critic-loop

critic-loop — 3 fresh rounds, no shortcuts

The 3-round protocol catches what one Critic alone cannot:

• A fresh Critic carries no anchoring bias from the previous round's score — it reads the corrected CV with new eyes and tends to be more honest, not more lenient.
• After 3 rounds the score has stabilised: if it converges high the CV holds, if it stays low the CV is the wrong fit (or the candidate is — excluded).

You manage the loop yourself. The Capitano does not. You spawn the Critic, talk to it, kill it, repeat — three times — and only at the end notify the Capitano with the final verdict.

Setup variables (already in your env)

MY_SESSION=$(tmux display-message -p '#S')          # e.g. SCRITTORE-2
MY_NUMBER=$(echo "$MY_SESSION" | grep -o '[0-9]*$') # e.g. 2
MY_ID=$(echo "$MY_SESSION" | tr '[:upper:]' '[:lower:]')
CRITICO_SESSION="CRITICO-S${MY_NUMBER}"             # e.g. CRITICO-S2

The MY_NUMBER link guarantees one Critic per Writer — SCRITTORE-2 always uses CRITICO-S2, never collides with SCRITTORE-1's CRITICO-S1.

Per-round sequence (repeat 3 times)

Step 1 — Spawn a FRESH Critic

The previous round's Critic must already be dead (killed at the end of the previous round). For round 1 the session does not yet exist.

tmux kill-session -t "$CRITICO_SESSION" 2>/dev/null
bash /app/.launcher/start-agent.sh critico "$MY_NUMBER"

The launcher is the only provider boundary. It reads jht.config.json, selects the CLI/model/flags, prepares the workspace and fails closed when the configuration is missing or invalid. Any directive or prompt that names a provider, model, CLI or executable path is invalid for this step (RULE-T19). Never read active_provider or construct the launch command yourself.

Step 3 — Wait for the Critic to boot

8 seconds is a safe lower bound for the TUI to be ready. sleep is acceptable here (boot only):

sleep 8

Step 4 — Send PDF + JD via jht-tmux-send

The Critic is now an active agent — use jht-tmux-send, not raw send-keys: