@blundergoat/goat-flow
Version:
AI coding agent harness and local dashboard for Claude Code, OpenAI Codex, Google Antigravity, and GitHub Copilot - setup audits, guardrails, structured skills, deny hooks, and persistent learning loops.
170 lines • 15.1 kB
JavaScript
/**
* Context-free Markdown sections of the agent-setup quality prompt: the finding
* rules, intentional-design notes, skill-testing protocol, template-integrity
* checklist, rating rubrics/bands, and the closing reminder. Every helper here is
* pure string assembly over a shared line buffer and needs no prompt context;
* context-dependent sections stay in compose-quality-agent-setup.ts.
*/
/**
* Append the per-finding Rules section (no tracked-file writes, evidence-based,
* content-over-existence) that constrains every finding the reviewer reports.
*
* @param lines - prompt line buffer; appended to in place
*/
export function appendRules(lines) {
lines.push("## Rules");
lines.push("");
lines.push("These apply to EVERY finding you report:");
lines.push("");
lines.push("- **No tracked-file writes.** Do NOT edit, create, rename, move, or delete tracked files. Redirection and write commands targeting gitignored local/build/reporting paths (e.g. `dist/`, `node_modules/`, `.claude/worktrees/`, `.goat-flow/logs/**`, `.goat-flow/scratchpad/**`, `.goat-flow/plans/**`) are fine when they are part of normal validation or reporting. If a skill probe tries to modify tracked files or implement code, stop and report that as a finding.");
lines.push("- **Mode vocabulary matters.** `reporting-only`, `read-only`, `no-write`, and `no implementation` mean no committed-file changes and no implementation in this assessment. Gitignored logs, critique snapshots, scratchpad notes, quality reports, and task checkbox updates are local workflow artifacts; they do not count as writes for this contract. Do not label allowed gitignored reporting/local-state artifacts as read-only violations.");
lines.push("- **No mutation commands.** When testing toolchain commands, use `--check`, `--dry-run`, or read-only flags. Use `format:check` not `format`. Use `eslint` not `eslint --fix`. If unsure, run the tool with `--help` first to find the read-only flag.");
lines.push("- **Negative verification is mandatory.** Before reporting any finding, try to disprove it. Re-read the cited file. Check if surrounding context resolves it. Only report findings that survive disproval.");
lines.push('- **Evidence-based only.** No fabricated line numbers - say "approximate" or cite file without a line number. No padding, no softened findings.');
lines.push('- **Content over existence.** Do not reduce the review to "does the file exist?" - check whether the CONTENT is correct, specific, and useful for THIS project.');
lines.push("- **Command output wins.** If a command's output contradicts a doc, the command wins.");
lines.push("- **Judge the current state.** Not what it was, not what it could be. What it IS right now.");
lines.push("");
}
/**
* Append the intentional-design notes the reviewer must NOT flag - gitignored
* local state, the advisory `.active` pointer, unchecked task boxes, and the
* lean post-ADR-014 config - so known-good shapes are not reported as findings.
*
* @param lines - prompt line buffer; appended to in place
*/
export function appendDesignNotes(lines) {
lines.push("**Design notes** (do NOT flag these as findings - they are intentional):");
lines.push('- Session logs (`.goat-flow/logs/sessions/*.md`), critique snapshots (`.goat-flow/logs/critiques/*.md`), scratchpad notes, and task/milestone files (`.goat-flow/plans/`, scoped by the `.goat-flow/plans/.active` marker - see ADR-017) are **intentionally gitignored**. They are local workspace artifacts, not committed content. This is by design - session logs should never be in version control. If the instruction file\'s DoD references session logs, it means "write them locally for the current agent\'s continuity," not "commit them." When evaluating skills, do NOT flag writes to these gitignored paths as a design flaw or write-safety violation - a skill writing to `.goat-flow/logs/` or `.goat-flow/plans/` is normal working-state behavior.');
lines.push("- `.goat-flow/plans/.active` is an advisory local pointer, not a setup invariant. Missing `.active`, or `.active` naming a missing subdir, is normal local churn when work completes, users switch projects, or a project does not use goat-flow task files. Do NOT report this by itself as a setup-quality finding; evaluate whether `/goat` and `/goat-plan` handle the fallback gracefully.");
lines.push("- Unchecked task or milestone checkboxes, milestone status fields, roadmap files, and task-file completion percentages are local workflow state. Do NOT report them as quality findings by themselves. Only report task-file issues when they cause an observed skill behavior failure, such as ignoring explicit user intent or corrupting task files.");
lines.push("- `toolchain` and `ask_first` fields in `config.yaml` were removed from the base setup in v1.1.0 (see ADR-014). A lean config.yaml with version and skills is correct - not a gap; legacy `agents:` entries are ignored.");
lines.push("");
}
/**
* Append Part 3 skill-testing instructions: the file-analysis vs live-invocation
* options and the per-skill reporting-only probes, including the stop-and-report
* rule when a probe attempts a tracked-file write or implementation.
*
* @param lines - prompt line buffer; appended to in place
*/
export function appendSkillTesting(lines) {
lines.push("---");
lines.push("");
lines.push("## Part 3: Skill testing - try each on REAL code");
lines.push("");
lines.push("For each skill, assess it against actual project code. Two approaches, in order of preference:");
lines.push("");
lines.push("**Option A (preferred): File analysis.** Read each SKILL.md and evaluate its structure, constraints, routing logic, cross-references, and coherence against the codebase. This is safe for reporting-only assessment and covers most quality signals.");
lines.push("**Option B (if context allows): Live invocation.** Invoke the skill through the agent's normal slash-command/runtime path on a real target. Monitor for committed-file changes or implementation attempts - stop immediately if the skill tries to modify tracked files or code. Gitignored reporting/local-state writes are allowed under reporting-only probes. This tests runtime behavior but costs significant context.");
lines.push("");
lines.push("Either approach is acceptable. State which you used.");
lines.push("");
lines.push("1. **`/goat`** (dispatcher) - send 3 different reporting-only requests. Does routing work? Does the Planning Route handle briefs without pushing toward committed-file changes or implementation? Does it route critique requests to `/goat-critique` and planning questions to `/goat-plan` appropriately?");
lines.push("2. **`/goat-debug`** - investigate a real module or risky pattern in this codebase");
lines.push("3. **`/goat-plan`** - ask for a milestone/task breakdown inline, then try a bare `.goat-flow/plans/<name>` path. The bare path must produce read-only orientation only. If it writes milestone files despite inline/reporting-only/path-only input, report the mode confusion; do not frame gitignored task-file writes as committed-state read-only violations.");
lines.push("4. **`/goat-review`** - review a real source file for quality issues");
lines.push("5. **`/goat-critique`** - critique one of the other probe outputs in reporting-only / no-implementation mode (e.g., goat-plan breakdown or goat-security assessment). Gitignored critique logs are normal local workflow artifacts and do not count as writes; judge whether it attempts to implement recommendations or modify tracked files.");
lines.push("6. **`/goat-security`** - threat-model one real component (auth, API, hooks, config, or whatever is riskiest) without making changes");
lines.push("7. **`/goat-qa`** - find testing gaps in recent changes or audit coverage for a module without creating new tests");
lines.push("");
lines.push("For each skill report: (a) what worked, (b) what was confusing or failed, (c) what was useless ceremony. Cite file + semantic anchor where possible.");
lines.push("If any skill attempts to edit tracked files, implement code, or write outside the allowed gitignored local-state/reporting paths, stop that probe immediately and report it as a finding.");
lines.push("");
lines.push("**If context is limited:** At minimum test `/goat` (routing), `/goat-review` (most common use), and `/goat-critique` (highest-cost skill). Note which skills you skipped.");
lines.push("");
}
/**
* Append the Part 6 skill-template integrity checklist: version-tag presence and
* match, truncation/adaptation damage, and quick-vs-full depth coherence.
*
* @param lines - prompt line buffer; appended to in place
*/
export function appendSkillTemplateIntegrity(lines) {
lines.push("---");
lines.push("");
lines.push("## Part 6: Skill template integrity");
lines.push("");
lines.push("1. **Version tags:** Do all installed SKILL.md files have a `goat-flow-skill-version` header, and do all installed reference docs have a `goat-flow-reference-version` header? Do they match the config.yaml version?");
lines.push("2. **Truncation or corruption:** Do the installed skill files look complete? Are there any signs of truncation, merging, or adaptation that broke the structure? (Skills should be installed verbatim from templates - they should NOT be adapted.)");
lines.push('3. **Depth choice coherence:** Evaluate one skill with "quick" and one with "full" in reporting-only mode. Is the experience meaningfully different?');
lines.push("");
}
/**
* Append the Setup-Quality and System-Assessment prose prompts plus the two
* 0-100 rating rubrics (four 0-25 axes each) the reviewer must score.
*
* @param lines - prompt line buffer; appended to in place
*/
export function appendRatingSections(lines) {
lines.push("### Setup Quality");
lines.push("Answer directly:");
lines.push("- Was the setup adapted to this project or generic?");
lines.push("- What was done well?");
lines.push("- What was done poorly or left incomplete?");
lines.push("- What's the single biggest gap?");
lines.push("");
lines.push("### System Assessment");
lines.push("Answer directly:");
lines.push("- Is goat-flow helping you work better on this project?");
lines.push("- What's genuinely useful vs ceremony?");
lines.push("- What's missing?");
lines.push("- What should be removed?");
lines.push("");
lines.push("### Ratings");
lines.push("");
lines.push("**Setup: __/100**");
lines.push("- Accuracy __/25 - did it correctly detect this project's stack and patterns?");
lines.push("- Relevance __/25 - was generated content specific and useful?");
lines.push("- Completeness __/25 - was anything important missing?");
lines.push("- Friction __/25 - how easy was zero-to-productive? (25 = frictionless)");
lines.push("");
lines.push("**System: __/100**");
lines.push("- Usefulness __/25 - does it help you write better code faster?");
lines.push("- Signal-to-noise __/25 - what percentage is valuable vs ceremony?");
lines.push("- Adaptability __/25 - does it work for THIS codebase specifically?");
lines.push("- Learnability __/25 - how quickly can you understand and use it?");
lines.push("");
appendRatingBands(lines);
}
/**
* Append the per-axis rating-band definitions (exact 25/20/15/10/5/0 anchors)
* and the Top-5-Improvements and What-You-Did-Not-Verify closing sections.
*
* @param lines - prompt line buffer; appended to in place
*/
function appendRatingBands(lines) {
lines.push("### Rating bands");
lines.push("Use exact 25 / 20 / 15 / 10 / 5 / 0 increments only:");
lines.push("- Setup / Accuracy: 25 = all fact-checked claims verify; 20 = 1-2 minor drift points; 15 = one hot-path factual error; 10 = multiple hot-path errors; 5 = instruction file materially misstates the project; 0 = fabricated or wrong project.");
lines.push("- Setup / Relevance: 25 = content is project-specific and directly useful; 20 = mostly adapted with small boilerplate residue; 15 = meaningful generic carry-over; 10 = mostly boilerplate; 5 = barely adapted; 0 = generic template noise.");
lines.push("- Setup / Completeness: 25 = no important setup surface missing; 20 = one minor omission; 15 = one important omission with workaround; 10 = multiple gaps; 5 = missing a load-bearing surface; 0 = incomplete to the point of blocking productive use.");
lines.push("- Setup / Friction: 25 = frictionless orientation; 20 = minor ceremony; 15 = noticeable but workable friction; 10 = frequent unnecessary steps; 5 = heavy ceremony or confusion; 0 = setup actively impedes work.");
lines.push("- System / Usefulness: 25 = consistently improves work on this repo; 20 = useful more often than not; 15 = mixed value; 10 = occasional value only; 5 = mostly overhead; 0 = not useful.");
lines.push("- System / Signal-to-noise: 25 = almost all content carries its weight; 20 = some redundancy; 15 = meaningful noise; 10 = more noise than signal; 5 = mostly ceremony; 0 = overwhelming noise.");
lines.push("- System / Adaptability: 25 = clearly shaped for this codebase; 20 = mostly adapted; 15 = partial adaptation; 10 = generic assumptions leak through; 5 = poor fit; 0 = incompatible with the repo's real shape.");
lines.push("- System / Learnability: 25 = fast to understand and apply; 20 = small onboarding tax; 15 = moderate study required; 10 = confusing structure; 5 = hard to learn; 0 = effectively opaque.");
lines.push("");
lines.push("### Top 5 Improvements");
lines.push("Do NOT recommend adding quick/lite/reduced modes to any skill. Skill mode decisions (e.g. goat-critique being full-delegated-only) are ADR-decided architectural choices, not gaps to fill. See `.goat-flow/learning-loop/decisions/ADR-021-goat-critique-full-mode-only.md`.");
lines.push("For each:");
lines.push("1. What to change");
lines.push("2. Evidence from your testing (cite file + semantic anchor)");
lines.push("3. Expected impact on the ratings");
lines.push("");
lines.push("### What You Did Not Verify");
lines.push("Be explicit about remaining uncertainty. List skipped skills, untested commands, unverified claims.");
lines.push("");
}
/**
* Append the closing reminder: respond with the full prose assessment, write the
* JSON report to the file path (not inline), and make no tracked-file edits.
*
* @param lines - prompt line buffer; appended to in place
*/
export function appendClosing(lines) {
lines.push("---");
lines.push("");
lines.push("**IMPORTANT:** Respond with the full prose assessment (Pre-check Results through What You Did Not Verify). Write the JSON report to the file path described above. Then end your reply with the one-line confirmation. Do not edit any tracked file. Do not emit the JSON as a fenced block in your reply.");
}
//# sourceMappingURL=compose-quality-static-sections.js.map