agentmc15/polytropos

goliath

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.

architect

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…

execute

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…

context-weight

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…

graphify

Build and read a local knowledge graph of a repo with the external graphify CLI — symbol lookup with file:line precision, call structure, impact analysis, and an architect-grounding brief, all offline. Use when the user wants repo structure analyzed or…

update

Check and refresh everything this repo installs into its harnesses — Claude plugin cache staleness, Copilot and Codex bundle drift, generated pricing mirrors, and docs snapshot freshness. Use when the user asks to update or refresh the plugin or a harness…

architect

Do the expensive planning once — deep-plan a complex task and write an execution kit (PLAN.md + TASKS.md with model-pinned, self-contained briefs) under tasks/kits/<slug>/ for the execute driver to dispatch on cheaper models. Use for "architect this", "plan…

bench-routing

Compare benchmark priors with Codex-dispatchable roles while preserving benchmark uncertainty. Use to rank candidates, inspect Codex role recommendations, or run the synthetic routing demo.

context-weight

Analyze how Codex context grows, what resident instructions cost, and where fidelity stops. Use for context-window pressure, growth curves, loaded-surface audits, or compaction questions.

doctor

Diagnose Polytropos Codex plugin, agent, skill, prompt, and managed-install state without changing it. Use for setup, upgrades, stale paths, missing skills, or install conflicts.

effort

Control the GPT-5.6 reasoning-effort dial per run — pick the right level, apply it, and step it up only on failure evidence. Use when the user asks to raise/lower reasoning effort, run at the deepest level, or make a run think harder or cheaper.

escalate

Run one task on the cheapest sufficient GPT-5.6 tier behind a machine-checkable success check, escalating up the tiers — frontier last — only if the check fails. Use for "try it cheap first, fall back to the top model if it doesn't work" or an…

execute

Execute a prepared Codex kit task-by-task with dependency checks, independent verification, and phase review. Use after architect creates tasks/kits/<slug>, or when asked to run or continue a kit.

frontier-check

Decide whether a task is worth the frontier-tier model in the GPT-5.6 family versus a mid or cheap candidate, and how to run it optimally — reasoning effort, task spec, refusal fallbacks. Use when the user asks "is the top model worth it here" or how to get…

journal

Generate the daily work journal — collect today's AI usage across Claude Code, Copilot CLI, and Codex CLI plus git activity into a digest, then write the narrative, technical, and next-day-plan summaries. Use when the user asks for their work journal, daily…

memory

Recall a small, relevance-gated set of private local facts and review their staleness. Use when durable project context may help but context quality and privacy must stay bounded.

route

Pick the right Codex model for a task and estimate its burn before running it. Use when unsure which GPT-5.6 tier a task needs, what it will cost (API) or burn (subscription), or how to run it fast.

usage

Analyze historical Codex CLI activity from local session logs, read-only — honestly unpriced or labeled-proxy. Use when the user asks what they've used, burned, or spent in Codex.

repo-bench

Benchmark candidate Claude models on a chosen repo's own real work — mines issue-fix pairs or generates red-validated mutation-repair tasks, dispatches candidates in leak-proof sandboxes, grades with four independent oracles, and renders a re-tiered…

route

Pick the right Claude model for a task and estimate its cost before running it. Use when the user asks which model to use, whether a task needs Fable/Opus, what a task will cost, or wants a task dispatched to a cheaper model. Args are the task description,…

architect

Do the expensive planning once — deep-plan a complex task and write an execution kit (PLAN.md + TASKS.md with model-pinned, self-contained briefs) under tasks/kits/<slug>/ for the execute driver to dispatch on cheaper models. Use when the user says "architect…

bench-routing

Decide whether a new or higher model should replace what a role currently runs on — a benchmark-informed routing recommendation. Use when the user asks "should we upgrade X to Y for this role" or wants a benchmark-backed routing check.

budget

Run an execution kit on a cheaper ladder of models when the AI-Credits budget is tight — one-tier-lower dispatch with honest actual-vs-standard cost reporting. Use when the user says budget mode, run this cheaply, low on credits, or wants the savings measured.

context-weight

Reach for this when your context is huge, cache reads are high, you're asking should I compact, or you want to know what filled the window — it explains what this skill can and cannot do, ranks prevent/prune/measure in priority order, and gives the checkpoint…

effort

Control the reasoning-effort dial for Copilot models — Copilot's per-model "Reasoning" setting, covering which models have it, how to set it, and when to turn it up or down. Use when the user asks to raise/lower reasoning effort, run at extra-high, or make a…

escalate

Run one task on the cheapest sufficient model behind a machine-checkable success check, escalating to a stronger tier — frontier last — only if the check fails. Use for "try it cheap first, fall back to the top model if it doesn't work".

execute

Run an execution kit under tasks/kits/<slug>/ — drive bin/copilot_execute.py task by task, verify each result, and climb the pricing tiers only on failure. Use when the user says to execute, continue, or resume a kit or plan.

frontier-check

Decide whether a task is worth the harness's frontier-tier model versus a strong or mid model, and how to run it optimally — effort, task spec, refusal fallbacks. Use when the user asks "is the top model worth it here" or how to get the most out of it.

journal

Generate the daily work journal — collect today's AI usage across Claude Code, Copilot CLI, and Codex CLI plus git activity into a digest, then write the narrative, technical, and next-day-plan summaries. Use when the user asks for their work journal, daily…

lessons-loop

Capture a durable lesson every time the human corrects the agent — or a task escalates models — so the same mistake doesn't recur. Use immediately after any user correction and after any model escalation. Also use at session start to load relevant past…

route

Pick the right Copilot model for a task and estimate its cost in AI Credits before running it. Use when the user asks which model to use, what a task will cost, whether a cheaper model would do, or how much of their plan allowance a job will burn.

usage

Analyze historical Copilot CLI spend from local session logs — spend by model and session in USD and AI Credits, read-only. Use when the user asks what they've spent, which models they've been using, or where they could save.

bench-routing

Rank published benchmark entries (Artificial Analysis Intelligence Index), recommend a model+effort per orchestration role given what each harness can actually dispatch, and check that recommendation against this repo's OWN measured routing outcomes. Use when…

cost-report

Analyze historical Claude Code usage from local transcripts — spend by model, most expensive sessions, and sessions where a cheaper model would have sufficed. Use when the user asks what they've spent, which model they've been using, or where they could save.

escalate

Run one task on the cheapest sufficient model behind a machine-checkable success check, and automatically escalate to a Fable 5 subagent (carrying the failure evidence) only if the check fails. Use when the user wants "try it cheap first, fall back to Fable…

fable-check

Decide whether a task is worth running on Fable 5 (vs Opus 4.8 / Sonnet 5) and how to run it optimally — effort level, task spec, refusal fallbacks. Use when the user asks "should this run on Fable", "is Fable worth it", or how to get the most out of Fable 5.

journal

Generate the daily work journal — collect today's AI usage across Claude Code, Copilot CLI, and Codex CLI plus git activity into a digest, then write the narrative, technical, and next-day-plan summaries. Use when the user asks for their work journal, daily…

memory

Remember durable facts across sessions and recall only the few relevant ones under a strict budget — never bulk-injected. Use when the user says "remember this", asks "what do you remember about…", wants a preference or decision saved, or when a stored fact…

setup

Install the polytropos statusline (current model, live session cost, context usage, rate-limit burn) into the user's Claude Code settings. Use when the user asks to set up or enable the cost statusline.