[FEATURE] Enable prompt suggestions on Vertex AI deployments

Resolved 💬 2 comments Opened Feb 21, 2026 by yyamaguchi93 Closed Mar 22, 2026

Feature Request

Prompt suggestions are architecturally disabled on Vertex AI deployments. The fast model lane (used for suggestion generation) is hardcoded off when CLAUDE_CODE_USE_VERTEX=1.

Evidence from debug log

1. Fast mode gate (1,335 occurrences per session):

[DEBUG] Fast mode unavailable: Fast mode is not available on Bedrock, Vertex, or Foundry

2. Token counting API fails on Vertex proxy:

[ERROR] API error: 400 {"type":"error","error":{"type":"invalid_request_error",
"message":"output_config.format: Extra inputs are not permitted"}}
[DEBUG] countTokensWithFallback: API returned null, trying haiku fallback (161 tools)

3. No suggestion lifecycle entries at all — the suggestion system never attempts to run. Zero log entries for suggestion gating, skipping, or generation.

Environment

  • Claude Code v2.1.x (latest)
  • CLAUDE_CODE_USE_VERTEX=1
  • CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=1 (set in settings.json)
  • CLOUD_ML_REGION=global
  • macOS, darwin

Expected behavior

When CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=1 is set, prompt suggestions should work on Vertex deployments, possibly using the same model (e.g., Sonnet) at reduced max_tokens for suggestion generation if Haiku isn't available through the proxy.

Current behavior

Suggestions silently never appear. The only indication is 1,335+ "Fast mode unavailable" debug entries per session. No user-facing error or config warning.

Suggested improvements

  1. Enable suggestions on Vertex — use the primary model with a lightweight prompt instead of requiring a separate fast model
  2. Surface a warning — if suggestions are enabled but fast mode is unavailable, log a visible warning (not just DEBUG) explaining why suggestions won't work
  3. Document the limitation — the settings reference doesn't mention that CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION is ignored on Vertex

Related issues

  • #23859 — Timing race condition also prevents suggestion display (closed as dup of #15019)
  • #24121 — Prompt cache utilization at 0 due to A/B flags
  • #20892 — CLOUD_ML_REGION missing from settings docs

View original on GitHub ↗

This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗