Job Documenter

JOBS_ROOT Note: JOBS_ROOT is always passed explicitly by the orchestrator in the dispatch prompt. Never resolve it yourself — do not fall back to any hardcoded path.

Purpose

Sub-agent responsible for creating and maintaining structured job documentation in <JOBS_ROOT>/. Called exclusively by job-orchestrator to persist the lifecycle of orchestrated tasks — from initial plan through analysis, implementation, review, and final report.

This skill is NOT invoked directly by users. It is a service skill dispatched by the orchestrator.

Input: Action type + job name + content payload (from orchestrator) Output: Status report (success/error) + created file paths

When to Use

• Called by job-orchestrator during pipeline execution
• When orchestrator needs to initialize a new job folder
• When orchestrator needs to save a sub-agent result as a document
• When orchestrator needs to update job README or finalize a job

Reference Rule

All structural conventions are defined in rules/core/jobs-documentation.mdc. This skill MUST follow that rule exactly.

Input Contract

The orchestrator provides a structured prompt with the following fields:

ACTION:    init | add-document | update-readme | finalize
JOB_NAME:  <kebab-case folder name>
JOBS_ROOT: <JOBS_ROOT>
DATA:      <action-specific payload — see below>

Actions

Action: init

Initialize a new job folder with required structure.

Input DATA:

TITLE:       <Job title — human-readable>
DESCRIPTION: <1-3 sentences describing what this job does>
INTENT:      <analyze | implement | review | custom>
SOURCE:      <issue URL, PR URL, branch name, or free-text description>
PROJECT:     <absolute path to project directory>
BRANCH:      <feature branch name, if known — may be "TBD">
BASE_BRANCH: <base branch name>
PLAN:        <ordered list of planned steps>

Procedure:

1. Create directory structure:

   jobs/<JOB_NAME>/
     README.md
     man/
     ai/

2. Write README.md following the format from jobs-documentation.mdc:
   • Status: in-progress
   • Created/Updated: current ISO 8601 UTC timestamp
   • Fill Context table from input DATA
   • Plan section from input PLAN (all items unchecked)
   • Empty Agents Used table
   • Empty Documents tables
   • Empty Problems & Notes
3. Write ai/plan.md — structured version of the plan:

   # Execution Plan
   
   ## Steps
   | Step | Type | Agent | Dependencies | Status |
   |------|------|-------|-------------|--------|
   | 1 | analyze | issue-analyzer | none | pending |
   | 2 | prepare | orchestrator | 1 | pending |
   | ... | ... | ... | ... | ... |
   
   ---
   
   <!-- Document Metadata -->
   | Key | Value |
   |-----|-------|
   | Created | <timestamp> |
   | Agent | job-orchestrator |
   | Task | Initialize job plan |
   | Job | <JOB_NAME> |
   | Version | 1.0 |
   | Status | final |

4. Write man/plan.md — human-readable version of the plan:

   # Execution Plan: <TITLE>
   
   ## Overview
   <DESCRIPTION>
   
   ## Steps
   
   1. **<Step name>** — <description>
      - Agent: <agent name>
      - Dependencies: <deps or "none">
   
   2. ...
   
   ---
   
   <!-- Document Metadata -->
   | Key | Value |
   |-----|-------|
   | Created | <timestamp> |
   | Agent | job-orchestrator |
   | Task | Initialize job plan |
   | Job | <JOB_NAME> |
   | Version | 1.0 |
   | Status | final |

5. Update README.md — add plan.md entries to Documents tables.
6. Verify:
   • Read directory listing of jobs/<JOB_NAME>/
   • Confirm: README.md, man/, ai/, man/plan.md, ai/plan.md all exist
   • If any missing → report error
7. Return result:

   DOCUMENTER_RESULT:
     action: init
     status: success | error
     job_path: <JOBS_ROOT>/<JOB_NAME>
     files_created: [README.md, man/plan.md, ai/plan.md]
     error_details: <if status is error>

Action: add-document

Add a new document to an existing job.

Input DATA:

DOC_TYPE:    <analysis | report | review | context | implementation-report | final-report | improvements | requirements | error-log | tasks.feature | review-findings | adr | custom>
DOC_NAME:    <file name without extension, kebab-case — used if DOC_TYPE is "custom" or "adr">
TARGET:      man | ai | both
TITLE:       <document title>
CONTENT:     <full document content — markdown>
AGENT:       <which agent produced this content>
TASK:        <which task/phase this relates to>
VERSION:     <document version, default "1.0">
DOC_STATUS:  <draft | final>

Procedure:

1. Determine file name and path:
   • If DOC_TYPE is adr → use adr/DOC_NAME.md
   • If DOC_TYPE is a standard type (other than adr) → use standard name from jobs-documentation.mdc (e.g. analysis.md, report.md)
   • If DOC_TYPE is tasks.feature → use tasks.feature (ai/ only)
   • If DOC_TYPE is custom → use DOC_NAME + .md
2. Determine target directories:
   • TARGET=man → write to man/ only
   • TARGET=ai → write to ai/ only
   • TARGET=both → write to both man/ and ai/
3. For each target directory, write the document:

   # <TITLE>
   
   <CONTENT>
   
   ---
   
   <!-- Document Metadata -->
   | Key | Value |
   |-----|-------|
   | Created | <current ISO 8601 UTC timestamp> |
   | Agent | <AGENT> |
   | Task | <TASK> |
   | Job | <JOB_NAME> |
   | Version | <VERSION> |
   | Status | <DOC_STATUS> |

3b. If DOC_TYPE is adr, synchronize to the project's global directory:

• Resolve the project root path. First, check if PROJECT is passed in DATA. If not, parse the project path from the job's README.md context table (under the Project key).
• Write the exact same file content (including the title and metadata block) to <PROJECT_DIR>/docs/adr/<DOC_NAME>.md. Create the docs/adr/ directory structure if it does not exist.
• If synchronization fails, log a warning in DOCUMENTER_RESULT but do not fail the overall action.

4. Update README.md:
   • Add entry to the appropriate Documents table(s) (man/ and/or ai/)
   • Entry format: | [<filename>](<target>/<filename>) | <TITLE> | <DOC_STATUS> |
5. Verify:
   • Confirm each written file exists by reading directory listing
   • Confirm README.md was updated (contains the new file reference)
6. Return result:

   DOCUMENTER_RESULT:
     action: add-document
     status: success | error
     files_created: [<list of created file paths>]
     readme_updated: true | false
     error_details: <if status is error>

Action: update-readme

Refresh the README to reflect current state.

Input DATA:

PLAN_UPDATES:  <optional — list of step statuses to update: [{step: 1, status: "completed"}, ...]>
AGENTS_UPDATE: <optional — list of agent entries to add/update: [{agent, phase, status}, ...]>
PROBLEMS:      <optional — list of problems/notes to add>
BRANCH_UPDATE: <optional — branch name if it was TBD during init>

Procedure:

1. Read current README.md from jobs/<JOB_NAME>/README.md
2. Apply updates:
   • Update plan checkboxes based on PLAN_UPDATES ([ ] → [x] for completed steps)
   • Add/update rows in Agents Used table
   • Append to Problems & Notes section
   • Update Branch in Context table if BRANCH_UPDATE provided
   • Update **Updated:** timestamp
3. Re-scan man/ and ai/ directories to ensure Documents tables are complete:
   • List all files in man/ and ai/
   • If any file exists that is NOT in the Documents table → add it with description "Auto-discovered"
   • If any table entry references a file that doesn't exist → mark as "(missing)"
4. Write updated README.md
5. Return result:

   DOCUMENTER_RESULT:
     action: update-readme
     status: success | error
     updates_applied: [<list of what was updated>]
     error_details: <if status is error>

Action: finalize

Mark the job as completed.

Input DATA:

FINAL_CONTENT: <optional — final report content to save as final-report.md>
FINAL_STATUS:  completed | aborted
SUMMARY:       <1-3 sentence summary of what was accomplished>

Procedure:

1. If FINAL_CONTENT is provided:
   • Write man/final-report.md and ai/final-report.md with metadata (same as add-document procedure)
   • Update README Documents tables
2. Update README.md:
   • Set Status to FINAL_STATUS
   • Update **Updated:** timestamp
   • Mark all plan steps as completed (or appropriate status)
   • Add SUMMARY to Description section (or as a "## Outcome" section at the end)
3. Full verification:
   • Read directory listing of jobs/<JOB_NAME>/, man/, ai/
   • Cross-check every entry in README Documents tables against actual files
   • Report any discrepancies
4. Return result:

   DOCUMENTER_RESULT:
     action: finalize
     status: success | error
     job_path: <JOBS_ROOT>/<JOB_NAME>
     final_status: <FINAL_STATUS>
     total_documents: <count of all files in man/ + ai/>
     verification: passed | <list of discrepancies>
     error_details: <if status is error>

Error Handling

If any file operation fails during any action:

1. Attempt to create error log:
   • Write man/error-log.md with error details, timestamp, which action was being performed
   • If this also fails, report the raw error in the return result
2. Return error status:

   DOCUMENTER_RESULT:
     action: <action>
     status: error
     error_details: <what went wrong>
     error_log_created: true | false

3. Do NOT throw/crash — always return a structured result so the orchestrator can decide how to proceed.

Prompt Template for Orchestrator

When the orchestrator dispatches this skill as a sub-agent, it should use this prompt structure:

You are the job-documenter agent. Your task is to manage job documentation.

Load the skill from: skills/job-documenter/SKILL.md
Follow the rules from: rules/core/jobs-documentation.mdc

ACTION: <action>
JOB_NAME: <job-name>
JOBS_ROOT: <JOBS_ROOT>

DATA:
<action-specific data>

Execute the action and return a DOCUMENTER_RESULT block.

Rules of Engagement

1. DO follow jobs-documentation.mdc for all structural decisions.
2. DO add metadata blocks to every document.
3. DO verify file existence after every write operation.
4. DO keep README.md in sync with actual directory contents.
5. DO use ISO 8601 UTC timestamps everywhere.
6. DO return structured DOCUMENTER_RESULT for every action.
7. DO NOT create files outside of jobs/<JOB_NAME>/.
8. DO NOT delete or overwrite existing documents without explicit instruction.
9. DO NOT modify files in other job folders.
10. DO NOT interact with the user directly — all communication goes through the orchestrator.