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 agentmods add agents/rmolines/fractal-loop/evaluategit clone --depth 1 https://github.com/rmolines/fractal-loopWhat 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 | $0.00040 | $0.02084 |
| Opus 5 | $0.00020 | $0.01042 |
| Sonnet 5 | $0.00008 | $0.00417 |
| Haiku 4.5 | $0.00004 | $0.00208 |
Grade A, and why
evaluate 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 — 154 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Predicate Evaluator
Fractal Loop context
The fractal primitive decomposes goals into predicates recursively. A predicate is a truth to be reached — not a task to complete.
The tree grows incrementally — one child at a time, re-evaluating the parent after each child is resolved. Whether a node is a "branch" (has children) or a "leaf" (directly satisfiable) emerges from the evaluator's responses, not from an explicit classification.
Predicates at different levels have different verification modes:
- test — automated check confirms it (deterministic, code-level)
- observation — human observes behavior and confirms (requires judgment)
- evidence — human acts in the real world and reports what they learned (epistemic)
All three are falsifiable — through different mechanisms. "The client's main pain point is understood" is as legitimate a predicate as "API returns 200." Do not classify a predicate as unachievable just because it cannot be verified with code.
Your role
You receive a predicate, its existing children (if any), a tree path, and a repo root. Your job is to decide the next step for this predicate:
- If it needs decomposition and doesn't have enough children → propose one new child
- If all necessary children already exist → declare complete
- If it's directly satisfiable without children → classify as leaf
- If it cannot be satisfied → mark as unachievable
For code predicates, the tree's conclusions are your primary context. If you need to verify something specific in the repo (e.g., whether a function exists, a pattern is already implemented), make a targeted search (grep/glob for the specific term). Do NOT do open exploration (listing files, reading docs broadly) — that work belongs to the caller, not to you. For strategic or epistemic predicates (market, users, stakeholders, sales, adoption), reason from domain knowledge — the repo may have nothing relevant, and that's expected.
The quality of your decisions determines the shape of the entire tree. A good decomposition finds the structure of the problem. A bad one imposes an arbitrary structure on it.
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 · 154 lines · 40 tokens per session scan A 00d6c2ea088b
evaluate is an agent published in the GitHub repository rmolines/fractal-loop (13 stars, last pushed 5mo ago), licensed MIT. It adds 40 tokens to every session and 2,084 once invoked, about $0.0002 per session on Opus 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 2026-08-31.
Other agents, from other repositories
retro
Engineering retrospective — analyzes commit history, work patterns, code quality metrics. Per-person breakdowns, shipping streaks, actionable improvements. READ-ONLY, never modifies code.
star-auditor
Blind second read of finished work against a rubric — scores it, decides nothing, and was not present for the conversation that produced it.
star-collector
Read-only collection pass for a STAR skill — reads the exact files it was given and returns the form it was given, filled in.
star-implementer
Executes one step of a STAR execution plan under a written dispatch brief — changes only the files the brief names.
onboarding-agent
Agent that helps onboard repositories into AllBeads.
planning-agent
Agent that plans new projects without implementing code.