cortex-level

Level

Places, organizes, modifies, and reviews level content using the Level Designer agent.

Mode Detection

Request	Mode	Agent turns
"Place", "Add", "Spawn", "Build", "Move", "Delete", "Organize", "Adjust lighting", multi-actor layout, scene construction	Create/Modify	25
"Review", "Audit", "Check", "Analyse", "What actors", "List", spatial/organization questions	Review/Analyze	15
Ambiguous	Default to Review	15

---

Create / Modify Mode

1. Launch Level Designer Agent

Use the Task tool with subagent_type: "cortex-toolkit:level-designer" and max_turns: 25 to delegate the work.

Structure the prompt using the 3-phase directive:

Make the following level changes using the 3-phase methodology:

**Request:** [user's request verbatim]
**Prefetched state:** [embed the main-thread `prefetched_state` block here before launching]

MANDATORY WORKFLOW:
0. VERIFY: call `get_info` to confirm MCP connectivity. If it fails, invoke `cortex-editor`.
1. Use `prefetched_state` first. Do not re-fetch the same baseline unless required for the next step.
2. Read `.cortex/domains/level.md` for level conventions
3. PLAN: call `get_info`, then `list_actors` or `find_actors` to understand current state.
   Design the complete `operations[]` array before touching anything.
4. Issue independent discovery reads in parallel.
5. BATCH: call `level_compose` once with the full spec.
   - Use `stop_on_error: true` if any op references `$ops[...]` from another op in the batch
   - Use `stop_on_error: false` for independent bulk modifications
   - Pass `expected_fingerprint` on each mutation guarded by `prefetched_state`
6. VERIFY: check `completed_steps == total_steps`.
   If false, diff `spawned_actors` against plan, call `find_actors` if needed,
   then construct a MINIMAL fix batch for the gap only.
   Maximum one retry batch. If retry fails, stop and report.

TOOL SELECTION RULE:
- level_compose: 2+ spawns OR 3+ existing actor modifications OR any spawn+configure chain
- Individual tools: 1-2 existing actors with a single change (quick corrections only)

PROHIBITED: Do NOT skip the Plan phase for multi-actor work. Do NOT call get_actor
for each spawned actor to verify — check completed_steps first.

2. Agent Workflow

The Level Designer agent will:

1. Read .cortex/domains/level.md
2. Plan — inspect level state, design full operations[] spec
3. Batch — call level_compose once
4. Verify — check result, apply minimal fix if needed
5. Report summary

3. Review Agent Results

The agent returns from level_compose:

• success: true/false
• actor_count + spawned_actors: actors created or duplicated
• completed_steps / total_steps: batch progress
• failed_steps[]: per-failure with op_id, command, error_code, error

---

Review / Analyze Mode

1. Launch Level Designer Agent

Use the Task tool with subagent_type: "cortex-toolkit:level-designer" and max_turns: 15 to delegate the review.

Pass context about what to review:

Review the current level and provide a report:

**Prefetched state:** [embed the main-thread `prefetched_state` block here before launching]

1. Use `prefetched_state` first; only fetch missing data
2. Use `get_info` for level overview (name, actor count, world type, sublevels, is_world_partition)
3. Use `list_actors` to enumerate actors (paginate if needed)
4. Run independent reads in parallel where possible
5. Use `get_bounds` to understand spatial layout
6. Check folder organization — are actors organized logically?
7. Check for common issues:
   - Actors without folders
   - Duplicate labels
   - Actors at origin that shouldn't be
   - Unloaded sublevels
   - Missing data layer assignments (if World Partition)
8. Summarize findings with recommendations

2. Agent Workflow

The Level Designer agent will:

1. Query level info and actor lists
2. Analyze organization patterns
3. Check spatial distribution
4. Report findings and recommendations

3. Review Agent Results

The agent returns:

• Level overview (name, actor count, world type, is_world_partition)
• Actor breakdown by class and folder
• Spatial bounds information
• Organization issues found
• Recommendations for improvement

---

Handling Agent Results

If the agent's response includes a Status line:

• completed — present results to the user.
• blocked / partial — surface what was done, what remains, and what blocked it. For Create/Modify mode, warn the user that level changes may be incomplete.

If the agent's response has no Status line (e.g., turn limit reached mid-response), treat as partial — summarize whatever the agent produced and note the work may be incomplete.