user approval after explaining the changed trajectory.
- Report reproducible commands with their working directory, configuration, run/seed, build mode, exit status, and artifact paths when applicable.
Unless the user directly requested the documentation change, ask before editing AGENTS.md or a skill to record a lesson. Propose the target, intended text, and why it is reusable. Store project-wide policy here, workflows in the owning SKILL.md, and detailed material in that skill's references/; do not persist one-off facts, guesses, logs, or guidance already documented nearby.