Per-repo learning layer for AI coding model routers: learns which task classes can safely run on a cheaper tier. Native for Claude Code, Codex, Antigravity · MCP server · zero deps.
This repository also configures its own agents. See what tierdecay tells them →
Skill Claude Code
Shared implementation discipline for executor agents — conventions, testing, diff hygiene, reporting. Preloaded into executor and heavy-executor via their skills frontmatter.
Skill Claude Code
Complexity-based routing of tasks across four model tiers — frontier (T3), strong workhorse (T2), fast executor (T1), cheapest recon (T0). Use when planning any multi-step work, when deciding which subagent should execute a task, when writing a dispatch brief, or when an executor fails and escalation is needed.
Skill Claude Code
Repo-specific execution patterns distilled from higher-tier solutions via the tier-decay protocol. Preloaded into executor and heavy-executor at startup. Written ONLY by the main thread — executors read, apply, and report; they never edit this file.
Skill Claude Code
Post-completion distillation protocol that converts high-tier solutions into cheaper-tier playbook entries and recalibrates routing from real outcomes. Use after completing any [T2] or [T3] task, when logging a task outcome to the routing ledger, when a playbook probe passes or fails, or when a class keeps escalating.
Skill Claude CodeCodex
Close a task under the TierDecay protocol — append the ledger row, update playbook hits, quarantine on failure, apply the confidence-gated decay, and distill a reusable entry after a high-tier success. Orchestrator only; executors never run this.
Skill Claude CodeCodex
Execution discipline for TierDecay executor agents — minimal diffs, conventions, proof by tests, playbook feedback, and the rule that executors never touch routing state.
Skill Claude CodeCodex
Route a coding task to the tier this repository has earned for its class (TierDecay). Use when planning multi-step work, choosing which agent, model, or reasoning effort runs a task, writing a dispatch brief, or escalating after failures.
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: