The essence of AI Coding is Context Management. This project establishes a balance between the probability of LLMs and the certainty of software engineering through standards, Harness, and Skills.

A complete AI behavior governance framework with four-layer synergy:

1. Context Engineering — AGENTS.md defines the reasoning framework, loop engineering discipline, 9-step process, coding philosophy, and security rules.
2. Harness Layer — harness/hooks/ executable checks upgrade key rules from "textual constraints" to "verifiable gates" (agent-executed per AGENTS.md §7, no platform hook registration).
3. Skills Layer — Reusable capability packages loaded on demand (Workflow / Debugging / Security / Architecture Review / Code Review).
4. SSoT Documentation — docs/STATUS.md records context, docs/tasks.json tracks progress.

Tool-agnostic: single AGENTS.md + harness/ work with any coding agent (Claude Code, Gemini CLI, Codex CLI, pi, Cursor, …). No per-tool config directories.

# Execute in the project root directory
curl -sSL https://raw.githubusercontent.com/rayjun/ai-coding-context/main/install.sh | bash

# Preview files to be written (dry run)
curl -sSL .../install.sh | bash -s -- --dry-run

Automatic .bak backup before overwriting; automatic rollback on download failure.

Reasoning framework, task complexity levels, coding philosophy, loop engineering discipline, 9-step development process, security rules, and documentation maintenance. The process is scaled via a trivial/moderate/complex × 9-step matrix to avoid a one-size-fits-all approach. Tool-agnostic: no per-tool config dirs; hooks are executed by the agent on its own per §7 (see harness/hooks/).

---
paths:
  - "harness/hooks/**/*.sh"
---
# Hook Development Rules
- All hook scripts must start with `set -euo pipefail`
- ...

Referenced by AGENTS.md §8; paths frontmatter is documentation only (no platform auto-injection).

• /review — Review current branch diff (dynamically detects base branch)
• /status — Project health check (STATUS.md freshness + tasks consistency + git status)
• /fix-issue <n> — Systematically fix from GitHub issue ID

Kept as reference implementations. Slash-command auto-registration is platform-specific; any agent can run these as documented procedures instead.

6 project-level skills (workflow-management / plan-review / code-review / investigate / careful-ops / obsidian-writer) + 1 eval-runner tool. Each SKILL.md contains EVAL 1..N quality assessment blocks, which can be automatically scored using eval-runner.py. code-review enforces tool-evidence: every review must cite actual compiler / lint / test output, not just code reading.

obsidian-writer is a demo skill showcasing the per-vault AGENTS.md rule pattern; this repo itself does not use it. Forks can drop it from install.sh if not needed.

Independent context windows for tasks that would otherwise pollute the main thread:

• plan-reviewer.md — runs the 5-dimension architectural review (data flow / concurrency / interface / testing / operability) in isolation. References plan-review skill as the rule SSoT, does not duplicate.
• code-reviewer.md — runs the 5-dimension code review (security / correctness / maintainability / test coverage / requirement fit) in isolation, with mandatory tool-evidence (compiler / lint / test output). References code-review skill as the rule SSoT, does not duplicate.
• retro-writer.md — extracts self-correction patterns ("expected X, actually Y") from session transcripts and appends to docs/lessons.md.

Skills define rules; agents execute them. complex tasks recommend Agent(subagent_type="plan-reviewer") / Agent(subagent_type="code-reviewer") (platform subagent API, or read the files directly); trivial / moderate use the skills directly.

Inspired by Kiro / spec-kit. The requirements.md (EARS) → design.md → tasks.md triplet is an optional alternative to docs/plans/ for cross-PR / multi-week features. AGENTS.md §6 is unchanged. See docs/specs/README.md for when to use spec vs plan.

Shared Libraries: lib/json-extract.sh, lib/session-dir.sh, lib/task-summary.py, lib/danger-patterns.sh (+ 25 unit tests), lib/pretool-response.sh.

• PreToolUse deny uses hookSpecificOutput.permissionDecision (not the deprecated {decision:"deny"} + exit 2)
• PostToolUse Bash's JSON schema lacks exit_code; uses tool_response.{stdout,stderr,interrupted} heuristic determination
• harness/rules/*.md's paths frontmatter is the official native loading mechanism

✅ Required · ◯ Optional · ➖ Skippable

See AGENTS.md §2 §6 for details.

• Harness Philosophy: 12 Factor Agents
• Skills Framework: 本仓库自带 harness/skills/（plan-review / code-review / investigate / careful-ops / workflow-management），无外部依赖
• Spec Origin: AGENTS.md adapted from Xuanwo's AI Context Gist
• Official Docs: Claude Code Hooks · Skills · Memory