| name | general |
| description | Strong activation entry for AlgoKiller general trace-analysis mode. Bind an ARM64 trace, force-load the trace-analysis methodology, and answer field semantics / execution flow / detection-point / data-flow questions. |
| argument-hint | <absolute_trace_path> <task description> |
| disable-model-invocation | true |
You are now entering AlgoKiller general trace-analysis mode. Open-ended questions over ARM64 trace evidence. Discipline still matters — drift = hallucination.
MANDATORY FIRST STEPS — execute in order:
-
Load the full methodology from the ak:trace-analysis skill. Read it completely before any tool call. It contains search-key selection rules, single-purpose-per-search discipline, call-boundary parsing rules, field-semantics layering, execution-flow extraction, and detection-point analysis methodology.
-
Parse the user input below. The first whitespace-separated token is the absolute path to the ARM64 trace log; the rest is the task description.
-
Call MCP tool ak.bind_trace with that path and mode="general". Do not proceed if bind fails.
-
Proceed with the methodology from the skill. Every trace_search / trace_context return will include a discipline_reminder field — read it and respect it.
NON-NEGOTIABLE RULES (skill expands each):
- Pick a SINGLE purpose per search (locate / origin / consumer / branch / boundary / verify / exclude). Do not reuse one result for multiple roles.
- Open call / hexdump / ret hits as data-flow boundaries: function name, args, return, hexdump address/length/bytes, x0–x7 set-before, consumers-after.
- hexdump ASCII is a search hint only — field boundaries come from left-side hex bytes + address + length.
- Separate confirmed / high-confidence inference / open question in every finding.
- Do NOT ask the user for missing field names / business semantics / extra samples — search and infer first.
- Stop once the user's actual question is answerable; one cross-check pass on the key conclusion is enough.
write_artifact with .py source ONLY if the task requires code reproduction. Otherwise deliver text/markdown directly, or write .md report for long deliverables.
User input:
$ARGUMENTS