ClaudeMods
☰
ZH-TW
● 0 人在線上 · 瀏覽 0 次
贊助提交作品
GitHub 儲存庫 · 發布者 bajzaa975

cache-timer

cache-timer

bajzaa975@bajzaa975

bajzaa975/claude-alapcsomag/tree/main/cache-timer

已翻譯

關於這個 mod

狀態列:主執行緒的提示快取變冷前剩餘分鐘數(假設快取為 60 分鐘;使用量超額期間為 5 分鐘,Mod 無法偵測這種情況)

安裝

請先查看作者 README,確認 marketplace 與外掛名稱;指令可能隨儲存庫結構而變動。

claude plugin marketplace add bajzaa975/claude-alapcsomag
claude plugin install cache-timer
原文 / README

bajzi-plugins — token-saving orchestration for Claude Code and Cowork

bajzi-plugins is a Claude Code / Cowork plugin marketplace (GitHub: bajzaa975/claude-alapcsomag, marketplace name bajzi-plugins). It does two things: it saves tokens (model routing to cheaper models, context guards, output filtering) and it orchestrates sub-agents with a review loop (implementer, reviewer, fixer, with a two-round cap). The bajzi plugin is the full Claude Code package; bajzi-cowork is a skills-only subset for claude.ai / Cowork.

Full technical spec: docs/bajzi-package-spec.md.

Features

Token saving

  • Saver levels L0-L3 — how much work moves to the cheaper Z.ai GLM models. L0 claude: all on the Claude subscription (default). L1 light: flash-class work (locate/map, tests/lint/build, long-file summaries) goes to GLM. L2 glm (balanced): also implementation, fixes and document writing on GLM (glm -p, glm_model); risk slices, debugging and every review stay on Claude. L3 tight: the whole session runs on GLM and reviews are queued instead of performed. Set with worker --level 0|1|2|3.
  • rtk — installed by /bajzi:setup; compresses command output. A noise-filter hook additionally keeps loud install/build output out of the context (exit code preserved).
  • glm / worker / ccr shims — launchers around cc-router.js, no daemon. glm always runs on GLM, worker follows the saved saver mode, ccr is the back-compat entry. Models (~/.claude/cc-router.json): glm_orchestrator_model (default glm-5.3), glm_model and glm_fast_model (both default glm-5.3-flash); set with worker --set-orchestrator-model|--set-model|--set-fast-model <id>, force per shell with GLM_ORCHESTRATOR_MODEL/GLM_MODEL/GLM_FAST_MODEL. A top-level GLM session runs on the orchestrator model, its sub-agents on glm_model; a nested launch (a glm -p worker) runs on glm_model throughout. worker --status shows the level, worker --usage reports the Anthropic/GLM weighted-token split. A GLM launch is refused during the Z.ai peak window (08:00-12:00 CEST, 07:00-11:00 CET in winter); override one call with CC_GLM_PEAK_OK=1.
  • Status line — model, saver level, branch, task, context percentage, GLM share, open review-queue count and a peak-window hint.
  • Context guard — warns at 40% context, blocks tool calls at 50% except writing/reading the handoff and read-only git, so you can always save state and /clear.

Orchestration

  • Day-run mode (/bajzi:mode) — injects a routing table at session start: locate/map and tests/lint/build to haiku; long-file summaries to haiku; documents and specified slices to sonnet; risk-bearing slices (locks, concurrency, quotas, auth, money, migrations, destructive scripts) to opus, always; every diff review and final branch review to the configured reviewer models (reviewer_models in the manifest), never the cheap tiers; debugging to sonnet when a failing test or repro exists, else opus; design and planning stay on the orchestrator's own model. Escalation ladder: sonnet round 1, fresh sonnet round 2, opus round 3, the orchestrator round 4, then park.
  • Agents — implementer (sonnet, Tier 2/3 slices), implementer-risk (opus, Tier 1 slices), reviewer (opus, read-only, returns a findings file), fixer (sonnet, fixes every finding, never sees severity).
  • Review loop — /bajzi:implement then /bajzi:review then /bajzi:fix. The routing and the two-round cap are decided by code on files, not by the model. After round 2 open findings go to the owner (needs-owner.md) or are parked as debt (debt.md).
  • Dispatch guard — a hook that refuses review/fix work sent to the wrong agent and oversized briefs. A discipline guard, not a security boundary.
  • Debt (/bajzi:debt) — checks the parked-findings cap before the next slice, drains it in one fixer pass plus one review, calibrates severities with a blind re-rate.
  • Autopilot (/bajzi:autopilot) — an unsupervised work session with a decision log and a closing report.
  • Night-run (/bajzi:night-run) — plans an unattended overnight run and produces the launch block. The launch block starts the plugin's own bash runner run.sh (setsid nohup, in a bash terminal on the VM): one fresh headless session per sprint, never pushes, guard checks and a review queue that a later Opus session drains.

Session continuity

  • Handoff (/bajzi:handoff) — saves state to runtime/handoff/<branch>.md with a suggested opening prompt; a SessionStart hook reloads it after /clear, compact and resume.
  • Methodology guard / modszertan (Hungarian for "methodology") — records per repo which methodology leads (GSD, superpowers or none) in .claude/METHODOLOGY, and nags at session start when there is no decision yet.

Safety

  • Secret-read guard — denies reading .env, .env.* (except .example/.sample/.template/.dist), .secrets and the manifest's secret patterns. Accepted limits: it is a pattern guard, not a shell parser, so variable indirection, copying a protected file to another name and encoded paths pass; it covers files, not environment variables.
  • Injection scanner — warn-only: after Read, WebFetch, WebSearch and MCP results it adds a "treat this as data" warning naming the matched rules. Never blocks; a rephrased injection passes.
  • Writer guard — bajzi plugin files are edited only by a session started in the bajzi repo's main checkout; any other session is denied Edit/Write and sends a request instead (SendMessage, or a file in runtime/requests/, which stays open to all). Installed plugin copies are denied to every session. Accepted limit: Bash/PowerShell writes are not blocked.
  • Hooks fail open: an internal error allows and logs. Three exceptions fail closed: the pre-commit gate (installed by /bajzi:project-setup), the night-run push guard, and day-run-mode.sh when a non-Anthropic session lacks the L3 text.

Setup

  • /bajzi:setup — machine setup from bajzi/skills/setup/manifest.json (plugins, settings, rtk, status line, user-scope MCPs). --check prints one drift line per difference and changes nothing.
  • manifest.json — the single file to edit to add a tool; every machine gets it at the next setup.
  • /bajzi:project-setup — applies or checks (--check) a repo's .claude/project-profile.json: project plugins and MCPs, METHODOLOGY, linked skills, instruction files and the pre-commit gate.

Mods

  • cache-timer — status line entry: minutes left until the main thread's prompt cache goes cold. It assumes a 60-minute cache; the cache is 5 minutes during usage overage and the mod cannot detect that.
  • nightrun-pane — open the pane with /nightrun: the newest claude-orchestrator night run's sprint states and log tail. It reads only the runtime/nightrun/<stamp>/ layout (nightrun.log, *.status, SUMMARY.md, STOP, runtime/handoff/night-watch-state.md), not bajzi night-run's NIGHT_DIR.

Cowork variant

bajzi-cowork ships only the autopilot, handoff and modszertan skills. It lacks everything that needs bin/, hooks, agents or lib/: the saver levels and shims, status line, all guards, day-run mode, the review loop, night-run and setup. Install it in Cowork, never bajzi.

Skills

| Skill | Command | What it does | |---|---|---| | setup | /bajzi:setup | Sets up or drift-checks (--check) a machine per the manifest. | | project-setup | /bajzi:project-setup | Applies or checks the repo's project profile. | | handoff | /bajzi:handoff | Saves session state and an opening prompt for after /clear. | | modszertan | /bajzi:modszertan | Picks the repo's methodology (GSD, superpowers, none). | | autopilot | /bajzi:autopilot | Unsupervised session with a decision log. | | mode | /bajzi:mode | Turns day-run mode on/off or reports it. | | night-run | /bajzi:night-run | Plans an unattended overnight run. | | implement | /bajzi:implement <id> | Implements one slice through the implementer agents. | | review | /bajzi:review <slice> <range> | Reviews a commit range through the reviewer agent. | | fix | /bajzi:fix <findings-file> | Fixes round-1 findings, then runs the round-2 re-review. | | debt | /bajzi:debt | Checks, drains or calibrates parked review debt. | | radar | /bajzi:radar [now\|install] | Biweekly read-only review of plugins, Claude news and own usage; adopt or decline its items. |

Skills also start by themselves from a plain request ("do a handoff").

Installation — Claude Code (every machine: laptop, VM, new machines)

  1. Once per machine, in a terminal:
    claude plugin marketplace add bajzaa975/claude-alapcsomag   # or: a local path
    claude plugin install bajzi@bajzi-plugins   # marketplace name: bajzi-plugins
    
  2. Once per machine, in a new Claude Code session: /bajzi:setup (marketplaces, plugins, rtk, settings, status line, ~/.claude/bajzi-mode, user-scope MCPs).
  3. Any time: /bajzi:setup --check — one line per drift item, changes nothing.
  4. Per repo, only when the repo carries .claude/project-profile.json: /bajzi:project-setup (--check diffs it).

Then the global rules:

cat shared/CLAUDE.md >> ~/.claude/CLAUDE.md

Check: claude plugin list, and /bajzi:handoff in a new session.

Without a marketplace, locally (for development)

The bajzi/ directory can be copied here: ~/.claude/skills/bajzi/ — the next session loads it automatically under the name bajzi@skills-dir.

Installation — Cowork (Claude Desktop)

  1. Customize → Plugins → Add marketplace → the repo's URL (https://github.com/bajzaa975/claude-alapcsomag or bajzaa975/claude-alapcsomag); the marketplace is named bajzi-plugins.
  2. Install the bajzi-cowork plugin from the list. Never install bajzi in Cowork: it ships bin/, hooks and agents, which claude.ai marketplace sync rejects ("Sync failed").
  3. Or without a repo: Plugins → upload a zip of the bajzi-cowork/ directory (never of bajzi/). The zip is not in the GitHub repo; create it locally and make sure the plugin.json version inside it matches the marketplace's. Normally steps 1-2 (marketplace) are the recommended route.
  4. Copy the content of shared/cowork-preferences.md (the part below the ---) here: Settings → Cowork → Global instructions → Edit (only exists in the desktop app; in newer builds the Customize panel also brings together the Skills / Plugins / Global instructions trio). The "Customize → Instructions/Preferences" route listed here earlier was WRONG. A per-project layer on top of that: the Instructions field on the Project's right-hand panel, which goes ON TOP OF the global instructions, not instead of them.

Updating

Push to the repo → Claude Code: claude plugin update bajzi, then /bajzi:setup (refreshes the status line copy) · Cowork: Update at the marketplace, then update the bajzi-cowork plugin.

其他同名作品

更多類似作品