agentsclimarketplace

Goal loop compiler

Skill fxbillu/goal-loop-compiler/skills/goal-loop-compiler

Use when a non-trivial finite task needs a file-shaped, verifiable, stoppable Goal Loop package. Compile GoalPro-style intent, evidence, scope, state, validation, and human handoff into compact V4 artifacts. Do not use for recurring automation, monitoring, daily/weekly routines, or prompt-only GoalPro output.From its SKILL.md

Install
npx -y skills add fxbillu/goal-loop-compiler --skill goal-loop-compiler

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

6.5 KB, ~1.4k tokens by cl100k_base, as published. Nobody here has run it

Goal Loop Compiler compact-v4

Compile a finite intention into a compact execution package that an agent can execute, verify, pause, and explain to a human.

GoalPro is the semantic blueprint. V4 is the durable package authority. Native Goal is the package's deterministic runtime projection, not a second authored goal. Do not create Goal Prompt, Loop Prompt, or Launch Prompt artifacts.

Alignment

Classify before package creation:

  • one_shot: simple and low risk; handle directly without a package.
  • finite_goal: clear terminal result with verifiable evidence.
  • strategic: route, architecture, or long-term decision; require evidence.
  • repair: prior output ran off course; diagnose before compiling a fix.
  • governed: high-risk, release, permission, data, or production-adjacent work.
  • recurring_system: ongoing automation, monitoring, operations, or routine; do not compile a finite package.

Output a Goal Alignment Card before compiling. Include the route, package mode, Goal, Success Evidence, Boundaries, and Stop / Pause Rules.

Use confirmed_by_request only when all four fields are explicit, the user explicitly invokes this skill to compile, the work is safe and local, and no material field was inferred. Label that status in the card and continue.

Use literal_confirm_required and wait for confirm when a material goal, evidence, boundary, or stop rule was inferred; when the action is external, destructive, costly, permission-changing, or materially expands scope; or when the user explicitly requests a confirmation gate. A user-requested literal gate always wins.

Compiler Pipeline

After confirmation or confirmed_by_request:

  1. CLASSIFY: choose the lifecycle route.
  2. INTENT_INTAKE: derive surface request, real intent, dissatisfaction, strategic outcome, wrong-turn risk, and acceptance evidence.
  3. EVIDENCE_GATE: for strategic, governed, external-fact, platform-version, legal, pricing, standards, or research work, collect evidence that changes at least one contract field.
  4. INVENTORY: list affected instructions, templates, references, scripts, tests, and source-of-truth files before mutation.
  5. CONTRACT_COMPILE: create exactly the five default package files.
  6. CONTRACT_REVIEW: remove duplicate authority and run required review gates.
  7. LOOP_INITIALIZATION: initialize state, gaps, budget, and next focus.
  8. EXPRESSION_ECONOMY: remove repetition, never boundaries or evidence.

Package And Authority

Create only:

.goal/goals/<YYYY-MM-DD-task-slug>/
  GOAL_CONTRACT.md
  PLAN.md
  STATE.json
  VALIDATION.md
  CHANGE_BRIEF.md
  • GOAL_CONTRACT.md: durable semantic source: goal, intent, scope, non-goals, constraints, pause conditions, acceptance criteria, and evidence standard.
  • PLAN.md: derived execution strategy only; it never redefines semantics.
  • STATE.json: runtime status, evidence, gaps, controls, and review records.
  • VALIDATION.md: compile/completion gates and evaluator rubric.
  • CHANGE_BRIEF.md: human summary, never self-evidence; starts pending.

CONTINUATION_SUMMARY.md is optional, outside the package, and only after Continue or Pause. It must say that STATE.json remains authoritative.

Native Goal Handoff

The Compiler completes after it creates the package, runs compile validation, renders the Native Goal projection, reports its path and digests, and stops at the Executor checkpoint. This is a phase boundary, not a required new chat.

Run:

python3 scripts/validate_package.py --phase compile <goal-dir>
python3 scripts/validate_package.py --phase native-goal <goal-dir>

The Executor normally continues in the same Codex task after the user sends execute. It must pass the returned native_goal text unchanged to create_goal, then read the full contract before acting. Native Goal freezes contract semantics. If semantics need to change, Pause, recompile, re-render, and start a new Native Goal.

Use a separate Executor task only for blind review, permission or workspace isolation, or explicit user direction.

Review Gates

Use a fresh, read-only goal_loop_evaluator. Give raw artifacts, evidence, and a narrow question, never the builder's conclusion.

Independent contract and pre-Done review are required for strategic, repair, and governed packages. A drift review is required after no progress or a strategy change. Any drift record requires an independent pre-Done review.

For a stable ordinary finite_goal, use executor self-check, target evidence, and static completion validation; do not add a reviewer merely by default.

A revise or blocked verdict prevents progression. If a required independent review is unavailable, Pause. Record review provenance in STATE.json.verification_delta as described in references/loop-runtime.md.

Validation

python3 scripts/validate_package.py --phase compile <goal-dir>
python3 scripts/validate_package.py --phase review-digest <goal-dir>
python3 scripts/validate_package.py --phase native-goal <goal-dir>
python3 scripts/validate_package.py --phase completion <goal-dir>

Compile validation rejects missing or extra package files, prompt-only artifacts, semantic conflicts, missing evidence rules, invalid state controls, and recurring routes. An Intent that matches the surface request is a warning, not a failure.

Completion validation rejects unsupported claims, incomplete acceptance evidence, Done with active gaps or blockers, non-goal violations, malformed review records, and a final evaluator that does not check Change Brief truthfulness. It requires pre-Done review only when the route or drift policy requires it.

References

Read only when needed:

  • references/goal-quality.md: contract quality and evidence gates.
  • references/loop-runtime.md: state, native handoff, reviews, and finalization.
  • references/quality-checklist.md: compact-v4 delivery checks.

Stop Conditions

Pause or hand off when the route is recurring_system, the request is one_shot, required evidence is unavailable, completion requires a semantic contract change, validation cannot establish confidence, or the next decision needs human business judgment or high-risk permission.

Ask only for fields that change route, evidence, boundaries, or stop rules.

What ships with it: 11 files

51.4 KB alongside SKILL.md, 1 of them executable

agents/

scripts/

Keep looking

Skills are one crate of 326,144. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.