Autocalibrate: Tuning the Two-Stage Status Assessment Engine

Hive detects agent status in two stages: internal/domain/terminal/assess (Stage 1, stateless: pane content -> assess.State) feeds internal/domain/terminal/status (Stage 2, stateful: debounces assess.State into a published terminal.Status). This skill is the closed loop for tuning both against a corpus of committed, deterministic recordings — never against live judgment calls that can't be replayed.

Corpus Layout

• internal/domain/terminal/assess/testdata/<tool>/<scenario>.txt — single-frame fixtures for Stage 1 rule classification. Table in internal/domain/terminal/assess/fixtures_test.go.
• test/calibration/sequences/<name>.jsonl + <name>.expected.json — multi-frame recordings (the {ts, content, title, inMode} shape hive x assess watch --record produces) plus a sidecar of the published-status sequence hive x assess replay must reproduce exactly. Covers Stage 2 debounce timing (idle-confirmation delay, churn-as-working, approval immediacy, ...) that a single-frame fixture can't.
• test/calibration/scenarios/<name>.yaml — step/expectation scripts for live, in-container runs against a real pane (hive x assess scenario). Not wired into mise run cli:integration; run by hand.

The Tuning Loop

1. Find mismatches. Run the automated guardrail:

   go test ./cmd/hive/internal/commands/... -run TestCalibrationCorpus_ReplayMatchesExpected -v
   

   This replays every test/calibration/sequences/*.jsonl through a fresh assess.Engine + status.Tracker on a virtual clock (never wall time, never an ambient config file — status.DefaultOptions() only) and diffs the published-status sequence against its .expected.json sidecar. A failing subtest names the sequence and the exact frame index where the sequences diverge.