The Review That Refuses to Ghost You

Most AI review demos are confident for ninety seconds. Then production happens. A sandbox wedges on a subprocess hang. A stream stops mid-tool-call with no exit code. The worker process exits cleanly from its own perspective and publishes nothing. From your side of the merge queue, all of that looks identical: silence. Humans interpret silence as thinking. Usually it is not thinking. It is stuck.

The consequence is specific and measurable. Your team runs three reviews before trust erodes. A bot that goes quiet once gets mentally filed as unreliable. After the second incident, developers stop waiting for it. By the third, the product is decorative infrastructure. That is the failure mode nobody talks about in AI code review—not "it gave a wrong comment" but "it just stopped."

TWO GUARDS, TWO TIME HORIZONS

The architecture separates the problem by time scale. In-session stalls—where the review workspace is still alive but quiet—need a fast, structured response before the silence becomes permanent. Run-level silence—where nothing has hit the board for long enough that the job may already be dead—needs a different enforcer: one that acts at the platform level, marks the outcome explicitly, and starts recovery without waiting for a human to notice.

WHY DEEPSEEK V4 FLASH

Supervision is not synthesis. The supervisor does not write review comments. Its job is to look at traces, compare expected cadence against observed activity, and pick from a short action list with a tight time budget. That is a reasoning task, but latency is part of the specification. If the supervisor thinks at flagship cost and speed, every stall check becomes a budget event—and you end up rate-limiting the thing designed to prevent rate-limiting your reviews.

DeepSeek V4 Flash is Mixture-of-Experts at 284B total parameters with only 13B active per forward step. Hybrid attention handles long server log traces without burning the full parameter budget. And unlike prior generations, it comes with documented reasoning modes—you trade depth against latency on a dial, not by swapping providers. For a control loop that fires every few seconds, that is the right architectural profile.

BENCHMARKS

GDPval-AA from Artificial Analysis measures Elo-style performance on agentic tasks with shell and browser access—a closer proxy to PR review reality than memoized benchmarks: tools fail, recovery matters. The Claude rows below use Adaptive reasoning with Max effort—the configuration Artificial Analysis exercised when Sonnet 4.6 set a new GDPval-AA high watermark in February 2026 publications, and Opus 4.6 reached 1606 Elo shortly before Sonnet narrowly exceeded it inside overlapping confidence intervals. The DeepSeek rows pair with their Reasoning Max effort reporting from Artificial Analysis DeepSeek launch notes later that spring (1554 for V4 Pro, 1388 for V4 Flash); these are directional leaderboards—your repository is the final test.

This is exactly why Critique standardized on DeepSeek V4 Flash for session supervision—not because it defeats Sonnet head-to-head, but because the intelligence density per marginal dollar aligns with infra that must chatter constantly. Critique couples that choice with deterministic board traces and a fallback path so stalls never disguise themselves as polite thinking. Promo pricing on Flash keeps the supervisory lane cheap enough that every team can justify turning continuity on everywhere, while Pro inherits the heavier GDPval-AA numbers when you escalate synthesis lanes.

Below we zoom into strictly DeepSeek-family curves so you can see how reasoning modes intersect with token throughput—the detail that persuaded us to bias the supervisor toward Flash's reasoning High profile instead of blindly maxing reasoning budget.

WHAT YOU ACTUALLY SEE

Open a PR on a repository with Critique installed. Watch the live board. You will see beat events arriving every few seconds as the review workspace makes progress: files read, test probes run, findings staged. If the workspace goes quiet past the stall threshold, a supervisor decision entry appears on the board—labeled, timestamped, with the action taken and the reason. If a nudge fires and the workspace recovers, the beats resume. If the supervisor aborts cleanly, the board shows that too. Nothing is ambiguous.

If the whole run goes cold—no beats for long enough that recovery is moot—the platform watchdog fires independently. The run status updates to failed with an explicit message. Another board entry logs the decision. A fallback review starts automatically, routed through the GitHub API path instead of the sandbox, so the PR still gets a verdict before the merge window closes.

COMMON QUESTIONS