who-needs-me
Tells you which of your parallel AI coding sessions needs a human, and why.
About the project
I run seven AI coding sessions in parallel. This morning two of them stopped on a login error and four more queued behind them for five hours before anyone noticed. who-needs-me monitors each session remotely over SSH (read-only, nothing installed on the remote machine) and puts the ones that need a person at the top of a board, with the line of evidence. It works across machines: one capture process per machine, one board.
Three layers: deterministic rules for what code can decide (shell prompt, spinner, login error); a System One decision model (Gemini Decision 1) that answers five typed questions about the session’s last message; and a policy where deploy, spend and delete always go to a human.
Evaluated on 153 hand-labeled screens (89 distinct, 45 locked as a holdout). With the policy fixed in advance it caught 8 of 10 screens that needed a human; that result is published unchanged. After restoring the “escalate when unsure” rule it caught 10 of 10 with three more false alarms; that run is marked as post-hoc. A repeat run gave the same decision on 45 of 45 screens. Model latency 0.29 s median.
Built with
AI tools
Prize Categories Entered
- Submitted

Best Use of Gemini API
Two Gemini APIs. (1) The Gemini Decisions API (gemini-decision-1) is the core of the classifier: for each coding session, rules handle what code can decide, and for the rest the session's last message is sent to the Decisions API with five typed questions (one choice: how does this turn end; four yes/no: is it asking to deploy, spend, delete, or for someone to act on the machine). It returns a probability per option, with no text to parse, in about 0.29 s. Those probabilities drive the board: red for needs-you, yellow for have-a-look. On a 45-screen holdout it gave the same decision on 45 of 45 screens across two runs. (2) Gemini TTS (gemini-3.8-flash-tts) generated the narration of the demo video.
- Event prizeSubmitted
Best Overall Hack
A real problem from today: two of my seven AI coding sessions stopped on a login error and four more queued behind them for five hours unnoticed. who-needs-me monitors sessions remotely over SSH, across machines, read-only, and shows which one needs a person and why, with evidence quoted from the session. It is evaluated on a locked holdout and the results, including the misses, are in the repo.
- Event prizeSubmitted
Best Use of System One Models
The classifier is rules first, then a System One decision model (Gemini Decision 1) for the one thing rules cannot decide: what the session's last message is asking for. One call answers five typed questions and returns probabilities, so thresholds can be tuned and low confidence can be escalated. Rules alone caught 7 of 10 screens that needed a human; rules plus the decision model caught 8 of 10 with the policy fixed in advance, and 10 of 10 after restoring the escalate-when-unsure rule (that second run is marked post-hoc). Latency 0.29 s median; identical decisions on 45 of 45 screens on a repeat run.
- Event prizeSubmitted
Cleanest Use of a Software Factory
This is one part of a factory: the escalation gate that says when a human has to look, plus the measurement loop that keeps it honest. Named signals (session waiting on approval, session broken, session dropped to shell) and all thresholds live in one declarative file, factory.yaml. Deploy, spend and delete always go to a human. Every rule has a test (98 tests, ruff clean). 153 screens were labeled by hand, 45 were locked as a holdout with a recorded hash, and the first evaluation is published unchanged including its two misses. The README states who did what: the human labeled and set every rule for choosing thresholds; the coding agent wrote the code.