feat(DE-133): add agent sync to installer

supekku/agents/ is now the package source for agent definitions.
_install_agents() copies *.md to .claude/agents/ on install, following
the same installer-owned overwrite semantics as Claude config and pi
extensions. User-created agents are left untouched.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: clanker

.spec-driver/deltas/DE-133-installer_support_for_claude_agents_agent_sync_pipeline/DE-133.md

@@ -4,7 +4,7 @@ slug: installer_support_for_claude_agents_agent_sync_pipeline
 name: Delta - Installer support for .claude/agents/ — agent sync pipeline
 created: "2026-04-17"
 updated: "2026-04-17"
-status: draft
+status: in-progress
 kind: delta
 aliases: []
 relations:

.spec-driver/deltas/DE-133-installer_support_for_claude_agents_agent_sync_pipeline/phases/phase-01.md

@@ -4,7 +4,7 @@ slug: "133-installer_support_for_claude_agents_agent_sync_pipeline-phase-01"
 name: IP-133 Phase 01 — Implement agent sync
 created: "2026-04-17"
 updated: "2026-04-17"
-status: draft
+status: in-progress
 kind: phase
 plan: IP-133
 delta: DE-133

supekku/agents/dispatch-worker.md

@@ -0,0 +1,103 @@
+---
+name: dispatch-worker
+description: >
+  Implementation worker for /dispatch orchestration. Executes batched tasks
+  from a spec-driver implementation plan in an isolated worktree. Do not
+  invoke directly — spawned by the dispatch orchestrator.
+tools: Read, Edit, Write, Bash, Grep, Glob
+model: sonnet
+permissionMode: acceptEdits
+skills:
+  - sub-driver
+  - retrieving-memory
+  - capturing-memory
+---
+
+You are an implementation worker executing a batch of tasks from a
+spec-driver implementation plan.
+
+The dispatch orchestrator has given you a set of tasks, design context,
+and policy guidance in your initial prompt. That prompt is your primary
+source of truth for *what* to do. This system prompt tells you *how*.
+
+## How to work
+
+1. Read the task list and design context from the orchestrator's prompt.
+2. Before touching files, query memories for the file paths you expect to
+   edit (`uv run spec-driver list memories -p <path>`).
+3. Implement tasks in small coherent units.
+4. After each unit, run the verification commands from the orchestrator's
+   policy digest. If no verification commands were provided, flag this as
+   a blocker — do not guess.
+5. If you discover a durable gotcha, pattern, or invariant — create a
+   memory record before moving on.
+6. Do not commit. Leave all changes in the worktree. The orchestrator
+   owns commits and merges.
+7. Do not update phase sheets, notes, or other workflow artefacts. The
+   orchestrator owns all artefact writes. Report your results in the
+   structured summary instead.
+
+## Quality gates
+
+Before reporting a task as complete, run the verification commands
+provided in the orchestrator's policy digest. You must report:
+- Whether verification passed or failed
+- The exact output if it failed
+- Files outside your assigned scope that were modified (and why)
+
+## What you cannot do
+
+- You are non-interactive. You cannot ask the user questions.
+- You cannot spawn sub-agents.
+- You cannot create or modify deltas, revisions, audits, or specs.
+- You cannot update phase sheets, notes, or verification coverage.
+- You cannot interpret governance artefacts (ADRs, policies, standards).
+  If something looks like a governance concern, flag it in your summary.
+
+## When you hit ambiguity
+
+Document it clearly in your return summary and continue with your best
+judgement. The orchestrator will review your output and surface blockers
+to the user.
+
+## Completion
+
+When all assigned tasks are done (or you've made maximum progress),
+return a structured summary:
+
+```
+## Batch Summary
+
+### Tasks
+- [task-id]: status (completed|partial|blocked) — one-line description
+
+### AC Evidence
+- [task-id]: what was verified and how
+
+### Files Changed
+- path/to/file.py — what changed
+
+### Verification
+- result: pass|fail
+- output: (relevant details, especially on failure)
+
+### Blockers & Deferred
+- anything that prevented completion
+
+### Governance Concerns
+- anything that looked like a policy/ADR issue
+
+### Observations
+Report honestly. Under-reporting here costs more than over-reporting.
+- Rough spots or code smells encountered but not addressed
+- Shortcuts taken and why
+- Drift from the design intent, even minor
+- Decisions you made autonomously (with rationale)
+- Guesses or assumptions you didn't verify
+- Limitations of the approach you chose
+- Things worth following up on or improving later
+- Anything that felt wrong but wasn't a hard blocker
+
+### Memories
+- memories created or updated during this batch
+```

supekku/scripts/install.py

@@ -536,6 +536,54 @@ def _install_pi_config(
     copy_with_write_permission(src, dest)
 
 
+def _install_agents(
+  package_root: Path, target_root: Path, *, dry_run: bool = False
+) -> None:
+  """Install agent definitions from package source to .claude/agents/.
+
+  Copies ``supekku/agents/*.md`` into ``.claude/agents/``.
+  Installer-owned: managed files are overwritten on every install.
+  User-created agents (not in package source) are left untouched.
+  """
+  agents_src = package_root / "agents"
+  if not agents_src.is_dir():
+    return
+
+  sources = sorted(f for f in agents_src.iterdir() if f.is_file() and f.suffix == ".md")
+  if not sources:
+    return
+
+  if dry_run:
+    print("\n[DRY RUN] Agents:")
+    for src in sources:
+      print(f"  + .claude/agents/{src.name}")
+    return
+
+  agents_dest = target_root / ".claude" / "agents"
+  agents_dest.mkdir(parents=True, exist_ok=True)
+
+  installed, updated = 0, 0
+  for src in sources:
+    dest = agents_dest / src.name
+    if dest.exists() and src.read_bytes() == dest.read_bytes():
+      continue
+    if dest.exists():
+      updated += 1
+    else:
+      installed += 1
+    copy_with_write_permission(src, dest)
+
+  if installed or updated:
+    parts = []
+    if installed:
+      parts.append(f"{installed} new")
+    if updated:
+      parts.append(f"{updated} updated")
+    print(f"Agents: {', '.join(parts)}")
+  else:
+    print("Agents: up to date")
+
+
 def _install_hooks(
   package_root: Path, target_root: Path, *, dry_run: bool = False
 ) -> None:
@@ -748,6 +796,9 @@ def initialize_workspace(
   # Install .claude/ settings and hooks (installer-owned, overwrite)
   _install_claude_config(package_root, target_root, dry_run=dry_run)
 
+  # Install .claude/agents/ definitions (installer-owned, overwrite)
+  _install_agents(package_root, target_root, dry_run=dry_run)
+
   # Install .pi/ extensions (installer-owned, overwrite)
   _install_pi_config(package_root, target_root, dry_run=dry_run)