fstack

Personal Claude Code workflow skills, commands, and hooks.

fstack is my personal agent operating system for Claude Code: it preserves context across sessions, turns repeated workflow failures into skill improvements, and keeps agent work auditable through plans, handoffs, hooks, queues, eval cases, and human review gates.

What This Demonstrates

• Context continuity: long-running agent work can be paused, handed off, resumed, and verified against the current repository state.
• Closed-loop skill evolution: user corrections and repeated failures are captured as signals, consolidated into claims and eval cases, then reviewed before changing durable skill behavior.
• Auditable agent work: plans, handoffs, changelogs, queues, and validation reports make agent decisions inspectable instead of hidden in chat history.
• Human-gated automation: agents can propose changes, but durable workflow mutations pass through explicit review.

Installation

claude plugin install https://github.com/fredrick84823/fstack

Core Workflows

Context Continuity

research-codebase / create-plan
        |
        v
implement-plan
        |
        v
create-handoff
        |
        v
resume-handoff
        |
        v
verify current repo state before continuing

See docs/context-continuity.md and examples/handoff.md.

Skill Evolution Loop

User correction / repeated failure
        |
        v
<<GAP skill-name: description>>
        |
        v
Stop hook captures the signal
        |
        v
signal-queue.md + memory/signals.jsonl
        |
        v
skill-memory-reflect consolidates claims + eval cases
        |
        v
/improve proposes a SKILL.md update
        |
        v
human gate
        |
        v
skill rule updated + changelog

See docs/skill-evolution-loop.md, examples/improve-signal.md, examples/skill-memory-claim.md, and examples/eval-cases.json.

Design Principles

• Verify before continuing: resumed work validates repository state instead of trusting stale notes.
• Raw events are not decisions: hook-captured signals are evidence; reflected claims and eval cases are the review package.
• Human gate before mutation: skill rewrites require explicit approval before changing durable behavior.
• Precision over noise: not every complaint becomes a skill gap; false positive traps are part of eval design.
• Scope-aware mutation: user, project, and repo skill scopes prevent accidental cross-context edits.

Skills (21)

Skill	Description
beautiful-mermaid	Render professionally-styled Mermaid diagrams
brainstorming	Structured brainstorming coach
codex-brainstorm	Adversarial brainstorming with Codex
codex-cli-review	Review uncommitted changes via Codex CLI
create-handoff	Create handoff docs for session transitions
create-plan	Create detailed implementation plans
create-team-plan	Decompose plans for agent teams
deep-module	Enforce Deep Module principle for interface-first design
extract-skill	Extract and build skills from candidate queue
grill-me	Turn vague ideas into concrete PRDs through targeted questioning
heptabase-task-card	Create Sprint-ready Heptabase task cards
implement-plan	Execute phased plans with human checkpoints
implement-team-plan	Execute team-plans with native agent teams
improve	Skill self-evolution and gap tracking
personal-wiki-mine	Mine mechanism-first candidate claims from bounded thoughts/ and Heptabase sources
research-and-plan	Research codebase + plan in one step
research-codebase	Document codebase architecture via sub-agents
resume-handoff	Resume work from a handoff document
skill-memory-reflect	Consolidate improve memory into claims, eval cases, and worklists
start-day	Build a daily task board from local and external work signals
work-wrap-up	Commit + PR + progress sync + Slack notification

Commands (8)

Command	Description
/brief-mode	Enable terse response mode
/cf-verify	Cloud Function full test & verification
/clean-worktree	Safely clean up a merged git worktree
/heptabase-title	Generate Heptabase document titles
/refine-text	Tighten text, remove filler
/safe-pull	Pull from origin with auto stash/rebase
/start-worktree-task	Create a new git worktree for isolated work
/work-wrap-up	Shortcut to the work-wrap-up skill

Agents (8)

Sub-agents for the research → plan → implement workflow (from HumanLayer):

Agent	Description
session-ender	Session 收尾：識別 improve signals → 提示 /work-wrap-up
skill-verifier	Independently verify proposed skill rule changes and eval scenarios
codebase-analyzer	Deep analysis of specific codebase components
codebase-locator	Locate relevant files and entry points for a task
codebase-pattern-finder	Find existing patterns and conventions
thoughts-analyzer	Analyze context from thoughts/ docs
thoughts-locator	Locate relevant prior decisions in thoughts/
web-search-researcher	Research external context via web search

Hooks

Wires the improve skill's signal capture into session lifecycle:

• SessionStart → checks signal queue for pending improvements
• Stop → captures <<GAP skill: desc>> markers and skill candidates automatically

Post-install: Enable auto gap detection

Add the following to your ~/.claude/CLAUDE.md so Claude emits gap signals after skill execution:

# Skill Gap Detection

After executing any skill, if a gap or improvement opportunity is found, output at the end of the response (not inside a code block):
<<GAP skill-name: one-line description of the gap>>

This rule tells Claude to emit <<GAP>> markers that the Stop hook captures into signal-queue.md. Run /improve to process pending signals.

License

MIT © Fredrick