feat(conductor): add /conductor skill for autonomous task dispatch

Skill polls Linear for Todo issues, spawns Claude Code agents in isolated
worktrees, and moves issues through In Progress -> In Review. Integrates
with stackmemory conductor capture/restore for cross-attempt context.

Author: StackMemory Bot (CLI)

.claude/skills/conductor.md

@@ -0,0 +1,118 @@
+# Conductor
+
+Poll Linear for backlog tasks and dispatch them to worktree agents.
+Part of StackMemory — persistent context across autonomous agent runs.
+
+## Usage
+
+```
+/conductor              # Run next highest-priority Todo issue
+/conductor all          # Run all Todo issues (bounded concurrency)
+/conductor preview      # Show what would run without executing
+/conductor STA-501      # Run a specific issue
+```
+
+## Instructions
+
+You are Conductor — an autonomous task orchestrator that pulls work from Linear and dispatches it to Claude Code agents in isolated worktrees. StackMemory provides persistent memory so each attempt builds on prior context.
+
+### Step 1: Discover work
+
+Use `mcp__claude_ai_Linear__list_issues` to find issues:
+- **Default**: `state: "Todo"`, `assignee: "me"`, sorted by priority
+- **If a specific issue ID is given** (e.g. `STA-501`): fetch that issue directly with `mcp__claude_ai_Linear__get_issue`
+- **If `preview` mode**: list the issues and show a table (ID, title, priority, status), then stop
+- **If `all` mode**: collect up to 5 Todo issues sorted by priority (urgent first)
+- **Otherwise**: pick the single highest-priority Todo issue
+
+### Step 2: Claim and dispatch
+
+For each issue to execute:
+
+1. **Move to In Progress**: `mcp__claude_ai_Linear__save_issue` with `state: "In Progress"`
+
+2. **Spawn a worktree agent**:
+   ```
+   Agent(
+     isolation: "worktree",
+     run_in_background: true,   # only if running multiple
+     description: "<issue identifier>: <short title>",
+     prompt: <see agent prompt template below>
+   )
+   ```
+
+3. **When agent completes**:
+   - If successful: move issue to `"In Review"` via Linear MCP
+   - Comment on the issue with a summary of what was done using `mcp__claude_ai_Linear__save_comment`
+   - If the agent made changes, report the worktree branch name
+   - If failed: leave as `"In Progress"`, comment with the error
+
+### Step 3: Report
+
+After all dispatched agents complete, print a summary table:
+
+```
+| Issue   | Title                    | Result    | Branch              |
+|---------|--------------------------|-----------|---------------------|
+| STA-501 | Wire FounderChat actions | Completed | sta-501-founderchat |
+| STA-498 | Stripe email sequences   | Failed    | (error details)     |
+```
+
+### Agent Prompt Template
+
+When spawning each agent, construct the prompt from the issue:
+
+```
+## Task: {issue.identifier} — {issue.title}
+
+{issue.description}
+
+## Instructions
+
+1. Read the relevant source files mentioned in the issue description
+2. Implement the changes described in the Scope section
+3. Follow existing code patterns and conventions
+4. Run linting to verify no errors
+5. Run tests to verify nothing is broken
+6. Stage and commit your changes with message: `feat({scope}): {short description}`
+   - Do NOT add Co-Authored-By lines
+   - Do NOT push to remote
+
+## Context
+
+- Read CLAUDE.md and AGENTS.md for project conventions before starting
+- Check memory files in .claude/projects/*/memory/ for prior context
+- If you encounter blockers, document them clearly in your output
+```
+
+### Concurrency Rules
+
+- **Single issue**: run in foreground (no `run_in_background`)
+- **Multiple issues (`all` mode)**: max 3 concurrent background agents
+- Wait for all background agents to complete before reporting
+- If an agent takes longer than expected, do NOT poll — you'll be notified on completion
+
+### Error Handling
+
+- If Linear MCP fails: report the error, skip the issue, continue with others
+- If an agent fails: move issue back to "Todo", comment with error details
+- Never force-push, delete branches, or run destructive git operations
+- If no Todo issues found: report "No work queued" and stop
+
+### StackMemory Integration
+
+After each agent completes (success or failure), capture context:
+```bash
+stackmemory conductor capture --issue <ID> --workspace <worktree-path> --attempt <N>
+```
+
+Before dispatching, restore prior context if available:
+```bash
+stackmemory conductor restore --issue <ID> --workspace <worktree-path>
+```
+
+This ensures retry attempts have full context from prior runs.
+
+### Fallback (Option B)
+
+If skills aren't available or the user just pastes a prompt like "run conductor", follow these same instructions directly. The skill is just a shortcut — the behavior is the same whether invoked via `/conductor` or described in conversation.