77 lines
4.1 KiB
JSON
77 lines
4.1 KiB
JSON
{
|
|
"schema": "fable.research_map.receipt.v1",
|
|
"createdAt": "2026-07-05T14:01:00.000Z",
|
|
"taskLabel": "T3MP3ST-pxpipe-advisory-map",
|
|
"task": "Review T3MP3ST (autonomous red-teaming framework by elder-plinius) and pxpipe (context-to-image token compression by teamchong) for fable-agent applicability",
|
|
"source": {
|
|
"classification": "external-project-review",
|
|
"items": [
|
|
"https://github.com/elder-plinius/T3MP3ST",
|
|
"https://github.com/teamchong/pxpipe"
|
|
],
|
|
"contentVerified": true,
|
|
"liveVerification": "GitHub API metadata + raw README fetched at review time"
|
|
},
|
|
"externalProjectReviews": [
|
|
{
|
|
"repo": "elder-plinius/T3MP3ST",
|
|
"stars": 815,
|
|
"license": "AGPL-3.0",
|
|
"description": "Multi-agent offensive-security meta-harness for autonomous red-teaming",
|
|
"keyFindings": [
|
|
"Uses 'verify-claims' pattern (24/24 green) — every headline recomputed from committed data. Identical to fable-agent's receipt-first philosophy.",
|
|
"Keyless — uses the AI coding agent already on the machine as backbone. Same as fable-agent's model-agnostic pipeline.",
|
|
"90.1% pass@1 on XBEN security benchmark across 104 challenges.",
|
|
"War Room UI + CLI interface for multi-agent kill chain (recon > exploit > report).",
|
|
"AGPL-3.0 license — copyleft, incompatible with MIT codebase. Cannot merge code."
|
|
],
|
|
"relevantPatterns": [
|
|
"verify-claims mechanism (re-derivable measurement from committed artifacts)",
|
|
"Receipt-first claim validation (same as fable-agent)",
|
|
"Keyless agent-agnostic design (same as fable-agent)"
|
|
],
|
|
"constraints": ["AGPL-3.0 license prevents code merging", "Security domain differs from agent harness domain"],
|
|
"safeToVendor": false,
|
|
"decision": "learn-only pattern — verify-claims mechanism validates fable-agent's approach"
|
|
},
|
|
{
|
|
"repo": "teamchong/pxpipe",
|
|
"stars": 2364,
|
|
"license": "MIT",
|
|
"description": "Cut Fable 5 token usage by rendering text context as images — ~59-70% lower token bill",
|
|
"keyFindings": [
|
|
"Renders dense text (code, JSON, tool output) as compact PNGs — ~3.1 chars per image-token vs ~1 char per text-token",
|
|
"Proven on real Claude Code traffic: 10/10 SWE-bench Lite both arms at -65% request size",
|
|
"Lossy for byte-exact content: 13/15 hex strings on Fable 5, 0/15 on Opus — silent confabulations on misses",
|
|
"Optional: profitability gate images only where math wins (calibrated on N=391 rows)",
|
|
"Escape hatch: subagents on non-allowlisted models pass through as text"
|
|
],
|
|
"relevantPatterns": [
|
|
"Token compression for dense pipeline context (repo maps, tool docs, system prompts)",
|
|
"Profitability gate (only image when compression wins) — applicable to fable-agent's runAgent callback",
|
|
"Lossy vs lossless routing — byte-exact content (hashes, IDs) stays text"
|
|
],
|
|
"constraints": [
|
|
"Lossy compression risks silent confabulations on repo map hashes/IDs",
|
|
"Requires rendering pipeline (Pillow/canvas) — add dependency",
|
|
"Fable-5-optimized; may not translate to DeepSeek/Gemma vision capabilities"
|
|
],
|
|
"safeToVendor": true,
|
|
"decision": "advisory — token compression pattern is relevant, but lossy nature requires careful routing"
|
|
}
|
|
],
|
|
"fableAgentAlignment": [
|
|
"T3MP3ST's verify-claims validates fable-agent's receipt-first approach — same philosophy, different domain",
|
|
"pxpipe's token compression could apply to fable-agent's pipeline: the runAgent callback sends repo maps + system prompts + tool docs — all dense text prime for compression",
|
|
"But: repo map contains file paths and structure — if compression is lossy on paths, the model sees wrong structure",
|
|
"Lossless routing (pxpipe's escape hatch) could keep byte-exact content in text and image only what's compressible"
|
|
],
|
|
"skipped": [
|
|
"actual integration or vendoring of either repo",
|
|
"benchmark comparisons",
|
|
"deploy actions"
|
|
],
|
|
"deployAttempted": false,
|
|
"autoPatchProven": false
|
|
}
|