rate-limit-guard

lemomo-ai/lemo-mod

核心：风格、中英文界面、消息编号和 /lemo-mod 面板，以及给其他 lemo mod 用的 $.lemo / Core: styles, zh/en interface, message numbers and the /lemo-mod panel, plus the $.lemo helpers other lemo mods use

OneWave-AI/claude-code-mods

Live session cost odometer above the prompt: a burning fuse, real-world comparisons, threshold alerts

Parais4/parais-mods

Zsemle, a pixel-art companion above the prompt (golden retriever, cat, slime, paperclip): watches every usage limit (50/75/90% warnings, pace forecast, resets, API errors, automatic model switch and compaction, cost), stops work at 95%, guards file writes and cheers long turns. English and Hungarian, terminal and desktop.

ucsandman/Agnostic-AI

COSTCLAW LIVE: per-request cost, cache decay and repeated-read detection while the session runs, and it serves the Nth identical Read of an unchanged file from its own cache instead of paying for it

cdeust/claude-mods

Automatic effort and model, and a quota that limits how many agents run, not how well one runs. Every request runs at the effort its grade calls for (zetetic-genius) and the main tool loop steps down the ladder. A subagent runs on the model its Agent call names, else the one its agent definition pins (read from the definition file, since agent.spawn carries only the call's), else the one the DESIGN block of its prompt names, else, for an inheriting type nobody designed, the rung of the dispatch table (hooks/dispatch-table.ts) for its task's class: one haiku completion grades the spawn prompt (routine, planned, bugfix, analysis, critical; a review or not), and the table gives model and effort (no haiku rung; a review runs on opus when complex, else on sonnet). A classifier that gives no class decides nothing: the spawn inherits and a POLICY row says why. Its effort is the Agent call's, else the one its definition sets, else DESIGN's, else the table's; an inheriting type the call gave a model runs every step at the effort of the turn's grade (main's spawns on a graded turn; a fork, a workflow agent or a teammate is never rewritten). No pressure, quota or context changes a model or an effort the orchestrator designed. Each spawn's requested, definition and decided model, what ran, the effort decided for its steps, the table's class and a short reason for each are kept per session. Quota pressure limits concurrency: the cap on subagents running at once falls from max_parallel_agents (5) at an empty five-hour window to 1 at quota_pressure_percent, and a spawn past it is refused with a message the orchestrator reads (wait for a running agent to finish). Only the main loop's own spawns, a fork of main included, are counted and refused: a subagent's nested spawn, a workflow script's agent() and a teammate are not. Main itself drops to sonnet under quota pressure, and doing implementation work in the main loop is warned about (POLICY row, toast), never blocked. A long streaming tool result is written whole to a file in the session's tool-results directory and the model reads a preview and the path (nothing is cut). Owns `phases`: the phase each loop is in (planning, analysis, design, implementation, verification, test, review, report), read off its tool calls or, for an agent, its type; on each run, the steps the stuck escalation raised (counted where each raised step landed, under enforce only); and the advisor calls of main with when they came. The tool loop of a phase the insights calibrated runs at that rung instead of medium (never above the request's grade). Starts in enforce; observe logs the decisions without applying them.

Arunjay4213/claude-mods

A per-turn cost and token ledger: a pinned status line with the session cost, the last turn's tokens and the cache hit ratio, and /ledger for a table of the recent turns with their model, tokens, cache reads and writes, duration and cost.