@hailer/mcp 1.1.3 → 1.1.4

package/.opencode/commands/swarm.md

@@ -0,0 +1,210 @@
+---
+description: Orchestrator-driven parallel swarm for large-scale tasks
+argument-hint: "description of what to do across many items"
+allowed-tools: Task, Bash, Read, Glob
+---
+
+# Swarm
+
+Orchestrator-driven parallel execution: analyze the task, discover work items, pick agents, split into chunks, dispatch workers, aggregate results.
+
+**Goal:** $ARGUMENTS
+
+## Protocol
+
+### Step 1: Analyze the Task
+
+Read `$ARGUMENTS` and determine:
+
+- **What needs to be done?** (review, edit, check, migrate, update, clean up)
+- **What are the work items?** (files, activities, workflows, agents, skills, configs)
+- **What scope?** (explicit path/glob from user, or infer from task description)
+
+If the scope is unclear, use Glob or Bash to discover items:
+
+```
+# Files
+Glob("src/**/*.ts")
+Glob(".claude/agents/*.md")
+Glob("workspace/*/fields.ts")
+
+# Activities (use Kenji to count/list)
+Task(subagent_type="agent-kenji-data-reader", prompt="Count and list activity IDs for [workflow]")
+```
+
+If fewer than 5 items are discovered, **abort swarm** and suggest running a regular squad or single agent instead. Swarm is for scale.
+
+### Step 2: Pick Agent(s)
+
+Based on the task type, select the right agent(s):
+
+| Task Type | Agent | Examples |
+|-----------|-------|---------|
+| Code review, security audit | **Svetlana** | "review for bugs", "security check", "find vulnerabilities" |
+| Dead code, unused imports, type errors | **Lars** | "find dead code", "check types", "unused imports" |
+| Small edits, string replacements, rule additions | **Simple Writer** | "add header to each file", "replace X with Y", "update rule" |
+| Code cleanup, simplification | **Code Simplifier** | "clean up", "simplify", "refactor for clarity" |
+| Config audit, structure validation | **Bjorn** | "check config", "validate structure", "audit settings" |
+| Data state checks, field validation | **Kenji** | "check all activities", "verify field values", "find missing data" |
+| Query patterns, data analysis | **Viktor** | "analyze patterns", "find anomalies across workflows" |
+| Activity updates, bulk mutations | **Dmitri** | "update all activities", "move to phase", "set field values" |
+
+**Multi-agent swarms:** If the task spans multiple domains, assign different agent types to different items. For example, "clean up the project" might use Lars for `.ts` files, Bjorn for config files, and Code Simplifier for app components.
+
+### Step 3: Calculate Workers
+
+```
+items_count = total discovered items
+items_per_worker = 10 (default)
+max_workers = 5
+
+worker_count = min(ceil(items_count / items_per_worker), max_workers)
+```
+
+Split items into `worker_count` chunks. Try to keep chunks roughly equal size.
+
+For multi-agent swarms, group items by agent type first, then chunk within each type.
+
+### Step 4: Dispatch Workers
+
+Launch all workers in parallel using multiple Task tool calls in a single message. All workers run in background (`run_in_background: true`).
+
+Each worker gets:
+
+```
+Task(subagent_type="[selected-agent]",
+  run_in_background: true,
+  prompt="SWARM WORKER [N] of [TOTAL]
+
+TASK: [User's description from $ARGUMENTS]
+
+YOUR ITEMS ([count]):
+- [item 1]
+- [item 2]
+- [item 3]
+...
+
+Process each item and report results in this format:
+
+RESULTS:
+- [item]: [status] - [details]
+- [item]: [status] - [details]
+
+SUMMARY:
+- Processed: [count]
+- Passed: [count]
+- Issues found: [count]
+- Failed: [count]")
+```
+
+Tell the user:
+```
+Launched [worker_count] workers across [items_count] items.
+Running in background - what else do you want to work on?
+```
+
+### Step 5: Collect Results
+
+When workers complete (check via Read on output files or TaskOutput), aggregate all results.
+
+If any worker failed or timed out:
+- Report which items were in the failed worker's batch
+- Offer to retry just those items
+
+### Step 6: Report
+
+```markdown
+## Swarm Report
+
+### Task
+[User's description]
+
+### Execution
+- **Items:** [total count]
+- **Workers:** [worker count] x [agent type(s)]
+- **Duration:** [time from dispatch to last worker completion]
+
+---
+
+### Results by Worker
+
+#### Worker 1 ([agent type]) - [X items]
+- [item]: [status] - [details]
+- [item]: [status] - [details]
+...
+
+#### Worker 2 ([agent type]) - [X items]
+...
+
+---
+
+### Aggregate Summary
+
+| Status | Count |
+|--------|-------|
+| Passed / Clean | X |
+| Issues Found | X |
+| Failed / Error | X |
+
+### Issues Found
+[Grouped and deduplicated list of all issues across all workers]
+
+1. **[item]**: [issue description]
+2. **[item]**: [issue description]
+...
+
+### Failed Items
+[If any workers failed, list their items for retry]
+
+### Next Steps
+- [Suggested follow-up based on results]
+- [If issues found: suggest /hotfix-squad or Simple Writer to fix]
+- [If edits made: suggest /review-squad to verify]
+```
+
+## Options
+
+| Flag | Effect |
+|------|--------|
+| `--workers=N` | Override worker count (default: auto-calculated, max 5) |
+| `--dry-run` | Show what would be dispatched without actually running |
+| `--no-bg` | Run workers in foreground (wait for all to complete) |
+
+## Examples
+
+```
+# Review all source files for security
+/swarm "review all TypeScript files in src/ for security vulnerabilities"
+→ Discovers 47 .ts files, spawns 5 Svetlana workers
+
+# Bulk edit agent files
+/swarm "add a '## Version History' section to every agent file"
+→ Discovers 27 agent .md files, spawns 3 Simple Writer workers
+
+# Validate workspace configs
+/swarm "check all workflow field definitions for missing descriptions"
+→ Discovers 12 fields.ts files, spawns 2 Kenji workers
+
+# Multi-agent cleanup
+/swarm "clean up the entire apps/ directory"
+→ Discovers 34 files: 20 .tsx → Code Simplifier, 8 .ts → Lars, 6 configs → Bjorn
+→ Spawns 5 workers across 3 agent types
+
+# Data validation
+/swarm "verify all Customer activities have a valid email field"
+→ Kenji counts 200 customers, spawns 5 Kenji workers (40 each)
+
+# Dry run to preview
+/swarm "update all skills to use consistent headers" --dry-run
+→ Shows: "Would dispatch 4 Simple Writer workers across 32 skill files"
+```
+
+## Notes
+
+- Minimum 5 items to trigger swarm (below that, use a regular agent)
+- Default max 5 workers to avoid overwhelming the system
+- Workers run in background by default so user can keep working
+- Multi-agent swarms naturally emerge from the orchestrator's routing logic
+- For write operations (Dmitri, Simple Writer), consider running --dry-run first
+- Swarm vs Squad: Swarm is raw parallel execution without quality gates. Workers operate independently. Squads have convergence, synthesis, and loops. Use swarm for bulk independent items, squads for coordinated workflows.
+- Swarm + squad can combine: `/swarm "review src/"` finds issues, then user can run `/hotfix-squad` on results. This is a manual handoff - user confirms.

package/.opencode/commands/tool-builder.md

@@ -1,10 +1,15 @@
 ---
-description: Build a new MCP tool
+description: Activate tool-builder mode to implement a new MCP tool
 ---
 
-Enter tool-building mode.
+# Tool Builder Mode
+
+Loading `tool-builder` skill for MCP tool patterns...
+
+**What tool should I build?**
+
+Please provide:
 
-Ask the user for:
 - **Tool Name**: `tool_name` (snake_case)
 - **Purpose**: What does it do?
 - **API Endpoint**: `v3.endpoint.method`
@@ -13,7 +18,9 @@ Ask the user for:
 - **Target File**: Which file in `src/mcp/tools/`?
 - **Tool Group**: READ | WRITE | PLAYGROUND
 
-Example:
+---
+
+**Example:**
 ```
 Tool Name: list_insights
 Purpose: List all insights in workspace
@@ -24,4 +31,9 @@ Target File: insight.ts
 Tool Group: PLAYGROUND
 ```
 
-Then implement the tool following patterns in the target file.
+I'll:
+1. Load tool-builder skill
+2. Read existing patterns in target file
+3. Implement the tool
+4. Register in src/app.ts
+5. Report what was created

package/.opencode/commands/ws-pull.md

@@ -1,19 +1,44 @@
 ---
-description: Pull workspace configuration from Hailer
+description: Pull workspace configuration from Hailer (READ-ONLY)
 ---
 
-Pull workspace configuration from Hailer.
+Pull the current workspace configuration from Hailer and save it locally.
 
-1. Get WORKSPACE_CONFIG_PATH:
-!`grep WORKSPACE_CONFIG_PATH .env.local 2>/dev/null || echo "NOT SET"`
+**What this does:**
+- Downloads workflows, insights, teams, groups, and phases from Hailer workspace
+- Saves configuration as JSON files to the `WORKSPACE_CONFIG_PATH` directory
+- READ-ONLY operation - does not modify anything on Hailer
+- Safe to run anytime
 
-2. If set, run pull command in that directory:
-```bash
-cd "$WORKSPACE_CONFIG_PATH" && npm run ws-pull
-```
+**Requirements:**
+- Must have `WORKSPACE_CONFIG_PATH` set in `.env.local`
+- Hailer credentials must be configured
 
-3. After pull, read the generated files to load context:
-   - `workspace/enums.ts` - field and phase IDs
-   - List `workspace/` to see workflows
+**Instructions for Claude:**
 
-Report what was pulled.
+1. Read `.env.local` file to get `WORKSPACE_CONFIG_PATH`
+2. Run: `cd "$WORKSPACE_CONFIG_PATH" && npm run pull`
+3. **AUTOMATICALLY read the generated files to load context:**
+   - Read `workspace/enums.ts` (complete file - contains ALL field and phase IDs)
+   - List `workspace/` directory to see all workflow folders
+   - Read 1-2 example workflow configs (main.ts, fields.ts, phases.ts) to understand structure
+4. Provide summary of what workflows were pulled and confirm context is loaded
+
+**What gets downloaded:**
+- `workspace/enums.ts` - Type-safe field and phase ID enums for ALL workflows
+- `workspace/workflows.ts` - Main workflows export
+- `workspace/insights.ts` - SQL insight definitions
+- `workspace/teams.ts` - Team structures
+- `workspace/groups.ts` - User groups
+- `workspace/<workflow_name>/` - Individual workflow configs (main.ts, fields.ts, phases.ts)
+
+**Why automatic reading matters:**
+- Loads all field/phase IDs into conversation context immediately
+- Enables local-first strategy without manual file reads
+- Saves 75% of API calls and hundreds of tokens on subsequent operations
+- You'll know the workspace structure instantly
+
+**Common use cases:**
+- Backup current workspace configuration
+- Review workflow structures locally
+- Prepare for making configuration changes

package/.opencode/commands/yolo-off.md

@@ -0,0 +1,17 @@
+---
+description: Disable yolo mode and restore normal permissions
+---
+# Disable Yolo Mode
+
+Run this command to disable yolo mode:
+
+```bash
+node .claude/scripts/yolo-toggle.cjs off
+```
+
+This will:
+- Remove the permissive permissions added by yolo mode
+- Restore normal confirmation prompts
+- Delete the .yolo-mode marker file
+
+Confirm when done: "Normal mode restored. Permission prompts re-enabled."

package/.opencode/commands/yolo.md

@@ -0,0 +1,82 @@
+---
+description: Run autonomously without asking permission
+argument-hint: <task to complete>
+---
+# AUTONOMOUS MODE
+
+Execute the following task without stopping to ask questions:
+
+**Task:** $ARGUMENTS
+
+## Step 1: Enable yolo mode
+
+Run this command to find and execute yolo-toggle:
+```bash
+node .claude/scripts/yolo-toggle.cjs on
+```
+
+## Step 2: Create restore point (if git available)
+
+If this is a git repository, run `/save "Before YOLO: $ARGUMENTS"` to create a restore point.
+If no git, skip this step silently.
+
+## Step 3: Gather context first
+
+Before writing any code:
+1. Read DEVELOPMENT.md to understand project state
+2. Read relevant PRD if one exists for this task
+3. Check workspace/ files if this is a Hailer project
+4. Understand existing patterns before creating new ones
+
+**Don't skip context gathering** - autonomous doesn't mean blind.
+
+## Step 4: Execute the task
+
+**Rules:**
+1. Make decisions yourself - don't ask for preferences
+2. Choose the most sensible option when multiple approaches exist
+3. Proceed through errors - try to fix them, only stop if truly stuck
+4. Complete the entire task before reporting back
+5. Use your best judgment for code style, naming, structure
+6. **ALWAYS delegate to agents** - Kenji reads, Dmitri writes, Giuseppe builds apps, etc.
+   Yolo = "agents work without permission prompts", NOT "orchestrator does everything".
+   The delegation-reminder hook ENFORCES this - it blocks direct tool use in yolo mode.
+
+**Boundaries:**
+- ONLY edit/create/delete files in the CURRENT PROJECT FOLDER
+- You MAY read other projects for reference/context
+- You MAY NOT edit, write, or modify anything outside this project
+
+**Still requires confirmation:**
+- Push commands (npm run push, fields-push, etc.) - safety hooks still active
+
+## Step 5: Context management
+
+The **context-watchdog hook** monitors context automatically during yolo mode. It will:
+- Warn you at ~20 agent calls (early heads-up)
+- Warn urgently at ~35 agent calls (save progress now)
+- Inject handoff instructions when auto-compaction fires (context IS full)
+- Block you from stopping after compaction until handoff is written
+
+**When you see a watchdog warning**, follow its instructions:
+1. Update DEVELOPMENT.md with current progress
+2. Write SESSION-HANDOFF.md with full state and next steps
+3. Run `/save` with progress description
+4. Tell user: "Context getting full. Run `/yolo` again to continue."
+
+**Don't wait for the watchdog** if you notice the task is large - save progress proactively at natural breakpoints.
+
+## Step 6: When complete
+
+1. Update DEVELOPMENT.md with what was done
+2. If implementing a PRD, update its status (Draft → Complete) and check off completed items
+3. Run `/save` with completion description
+4. Disable yolo mode:
+   ```bash
+   node .claude/scripts/yolo-toggle.cjs off
+   ```
+5. Report summary of what was accomplished
+
+---
+
+GO. Do the task end-to-end, then report what you did.

package/inbox/usage.jsonl

@@ -0,0 +1 @@
+{"ts":"2026-02-13T08:23:26.264Z","agent":"general-purpose","status":"unknown","project":"hailer-mcp","description":"Convert new agents to OpenCode"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hailer/mcp",
-  "version": "1.1.3",
+  "version": "1.1.4",
   "config": {
     "docker": {
       "registry": "registry.gitlab.com/hailer-repos/hailer-mcp"

package/.opencode/agent/nora.md

@@ -1,47 +0,0 @@
----
-description: Creates workflow nameFunction configurations for dynamic activity naming
-mode: subagent
-model: anthropic/claude-haiku-4-5
-tools:
-  read: true
-  glob: true
-  edit: true
-  bash: true
-  write: false
----
-
-I am Nora, Finnish name function specialist. Every activity deserves a meaningful name. SDK v0.8.4. Output JSON. Full stop.
-
-**EXCLUSIVE DOMAIN:** I am the ONLY agent who handles nameField, nameFunction, and nameFunctionVariables.
-
-## Handles
-- Add nameFunction to workflow main.ts files
-- Configure nameFunctionVariables with field dependencies
-- Enable/disable dynamic activity naming
-
-## Rules
-1. **NEVER FABRICATE** - Must call tools.
-2. **CRITICAL** - Pull OVERWRITES local changes. Push before pulling.
-3. Edit main.ts in workspace/[Workflow]_[id]/ directory.
-4. **NEVER run workflows-push** - Return command for orchestrator.
-5. Use enums from enums.ts - Never hardcode field IDs.
-6. **JSON ONLY** - Output closing brace, then STOP.
-
-## Critical Syntax
-nameFunction is ONLY the function BODY (what's between {}), NOT the full function!
-- WRONG: "function(dep) { return 'value'; }"
-- CORRECT: "return 'value'"
-
-## Patterns
-```javascript
-// Simple field
-nameFunctionVariables: { fieldName: { data: [FieldIds.field_name_abc], type: "=" } }
-nameFunction: "return (dep.fieldName || 'Unnamed')"
-
-// Linked field (forward link ">")
-nameFunctionVariables: { clientName: { data: [FieldIds.client_link, ClientFieldIds.company_name], type: ">" } }
-nameFunction: "return 'Project for ' + (dep.clientName || 'Unknown')"
-```
-
-## Protocol
-Output: { "status": "ready_to_push|error", "result": { "name_function_added": true, "variables": 0 }, "commands": [], "summary": "" }

package/.opencode/commands/install-plugin.md

@@ -1,16 +0,0 @@
----
-description: Install a plugin from local marketplace
----
-
-Install plugin "$ARGUMENTS" from the marketplace.
-
-Check if plugin exists:
-!`ls -d hailer-marketplace/$ARGUMENTS 2>/dev/null && echo "Found" || echo "NOT FOUND"`
-
-If found, copy files:
-- Copy agents: `hailer-marketplace/$ARGUMENTS/agents/*.md` → `.claude/agents/`
-- Copy skills: `hailer-marketplace/$ARGUMENTS/skills/*` → `.claude/skills/`
-- Copy hooks: `hailer-marketplace/$ARGUMENTS/hooks/*.cjs` → `.claude/hooks/`
-
-Run these copy commands and report what was installed.
-Then remind user: "Restart Claude Code to load: `claude -c`"

package/.opencode/commands/list-plugins.md

@@ -1,9 +0,0 @@
----
-description: List available plugins from local marketplace
----
-
-List the available plugins from the marketplace:
-
-!`cd hailer-marketplace && git pull origin main --quiet 2>/dev/null; jq -r '.plugins[] | "  \(.name)@\(.version) - \(.description)"' .claude-plugin/marketplace.json`
-
-Show the count and remind user they can install with `/install-plugin <name>`.

package/.opencode/commands/marketplace-setup.md

@@ -1,9 +0,0 @@
----
-description: Clone or pull the Hailer marketplace repo
----
-
-Set up the marketplace:
-
-!`if [ -d "hailer-marketplace/.git" ]; then cd hailer-marketplace && git pull origin main && echo "Pulled latest"; else git clone git@github.com:Bdolf/Hailer-Marketplace.git hailer-marketplace && echo "Cloned marketplace"; fi`
-
-Report the result.

package/.opencode/commands/publish-plugin.md

@@ -1,19 +0,0 @@
----
-description: Publish a plugin to the Hailer marketplace
----
-
-Publish plugin "$ARGUMENTS" to the marketplace.
-
-If argument is a path (contains `/`), use it as source.
-If argument is a name, look for `.claude/agents/agent-$ARGUMENTS.md`.
-
-Steps:
-1. Validate plugin exists
-2. Check marketplace for existing versions
-3. Create branch `publish/$ARGUMENTS-vX.X.X`
-4. Copy plugin files to `hailer-marketplace/$ARGUMENTS/`
-5. Update `marketplace.json`
-6. Commit and push
-7. Create PR
-
-Ask for version confirmation before proceeding.

package/.opencode/commands/uninstall-plugin.md

@@ -1,16 +0,0 @@
----
-description: Uninstall a plugin from .claude/ folder
----
-
-Uninstall plugin "$ARGUMENTS".
-
-If no argument provided, list installed plugins:
-!`ls -1 .claude/agents/agent-*.md 2>/dev/null | xargs -I{} basename {} .md | sed 's/^agent-//' | sort -u`
-
-If argument provided, remove:
-- `.claude/agents/agent-$ARGUMENTS*.md`
-- `.claude/skills/$ARGUMENTS/`
-- `.claude/hooks/$ARGUMENTS*.cjs`
-
-Report what was removed.
-Then remind user: "Restart Claude Code: `claude -c`"