メニュー
Run Copilot CLI work through the Goliath five-role pipeline, with the architect as planner and explicit model fallbacks for implementation, test authoring, verification, review, and red-team checks.
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/agentmc15/polytropos --skill goliathコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SKILL.md を表示中
Have Fable 5 do the expensive meta-work once — deep plan a complex task (codebase review, greenfield project, large migration) AND build the execution kit (task briefs, model-pinned subagents, skills, verification loops) so Opus/Sonnet can execute at near-Fable quality. Use when the user says "plan this with Fable", "architect this", "review this codebase", or escalates a complex task.
Run an execution kit produced by /polytropos:architect — loop through TASKS.md, dispatching each task to the kit's model-pinned subagents, verifying, and updating state. Use when the user says to execute/continue/resume a kit or plan. Args: the kit slug (optional if only one kit is active).
Reach for this when your context is huge, you're near 700K, cache reads are high, or you're asking should I compact — it explains what this skill can and cannot do, ranks prevent/prune/measure in priority order, and gives the checkpoint move before you compact.
| name | goliath |
| description | Run Copilot CLI work through the Goliath five-role pipeline, with the architect as planner and explicit model fallbacks for implementation, test authoring, verification, review, and red-team checks. |
Goliath is a Copilot CLI-only orchestration policy. Every run uses the architect as planner and the same five execution roles: implementer, test-author, verifier, reviewer/orchestrator, and red-team. It does not configure or invoke the Claude Code harness, and it does not silently substitute a different role or model.
Use the first model in each row that is available in the current Copilot /model picker and
documented by data/pricing.copilot.json. If a requested model is unavailable, try the next
model in that row. Model IDs must be resolved from the pricing data at run time; do not invent
IDs or hardcode prices.
| Role | Primary | Fallback 1 | Fallback 2 |
|---|---|---|---|
| Architect | Claude Fable 5 | GPT-5.6 Sol | — |
| Implementer | Grok 4.6 | Claude Sonnet 5 | — |
| Test-author | Gemini 3.7 Flash | Claude Sonnet 5 | — |
| Verifier | GPT-5.6 Luna | Gemini 3.7 Flash | — |
| Orchestrator / reviewer | GPT-5.6 Sol | Claude Opus 5 | — |
| Red-team | Grok 4.6 | Gemini 3.7 Flash | Claude Sonnet 5 |
The architect is always the planner and runs before the five execution roles. The orchestrator/reviewer is the Copilot-side coordination role. It owns sequencing, passes the architect's plan to the implementer and test-author, requires an independent verifier result, and then dispatches the red-team check before accepting the work.
Inspect the active roster before an expensive run:
python3 {{POLYTROPOS_ROOT}}/bin/copilot_pricing.py models --json
Confirm the selected IDs are also exposed by the current /model picker. If a primary is
absent, use that role's fallback order. If every candidate for a role is unavailable, stop
and report the missing role; do not substitute a different role's model.
The architect creates the durable plan and acceptance criteria before any execution role. Do not use the implementer, reviewer, or another role as the planner.
The implementer changes files against the approved plan. It must report changed paths, assumptions, and the exact validation it ran.
The test-author writes adversarial tests from the architect's plan and acceptance criteria. It may change tests only and must not implement the feature itself.
The verifier starts from fresh context and independently checks the acceptance criteria. It must rerun the relevant checks, never trust the implementer's success claim, and must report failures explicitly.
The orchestrator/reviewer compares the result with the plan, checks scope and integration, and coordinates the final review. It must not replace the architect as planner.
The red-team role is mandatory and attempts to break the implementation beyond the verifier's ordinary checks, focusing on edge cases, misuse, regressions, and unsafe assumptions. It must not rewrite tracked files while reviewing.
Accept the work only when all five execution roles report success and the architect's plan is satisfied. On failure, send the exact evidence back to the responsible role and rerun only the failed stage or its necessary dependency.
For an isolated role dispatch, use the selected model explicitly:
copilot -p "<self-contained role brief>" --model <resolved-model-id>
Resolve the display name to its current model ID with the pricing engine before dispatching; the examples above intentionally do not embed live pricing-key IDs.
Never invoke the real Copilot CLI from tests or verify commands. Use --dry-run, injected
runners, or synthetic fixtures when validating orchestration logic.
End with a compact ledger containing the architect planner plus all five execution roles, selected models, availability fallbacks used, files changed, verification commands and results, reviewer decision, red-team findings, and any unresolved blocker. Do not claim success when a required role could not run.
If {{POLYTROPOS_ROOT}} is still literal, the Copilot bundle is not installed. Install it with
python3 bin/harness_select.py install --harness copilot, then reload skills in Copilot CLI.