Resume an investigation that started in the dashboard's AI chat panel. The user pastes a `cat ~/test-dev/.claude/chats/<id>.md` command (or a URL); read the file, understand what scope was set and which tools were already called, then continue the investigation with full Claude Code capabilities (shell, edit, full reasoning budget). Use when the user references a "chat handoff" or pastes one of those commands without further explanation.
Inject HTTP errors, socket faults, or transport-layer drops on a player session via the harness CLI. Invoke when the user says "inject", "drop a 404", "fault on segments", "every N seconds", "kill the connection mid-body", "throttle", "reset faults", or otherwise wants to mutate the proxy's behaviour for one player. This skill owns the English-to-CLI-flag translation including the ambiguous-phrase tables; resolve disambiguation BEFORE running any harness command.
Wrapper for the harness CLI. Read before invoking any `harness` command — surfaces flag-name traps, output-contract gotchas, and the canonical operator-label propagation flow. Sibling skills (triage, investigate, forensics, fault, shape, finding) all shell out to harness; this is the source of truth for HOW.
Capture a discovery into the .claude/findings/ library, or recall existing findings related to the current investigation. Invoke when the user says "save that", "record this finding", "this is worth remembering", "what do we know about X", "have we seen this before", or after a successful `investigate` / `forensics` run that reached a tagged hypothesis. The library is the project's memory across sessions — write to it when a non-obvious cause has been confirmed or strongly suspected.
Multi-event causal analysis — "why does X keep happening", "compare A vs B", "find the pattern across these failures", "is this related to a change". Gathers events/network/samples + relevant past findings + applicable standards, then dispatches to the `playback-forensics-expert` subagent for hypothesis generation. Invoke when one event isn't enough (use `investigate` for single events) and the question is "why" not "what". Returns a tagged hypothesis plus a suggested test to confirm or refute.
Deep-dive ONE specific event on a player — pull network log, samples, and player state from a focused time window around the event, then state an explicit hypothesis about cause. Invoke when the user says "investigate the X at Y", "why did Z happen at T", "what was happening around <timestamp>", or after `triage` has surfaced a specific event worth chasing. Output is timeline + evidence + hypothesis tagged confirmed/refuted/needs-test.
Set bandwidth cap, delay, packet loss, or pattern-based ramp on a player session via the harness CLI. Invoke when the user says "throttle to X Mbps", "add N ms delay", "X percent loss", "ramp up over 30s", "drop-and-recover pattern", "shape the network", or otherwise wants kernel-level traffic shaping. This skill handles unit normalisation and routes ramp-style asks to the `abr-sweep` procedure.
Quick health check at the *menu* level — either scan recent plays for ones needing attention (no target given, or "what's broken right now"), or drill the named player's current session for state + recent failures. Surfaces 3-6 worst signals with timestamps and a suggested next skill. Invoke when the user says "triage X", "what's wrong with X", "how's X doing", "any plays needing attention", "scan plays", "what's broken right now", or asks for a summary of recent failures. Does NOT drill into individual events (use `investigate` for that).