tools: expose a2a_pack source to the inner LLM loop

list_a2a_pack(subdir) / read_a2a_pack(path) let the builder browse the
installed SDK at runtime — the same package the scaffolded agents
import. Path-traversal guarded; reads capped at 64 KiB.

Why: the system prompt only carries an abridged example. When the LLM
needed to recall what RunContext exposes (ctx.workspace API, ctx.ask
vs ctx.collect, runtime.apt_packages…) it would guess. Now it can
read the actual source.

Prompt updated to point at the new tools.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

agent_builder/builder.py

@@ -94,6 +94,23 @@ Your tools:
   - cp_deploy_tarball(name, version="0.1.0", public=True)
                                   — ship it to the platform; returns the
                                     public URL
+  - cp_refresh_agent(name)        — force the control plane to re-fetch
+                                    the agent's live card after an
+                                    out-of-band redeploy
+  - list_a2a_pack(subdir="")      — browse the installed a2a_pack SDK
+  - read_a2a_pack(path)           — read SDK source. Use these when
+                                    you're unsure what's exposed. The
+                                    code you scaffold runs on this
+                                    exact SDK, so reading it is the
+                                    authoritative reference — start
+                                    with ``read_a2a_pack("agent.py")``
+                                    for ``@skill`` semantics,
+                                    ``context.py`` for ``RunContext``
+                                    (workspace, sandbox, emit_progress,
+                                    ask, collect, request_scope),
+                                    ``runtime.py`` for ``AgentRuntime``
+                                    fields (apt_packages, pricing,
+                                    egress, etc.).
 
 Discipline: