ORCH-044 closes two blind spots that let a single de-authenticated agent
stall the shared queue for all projects:
P1 — preflight auth gate. `claude --version` answers even when logged out,
so version-only preflight was blind to auth. Adds a token-free, network-free
check of <AGENT_HOME>/.claude/.credentials.json: missing/unreadable/no-oauth
or an expired `claudeAiOauth.expiresAt` (epoch ms, vs now + skew) => preflight
FAIL; absent expiry => OK (no false positives). Result is cached on the same
preflight_cache_ttl. Post-factum safety net: launcher detects auth markers
("not logged in" / "/login" / "unauthorized" / 401) in the run log and resets
the preflight cache so the next tick re-evaluates auth. Auth failure is a gate,
not a transient — it does not spin the circuit breaker. Emergency toggle
ORCH_PREFLIGHT_CHECK_AUTH=false restores version-only behaviour.
P3 — empty log / no result-JSON => job failed. exit_code==0 with an empty or
JSON-less run log no longer counts as success: a separate result_ok flag gates
stage advance + usage comments, fires a Telegram alert, and routes the job
through the normal transient/permanent failure path (exit_code integrity in
agent_runs preserved).
Scope: P2 (--effort) is intentionally excluded and tracked in ORCH-50.
New settings: ORCH_PREFLIGHT_CHECK_AUTH, ORCH_CLAUDE_CREDENTIALS_PATH,
ORCH_AUTH_EXPIRY_SKEW_SECONDS. Docs updated (INFRA.md, internals.md, CHANGELOG).
Refs: ORCH-044
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
