跳到正文
cccskills
免费在 GitHub 公开

frontmatter-parsing

YAML frontmatter parsing and manipulation for .planning/ documents. Provides read, write, update, query, and validation operations on frontmatter blocks in GSD markdown artifacts.

中文简介正在准备中。当前显示原文说明。

查看安装方法

包含的文件(2)

  • SKILL.md7.0 KB
  • README.md1.7 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

status: in-progress phase: 72 wave: 1 depends_on: [] files_modified: ["src/auth/oauth.ts", "src/auth/tokens.ts"] created: 2026-03-02 updated: 2026-03-02

Plan content below...


## Capabilities

### 1. Parse Frontmatter

Extract frontmatter from a markdown file into structured data:

```yaml
# Input file: .planning/phase-72/PLAN-1.md
---
status: planned
phase: 72
plan_number: 1
wave: 1
depends_on: []
files_modified:
  - src/auth/oauth.ts
  - src/auth/tokens.ts
  - src/middleware/auth.ts
task_count: 4
created: 2026-03-02
gap_closure: false
---

Parsed result:

{
  "status": "planned",
  "phase": 72,
  "plan_number": 1,
  "wave": 1,
  "depends_on": [],
  "files_modified": ["src/auth/oauth.ts", "src/auth/tokens.ts", "src/middleware/auth.ts"],
  "task_count": 4,
  "created": "2026-03-02",
  "gap_closure": false
}

2. Extract Specific Fields

Read individual fields without parsing the entire frontmatter:

get_field(.planning/phase-72/PLAN-1.md, "wave") -> 1
get_field(.planning/phase-72/PLAN-1.md, "status") -> "planned"
get_field(.planning/phase-72/PLAN-1.md, "files_modified") -> ["src/auth/oauth.ts", ...]

3. Update Fields

Update individual frontmatter fields without modifying body content:

update_field(.planning/phase-72/PLAN-1.md, "status", "executed")
update_field(.planning/phase-72/PLAN-1.md, "wave", 2)
update_field(.planning/phase-72/PLAN-1.md, "updated", "2026-03-02")

Uses Edit tool to surgically replace only the target field line.

4. Add New Fields

Add fields to existing frontmatter:

add_field(.planning/phase-72/PLAN-1.md, "executed_at", "2026-03-02T14:30:00Z")
add_field(.planning/phase-72/PLAN-1.md, "executor_agent", "gsd-executor")

Inserts new field before the closing --- delimiter.

5. Remove Fields

Remove fields from frontmatter:

remove_field(.planning/phase-72/PLAN-1.md, "gap_closure")

6. Query Across Files

Find documents matching frontmatter criteria:

query(directory: ".planning/phase-72/", field: "wave", value: 1)
-> [".planning/phase-72/PLAN-1.md", ".planning/phase-72/PLAN-2.md"]

query(directory: ".planning/", field: "status", value: "planned", recursive: true)
-> [".planning/phase-72/PLAN-1.md", ".planning/phase-73/PLAN-1.md"]

Uses Grep to search for field patterns, then validates matches.

7. Validate Frontmatter

Validate frontmatter against document type schema:

# PLAN.md required fields
required:
  - status       # planned|executing|executed|verified
  - phase        # integer or decimal
  - plan_number  # integer
  - wave         # integer (execution order group)
  - task_count   # integer

# PLAN.md optional fields
optional:
  - depends_on      # array of plan references
  - files_modified  # array of file paths
  - gap_closure     # boolean
  - created         # date
  - updated         # date
  - executed_at     # datetime

Document type schemas:

  • PLAN.md: status, phase, plan_number, wave, task_count
  • SUMMARY.md: phase, status, tasks_completed, commits
  • RESEARCH.md: phase, status, approach_count, recommended
  • STATE.md: last_updated, session_count, current_milestone
  • CONTEXT.md: phase, decisions_count, preferences

8. Batch Operations

Update frontmatter across multiple files:

batch_update(
  files: [".planning/phase-72/PLAN-1.md", ".planning/phase-72/PLAN-2.md"],
  field: "status",
  value: "executed"
)

Tool Use Instructions

Parsing Frontmatter

  1. Use Read to load the target file
  2. Extract content between first --- and second ---
  3. Parse YAML content into key-value pairs
  4. Return structured object

Updating a Field

  1. Use Read to load the file
  2. Locate the target field line within frontmatter
  3. Use Edit with the exact field line as old_string
  4. Replace with new field value line
  5. If field is multi-line (arrays), handle properly

Querying Across Files

  1. Use Glob to find candidate files in the target directory
  2. Use Grep to search for the field pattern (e.g., ^wave: 1$)
  3. For each match, verify it's within frontmatter (not body content)
  4. Return list of matching file paths

Validating Frontmatter

  1. Use Read to load the file
  2. Parse frontmatter
  3. Check required fields exist for the document type
  4. Validate field types (string, integer, array, boolean, date)
  5. Return validation result with any errors

Process Integration

  • plan-phase.js - Read/write plan frontmatter (wave, depends_on, files_modified)
  • execute-phase.js - Update plan status as execution progresses, read wave assignments for parallel execution
  • audit-milestone.js - Query all plans by status for completion checks
  • research-phase.js - Write research frontmatter (approach_count, recommended)
  • discuss-phase.js - Read phase metadata from roadmap sections

Output Format

{
  "operation": "parse|get|update|add|remove|query|validate|batch",
  "status": "success|error",
  "file": ".planning/phase-72/PLAN-1.md",
  "field": "status",
  "previousValue": "planned",
  "newValue": "executed",
  "frontmatter": {},
  "queryResults": [],
  "validation": {
    "valid": true,
    "errors": [],
    "warnings": []
  }
}

Configuration

SettingDefaultDescription
frontmatterDelimiter---YAML frontmatter delimiter
strictValidationfalseFail on unknown fields
autoUpdateTimestamptrueAuto-update updated field on writes

Error Handling

ErrorCauseResolution
No frontmatter foundFile missing --- delimitersAdd frontmatter block or skip
YAML parse errorMalformed YAML in frontmatterFix YAML syntax (check indentation, colons, quotes)
Field not foundRequested field not in frontmatterReturn null, optionally add field
Type mismatchField value does not match expected typeCoerce if possible, error if not
Edit collisionField line not unique in fileInclude surrounding context for uniqueness

Constraints

  • Frontmatter must be valid YAML between --- delimiters at the top of the file
  • Field updates must not modify body content below the frontmatter
  • Array fields must use YAML list syntax (either inline [a, b] or block - a\n- b)
  • Date fields must use ISO 8601 format
  • Boolean fields must use true/false (not yes/no)
  • Query operations are read-only; they never modify files
  • Batch operations are atomic per file (each file update succeeds or fails independently)

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

同一仓库的技能

Provide implementations of advanced data structures

中文简介正在准备中。当前显示原文说明。

a5c-ai/babysitter18622026年9月17日 更新

Fresh adversarial code review with binary PASS/FAIL verdicts, evidence citations, and anchoring bias prevention via fresh reviewer spawning.

中文简介正在准备中。当前显示原文说明。

a5c-ai/babysitter18622026年9月17日 更新

WASM-based instant code transforms for simple tasks, achieving 352x speedup over LLM inference with zero cost.

中文简介正在准备中。当前显示原文说明。

a5c-ai/babysitter18622026年9月17日 更新

Coordinate Crew (persistent) and Polecat (transient) agents using Gas Town's hook-based work distribution and GUPP principle.

中文简介正在准备中。当前显示原文说明。

a5c-ai/babysitter18622026年9月17日 更新

Generate visual representations of algorithm execution

中文简介正在准备中。当前显示原文说明。

a5c-ai/babysitter18622026年9月17日 更新

Hierarchical coordination and drift detection with frequent checkpoints, shared memory coherence validation, role specialization enforcement, and short task cycles.

中文简介正在准备中。当前显示原文说明。

a5c-ai/babysitter18622026年9月17日 更新

查看 a5c-ai 的全部技能

报告此技能的问题