fix: surface actionable hint when Claude model unavailable on Vertex

[vildursn]

Context

When running zeroshot with CLAUDE_CODE_USE_VERTEX=1, the Claude CLI resolves model shorthands (sonnet, haiku) to versioned Vertex model names (e.g. claude-sonnet-4-5@20250929) that may not be enabled on the user's Vertex deployment. The resulting 404 error was swallowed into a raw JSON blob with no guidance on how to fix it.

Changes

• Added detectVertexModelError() to parse the Claude CLI result JSON and identify Vertex model 404 errors
• logTaskAttemptFailure now prints an actionable fix with the exact zeroshot settings set command when this error is detected
• Vertex model errors short-circuit the retry loop immediately (like HookExecutionError) — the model being unavailable is deterministic, so retrying wastes time

Alternatives considered

Option B — Skip --model on Vertex: When CLAUDE_CODE_USE_VERTEX=1 is set and no levelOverrides are configured, pass no --model flag and let the CLI use the deployment default. Avoids the error entirely but surrenders model-level control.

Option C — Inherit the parent session model: When running under Vertex, the model the parent Claude Code session is using is guaranteed to work. Zeroshot could detect CLAUDE_CODE_USE_VERTEX=1 and default to that model. The challenge is there's no clean way to read which model the parent session resolved to — ANTHROPIC_MODEL in the environment is empty even when a model is active.

🤖 Generated with Claude Code