Claude Skills by tools-only

**Problem:** Variables lost after hibernation **Cause:** DO auto-hibernates when idle; in-memory state not persisted **Solution:** Use `ctx.storage` for critical data, `ws.serializeAttachment()` for per-connection metadata

**Cause:** Eventual consistency means writes may not be immediately visible in other regions **Solution:** Don't read immediately after write; return confirmation without reading or use the local value you just wrote. Writes visible immediately in same location, ≤60s globally

**Cause:** Mutating state directly or not calling `setState()` after modifications **Solution:** Always use `setState()` with immutable updates:

Use this snippet to register the Figma MCP server in your agent's configuration as a streamable HTTP server with bearer auth pulled from your env.

| File | Max Tokens | ~Words | Warning At | | --------------- | ---------- | ------ | ----------- | | PROJECT.md | 2,000 | 1,200 | 1,600 (80%) | | ROADMAP.md | 3,000 | 1,800 | 2,400 | | STATE.md | 10,000 | 6,000 | 7,000 (70%) | | spec.md | 5,000 | 3,000 | 4,000 | | design.md | 8,000 | 4,800 | 6,400 | | tasks.md | 10,000 | 6,000 | 8,000 | | STACK.md | 2,000 | 1,200 | 1,600

Comprehensive intervention, onboarding, renewal, expansion, and escalation playbooks for SaaS customer success management. ---

Complete methodology for multi-dimensional customer health scoring in SaaS customer success. ---

Best practices for pipeline management including stage definitions, conversion benchmarks, velocity optimization, and inspection cadence. ---

A comprehensive guide for Sales Engineers planning, executing, and evaluating proof-of-concept engagements.

Markdown documentation contained example TODO/FIXME/HACK markers, which the scanner counted as new debt and caused baseline checks to fail. We now exclude `.md` files from scanning and update the scheduled workflow to flag any debt and report the correct timestamp.

Create a git commit

List all git worktrees with enhanced status information

**Status:** ✅ Complete **Date:** February 5, 2026

统一数据调度系统，负责定时数据同步、智能补数和缺失检测的自动化调度。

数据管理模块负责数据源监控、同步任务管理、数据质量检测和插件管理。

提供全局统一的交易日历服务，将交易日历作为配置文件管理，而非数据库数据，简化交易日期的查询和管理。

系统 SHALL 提供多Agent竞技场的完整生命周期管理，包括创建、启动、暂停、恢复和销毁。

The system SHALL execute data synchronization tasks asynchronously via a Redis-backed queue and dedicated worker processes. The HTTP API server SHALL NOT execute data synchronization work inline as part of request handling.

数据管理模块负责数据源监控、同步任务管理、数据质量检测和插件管理。提供基于交易日历的数据缺失检测、批量同步任务创建与执行跟踪、插件状态查询与详情展示、AI日志诊断等能力。

``` Severity: High

Show current task plan status and phase summary

Integrate Ralph async control commands into flow-next-tui, allowing users to monitor and control Ralph runs from the terminal UI.

Add Codex CLI as alternative review backend for cross-model code reviews. Enables Windows/Linux users (who can't use RepoPrompt) to get multi-model review quality gates.

Tab key cycles between Ralph runs in current project. Users may have multiple runs (interrupted, retried, different epics).

Evaluate OpenProse (prose.md) as a potential replacement or enhancement for Ralph's bash loop orchestration. OpenProse is a DSL where the LLM interprets "prose programs" as a virtual machine, enabling smarter agent coordination without external frameworks.

Create memory-scout subagent for retrieving relevant memory during planning and work phases. One agent, two call sites:

Document memory system in relevant places. Key point: memory is a flow-next feature (not Ralph-specific), works in manual + Ralph modes.

This document captures the philosophical foundation of the namshub toolkit — the mental model that makes skills, hooks, memory, and health feel like ONE system rather than four bolted together.

- Updated `parse_id()` regex: `fn-(\d+)(?:-[a-z0-9]{3})?(?:\.(\d+))?` - Updated `epic_id_from_task()` to preserve suffix via string split - Verified with tests for legacy and new formats - Commits: 75be01ae97efbede5cdf1b0b49e52ad835e5a551 - Tests: inline python test - PRs:

The Amnesiac Architect What Twenty-Four AI Agents Learned About Teaching Machines to Remember [PAUSE] This is a story about forgetting. About software that writes other software, brilliantly, and then has no memory of having done so. About twenty-four artificial minds convened to diagnose a problem that turned out to be their own. And about the solution they arrived at, which was far smaller than anyone expected and far stranger than anyone planned. If you build things with AI coding agents, ...

The `config/settings.json` file defines global settings and hook configurations for Claude Code.

- Updated `TASK_ID_PATTERN` regex for fn-N-xxx format - Updated `validateTaskId` error message - Updated epic pattern in `getRunEpics()` - All 379 TUI tests pass - Commits: 4891cb9a6deb193c4d9d5345e4f839757f8d2dea - Tests: bun test - PRs:

*Complete identifier reference for all YARA rule conventions, errors, and recommendations. Combines yaraQA automated checks with style guide and performance manual review items.*

Add `flowctl status [--json]` command to show epic/task counts plus active Ralph run info.

Add `flowctl task reset <task_id> [--cascade]` to reset task status to todo.

Add CI tests for all new async control commands.

Added plan-sync step 3e to phases.md between verify completion and loop, with explicit done-status gate and Task tool usage. Renamed Loop to 3f. - Commits: e9e2a22, 4a04d73c336cd725cccb2dec516bdb3da02b1f12 - Tests: - PRs:

Create a new `/flow-next:sync` skill that manually triggers the plan-sync agent to update downstream task specs.

Update the existing plan-sync agent to support a `DRY_RUN` mode where it reports proposed changes without actually editing files.

Update Ralph templates to support codex as review backend option.

PocketPaw integrates with Google services via OAuth 2.0. Currently supported: Gmail and Google Calendar.

Update hook enforcement for completion review receipts.

Update `/flow-next:work` skill to handle `status=completion_review` from selector.

Update all regex patterns that parse epic/task IDs to support variable-length slugs while maintaining backwards compatibility with existing formats.

YAML frontmatter contains `name` and `description` fields. **These are the only fields Claude reads to decide when to use a skill.**

Update all error messages and CLI help text that reference ID format to show the new slug-based format.

Create HelpOverlay component showing keybindings.

Create theme system with dark (default) and light variants.

Define TypeScript types/interfaces for all data structures.

These are the release notes for gptme version v0.2.1.