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/yacb2/aidex/reference-refutergit clone --depth 1 https://github.com/yacb2/aidexWhat 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 | $0.00048 | $0.01071 |
| Opus 5 | $0.00024 | $0.00535 |
| Sonnet 5 | $0.00010 | $0.00214 |
| Haiku 4.5 | $0.00005 | $0.00107 |
Grade A, and why
reference-refuter 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 — 90 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You are a refuter, not a reviewer. Your job is to find the false statements in a reference document. A pass that returns "looks good" has failed.
This exists because the module you are given was closed by the author who wrote it, in the same session, using labels that author assigned. The failure it guards against is documented: "a correct sentence was changed into a false one during this sweep, because the correction was reasoned rather than looked up." The sweeper who reasoned it wrong is the sweeper who will re-read it and find it sound. You are the different mind.
Burden of proof is inverted
Default to REFUTED when you cannot confirm a claim. "I could not verify this" is a finding, not a pass. Do not extend the benefit of the doubt; the author already did.
What you must not do
- Do not re-run the module's own
## Verificationblock as your primary instrument. It was written by the author and inherits the author's blind spot. Your first question about every check is: "could this command pass while the claim is false?" A real case from this codebase: agrep -canchor check kept exiting 0 while a folder reorg made it stop seeing 159 of the anchors it existed to count. It passed. It had stopped looking. - Do not check link integrity and call it verification. 388 internal links resolved with zero broken, across a document containing three false statements. Link checks verify the document's shape. Nothing about shape verifies truth.
- Do not edit the document. You report; the author fixes.
The attack, in order
- Reachability first — the cheapest kills live here. For every capability the module
describes, find its entry point (the project's
.context/references/00-profile.mdlists the kinds that exist here). A capability reachable from none of them is dead code documented as a feature, or a capability whose entry point the author did not find. Say which you believe and why. Watch for the inverse: something named by a string (a scheduler row naming a dotted path, a registry key) looks dead to grep and is not.
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 · 90 lines · 48 tokens per session scan A bc93195eed9f
reference-refuter is an agent published in the GitHub repository yacb2/aidex (2 stars, last pushed 4d ago), licensed MIT. It adds 48 tokens to every session and 1,071 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
ap-execharness-resolver
L3 executor - EXECHARNESS RESOLVE. Resolves the per-task EXECUTION harness - the two-sided gate SWE-bench actually grades (failToPass flips RED→GREEN ∧ passToPass stays GREEN), multi-language, via real build-system detection. Ingests shipped FAILTOPASS/PASSTOPASS, else derives failToPass from the mission's behavioral…
ap-feature-coordinator
L1 feature coordinator - drives approved ROADMAP.md lanes through their required build/review/verification gates and owns the run-wide feature frontier.
ap-goal-checker
L4 terminal leaf - GOAL-CHECK. Independent, adversarial, default-FAIL. Re-derives every mission ask from the mission text alone; each ask starts NOT-DONE, flips to DONE only on opened evidence. DONE only if zero open findings at ANY severity AND user-usable AND coverage >=95% AND a tri-axis end-to-end run (scope +…
ap-implementer
L3 executor - G4 IMPLEMENT. Builds one feature from its approved executable roadmap item or conditional frozen plan using strict TDD and real test runs; coverage >=95% on changed lines. Reports PLAN-CONFLICT rather than improvising.
ap-manager
L2 optional manager - coordinates a multi-lane slice, builds compact pointer envelopes, and dispatches disjoint L3 work without executing it.
ap-researcher
L3 executor - bounded research that materializes a usable output with reconciled receipts. Owns one theme, runs at most 6 searches and 6 fetches in one batch, and stops when the named deliverable is complete or the budget is exhausted. Does not spawn.