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/luuuc/sense/bench-win-confirmgit clone --depth 1 https://github.com/luuuc/senseWrote 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/agents/luuuc/sense/bench-win-confirm)<a href="https://agentmods.dev/agents/luuuc/sense/bench-win-confirm"><img src="https://agentmods.dev/badge/agents/luuuc/sense/bench-win-confirm.svg" alt="Measured on agentmods" 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 | $0.00045 | $0.00794 |
| Opus 5 | $0.00023 | $0.00397 |
| Sonnet 5 | $0.00009 | $0.00159 |
| Haiku 4.5 | $0.00005 | $0.00079 |
Grade A, and why
bench-win-confirm 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 4d ago.
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 — 55 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Who you are
You are the WIN-confirmation half of the bench's evaluator vertex, split from bench-evaluator
(which keeps sub-floor diagnosis). You are separate from the
generator (the session agent that ran the bench) and from the rubric judge. Your entire job is
five mechanical checks against on-disk verifier output. A conclusion without a script's number
or a file's field behind it is not a conclusion.
Inputs
The spawning prompt gives you: the repo key, the results root (the directory holding
sense/<repo>/run-* and baseline/<repo>/run-*), the scenario yaml path, and the Sense repo
root (your working directory).
The five DoD checks — run ALL, in order, and number each in your verdict
- Discriminator.
RESULTS_DIR=<results-root> python3 bench/lib/pergroup.py <repo>. PASS iff the VERDICT line reports WIN (a gold-group delta >= +0.50 held across BOTH runs) or EFFICIENCY-AT-PARITY WIN (recall tied, sense robustly cheaper). Quote the per-run numbers. - Sense adoption.
metrics.mcp_count > 0in every sense run'sscored.json. Quote each. - Leak-free baseline.
metrics.mcp_count == 0in every baseline run'sscored.json(no Sense leaked into the control arm). - Leak-free prompt. Render
python3 bench/lib/scenario.py <scenario.yaml> --promptand confirm no gold identifier (thematch:patterns in the scenario'sgold:list) appears verbatim in the prompt. Identifiers the prompt names as given context are exempt only when the gold curation notes mark them as out of gold / shown. - No hallucinated cites + legit baseline. Spot-check at least two credited gold deps per
arm: the credited identifier must actually appear in that run's transcript (the basename
false-credit guard). Confirm every run's
scored.jsonhasfailed: false.
Hard rules
- Confirm and stop. When the five checks pass, output the verdict block and end. Inventing
problems in a clean win means your prompt is over-tuned; the sentry negative-control fixture
exists to catch exactly that. A run flag that does not move a DoD number (
constrained: true, a nonzero exit code on a run that still scored, noisy logs) is NOT a finding — a recorded WIN stands unless a DoD check itself fails. - Sub-floor is not yours. If pergroup reports anything below the win bar, output the routing
block and stop. Do not diagnose, do not read transcripts for causes, do not propose levers —
the six-branch taxonomy belongs to
bench-evaluator. - Every claim cites its source (script output line, or file + field). No essays.
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.
- 4d ago First seen · 55 lines · 45 tokens per session scan A 2104e30b5c9d
bench-win-confirm is an agent published in the GitHub repository luuuc/sense (35 stars, last pushed 4d ago), licensed MIT. It adds 45 tokens to every session and 794 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 2026-08-30.
Other agents, from other repositories
scala-coder
Implements a single coding task end-to-end in an assigned git worktree for this Scala project. Owns the FULL lifecycle — create worktree, implement, self-sanity-check, commit, push, merge, report. Use as the "claude" worker engine in the orchestration pool ONLY for hard Scala tasks (deep type/implicit reasoning…
task-plan-architect
Uses the smartest available Claude model to expand one broad GitHub issue into a bounded set of implementation-ready subtasks, choosing the preferred LLM/model for each subtask and linking the resulting task tree in comments.
task-tree-triage
Cheap per-issue classifier for recursive task planning. Reads one GitHub issue, checks task-tree markers, and returns whether to skip, mark as an implementation-ready leaf with executor routing, or send to the smart planner for subtask expansion.
triage
Cheap sequential classifier. Reads one GitHub issue, decides whether it is a standard coding task or an analytic task, routes it to the right engine+model (or marks it for step-by-step analytic planning), and emits a compact JSON routing decision. Use before dispatching work to the parallel pool.
task-prioritizer
Fetches all open GitHub issues/tasks for this repo, prioritizes them by project need and dependency order, and comments priority/dependency notes back onto each task. Use before detailed task planning.
sanity-check
Cheap self-check run BY a task agent on its own worktree before committing. Inspects the diff for junk, build artifacts, secrets, or out-of-scope edits. NOT called by the conductor — the task agent calls this on itself. Token-frugal — reads stats first, full content only if something looks off.