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 uqeu/estelle-cli --skill grounded-generategit clone --depth 1 https://github.com/uqeu/estelle-cliWrote 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/uqeu/estelle-cli/grounded-generate)<a href="https://agentmods.dev/skills/uqeu/estelle-cli/grounded-generate"><img src="https://agentmods.dev/badge/skills/uqeu/estelle-cli/grounded-generate/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/uqeu/estelle-cli/grounded-generate"><img src="https://agentmods.dev/badge/skills/uqeu/estelle-cli/grounded-generate.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 | $0.00000 | $0.00983 |
| Opus 5.5 | $0.00000 | $0.00393 |
| Sonnet 5.5 | $0.00000 | $0.00197 |
| Haiku 4.5 | $0.00000 | $0.00098 |
Grade A, and why
grounded-generate 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 1mo 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 — 51 lines — stays where its author put it; the contents beside it link to each section on GitHub.
grounded-generate
Estelle's anti-hallucination generator — dead center of the thesis. Where verify-gate checks one finished output, this
samples K candidate solutions and K self-tests, executes them against each other, keeps only the survivors, and selects by
consensus — then proves every imported package is real. It's built from generate_candidates / self_consistency
(agent/candidates.py), propose_tests (agent/testgen.py), the real-suite filter run_and_repair_suite /
parse_suite_output, the reproducible patch critic (agent/critic.py), and the two grounding gates — repo-scoped
(GroundingReport.is_grounded) and library-scoped third_party_violations (agent/third_party.py). Grounded in CodeT's
dual-execution consensus (2207.10397), the AlphaCodium flow (2401.08500), and the package-hallucination / slopsquatting
work (2406.10279).
When to run
When generating code you'll actually run — a function, a script, a fix — where a single greedy sample might invent an API or
import a package that doesn't exist. Reach for it especially when the answer pulls in third-party libraries: a hallucinated
package name is the exact supply-chain surface attackers pre-register (slopsquatting), so an unverified import is a security
risk, not just a bug.
Procedure
- Sample K candidate solutions.
generate_candidates(ask, prompt, K)(agent/candidates.py), varying temperature for real diversity and routing the work throughroute()(serve/model_router.py). One greedy sample is a guess; K gives consensus something to measure. - Sample K self-tests.
propose_tests/build_testgen_prompt(agent/testgen.py) writes tests for the ADDED code (added_code), grounded against real repo symbols so the tests themselves can't invent helpers. - Execute the K×K grid. Run each candidate solution against each self-test in a sandbox;
parse_suite_outputdecides a pair "agrees" only when the pytest summary is truly green (no failed/error) — a clean import or partial run never counts. - Execution-filter both sides. Drop every candidate that passes no test and every test no candidate passes — a bad test is as misleading as a bad solution, so CodeT filters both. Only mutually-passing survivors move forward; if nothing survives, the request is unanswerable as posed, not silently answered.
- Select by dual-execution consensus. The candidate that agrees with the most tests and clusters with the most peers wins:
majority/self_consistencyfor convergent answers, withselect_bestbreaking ties toward the grounded candidate (fewest ungrounded APIs). Consensus over execution, not an unverified judge's opinion. - Gate the repo APIs. Run the deterministic grounding gate (
GroundingReport.is_grounded) on the winner — every repo-scoped symbol it references must actually be defined in this codebase; a hallucinated repo API is rejected outright. - Verify every imported package exists (anti-slopsquatting). Run
third_party_violations(agent/third_party.py) over the allowlisted, actually-installed libraries: arequests.fetch()that isn't a real method, or animportof a package that isn't installed, is flagged. Never ship an answer that imports a nonexistent package — that is the slopsquatting attack, not a typo. - Record the catch and gate. Any hallucination caught →
learn_from_grounding/ExperienceStore.record(agent/reflect.py) so the next similar ask recalls the lesson, then hand the survivor toverify-gatebefore it's called done.
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.
- 1mo ago First seen · 51 lines · 0 tokens per session scan A 4a95d8554cb1
grounded-generate is a skill published in the GitHub repository uqeu/estelle-cli (0 stars, last pushed 14d ago), licensed Apache-2.0. It costs nothing until one of its globs matches a file; then it loads 983 tokens. 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 2026-09-12.
Other skills, from other repositories
remove-ai-slops
Removes AI-generated code smells from branch changes or an explicit file list behind regression tests. Use when the user asks to clean up, deslop, or remove AI-slop patterns from recent changes.
review-work
Post-implementation gate review: run manual QA on the real surface yourself, then launch ONE gate reviewer (never a panel) to audit goal, constraints, code quality, security, missed context, and QA evidence. Use before a PR handoff or when the user explicitly asks to review completed work.
include-test-files-that-assert-on-behavior-being-changed-in-decl
When delegating a task affected by this skill, include.
check
Auto-fix lint and formatting, then report types and tests. Run before every commit.
code-guidelines-go
Go 1.24–1.27 coding guidelines for the dimetron/pi-go AI agent runtime. Use this skill whenever writing, reviewing, or refactoring ANY Go code in pi-go. This covers idiomatic style, error handling, concurrency, project layout, testing (table-driven, fuzz, benchmarks, synctest), new stdlib usage, golangci-lint v2…
vhs-e2e-gif
Record a test run, a TUI session, or any terminal command as a GIF with VHS and attach it to a GitHub PR as a release-hosted asset, never a repo commit. Use when asked to record an e2e run, demo a fix on a PR, attach a GIF or screen recording to a pull request, show a test passing visually, or produce a terminal…