feat: relay daemon skeleton — queue, dispatch, conversation, ntfy (#1)

First-PR scope from #1. Single-process Python daemon that relays
between Claude Code instances and chat-Claude (Anthropic API).

Components:

* relay.config — .env + config.yaml loader. Auto-generates ntfy
  topic on first run and persists it back to .env.
* relay.state — atomic file I/O via tempfile + rename, advisory
  flock at state/.lock to enforce single-instance.
* relay.conversation — append-only history with summarization.
  Triggers a summarize call when total chars exceed
  HISTORY_CHAR_CAP (default 400k); replaces history with the
  summary plus the most recent 10 turns.
* relay.anthropic_client — SDK wrapper. Marks the system prompt
  cacheable (5-min ephemeral cache); concatenates text blocks;
  estimates per-call cost from the Anthropic price table with
  cache-write/read accounted for.
* relay.queue — JSON envelope intake; oldest-by-mtime;
  malformed envelopes moved to queue/.rejected/.
* relay.dispatch — one-input-at-a-time per session
  (dispatch/<session_id>/input.txt). Won't overwrite a pending
  dispatch; queues internally and waits for CC to delete.
* relay.ntfy — best-effort POST to https://ntfy.sh/<topic>;
  failures logged but never block the main loop.
* relay.daemon — main loop. Polls jc_input.txt (priority) then
  queue/. Detects [NEEDS-JC] in the first 200 chars of any
  response and pauses dispatch until JC writes jc_input.txt.
  JC override supports @session-N: prefix for direct dispatch
  without an API call.
* relay.__main__ — CLI: relay run / relay status / relay topic.

Tests: 57 unit tests pass (config, state, conversation, queue,
dispatch, anthropic_client, ntfy, full daemon loop with a fake
client). One real-API smoke test marked real_api, opt-in via
pytest -m real_api; skips cleanly on credit-balance errors.

Out of scope for this PR (deferred to follow-ups): Flask status
endpoint, multi-session config in production, exponential
backoff, systemd unit, cost-tracking aggregation.

Closes #1.

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

tests/test_real_api.py

@@ -0,0 +1,57 @@
+"""Live-API smoke test.
+
+Runs only when ANTHROPIC_API_KEY is set in the env (or .env). Sends one
+real turn, verifies we get back a non-empty assistant text and usage
+counts. Marked ``real_api`` so it is excluded from the default suite —
+opt in with ``pytest -m real_api``.
+"""
+
+from __future__ import annotations
+
+import os
+
+import pytest
+from dotenv import dotenv_values
+
+from relay.anthropic_client import AnthropicClient
+from relay.config import REPO_ROOT
+
+
+def _api_key_available() -> str | None:
+    env_values = dotenv_values(REPO_ROOT / ".env")
+    return (
+        env_values.get("ANTHROPIC_API_KEY") or os.environ.get("ANTHROPIC_API_KEY") or ""
+    ).strip() or None
+
+
+@pytest.mark.real_api
+@pytest.mark.skipif(_api_key_available() is None, reason="ANTHROPIC_API_KEY not configured")
+def test_real_round_trip_against_haiku() -> None:
+    """One round-trip against the Haiku model — cheapest available, ~$0.0001 per call.
+
+    Skips (rather than fails) on credit-balance errors so the test is
+    honest about why it didn't run. Real authentication errors (401)
+    still fail loudly.
+    """
+    import anthropic
+
+    api_key = _api_key_available()
+    assert api_key is not None
+    client = AnthropicClient(
+        api_key=api_key, model="claude-haiku-4-5-20251001", max_output_tokens=64
+    )
+    try:
+        result = client.send(
+            system_prompt="You are a terse echo. Reply with exactly 'pong' and nothing else.",
+            messages=[{"role": "user", "content": "ping"}],
+        )
+    except anthropic.BadRequestError as exc:
+        msg = str(exc).lower()
+        if "credit balance" in msg or "billing" in msg:
+            pytest.skip(f"Anthropic API key has no credit balance: {exc}")
+        raise
+
+    assert result.text.strip().lower() == "pong"
+    assert result.input_tokens > 0
+    assert result.output_tokens > 0
+    assert result.estimated_cost_usd > 0