[Arai-Y-SLC]

Summary

Two Claude 5 changes in one PR, so they ship as a single version bump (0.0.92 → 0.0.93, rebased on main after #4106 and #4095). Each is its own commit if you prefer to review them separately.

Fixes #4044. Follow-up to #4059 (closed by #4097).

1. Thinking Display option (as offered in #4059 (comment))

Claude 5 models default to thinking.display: "omitted", so the stream carries nothing while the model thinks. #4097 raised the bedrock-runtime read timeout to 3600 s, so botocore no longer gives up, but the connection still sits idle for the whole thinking phase. Networks with a fixed idle timeout (VPC endpoints, NAT gateways and NLBs are 350 s, per AWS) can still drop it. This adds the thinking_display option the Anthropic plugin already has.

• New parameter Thinking Display (omitted / summarized, default omitted) on the "Anthropic Claude 5" family, right after Effort.
• omitted or unset: the request is exactly what 0.0.92 sends today.
• summarized: adds "display": "summarized" to the adaptive thinking config. Reasoning summaries stream as <think> content, which the existing stream handler already renders. This also works with Effort unset (sends {"type": "adaptive", "display": "summarized"} with no output_config).

2. Claude Haiku 5.5 (anthropic.claude-haiku-5-5), following the pattern of #3926 and #4009

• Option in the Claude 5 family, CLAUDE5_PROFILE_PREFIXES us./eu./au./global. (the model card also lists jp., left unwired as for Opus 5.5), cache minimum 512 / 4 checkpoints (model card), pricing file at the Global rate of $0.10 / $0.50 per 1M tokens (prompts up to 100K; the long-context rate of $0.50 / $2.50 cannot be expressed in the flat schema, noted in a comment), README (en / zh).
• Because Haiku 5.5 profile IDs are now recognised as Claude 5 profiles, custom models on them get the Claude 5 parameter rules instead of the legacy ones, which is what [Bedrock] Disable "Temperature" parameter by default for Claude Haiku 5.5 (same as Sonnet 5.5) #4044 asks for. fix(bedrock): avoid unsupported inference profile JSON prefill #4095 (now merged) handles the same custom-profile case with a Haiku-specific filter; the two work together. With Haiku 5.5 registered as a Claude 5 model, its profile IDs now also request /stop_details like the other Claude 5 models, so one assertion in the fix(bedrock): avoid unsupported inference profile JSON prefill #4095 test is updated for that case (the application inference profile ARN case is unchanged), and its README note now mentions the predefined model.

Case (Dify 1.14.2 → plugin daemon → live Bedrock, geographic profile, stream unless noted)	Before	After
Haiku 5.5 in the Claude 5 family	❌ not selectable	✅ us.anthropic.claude-haiku-5-5, answers in stream and non-stream
Haiku 5.5, effort high, long arithmetic task, summarized	—	longest silence 2.7 s (vs 26.7 s with omitted), 2,017 chars of summary, same answer
Sonnet 5.5, effort high, same task	longest silence 28.6 s (whole thinking phase), no reasoning	omitted: unchanged / summarized: longest silence 3.1 s, 2,030 chars of summary, same answer
Haiku 5.5, Effort unset, summarized	—	longest silence 2.6 s, 1,817 chars of summary
Haiku 5.5, summarized, 2-round tool loop (stream / non-stream)	—	✅ tool_use → tool_use → end_turn, 1173 (both modes)
Haiku 5.5 system cache checkpoint, two calls	—	✅ write 3,917, then read 3,917
Effort set, Thinking Display unset or omitted	✅	✅ unchanged request (unit test asserts equality)
Sonnet 5.5 omitted, Claude 4.6 Sonnet, embeddings, rerank	✅	✅ unchanged
Unit tests	295 passed	✅ 322 passed (27 new; 30 fail with the source reverted to main)

The silence grows with the thinking time. In a knowledge-pipeline job on the same deployment, Sonnet 5.5 at the default effort was silent for 291 s before the first event, close to the 350 s limit; with summarized the longest gap was 3.1 s.

Release Notes

• Add Claude Haiku 5.5 to Anthropic Claude 5 (global and geographic profiles, prompt caching, $0.10/$0.50 per 1M tokens at the global rate for prompts up to 100K tokens). Custom models on Haiku 5.5 inference profiles get the Claude 5 parameters.
• Add a Thinking Display option to Anthropic Claude 5 models. Choose summarized to stream reasoning summaries while the model thinks; this also keeps long-thinking requests from going idle behind proxies or VPC endpoints with an idle timeout. The default (omitted) leaves requests unchanged.

Change Type

• Documentation / non-plugin change
• Non-LLM plugin (tools, extensions, datasource, etc.)
• LLM plugin

Screenshots / Videos

Before	After
No Haiku 5.5 option; no Thinking Display, nothing streams until thinking ends	Haiku 5.5 in Bedrock Model; Thinking Display below Effort, and with summarized reasoning summaries appear while the model thinks

LLM Plugin Checklist

Areas affected by this change (check all that apply)

• Message flow (system messages, user ↔ assistant turn-taking)
• Tool interaction flow (multi-round usage, Agent App and Agent Node)
• Multimodal input (images, PDFs, audio, video, etc.)
• Multimodal output (images, audio, video, etc.)
• Structured output (JSON, XML, etc.)
• Token consumption metrics
• Other LLM functionality (reasoning, grounding, prompt caching, etc.)
• New models / model parameter fixes

Version

• Bumped top-level version in manifest.yaml (not the one under meta): 0.0.92 → 0.0.93
• dify_plugin>=0.10.2,<0.11.0 is declared in pyproject.toml and locked in uv.lock (unchanged)

Testing

• Local deployment — Dify version: 1.14.2

• SaaS (cloud.dify.ai)

• uv sync --all-groups --frozen, Python 3.12: 322 passed (on main + this branch).

  • tests/test_claude5_thinking_display.py (new, 15 tests): unchanged request with Effort and no display, omitted, summarized with and without Effort, sampling params dropped, legacy reasoning_type not mixed in, no leak of thinking_display into inferenceConfig, the YAML rule, and pass-through to the Claude 4.8 Opus refusal fallback.
  • Haiku 5.5 rows added to the Claude 5 model id, cache and llm param tests, plus regression tests for custom models on us./eu./au./global. Haiku 5.5 profiles ([Bedrock] Disable "Temperature" parameter by default for Claude Haiku 5.5 (same as Sonnet 5.5) #4044).
  • With the source reverted to main and the new tests kept: 7 of the thinking-display tests, 22 of the Haiku 5.5 tests and the updated fix(bedrock): avoid unsupported inference profile JSON prefill #4095 assertion fail.

• Live rows above: the plugin installed in a self-hosted Dify 1.14.2, invoked through the plugin daemon the way workflow nodes call it. Both parameters appear in the model settings panel after the plugin update.

Notes: the Bedrock docs say a model may accept summarized and still return no thinking text, so it is a best-effort option. The summary adds to the output Dify shows (with the "separate" reasoning format it goes to reasoning_content, not the answer text).

🤖 Generated with Claude Code

[crazywoola]

Hi @Arai-Y-SLC, thanks for adding Claude Haiku 5.5 and the thinking display option!

Heads-up about models/bedrock/manifest.yaml: the consolidated dependency refresh (#4106–#4111) bumped it to 0.0.92 on main, and #4095 from another author also targets this plugin. Since a plugin can only be released one version at a time, please rebase on latest main and bump the version to the next unused one after the preceding PR lands, so the releases don't collide.

Thanks again for your contribution! 🙏

[Arai-Y-SLC]

Thanks @crazywoola. Merged the latest main (with #4106 and #4095) and bumped manifest.yaml to 0.0.93, the next unused version (4dc7eb0). With Haiku 5.5 now a predefined Claude 5 model, one assertion in the #4095 test is updated for the profile-ID case (it now requests /stop_details like the other Claude 5 models; the application inference profile ARN case is unchanged). All 322 tests pass, and the PR description is updated to match.

Re-ran the live checks on the merged branch (4dc7eb0: main with #4106 and #4095, boto3 1.43.118), in a fresh uv sync --frozen environment against live Bedrock:

Case	Result
Haiku 5.5, effort high, long task	omitted: longest silence 19.7 s / summarized: 2.5 s, 3,537 chars of summary, same answer
Sonnet 5.5, effort high, same task	omitted: 23.9 s / summarized: 2.8 s, 1,705 chars of summary, same answer
Haiku 5.5, Effort unset, summarized	longest silence 2.7 s, 1,288 chars of summary
Haiku 5.5, effort medium, Thinking Display unset	✅ unchanged request, answer returned
Haiku 5.5 with stale temperature / top_k saved (#4095 filter)	✅ answer returned
2-round tool loop, stream / non-stream	✅ 1173 in both
System cache checkpoint, two calls	✅ write 3,931, then read 3,931
Embeddings (cohere.embed-multilingual-v3, titan-embed-text-v2), rerank (cohere.rerank-v3-5)	✅ 1024 dims, scores 0.357 / 0.051

Same behaviour as before the merge; the silence numbers vary run to run with the thinking length.