mcpField
L
ln-403-task-rework

ln-403-task-rework

github.com/levnikolaevichDevelopment584371 installsUpdated 2d ago
Repository

About ln-403-task-rework

A single-task rework skill that applies reviewer feedback and returns the task for review. It follows the configured tracker provider, limits changes to the selected task, and leaves changes uncommitted.

Key features

  1. 01Load task, comments, and parent Story
  2. 02Map review feedback to focused fixes
  3. 03Run project quality checks
  4. 04Update task status and kanban
  5. 05Add handoff summary and root-cause classification

Use cases

  1. 01Fixing a task marked To Rework
  2. 02Addressing reviewer comments without expanding scope
  3. 03Returning completed fixes for review

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/levnikolaevich/claude-code-skills --skill ln-403-task-rework

After installing, type/ln-403-task-reworkin Claude Code to confirm it loaded.

Bundled files

References and scripts installed along with SKILL.md.

references/
agent_skill_roots_contract.md
audit_scoring.md
audit_summary_contract.md
audit_worker_core_contract.md
code_efficiency_criterion.md
coordinator_summary_contract.md
environment_state_contract.md
goal_articulation_gate.md
input_resolution_pattern.md
mcp_tool_preferences.md
problem_solving.md
provider_file.md
provider_github.md
provider_linear.md
references/scripts/coordinator-runtime/lib/
artifacts.mjs
cli-helpers.mjs
core.mjs
loop-health.mjs
runtime-constants.mjs
schemas.mjs
validate.mjs
worker-guards.mjs
worker-runtime.mjs
references/scripts/planning-worker-runtime/
cli.mjs
references/scripts/planning-worker-runtime/lib/
guards.mjs
phases.mjs
store.mjs
references/scripts/task-plan-worker-runtime/
cli.mjs
references/scripts/task-plan-worker-runtime/lib/
guards.mjs
phases.mjs

Details

From the project README

Paths: File paths (references/, ../ln-*) are relative to this skill directory.

Task Rework Executor

Type: L3 Worker

Executes rework for a single task marked To Rework and hands it back for review.

Purpose & Scope

  • Load full task, reviewer comments, and parent Story; understand requested changes.
  • Apply fixes per feedback, keep KISS/YAGNI, and align with guides/Technical Approach.
  • Update only this task: To Rework -> In Progress -> To Review; no other tasks touched.

Hex-line acceleration (if available): Use outline(file_path) before reading large code files. Use read_file() for discovery and read_file(edit_ready=true, verbosity="full") before any edit that needs revision and checksums. After edits: edit_file(base_revision=rev) → verify(checksums). Use changes() to show what was fixed.

Inputs

Use read_file() and edit_file() as the primary path for code/config/script/test files during rework. Keep read_file() discovery-first by default; request edit_ready=true, verbosity="full" only when you are about to reuse its revision/checksum protocol. Built-in Read/Edit are fallback only when hex-line is unavailable.

InputRequiredSourceDescription
taskIdYesargs, parent Story, kanban, userTask to rework

Resolution: Task Resolution Chain. Status filter: To Rework

Task Storage Mode

MANDATORY READ: Load references/environment_state_contract.md, references/storage_mode_detection.md, references/tracker_provider_contract.md, and references/input_resolution_pattern.md

Extract: task_provider = Task Management → Provider (linear | github | file). Operations stay provider-agnostic in this skill — see references/tracker_provider_contract.md for the canonical operation set and provider_*.md for transport binding.

Tracker operations used by this skill: getTask, getStory, listComments, updateStatus (To Rework → In Progress → To Review), addComment. Transport per provider lives in references/provider_file.md, references/provider_github.md, references/provider_linear.md.

MANDATORY READ: Load references/mcp_tool_preferences.md — ALWAYS use hex-line MCP for code files when available. No fallback to standard Read/Edit unless hex-line is down.

Workflow (concise)

  1. Resolve taskId: Run Task Resolution Chain per guide (status filter: [To Rework]).
  2. Load task: Read task and review notes via the configured tracker provider (getTask, listComments); fetch parent Story via getStory. 2b) Goal gate: MANDATORY READ: Load references/goal_articulation_gate.md — State REAL GOAL of this rework (what was actually broken, not "apply feedback"). Combine with 5 Whys (references/problem_solving.md) to ensure root cause is articulated alongside the rework goal. NOT THE GOAL: a superficial patch that addresses the symptom, not the cause.
  3. Plan fixes: Map each comment to an action; confirm no new scope added.
  4. Implement: MANDATORY READ: Load references/code_efficiency_criterion.md — Follow task plan/checkboxes; address config/hardcoded issues; update docs/tests noted in Affected Components and Existing Code Impact. Before handoff, verify 3 efficiency self-checks.
  5. Quality: Run typecheck/lint (or project equivalents); ensure fixes reflect guides/manuals/ADRs/research.
  6. Root Cause Analysis: Ask "Why did the agent produce incorrect code?" Classify: missing context | wrong pattern | unclear AC | gap in docs/templates. If doc/template gap found → update the relevant file (guide, template, CLAUDE.md) to prevent recurrence.
  7. Handoff: Set task to To Review via the configured tracker provider (updateStatus); move it in kanban; add summary comment referencing resolved feedback + root cause classification.

Critical Rules

  • Single-task only; never bulk update.
  • Do not mark Done; only To Review (the reviewer decides Done).
  • Keep language (EN/RU) consistent with task.
  • No new tests/tasks created here; only update existing tests if impacted.
  • Do NOT commit. Leave all changes uncommitted — the reviewer reviews and commits.

Runtime Summary Artifact

MANDATORY READ: Load references/coordinator_summary_contract.md, references/worker_runtime_contract.md, references/task_worker_runtime_contract.md

Shared contract:

  • emit summary_kind=task-status
  • standalone mode omits runId and summaryArtifactPath
  • managed mode passes both runId and exact summaryArtifactPath before the worker writes its validated summary

Monitor (2.1.98+): When verification commands expected >30s, use Monitor. Fallback: Bash(run_in_background=true).

Definition of Done

  • Task and review feedback fully read; actions mapped.
  • Fixes applied; docs/tests updated as required.
  • Quality checks passed (typecheck/lint or project standards).
  • Root cause classified (missing context | wrong pattern | unclear AC | doc gap); doc/template updated if gap found.
  • Status set to To Review; kanban updated; summary comment with fixed items + root cause.
  • Runtime summary artifact written to the shared task-status location.

Reference Files

  • Environment state: references/environment_state_contract.md
  • Storage mode operations: references/storage_mode_detection.md
  • [MANDATORY] Problem-solving approach: references/problem_solving.md
  • Kanban format: docs/tasks/kanban_board.md

Version: 3.0.0 Last Updated: 2025-12-23

Comments

Markdown supported