@dreb/coding-agent 2.71.0 → 2.72.0

package/docs/agent-models.md

@@ -54,7 +54,7 @@ Configure it in `settings.json` (global file, with the standard project-level ov
 
 The `subagent` tool accepts an optional `thinking` value in single mode, at the top level for parallel/chain inheritance, or on an individual task/step. Per-task values win over the top-level value.
 
-Supported values are `off`, `minimal`, `low`, `medium`, `high`, `xhigh`, and `max`. Validation happens after the child model resolves: non-`off` levels fail for non-reasoning models, `xhigh` requires advertised xhigh support, and `max` requires a max-capable model (currently the GPT-5.6 family including Sol, Terra, and Luna, and the GPT-6 family including Astra). `max` does not replace `xhigh`. Codex `ultra` additionally enables client-side multi-agent orchestration, so it is not accepted or sent as a raw effort. Omit the field to preserve the child's normal default/settings behavior. The child's actual effective level is reported in its `agent_start` event, result metadata, and `background_agent_end` event.
+Supported values are `off`, `minimal`, `low`, `medium`, `high`, `xhigh`, and `max`. Validation happens after the child model resolves: non-`off` levels fail for non-reasoning models, `off` fails for models whose thinking cannot be turned off (Claude Opus 5.5 — use `minimal` or `low`), `xhigh` requires advertised xhigh support, and `max` requires a max-capable model (currently the GPT-5.6 family including Sol, Terra, and Luna, the GPT-6 family including Astra, and Claude Opus/Sonnet 4.6+ and Haiku 5.5 on Anthropic, GitHub Copilot, and Bedrock). `max` does not replace `xhigh`. Codex `ultra` additionally enables client-side multi-agent orchestration, so it is not accepted or sent as a raw effort. Omit the field to preserve the child's normal default/settings behavior. The child's actual effective level is reported in its `agent_start` event, result metadata, and `background_agent_end` event.
 
 ## Evidence-based Routing Guide
 

package/docs/models.md

@@ -346,7 +346,7 @@ For providers with partial OpenAI compatibility, use the `compat` field.
 | `openRouterRouting` | OpenRouter routing config passed to OpenRouter for model/provider selection |
 | `vercelGatewayRouting` | Vercel AI Gateway routing config for provider selection (`only`, `order`) |
 
-The normalized scale is `minimal`, `low`, `medium`, `high`, `xhigh`, and `max` (`off` disables reasoning at the session layer). Native normalized `max` is model-aware and currently supported by GPT-5.6 aliases and Sol/Terra/Luna variants, plus the GPT-6 family (Astra and future variants); unsupported defaults fall back `max` → `xhigh` → `high`. Existing provider mappings remain independent—for example, Claude can still map dreb's `xhigh` to Anthropic's provider-native `max`. Codex `ultra` also enables local multi-agent orchestration and is therefore not a raw effort value.
+The normalized scale is `minimal`, `low`, `medium`, `high`, `xhigh`, and `max` (`off` disables reasoning at the session layer). Native normalized `max` is model-aware and currently supported by GPT-5.6 aliases and Sol/Terra/Luna variants, the GPT-6 family (Astra and future variants), and Claude Opus/Sonnet 4.6+ and Haiku 5.5 on Anthropic-native transports (`anthropic-messages`, including GitHub Copilot, and Bedrock); unsupported defaults fall back `max` → `xhigh` → `high`, and an unsupported `xhigh` falls back to `max` where the model has it (Claude Opus/Sonnet 4.6), otherwise `high`. On Anthropic-native transports Claude `xhigh` and `max` are sent as Anthropic's native efforts of the same name. Capability detection is by model id, so custom `models.json` Claude entries (including `anthropic/claude-*` ids on an Anthropic-compatible endpoint) get the same behavior. When reasoning is off, dreb sends each Claude model's own off config: `thinking: {type: "disabled"}` for most models, `{type: "between_tools"}` for Sonnet 5.5, and no `thinking` field or effort for Opus 5.5 (its thinking cannot be turned off, so the session offers `minimal` as its lowest level). Codex `ultra` also enables local multi-agent orchestration and is therefore not a raw effort value.
 
 `qwen` uses top-level `enable_thinking`. Use `qwen-chat-template` for local Qwen-compatible servers that require `chat_template_kwargs.enable_thinking`. When reasoning is enabled, dreb sends the mapped effort as a top-level `reasoning_effort`; Qwen3.8+ models default-map dreb's levels onto the three native tiers (`minimal`/`low` → `low`, `medium` → `medium`, `high`/`xhigh`/`max` → `xhigh`).
 

package/docs/providers.md

@@ -8,6 +8,7 @@ dreb supports subscription-based providers via OAuth and API key providers via e
 - [Subscriptions](#subscriptions)
 - [API Keys](#api-keys)
 - [Auth File](#auth-file)
+- [Claude Thinking Levels](#claude-thinking-levels)
 - [Cloud Providers](#cloud-providers)
 - [Custom Providers](#custom-providers)
 - [Resolution Order](#resolution-order)
@@ -37,6 +38,14 @@ Use `/logout` to clear credentials. Tokens are stored in `~/.dreb/agent/auth.jso
 - If you get "model not supported", enable it in VS Code: Copilot Chat → model selector → select model → "Enable"
 - GPT-6 models, including `github-copilot/gpt-6-astra`, use the Responses API. Astra supports `xhigh` and native `max` thinking; access depends on the subscription and enabled models.
 - For adaptive Claude models such as Opus 4.8, readable thinking requires `thinkingDisplay: "summarized"` in SDK calls; enabling thinking alone can return only opaque signatures. See [thinking display](../../ai/README.md#controlling-thinking-display-anthropic-adaptive-models).
+- Copilot Claude Opus, Sonnet, and Haiku (4.x and 5.x) use Copilot's Anthropic Messages endpoint (`/v1/messages`) and follow the same [per-model Claude thinking rules](#claude-thinking-levels) as the Anthropic API: native `xhigh`/`max` effort, `between_tools` for Sonnet 5.5 when thinking is off, and no `off` level for Opus 5.5. Copilot Fable still uses the OpenAI Completions API, which receives no reasoning field.
+- **Model discovery:** at login and on every Copilot token refresh, dreb reads Copilot's `/models` for your account and stores the result with the OAuth credential. Once discovery has succeeded:
+  - Built-in Copilot models your plan or organization doesn't offer (or has disabled) are hidden from `/model`. Custom `models.json` entries stay.
+  - A model whose built-in API endpoint Copilot doesn't serve switches to one it does (`/responses`, then `/chat/completions`).
+
+  Thinking levels and limits still come from dreb's built-in catalog, and models released after your dreb version appear after the next update (or can be added through `models.json`).
+
+  If discovery fails, the last discovered list (or the built-in catalog) is used. The model list updates when the model selector opens or dreb restarts.
 - A registry context window can be lower than the endpoint's hard limit. Accepting a larger request does not change dreb's configured window. Usage-based overflow detection includes uncached input, cache reads, and cache writes.
 
 ### Google Providers
@@ -133,6 +142,21 @@ The `key` field supports three formats:
 
 OAuth credentials are also stored here after `/login` and managed automatically.
 
+## Claude Thinking Levels
+
+On Anthropic-native transports — the Anthropic Messages API (Anthropic, GitHub Copilot, opencode, and custom Anthropic-compatible endpoints) plus Amazon Bedrock — dreb picks Claude thinking and effort fields per model. Capability detection is by model id, so custom `models.json` entries behave the same as built-in ones.
+
+| Model | Thinking on | dreb `off` sends | `off` offered | `xhigh` | `max` |
+|---|---|---|---|---|---|
+| Opus/Sonnet/Haiku 4.5 and older | budget (`enabled`) | `disabled` | ✓ | ✗ | ✗ |
+| Opus 4.6, Sonnet 4.6 | adaptive + effort | `disabled` | ✓ | ✗ (uses `max`) | ✓ |
+| Opus 4.7/4.8, Opus 5, Sonnet 5 | adaptive + effort | `disabled` | ✓ | ✓ | ✓ |
+| Opus 5.5 | adaptive + effort | no `thinking` field or effort (model default) | ✗ (uses `minimal`) | ✓ | ✓ |
+| Sonnet 5.5 | adaptive + effort | `between_tools` | ✓ | ✓ | ✓ |
+| Haiku 5.5 | adaptive + effort | `disabled` | ✓ | ✓ | ✓ |
+
+Effort mapping: `minimal`/`low` → `low`, `medium` → `medium`, `high` → `high`, `xhigh` → `xhigh`, `max` → `max`. No effort is sent while thinking is off. Background calls that send no reasoning (compaction, tab titles, an arbiter with no `thinking` set) to Opus 5.5 therefore run at the model's default effort. Claude served through OpenAI-compatible APIs (OpenRouter, Copilot Fable) keeps its existing mapping.
+
 ## Cloud Providers
 
 ### Azure OpenAI
@@ -171,6 +195,8 @@ Also supports ECS task roles (`AWS_CONTAINER_CREDENTIALS_*`) and IRSA (`AWS_WEB_
 dreb --provider amazon-bedrock --model us.anthropic.claude-sonnet-4-20250514-v1:0
 ```
 
+Claude thinking on Bedrock follows the same per-model rules as the Anthropic API (see [Claude thinking levels](#claude-thinking-levels)). When reasoning is off, Claude 5-generation models (Opus 5, Sonnet 5/5.5, Haiku 5.5), which think by default, receive an explicit off config in `additionalModelRequestFields` (`disabled`, or `between_tools` for Sonnet 5.5); Opus 5.5 gets no thinking field, and older Claude models keep the previous behavior of omitting it.
+
 Prompt caching is enabled automatically for Claude models whose ID contains a recognizable model name (base models and system-defined inference profiles). For application inference profiles (whose ARNs don't contain the model name), set `AWS_BEDROCK_FORCE_CACHE=1` to enable cache points:
 
 ```bash

package/docs/rpc.md

@@ -514,7 +514,7 @@ Set the reasoning/thinking level for models that support it.
 
 Levels: `"off"`, `"minimal"`, `"low"`, `"medium"`, `"high"`, `"xhigh"`, `"max"`
 
-Note: `"xhigh"` is supported by GPT-5.2 through GPT-5.6 model families, the GPT-6 family (Astra and future variants), Claude Opus 4.6–4.x and Claude 5 families (where it maps to adaptive effort `"max"`), and Kimi Code K3. The separate normalized `"max"` tier is model-aware and currently supported by GPT-5.6 aliases and Sol/Terra/Luna variants, plus the GPT-6 family.
+Note: `"xhigh"` is supported by GPT-5.2 through GPT-5.6 model families, the GPT-6 family (Astra and future variants), Claude Opus 4.7+, Sonnet 5+ and Haiku 5.5 on Anthropic, GitHub Copilot and Bedrock (sent as native effort `"xhigh"`), Claude Opus 4.6+ and Sonnet 5+ on OpenRouter, Kimi Code K3, and Qwen 3.8+. The separate normalized `"max"` tier is model-aware and currently supported by GPT-5.6 aliases and Sol/Terra/Luna variants, the GPT-6 family, and Claude Opus/Sonnet 4.6+ and Haiku 5.5 on Anthropic, GitHub Copilot and Bedrock. Unsupported levels are clamped (`xhigh` → `max` on Claude Opus/Sonnet 4.6, otherwise `high`; `max` → `xhigh` → `high`). `availableThinkingLevels` in `get_state` lists only the levels the current model supports and may omit `"off"`: Claude Opus 5.5 cannot turn thinking off, so its list starts at `"minimal"` and `set_thinking_level` with `"off"` lands on `"minimal"` (the saved default is left unchanged).
 
 Response:
 ```json

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@dreb/coding-agent",
-	"version": "2.71.0",
+	"version": "2.72.0",
 	"description": "Coding agent CLI with read, bash, edit, write tools and session management",
 	"type": "module",
 	"drebConfig": {