Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add knewstimek/agent-tool --skill agent-tool-reversinggit clone --depth 1 https://github.com/knewstimek/agent-toolWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/knewstimek/agent-tool/agent-tool-reversing)<a href="https://agentmods.dev/skills/knewstimek/agent-tool/agent-tool-reversing"><img src="https://agentmods.dev/badge/skills/knewstimek/agent-tool/agent-tool-reversing/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/knewstimek/agent-tool/agent-tool-reversing"><img src="https://agentmods.dev/badge/skills/knewstimek/agent-tool/agent-tool-reversing.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00106 | $0.01325 |
| Opus 5.5 | $0.00042 | $0.00530 |
| Sonnet 5.5 | $0.00021 | $0.00265 |
| Haiku 4.5 | $0.00011 | $0.00133 |
Grade A, and why
agent-tool-reversing scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 96 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Reverse engineering with agent-tool
Order of work and judgment calls for the analyze (static), memtool (live
memory) and debug (DAP debugger) tools. Parameters are not repeated here; read
them from the server so they match the installed version:
agent_tool_help(topic="analyze") index of every operation
agent_tool_help(topic="analyze/decompile") one operation in full
agent_tool_help(topic="memtool") agent_tool_help(topic="debug")
If analyze is not exposed as a direct tool, go through toolbox:
toolbox(operation="describe", tool="analyze", compact=true, tool_operation="decompile"),
then toolbox(operation="call", tool="analyze", arguments={...}).
1. Orient, once per binary
pe_info/elf_info/macho_info: architecture, image base, sections, imports.entropy,overlay_detect: a code section above 7 bits/byte is packed or encrypted. Static analysis of it is meaningless; read the unpacked code from the running process with memtool instead.- Know what names you have. A PDB beside a PE, embedded DWARF and IL2CPP
metadata are applied automatically. Output full of
FUN_...means no debug info: navigate by strings, imports and RTTI (vtable_scan).
2. Find the code from something observable
Do not read from the entry point. Start from evidence:
| You have | Call |
|---|---|
| A message, path or format string | xref target_text="..." |
| An API the code must call | IAT slot VA from pe_info, then xref target_va= |
| A magic number, error code, packet id | instruction_search immediate=, add call_target= to see which calls receive it |
| A struct field | xref field="Class::member" (debug types) or instruction_search displacement= |
| A C++ class | vtable_scan, then rtti_dump and struct_layout on its vtable |
| A crash or trace address | function_at |
3. Read the code
- Search and xref hits already name their function (
; in sub_140008070+0x47): use that start.sub_<hex>andFUN_<hex>are generated names, not symbols; pass the hex address (va="0x140008070"). - A bare address (crash, trace, pointer) goes through
function_atfirst. Build only onconfidenceexact or high;heuristic-misalignedis a wrong start. decompileis the default way to read. Pass related functions together (up to 16 in oneva, comma-separated) instead of one call each; symbols work where addresses do (va="Player::TakeDamage").disassemblewhen exact instructions matter: patch sites, byte signatures (show_bytes=true), flags and calling-convention details, or when the C looks wrong.call_graphbefore decompiling neighbours blindly;type_infofor a type the C names but did not print.- Imports print by name (
CreateFileW(...)), also delay-loaded ones. An import by ordinal is named from the DLL beside the binary or in the Windows system folder;Ordinal_<n>remains when that DLL is not on this machine (pe_infoshows which DLL it is). - When decompile struggles:
halt_baddata(), "flow out of bounds" or "Flow exceeded maximum allowable instructions" mean obfuscation. Establish the real branch directions (debugger trace or disassembly) and pass them asassume_branches.- C that stops right after a call that should return:
noreturn_verify=true. - A timeout: retry that one function with a larger
timeout_sec.
- To search across every function, build a corpus once with the CLI
(
agent-tool decompile-all, run in the background through procexec) and grep the JSONL, rather than decompiling function by function.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 96 lines · 106 tokens per session scan A d31ff78173b6
agent-tool-reversing is a skill published in the GitHub repository knewstimek/agent-tool (26 stars, last pushed yesterday), licensed MIT. It adds 106 tokens to every session and 1,325 once invoked, about $0.0004 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-11-15.
agent-tool-reversing compared with skills from other repositories
| Mod | Type | Stars | Last push | Licence | Scan | Tokens | Works with |
|---|---|---|---|---|---|---|---|
| agent-tool-reversingknewstimek/agent-tool | Skill | 26 | yesterday | MIT | A | 106 | Claude Code, Codex |
| systematic-debuggingjnMetaCode/superpowers-zh | Skill | 8.5k | today | MIT | A | 24 | Claude Code |
| systemic-issue-triageGentleman-Programming/gentle-ai | Skill | 9.2k | today | MIT | A | 57 | Claude Code, Codex |
| issue-root-resolutionGentleman-Programming/gentle-ai | Skill | 9.2k | today | MIT | A | 45 | Claude Code, Codex |
| diagnosing-superpowersjnMetaCode/superpowers-zh | Skill | 8.5k | today | MIT | A | 111 | Claude Code |
| rdd-defect-workflowGentleman-Programming/gentle-ai | Skill | 9.2k | today | MIT | A | 38 | Claude Code, Codex |
| memorix-troubleshootingAVIDS2/memorix | Skill | 938 | 5d ago | Apache-2.0 | A | 36 | Claude Code |