ci(pick-model): choose the model with one TypeSafe Choice, claude as fallback

The picker ran a full claude -p in plan mode per selection and parsed JSON out of prose. With TYPESAFE_API_KEY set it now sends the task and the catalog rows as state and asks one Choice over the model ids, each with its catalog row as criterion plus NONE; the id is validated against the pristine catalog as before. No key or a failed request falls through to the claude picker unchanged. Measured on the two-model fixture (Opus 5 / Haiku 4.5): issue-fix and security-fix -> Opus 0.99, alert triage -> Opus 0.62, wiki audit -> Haiku 0.94, update audit -> Haiku 0.46, distill -> Haiku 0.63; about a second per pick. Dry-run test added to test-model-selection.sh. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Merge pull request 'ci(wiki-maintenance): rebase onto main before pushing, retry three times' (#892 ) from fix/wiki-maintenance-rebase-push into master
2026-09-21 09:42:39 +02:00 · 2026-09-21 06:41:40 +00:00
2 changed files with 42 additions and 0 deletions
@@ -33,6 +33,35 @@ none() { log "$1"; exit 0; }            # emit nothing -> caller uses CLI defaul
 TABLE="$(sh "$HERE/fetch-models.sh" --table || true)"
 [ -n "$TABLE" ] || none "no catalog available — caller will use the CLI default model"

+# First choice: one TypeSafe (System One) Choice over the catalog. The task and
+# the catalog rows are state, every model id is an option with its catalog row
+# as the criterion, and the answer is typed — no prompt, no JSON parsing, and a
+# request that takes a fraction of a second instead of a claude run. The id is
+# still validated against the pristine catalog. Falls through to the claude
+# picker when the key is absent or the request fails.
+if [ -n "${TYPESAFE_API_KEY:-}" ]; then
+  REQ="$(printf '%s\n' "$TABLE" | jq -R -s --arg task "$GUIDANCE" '
+    split("\n") | map(select(length > 0) | split("\t") | {id: .[0], name: .[1], info: (.[2:] | join(", "))})
+    | {model: "jev-latest",
+       state: {task: $task, models: .},
+       questions: {model: {
+         type: "choice",
+         instructions: "Which model in `models` should run `task`? Match the model to the task: not the largest or newest flagship for trivial or mechanical work, not a minimal model for hard, multi-step or security-sensitive work; prefer a newer model when otherwise comparable. Consider every listed model, including unfamiliar ones. Pick NONE only if no listed model fits.",
+         criteria: ((map({key: .id, value: "\(.name); \(.info)"}) | from_entries) + {NONE: "No listed model fits the task"})}}}')"
+  if [ -n "${PICK_MODEL_TYPESAFE_DRYRUN:-}" ]; then printf '%s\n' "$REQ"; exit 0; fi
+  RESP="$(printf '%s' "$REQ" | curl -sS --max-time 30 -X POST https://api.typesafe.ai/v1/systemone \
+           -H "Authorization: Bearer $TYPESAFE_API_KEY" -H 'Content-Type: application/json' --data-binary @- || true)"
+  CHOICE="$(printf '%s' "$RESP" | jq -r '.answers.model.choice // empty' 2>&1 | head -1)"
+  CONF="$(printf '%s' "$RESP" | jq -r '.answers.model.confidence // empty' 2>&1 | head -1)"
+  VALID="$(printf '%s\n' "$TABLE" | sh "$HERE/valid-model.sh" "$CHOICE")"
+  if [ -n "$VALID" ]; then
+    log "typesafe chose '$VALID' (confidence ${CONF:-?}) for: $GUIDANCE"
+    printf '%s\n' "$VALID"
+    exit 0
+  fi
+  log "typesafe picker returned '${CHOICE:-<empty>}' (rejected, NONE, or request failed) — falling back to the claude picker"
+fi
+
 PROMPT="$(cat <<EOF
 You are a model-selection agent for an automated CI job. From the models
 currently available (listed below with their capabilities), choose the single
@@ -84,6 +84,19 @@ else
  printf 'FAIL  recorded fixture: opus=[%s] haiku=[%s]\n' "$opusline" "$haikuline"; fails=$((fails + 1))
 fi

+
+# TypeSafe path of pick-model.sh: the request offers exactly the catalog ids plus
+# NONE as options, carries the task as state, and never sends anything else.
+printf '%s' '{"data":[{"type":"model","id":"claude-opus-5","display_name":"Claude Opus 5","created_at":"2026-07-24T00:00:00Z","max_input_tokens":1000000,"max_tokens":128000},{"type":"model","id":"claude-haiku-4-5-20251001","display_name":"Claude Haiku 4.5","created_at":"2025-10-15T00:00:00Z","max_input_tokens":200000,"max_tokens":64000}],"has_more":false}' > "$FX"
+req="$(FETCH_MODELS_BODY_FILE="$FX" TYPESAFE_API_KEY=dry PICK_MODEL_TYPESAFE_DRYRUN=1 sh "$HERE/pick-model.sh" "a task; with \"quotes\" and $ signs" 2>/dev/null || true)"
+keys="$(printf '%s' "$req" | jq -r '.questions.model.criteria | keys | join(",")' 2>/dev/null || true)"
+task="$(printf '%s' "$req" | jq -r '.state.task' 2>/dev/null || true)"
+if [ "$keys" = "NONE,claude-haiku-4-5-20251001,claude-opus-5" ] && [ "$task" = 'a task; with "quotes" and $ signs' ]; then
+  printf 'ok    typesafe request offers catalog ids + NONE and carries the task verbatim\n'
+else
+  printf 'FAIL  typesafe request: keys [%s] task [%s]\n' "$keys" "$task"; fails=$((fails + 1))
+fi
+
 if [ "$fails" -eq 0 ]; then
  echo "All model-selection tests passed."
 else