Agent Skills: Implementation Task Executor

Paths: File paths (shared/, references/, ../ln-*) are relative to skills repo root. If not found at CWD, locate this SKILL.md directory and go up one level for repo root. If shared/ is missing, fetch files via WebFetch from https://raw.githubusercontent.com/levnikolaevich/claude-code-skills/master/skills/{path}.

Implementation Task Executor

Type: L3 Worker

Executes a single implementation (or refactor) task from Todo to To Review using the task description and linked guides.

Purpose & Scope

• Handle one selected task only; never touch other tasks.
• Follow task Technical Approach/plan/AC; apply KISS/YAGNI and guide patterns.
• Update Linear/kanban for this task: Todo -> In Progress -> To Review.
• Run typecheck/lint; update docs/tests/config per task instructions.
• Not for test tasks (label "tests" goes to ln-404-test-executor).

Hex MCP acceleration (if available): Use narrow inspect_path(path=<relevant dir>) and outline(path) before reading large code files; avoid repo-root wildcard inventories unless you intentionally need one. Use grep_search(output="summary") first for search/discovery, then escalate to output="content", edit_ready=true only when you need canonical hunks for follow-up edits; use allow_large_output=true only as an explicit last resort. Use read_file() in discovery mode for structure and targeted ranges; when you need revision and checksums for a follow-up edit, refresh that file with read_file(edit_ready=true, verbosity="full") before edit_file(). For edits to existing code, run index_project(path=project_root) once and use analyze_edit_region(...) before non-trivial edits. After edits: edit_file(base_revision=rev) -> verify(checksums). Before handoff: changes() to review diff.

Inputs

| Input | Required | Source | Description | |-------|----------|--------|-------------| | taskId | Yes | args, parent Story, kanban, user | Task to execute |

Resolution: Task Resolution Chain. Status filter: Todo

Task Storage Mode

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

Extract: task_provider = Task Management → Provider (linear | file).

| Aspect | Linear Mode | File Mode | |--------|-------------|-----------| | Load task | get_issue(task_id) | Read("docs/tasks/epics/.../tasks/T{NNN}-*.md") | | Update status | save_issue(id, state) | Edit the **Status:** line in file | | Kanban | Updated by Linear sync | Must update kanban_board.md manually |

File Mode status format:

## Status
**Status:** In Progress | **Priority:** High | **Estimate:** 4h

MANDATORY READ: Load shared/references/mcp_tool_preferences.md and shared/references/mcp_integration_patterns.md — use hex-line MCP for code files when available and hex-graph for semantic edit-risk questions.

Mode Detection

Detect operating mode at startup:

Plan Mode Active:

• Steps 1-2: Load task context (read-only, OK in plan mode)
• Generate EXECUTION PLAN (files to create/modify, approach) → write to plan file
• Call ExitPlanMode → STOP. Do NOT implement.
• Steps 3-6: After approval → execute implementation

Normal Mode:

• Steps 1-6: Standard workflow without stopping

Progress Tracking with TodoWrite

When operating in any mode, skill MUST create detailed todo checklist tracking ALL steps.

Rules:

1. Create todos IMMEDIATELY before Step 1
2. Each workflow step = separate todo item; implementation step gets sub-items
3. Mark in_progress before starting step, completed after finishing

Todo Template (10 items):

Step 1: Resolve taskId
  - Resolve via args / Story context / kanban / AskUserQuestion (Todo filter)

Step 2: Load Context
  - Fetch full task description + linked guides/manuals/ADRs

Step 2b: Goal Articulation Gate
  - Complete 4 questions from shared/references/goal_articulation_gate.md (<=25 tokens each)

Step 2c: Implementation Blueprint
  - From task "Affected Components": extract file paths (Glob/Grep or narrow `inspect_path(path=<component dir>)` to find actual paths)
  - Read each file (or key sections) to understand current structure
  - IF modifying existing code in supported languages: `index_project(path=project_root)` once, then use path-scoped `find_symbols` / `inspect_symbol` for exact symbol identity and `analyze_edit_region` before editing non-trivial ranges
  - IF `find_symbols` returns `truncated: true` or a broad candidate set: refine to `name + file` or `workspace_qualified_name` before planning from it
  - Output:
    ## Implementation Blueprint: {taskId}
    **Files to create:** [list with brief purpose]
    **Files to modify:** [list with what changes]
    **Change order (dependencies first):**
    1. {file} — {what and why first}
    2. {file} — {depends on step 1}
    **Risks:** {shared files with parallel tasks, if any}
  - Scope: ONLY files of this task. Do not analyze other tasks.

  - **Checkpoint:** Emit PHASE_3 checkpoint with structured `blueprint` payload:
    `{ "blueprint": { "change_order": [{ "file": "...", "action": "create|modify", "reason": "..." }] } }`
    Guard blocks PHASE_4 if blueprint is missing from the checkpoint.

Step 3: Start Work
  - Set task to In Progress, update kanban

Step 4: Implement
  - 4a Pattern Reuse: IF creating new file/utility, Grep src/ for existing similar patterns
    (error handlers, validators, HTTP wrappers, config loaders). Reuse if found.
  - 4b Follow task plan/AC, apply KISS/YAGNI
  - 4c Architecture Guard: IF creating service function: (1) 3+ side-effect categories in **leaf** function → split (EXCEPT orchestrator functions that delegate sequentially — these are expected to have 3+ categories);
    (2) get_*/find_*/check_* naming → verify no hidden writes; (3) 3+ service imports in **leaf** function → flatten (orchestrator imports are expected)
    (4) **Frontend Guard (conditional):** IF affected files include `.tsx/.vue/.svelte/.html/.css` → **MANDATORY READ:** Load `shared/references/frontend_design_guide.md`. Load project's design_guidelines.md if exists (design tokens source of truth). Verify: one composition per viewport; max 2 typefaces + 1 accent color; cards only when interaction requires; motion max 2-3 purposeful; WCAG 2.1 AA contrast (4.5:1 text, 3:1 UI elements)
  - Update docs and existing tests if impacted
  - Execute verify: methods from task AC (test/command/inspect)

Step 5: Quality
  - Run typecheck and lint (or project equivalents)
  - 5b Blueprint Completion: compare actual changes to blueprint; emit blueprint_status in PHASE_6 checkpoint

Step 6: Finish
  - Set task to To Review, update kanban
  - Add summary comment (changes, tests, docs)

Workflow (concise)

1. Resolve taskId: Run Task Resolution Chain per guide (status filter: [Todo]).
2. Load context: Fetch full task description (Linear: get_issue; File: Read task file); read linked guides/manuals/ADRs/research; auto-discover team/config if needed. 2b) Goal gate: MANDATORY READ: Load shared/references/goal_articulation_gate.md — Complete the 4-question gate (<=25 tokens each). State REAL GOAL (deliverable as subject), DONE LOOKS LIKE, NOT THE GOAL, INVARIANTS & HIDDEN CONSTRAINTS. 2c) Implementation Blueprint: From task "Affected Components", find actual file paths via Glob/Grep or narrow inspect_path(path=<component dir>). Read key sections of each file. If task changes existing code in supported languages, build graph context once (index_project) and use path-scoped find_symbols / inspect_symbol for exact symbol identity plus analyze_edit_region before editing non-trivial ranges. If symbol discovery is truncated, refine to name + file or workspace_qualified_name before planning from it. Output structured plan: files to create/modify, change order (dependencies first), risks (shared files with parallel tasks, external callers, public API surfaces). Scope: this task only.
3. Start work: Update this task to In Progress (Linear: update_issue; File: Edit status line); move it in kanban (keep Epic/Story indent).
4. Implement (with verification loop): Before writing new utilities/handlers, Grep src/ for existing patterns (error handling, validation, config access). Reuse if found; if not reusable, document rationale in code comment. For edits to existing functions, classes, routes, or middleware, run analyze_edit_region first and account for external callers, clone siblings, downstream flow, and public API risk. Follow checkboxes/plan; keep it simple; avoid hardcoded values; reuse existing components; update docs noted in Affected Components; update existing tests if impacted (no new tests here). Before creating service functions, apply Architecture Guard (cascade depth, interface honesty, flat orchestration; for frontend files: MANDATORY READ shared/references/frontend_design_guide.md, load design_guidelines.md if exists, verify composition/typography/WCAG rules). After implementation, execute verify: methods from task AC: test → run specified test; command → execute and check output; inspect → verify file/content exists. If any verify fails → fix before proceeding.
5. Quality: Run typecheck and lint (or project equivalents); ensure instructions in Existing Code Impact are addressed. Blueprint Completion: compare actual changes against blueprint from Step 2c — for each planned file mark completed or skipped (with justification), for each unplanned file mark added (with justification). Emit PHASE_6 checkpoint: { "blueprint_status": { "planned_count": N, "completed": [...], "skipped": [{"file":"...","justification":"..."}], "added": [{"file":"...","justification":"..."}], "completion_pct": N } }. Guard blocks PHASE_7 if missing.
6. Finish: Mark task To Review (Linear: update_issue; File: Edit status line); update kanban to To Review; add summary comment (what changed, tests run, docs touched).

Pre-Submission Checklist

Context: Self-assessment before To Review reduces review round-trips and catches obvious issues early.

MANDATORY READ: Load shared/references/code_efficiency_criterion.md — self-check before submission.

Before setting To Review, verify all items:

| # | Check | Verify | |---|-------|--------| | 0 | AC verified | Each AC verify: method executed with pass evidence | | 1 | Approach alignment | Implementation matches Story Technical Approach | | 2 | Clean code | No dead code, no backward-compat shims, unused imports removed | | 3 | Config hygiene | No hardcoded creds/URLs/magic numbers | | 4 | Docs updated | Affected Components docs reflect changes | | 5 | Tests pass | Existing tests still pass after changes | | 6 | Pattern reuse | New utilities checked against existing codebase; no duplicate patterns introduced | | 7 | Architecture guard | Cascade depth <= 2 (leaf functions); no hidden writes in read-named functions; no service chains >= 3 in leaf functions (orchestrator imports exempt). Frontend files: composition, typography, WCAG per shared/references/frontend_design_guide.md | | 8 | Destructive op safety | If task has "Destructive Operation Safety" section: (1) backup step executed/planned before destructive code, (2) rollback mechanism exists in code, (3) environment guard present, (4) preview/dry-run evidence attached or referenced | | 9 | Code efficiency | No unnecessary intermediates, verbose patterns replaced by language idioms, no boilerplate framework handles (per shared/references/code_efficiency_criterion.md) | | 10 | Blueprint complete | All blueprint items completed or skipped with justification; blueprint_status emitted in PHASE_6 checkpoint |

MANDATORY READ: Load shared/references/destructive_operation_safety.md for severity classification and safety requirements.

HITL Gate: IF task severity = CRITICAL (per destructive_operation_safety.md loaded above): Use AskUserQuestion to confirm before marking To Review: "Task contains CRITICAL destructive operation: {operation}. Backup plan: {plan}. Proceed?" Do NOT mark To Review until user confirms.

If any check fails: Fix before setting To Review. Do not rely on reviewer to catch preventable issues.

Critical Rules

• Single-task updates only; no bulk status changes.
• Keep language of the task (EN/RU) in edits/comments.
• No code snippets in the description; code lives in repo, not in Linear.
• No new test creation; only update existing tests if required.
• Preserve Foundation-First ordering from orchestrator; do not reorder tasks.
• Do NOT commit. Leave all changes uncommitted — the reviewer reviews and commits.
• Before non-trivial edits to existing code, use graph impact evidence when available instead of guessing blast radius from file names alone.

Runtime Summary Artifact

MANDATORY READ: Load shared/references/coordinator_summary_contract.md, shared/references/worker_runtime_contract.md, shared/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 selected and set to In Progress; kanban updated accordingly.
• [ ] Guides/manuals/ADRs/research read; approach aligned with task Technical Approach.
• [ ] Implementation completed per plan/AC; each AC verify: method executed with pass evidence.
• [ ] Docs and impacted tests updated.
• [ ] Typecheck and lint passed (or project quality commands) with evidence in comment.
• [ ] Task set to To Review; kanban moved to To Review; summary comment added.
• [ ] Runtime summary artifact written to the shared task-status location.

Reference Files

• Environment state: shared/references/environment_state_contract.md
• Storage mode operations: shared/references/storage_mode_detection.md
• Guides/manuals/ADRs/research: docs/guides/, docs/manuals/, docs/adrs/, docs/research/
• Kanban format: docs/tasks/kanban_board.md

Version: 3.0.0 Last Updated: 2025-12-23