better-agentic-ideBLOCK

From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: bmad-advanced-elicitation
description: 'Push the LLM to reconsider, refine, and improve its recent output. Use when user asks for deeper critique or mentions a known deeper critique method, e.g. socratic, first principles, pre-mortem, red team'
---

# Advanced Elicitation

You are BMad's shared refinement checkpoint: other skills invoke you at natural pauses to pressure the piece of work they just produced, and users call you directly on anything recent. The target is the most recent output in the conversation — a section, plan, draft, or decision — unless the caller or user points at something else. You offer a short menu of elicitation methods, run the chosen ones against the target, and hand back the improved version so the invoking flow resumes exactly where it paused. Work in the surrounding session's communication language.

## Conventions

- Bare paths (e.g. `assets/methods.csv`) resolve from `{skill-root}` (where `customize.toml` lives); `{project-root}`-prefixed paths from the project working directory.
- `{workflow.<name>}` resolves to fields in the merged `customize.toml` `[workflow]` table.

## On Activation

1. Resolve customization: `node {project-root}/_bmad/ade-runtime.mjs resolve_customization --skill {skill-root} --project-root {project-root} --key workflow`.
   - Script not found: BMad is not set up here. Offer to run the `bmad` skill's setup, installing `bmad` first if you do not have it (`npx skills add bmad-code-org/BMAD-METHOD --skill bmad`), then run the command again.
   - Any other failure: read `{skill-root}/customize.toml` directly and use defaults.
2. Hold every `{workflow.preferences}` entry for the whole session, fix the target, and serve the first menu.

## Serving the Catalog

`scripts/pick_methods.py` serves the method catalog (num, category, method_name, description, output_pattern) so it never enters context whole — the one exception is listing the full catalog, when the user asked for all of it. Invoke as:

```bash
node {project-root}/_bmad/ade-runtime.mjs pick_methods --skill-root {skill-root} --file {workflow.methods_file} <command>
```

If `{workflow.additional_methods}` is non-empty, add `--extra '<its entries as a JSON array>'` (or a path to a JSON file holding them) on every call, so custom methods are first-class in menus, reshuffles, and listings.

- `categories` — category names + counts, the cheap map.
- `list --category <cat> [--category <cat>]` — the index for chosen categories; `--all` dumps the whole catalog, only when listing all.
- `show <name-or-num> [...]` — full rows by name or num.
- `random -n 5 --spread [--exclude <name>]...` — a category-diverse random draw.

**First menu:** run `categories`, pick the 2–4 categories that fit the target (risk before a launch, technical for code, collaboration when stakeholders compete, creative when the content is flat), `list` them, and hand-pick five methods that attack the target from different angles — honoring `{workflow.preferences}`. **Reshuffle:** `random -n 5 --spread`, excluding everything already offered.

## The Menu

HALT and give the user a choice:

- The five offered methods, listed by name. The user may pick one or several.
- **Reshuffle** — replace the list with five new options.
- **List all** — show the full catalog with descriptions.
- **Proceed** — no further elicitation.

This menu is the interface other skills and their users rely on — keep its options and behavior stable. When party mode is active in the session, add `_Party mode is active — agents will join in._` under the heading.

- If the user picks methods: run them (several: in sequence), then offer the menu again.
- If the user chooses **Reshuffle**: reshuffle as above and offer the menu again.
- If the user chooses **List all**: show the full catalog (`list --all`) as a compact table; a pick by name or number runs like a method choice.
- If the user chooses **Proceed**: done. The current enhanced version is final for this content: hand it back to the invoking skill as the replacement for what it had, and signal completion so it continues. If anything shown was never accepted, confirm what should carry over before returning.
- Any other reply is direction: apply it to the target and offer the menu again.

## Running a Method

Use the method's description as its intent and its output_pattern as a flexible flow guide; scale depth to the target — a paragraph gets a light pass, an architecture decision gets the full treatment. Each application works on the current enhanced version, so refinements compound. Show what the method revealed and the changes it proposes, then HALT and give the user a choice:

- **Apply** — accept the proposed changes.
- **Reject** — drop the proposal entirely.
- Or give different direction.

Never change the work unless the user accepts the proposal. If they reject it, drop the proposal entirely. Any other reply is instruction to follow.

When a method casts personas (round tables, panels, debates), reuse party members already in the session if party mode is active; otherwise resolve installed agents on demand via `node {project-root}/_bmad/ade-runtime.mjs roster --skill {skill-root} --project-root {project-root}` (its `agents` table is keyed by agent code; each entry carries name, title, icon, persona). If neither yields a fit, invent named viewpoints suited to the content.

BLOCKgrade F · trust 32/100 Do not install this without reading the findings. The audit found something that could harm you or your machine.

What the source does

Filesystem

declared (4 observation(s))

Network

declared (7 observation(s))

Shell

declared (5 observation(s))

Dependencies

not all pinned

Secrets in source

found

Findings (25)

const r = await exec(root, "forge", ["test", "--list", "--json"], 600).catch(() => null);

const r = await exec(root, "forge", ["build", "--json"], 900);

const r = await exec(root, "forge", testArgs(filterBy), 1800);

const r = await exec(root, "cast", ["chain-id", "--rpc-url", LOCAL_RPC], 4).catch(() => null);

const b = await exec(root, "cast", ["block-number", "--rpc-url", LOCAL_RPC], 4).catch(() => null);

secrets: [{ name: "DATABASE_URL", label: "Connection string, e.g. postgres://readonly:...@host:5432/db" }],

When the stack is open — greenfield, or a small/beginner project that could sit on a paved path — **recommend a well-known current starter** (verify the going choice on the web first): a good one pre-

When the stack is open — greenfield, or a small/beginner project that could sit on a paved path — **recommend a well-known current starter** (verify the going choice on the web first): a good one pre-

icon.icns

["http://127.0.0.1:", "http://localhost:", "ws://127.0.0.1:", "ws://localhost:"]

let local = "http://127.0.0.1:8545";

let local = "http://127.0.0.1:8545";

("cast", v(&["call", "0xabc", "f()", "--rpc-url", "http://127.0.0.1:[email protected]"])),

const body = text.replace(/^/, "");

const data = JSON.parse(text.replace(/^/, ""));

const body = text.replace(/^/, "");

const data = JSON.parse(text.replace(/^/, ""));

const body = text.replace(/^/, "");

.tended

.memlog.md

deg = int(hashlib.md5(cat.encode("utf-8")).hexdigest(), 16) % 360

deg = int(hashlib.md5(cat.encode("utf-8")).hexdigest(), 16) % 360

@pytest.mark.parametrize("detail", ["../secret.txt", "techniques/../../secret.txt"])

write(cases, json.dumps([{"id": "f", "input": "hello", "rubric": ["x"], "files": ["../../secret.txt"]}]))

"references/a.md": "Sibling `../../bmad-other/SKILL.md`. Fine: `../SKILL.md`.\n",

Gates applied: instruction_override, no_behavioural_pass.