feat(prompt): guide agents to hold for sub-agents, expect auto-notify, and use inter-agent messages

[andrei-hasna]

What & why

Codewith agents were spawning sub-agents and then busy-polling / spamming
"Waiting for agents" / "No agents completed yet". #364 (0.1.77) made sub-agent
completion push-based (wake_if_idle auto-notify: an idle parent is woken
with the child's result) and #377 delivers inter-agent messages into the
target's session, but the agent system prompt did not yet teach the model the
full Claude-Code-style coordination loop. This PR adds that guidance, merging
with (not duplicating) #364's existing text.

Where the prompt lives

The model-facing multi-agent guidance is not in prompt.md / the
gpt-5.* variants / models.json base_instructions (all verified to contain
zero multi-agent text), and templates/collab/experimental_prompt.md is dead
code (unreferenced anywhere). The live, model-independent "agent system prompt"
for multi-agent sessions is the v2 usage hints in codex-core config, which
every root and sub-agent inherits (no per-model-family variant carries this):

• codex-rs/core/src/config/mod.rs:508 DEFAULT_MULTI_AGENT_V2_ROOT_AGENT_USAGE_HINT_TEXT
• codex-rs/core/src/config/mod.rs:532 DEFAULT_MULTI_AGENT_V2_SUBAGENT_USAGE_HINT_TEXT

#364's existing text there was a single line: "Sub-agent completion is
push-based: after spawning, do other work or end your turn ... Do not busy-wait
with wait_agent." I replaced that single line (in both hints) with the fuller
coordination loop below. Tool-level guidance from #364 already lives in
codex-rs/core/src/tools/handlers/multi_agents_spec.rs (spawn_agent /
wait_agent v2 descriptions) and is left intact.

Guidance added (merged with #364)

1. Hold, don't poll. Call wait_agent once only when genuinely blocked on
   a specific result; never loop wait_agent/list_agents, and never emit
   repeated "waiting for agents" / "no agents completed yet" messages.
2. Expect auto-notify. Completion is push-based (feat(core): auto-notify parent on sub-agent completion (wake_if_idle) + honest wait + guidance #364): a finishing
   sub-agent's FINAL_ANSWER is delivered automatically and an idle parent is
   woken with the result, so it is safe to stop and wait.
3. Coordinate via inter-agent messages (fix(core,api): deliver inter-agent messages on chat-wire providers + surface mailbox-full to sender #377): send_message injects a
   message into the target agent's session without starting a turn;
   followup_task hands off a new task and wakes an idle target. Both work
   parent->child, child->parent, and peer-to-peer -- use them to steer an agent,
   delegate a sub-task, or answer a running agent's question.

Also updated create_send_message_tool's description
(codex-rs/core/src/tools/handlers/multi_agents_spec.rs:179) to state the
message is injected into the target agent's session and works in all directions.

Exact tool names referenced (real, verified in this codebase)

spawn_agent, wait_agent, list_agents, send_message, followup_task,
plus the FINAL_ANSWER completion notification. (No invented tools -- there is
no message_agent/get_status in this codebase.)

Verification (warm Blacksmith, codex-core)

• rustfmt --check: clean on both changed files (exit 0).
• just test -p codex-core: 2893 passed. Ran a clean origin/main baseline
  (patch stashed) vs. patched -- identical pass/fail set; the patched failure
  set is a strict subset of baseline. The 75 failures are pre-existing
  environmental sandbox/landlock/network/MCP/shell-snapshot integration tests;
  zero failures in config::tests or multi_agents. No prompt snapshot
  test covers these strings (config tests compare against the constants
  themselves; send_message test asserts params, not the description).
• clippy: the only red is pre-existing and unrelated, in
  code-mode/src/runtime/mod.rs:329 (crate not touched here). This change is
  pure string content and adds no lints.
• No *.md touched, so prettier is N/A.

Do not merge/tag from here -- hand-off for 0.1.78.