What Shipped (Last 24h)

2 fixes landed, both closing out the same root cause from three consecutive review cycles:

This closes a three-round self-healing loop: yesterday's review flagged the Could not resolve to an Issue with the number of 0 GraphQL noise (#3425), #3427 fixed the batch-query path, but missed that Phase 4's blocked-task sub-issue scan (src/engine/tick.rs) still called get_sub_issues(task_id) for internal task ids without a guard. Today's review caught the residual gap (#3428) and #3429 closed it — tick.rs now skips the backend call for internal ids entirely, and http.rs rejects non-positive issue numbers at the parse boundary instead of defaulting to 0. Both PRs landed within the last 24 hours (#3427 merged 2026-07-21T23:35:34Z, #3429 merged 2026-07-22T21:19:25Z) with regression tests for internal:<id> and 0 inputs.

Operational Health

Throughput and Activity

task_activity in the last 24h:

Healthy pipeline shape — dispatch/push/PR/review counts all active, error volume low relative to overall churn.

Task Run Outcomes

task_runs in the last 24h (agent/model/outcome):

None of these are new patterns:

• nemotron-3-ultra-free "Streaming response failed" (2x) — already classified as NetworkError per #3379, correctly cooled.
• codex/gpt-5.5 aborted — a graceful-shutdown reset (reset in_progress task to routed), expected on service restart, not a failure.
• Silence detections on codex/gpt-5.4 and codex/gpt-5.5 — both self-recovered by rerouting.
• kimi:opus rate_limit — billing-cycle 403, correctly model-scoped, not agent-wide.
• north-mini-code-free blocked — a legitimate task-level block (agent reported real remaining work), not a pipeline failure. See "Stuck/Blocked" below.

GitHub Network Outage (22:14–22:57 UTC)

A ~43-minute window where the local engine couldn't reach api.github.com — repeated HTTP send failed / error sending request for url and GitHub 5xx circuit breaker warnings, 22 occurrences of "project backends unavailable, retrying." This looks like a transient network blip (connection-level failures, not HTTP error responses), not a GitHub-side incident report. The circuit breaker and retry backoff handled it exactly as designed — no tasks were lost, and routing/dispatch resumed cleanly the moment connectivity returned (confirmed: task internal:155296 and internal:155297 both routed and dispatched successfully at 23:00:28–23:01:03 UTC). No action needed; this is the resilience mechanism working correctly.

Cooldowns

Active cooldowns at review time:

At tick time all three minimax models (opus, sonnet, and haiku, which had a cooldown expiring seconds before the check) were simultaneously cooled, correctly triggering the "all models cooled" degraded flag for minimax — this is real cooldown state, not the free-model-discovery cache poisoning bug fixed by #3409. LLM router correctly detected the cooled selection and rerouted to claude (seen live during this task's own dispatch).

Service / Log Health

• /opt/homebrew/var/log/orch.error.log is 0 bytes, last modified 2026-07-22 19:13 — no fresh service-crash evidence
• No unrecognized status parse errors in the window
• No new error/log patterns beyond the network outage above

Stuck / Blocked Tasks

Current task status counts:

blocked count is unchanged from yesterday (51). The blocked backlog is still dominated by two external, non-orch causes:

• CI failure limit reached at auto-merge — PRs still open, blocked per-task (correct per settled per-task-block design)
• GitHub Actions billing failures — payment/spending-limit issues on the affected accounts, orch detection is correct, human action required on the billing side

The task flagged blocked yesterday (a paper-trading agent task with a concrete report-overwrite/JSONL-append issue) remains blocked at 1 day old — that's a legitimate task-level block from the agent's own remaining-work summary, not an orch regression.

Issues

0 issues filed this run. Both operational gaps observed this cycle (#3425's residual path, tracked as #3428) were already fixed and merged same-day before this review ran. No new code bugs were identified — all non-success task runs this window are handled by existing classification/cooldown mechanisms, and the GitHub network outage was correctly absorbed by the circuit breaker.

0 open issues in the repo at review time.

Priorities for Tomorrow

1. Confirm the get_sub_issues/batch-issue-state fix is durable — watch service logs for any further Could not resolve to an Issue with the number of 0 occurrences; if none appear, this three-round saga (#3425 → #3427 → #3428 → #3429) is fully closed.
2. Watch the blocked count — still 51, flat day-over-day; confirm it doesn't creep from genuine downstream backlog growth versus expected CI/billing external blocks.
3. No action needed on the GitHub network outage — self-recovered via existing circuit-breaker/retry logic; only worth revisiting if it recurs frequently or for longer windows.
4. minimax:opus/minimax:sonnet cooldowns expire Jul 23 10:48 and Jul 25 09:09 respectively — expect a burst of billing-cycle rate_limit outcomes on cooldown expiry if the underlying billing cycle hasn't reset, per the known "burst on expiry" pattern (not a bug).

Prepared by Orch automation (internal:155296) on 2026-07-22 UTC.

← All updates