goal-harness

name	goal_harness
description	Goal-driven harness loop — keep working on a goal until /es, /er, /code_standards, and Codex plugin review all pass. All gates use adversarial subagents. Alias: /h

/goal_harness Skill

Purpose

Iterative goal-driven development with a 4-gate adversarial quality harness. The primary agent works on a goal, then validates the work through independent adversarial subagents that evaluate against specific standards.

Define a goal and iterate until 4 adversarial gates all pass:

1. /es — Evidence Standards
2. /er — Evidence Review
3. /code_standards — ZFC + ZFC-leveling + root-cause-first (3 lanes)
4. Independent Agent review — independent diff review

When to Use

• After implementing a feature or fix, before declaring done
• When code quality must be independently verified
• As the final step in /converge (the convergence loop)
• When /es and /er must both pass before merging

Cost-Aware Subagent Model Routing (Tiered Spectrum)

When spawning subagents for harness gates, route by task complexity across the model spectrum:

• Gate 1 /es (Evidence Format Check):
  • Codex: model: 'gpt-5.3-codex-spark' (reasoning_effort: medium)
  • Claude: model: 'haiku'
• Gate 3 /code_standards (ZFC, ZFC-leveling, Root-Cause Scans):
  • Codex: model: 'gpt-5.6-luna' (Fast 5.6 Tier)
  • Claude: model: 'haiku' / 'sonnet'
• Gate 2 /er (Evidence Review Synthesis):
  • Codex: model: 'gpt-5.6-terra' (Mid 5.6 Tier)
  • Claude: model: 'sonnet'
• Gate 4 Independent Agent Review (Deep Bug & Security Review):
  • Codex: model: 'gpt-5.6-sol' (Top 5.6 Tier for deep adversarial review)
  • Claude: model: 'sonnet' / 'opus'

This leverages the full Codex 5.6 model family (luna → terra → sol) alongside spark to match model capability directly to gate complexity. Real incident (2026-07-31): unrouted harness loop subagents running all 4 gates on gpt-5.6-sol + high reasoning burned >1B tokens.

Gate Details

Gate 1: /es (Evidence Standards)

Subagent (via the Agent tool) reads both evidence-standards skill layers and evaluates the diff and evidence artifacts. Returns: PASS / WARN / FAIL with file:line evidence.

Gate 2: /er (Evidence Review)

Subagent (via the Agent tool) runs evidence-review + evidence-standards synthesis against the diff and evidence artifacts (not the raw goal text). Subagent reviews: Map claims → artifacts, rate STRONG/WEAK/MISSING. Verdict: PASS / WARN / PARTIAL / FAIL / INCONCLUSIVE.

The evidence-review skill loads from the repo-root required path .claude/skills/evidence-review.md. If that path is not present, immediately abort Gate 2 and emit a missing-skill error — do not continue synthesis.

Normalization rule: WARN → PASS (document warnings). PARTIAL → FAIL (gaps remain). INCONCLUSIVE → FAIL (cannot confirm).

Gate 3: /code_standards (3 parallel adversarial lanes)

Lane	Skill Path	Checks
ZFC	~/.claude/skills/zero-framework-cognition/SKILL.md	No keyword routing, heuristic scoring, regex intent
ZFC-leveling	~/.claude/skills/zfc-leveling-roadmap/SKILL.md	Level-up fields, modal scoping, stale flag guards
Root-cause-first	~/.claude/skills/root-cause-first/SKILL.md	Root cause documented before adding protections

Dispatch all 3 in parallel. Any lane FAIL → overall /code_standards FAIL. WARN is acceptable but must be documented.

Gate 4: Independent Agent Review

Independent subagent reviews the full diff as a code reviewer. Looks for: bugs, anti-patterns, missing tests, security issues, dead code. Returns PASS or FAIL with file:line evidence.

Verdict Normalization

Raw verdict	Normalized	Harness treatment
PASS	PASS	Gate passes
WARN	PASS	Gate passes (document warnings)
PARTIAL	FAIL	Gate fails — evidence gaps remain
INCONCLUSIVE	FAIL	Gate fails — cannot confirm

Convergence

• 4/4 PASS (after normalization) → CONVERGED. Work is harness-verified.
• Any normalized-FAIL → fix blockers, increment, loop.
• Max iterations (10) → stop with partial report.
• Same score 2x → STALLED. Stop and report.

Subagent Isolation Rules

• Never skip a gate
• Never weaken a subagent's FAIL verdict
• Subagents are adversarial: only standard + diff, no primary agent optimism
• 3 /code_standards lanes run in parallel when possible
• Independent review must not share context with primary agent's own review

Environment Detection

When running in cmux (background/scheduled mode):

• Use scripts/goalexec-loop.sh for terminal-based loops (canonical)
• Use scripts/goalexec-loop-cmux.sh for cmux-aware loops (canonical)
• scripts/goal-loop.sh and scripts/goal-loop-cmux.sh are legacy aliases
• State directory: ~/.goal-loop/

Related Commands

• /goal — builtin Claude Code goal command
• /es — Evidence Standards (reference/display)
• /er — Evidence Review (adversarial synthesis)
• /code_standards — Coding standards dispatch (ZFC + ZFC-leveling + root-cause-first)
• /converge — Iterative goal achievement loop (formerly /goalexec)
• /converge_define — Define-only variant