pi-commandcode-provider 0.1.1-next.0 → 0.3.0

package/CHANGELOG.md

@@ -0,0 +1,28 @@
+# Changelog
+
+## 0.3.0 - 2026-05-28
+
+- Add OMP (Oh My Pi) provider compatibility: support `~/.omp/agent/auth.json` auth path, handle OMP's env-var-name-as-apiKey quirk, convert OMP system prompt arrays to text.
+- Close open thinking blocks before starting text or tool output to prevent event ordering issues when upstream omits `reasoning-end`.
+- Correct DeepSeek V4 Pro discount as permanent (no expiry), not time-limited.
+- Correct DeepSeek V4 Flash cache-read rate to $0.028/M and add xiaomi/mimo models to pricing table.
+- Upgrade pi dependencies from `@mariozechner` 0.72.0 to `@earendil-works` 0.75.5.
+- Move `pi-coding-agent` to optional peerDependencies.
+
+## 0.2.0 - 2026-05-27
+
+- Stream `reasoning-delta` events incrementally instead of buffering the full thinking block until `reasoning-end`. Emits `thinking_start`, `thinking_delta`, and `thinking_end` events as they arrive so the UI can show reasoning in real time.
+- Close open text blocks on `reasoning-start` and `reasoning-delta` so thinking and text never overlap in the output.
+- Add live display pricing (`MODEL_COSTS`) for known Command Code models. Cost falls back to zero for models not yet in the price table until the Provider API exposes pricing directly.
+- Fetch models from the Command Code Provider API at startup (inherited from upstream 0.1.1) and overlay the static cost table.
+
+## 0.1.1 - 2026-05-26
+
+- Align Command Code generate requests with CLI `0.27.2` headers and payload shape.
+- Support official Command Code CLI auth files using the `command-code` credential key.
+- Handle `reasoning-start` and ignore streamed `tool-result` events.
+- Cap generated `max_tokens` by the selected model and the Command Code output limit.
+
+## 0.1.0 - 2026-05-05
+
+- Initial public release.

package/CONTRIBUTING.md

@@ -0,0 +1,136 @@
+# Contributing
+
+Thanks for helping improve `pi-commandcode-provider`.
+
+This is an unofficial Command Code provider for pi. Keep changes small, tested, and easy to review.
+
+## Development setup
+
+```sh
+npm install
+npm test
+```
+
+Useful commands:
+
+```sh
+npm run typecheck
+npm run format:check
+npm run test:unit
+npm run test:models
+npm run test:oauth
+npm run test:abort
+npm run test:stream
+npm run test:pi-local
+```
+
+Before opening a PR, run:
+
+```sh
+npm test
+npm run format:check
+git diff --check
+```
+
+For release and npm smoke-test steps, see [RELEASE.md](RELEASE.md).
+
+## Pull request guidelines
+
+- Keep PRs focused on one problem or feature.
+- Add or update tests for behavior changes.
+- Update `README.md`, `CHANGELOG.md`, or `RELEASE.md` when user-facing behavior changes.
+- Avoid broad refactors unless the PR is specifically about refactoring.
+- Do not include API keys, tokens, real auth files, `.env` files, or other secrets.
+- Prefer documented/public Command Code API behavior. If compatibility with CLI behavior is needed, document why.
+- Make sure npm package contents still make sense when `package.json` `files` changes.
+
+## Testing pi integration changes
+
+For provider, auth, request-shape, or stream changes, test both local code and the package form when possible.
+
+Local extension smoke:
+
+```sh
+pi --no-extensions -e ./index.ts --list-models commandcode
+```
+
+Npm package smoke and isolated `/login` testing are documented in [RELEASE.md](RELEASE.md#test-the-npm-package-in-pi).
+
+## Commit message rules
+
+Use Angular-style Conventional Commits.
+
+Format:
+
+```txt
+<type>(<scope>): <subject>
+```
+
+Examples:
+
+```txt
+feat(auth): support Command Code CLI auth files
+fix(core): cap max tokens by selected model
+docs(release): document npm smoke testing
+test(stream): cover reasoning start events
+chore(release): publish 0.1.1
+```
+
+### Types
+
+Use one of these types:
+
+- `feat`: a new user-facing feature
+- `fix`: a bug fix
+- `docs`: documentation-only changes
+- `style`: formatting-only changes, no behavior change
+- `refactor`: code restructuring without behavior change
+- `perf`: performance improvement
+- `test`: adding or changing tests
+- `build`: package, dependency, or build-system changes
+- `ci`: CI workflow changes
+- `chore`: maintenance that does not fit another type
+- `revert`: revert a previous commit
+
+### Scopes
+
+Use a short lowercase scope. Prefer existing project areas:
+
+- `auth`
+- `oauth`
+- `core`
+- `models`
+- `stream`
+- `tests`
+- `docs`
+- `release`
+- `deps`
+- `ci`
+
+A scope is strongly recommended. If no scope fits, choose the closest project area instead of omitting it.
+
+### Subject line
+
+- Use imperative mood: `fix(auth): read oauth credentials`, not `fixed` or `fixes`.
+- Keep it concise.
+- Start lowercase after the colon.
+- Do not end with a period.
+
+### Body and footers
+
+Use a body when the reason is not obvious:
+
+```txt
+fix(core): cap max tokens by selected model
+
+Command Code can return models with lower output limits than the provider-wide cap.
+Clamp defaults to the selected model so requests do not exceed upstream limits.
+```
+
+Breaking changes must be marked with `!` or a `BREAKING CHANGE:` footer:
+
+```txt
+feat(api)!: switch to provider api endpoints
+
+BREAKING CHANGE: removes support for the legacy internal generate endpoint.
+```

package/README.md

@@ -36,6 +36,18 @@ Then reload pi:
 /reload
 ```
 
+### Oh My Pi
+
+```sh
+omp plugin install pi-commandcode-provider
+```
+
+Then restart OMP or run:
+
+```txt
+/reload
+```
+
 ## Setup
 
 Set your Command Code API key using one of these methods:
@@ -83,7 +95,7 @@ The official Command Code CLI auth shape is also supported:
 }
 ```
 
-Or use pi's auth file at `~/.pi/agent/auth.json`:
+Or use a pi/OMP auth file at `~/.pi/agent/auth.json` or `~/.omp/agent/auth.json`:
 
 ```json
 {
@@ -99,12 +111,21 @@ After installing and setting your API key, select a Command Code model in pi:
 /model deepseek/deepseek-v4-flash
 ```
 
-Any query will then use the Command Code API. You can list available models within pi:
+Any query will then use the Command Code API. You can list available models:
 
-```txt
-/models
+```sh
+pi -e index.ts --list-models   # or /models within pi
+omp -e index.ts --list-models
+```
+
+In OMP, use the provider-qualified model name:
+
+```sh
+omp -p "hello" --model commandcode/deepseek/deepseek-v4-flash
 ```
 
+OMP currently resolves `--provider commandcode --model ...` before extension providers are loaded, so prefer `--model commandcode/<model-id>`. <!-- TODO: remove this note once OMP fixes provider resolution order for extension-loaded providers -->
+
 ## Model discovery
 
 On startup, the provider fetches:
@@ -115,12 +136,22 @@ https://api.commandcode.ai/provider/v1/models
 
 For tests or local mocks, override it with `COMMANDCODE_MODELS_URL`.
 
-## Publish
+## Pricing
 
-```sh
-npm login
-npm publish --access public
-```
+Command Code does not yet expose model pricing through its Provider API. The provider ships a static cost table (`MODEL_COSTS` in `index.ts`) for known models so that pi can display per-model pricing.
+
+- Models present in `MODEL_COSTS` show their real per-million-token rates (including promotional deals like the DeepSeek V4 Pro 4× discount and Qwen 3.7 Max 2× discount).
+- Models **not** in the table fall back to zero cost. When the Provider API adds a `cost` field, the static table can be removed.
+
+To add or update a price, edit the `MODEL_COSTS` record in `index.ts` and update the corresponding test in `tests/test-pricing.ts`.
+
+## Contributing
+
+See [CONTRIBUTING.md](CONTRIBUTING.md) for development setup, PR expectations, and commit message rules.
+
+## Release
+
+See [RELEASE.md](RELEASE.md) for the prerelease, npm smoke-test, stable publish, git tag, and GitHub follow-up checklist.
 
 ## License
 

package/RELEASE.md

@@ -0,0 +1,192 @@
+# Release Process
+
+This project uses npm semver releases.
+
+Recommended flow:
+
+- publish prereleases with the `next` dist-tag
+- smoke-test the npm package directly in pi
+- publish stable releases with the `latest` dist-tag
+- commit and tag the stable release
+- comment on the related PR or issue after shipping
+
+## Prerelease flow
+
+Use `next` for beta/alpha/manual validation builds.
+
+```sh
+npm version prepatch --preid next --no-git-tag-version
+npm test
+npm run format:check
+npm pack --dry-run
+npm publish --tag next --access public
+```
+
+If npm asks for browser or OTP auth, run the publish command manually and complete the npm prompt.
+
+Verify the registry state:
+
+```sh
+npm view pi-commandcode-provider@next version dist-tags --json
+```
+
+Expected:
+
+- `next` points to the prerelease version
+- `latest` still points to the previous stable version
+
+## Test the npm package in pi
+
+Always test from npm, not the local checkout.
+
+### 1. Model discovery smoke test
+
+```sh
+PI_SKIP_VERSION_CHECK=1 \
+pi --no-extensions \
+  -e npm:pi-commandcode-provider@next \
+  --list-models commandcode
+```
+
+Expected:
+
+- provider `commandcode` appears
+- live Command Code models are listed
+
+### 2. Manual `/login` test with isolated pi config
+
+Use temporary pi config and session directories so the test does not touch your real pi auth.
+
+```sh
+export PI_CC_TEST_AGENT_DIR="$(mktemp -d)"
+export PI_CC_TEST_SESSION_DIR="$(mktemp -d)"
+
+export PI_CODING_AGENT_DIR="$PI_CC_TEST_AGENT_DIR"
+export PI_CODING_AGENT_SESSION_DIR="$PI_CC_TEST_SESSION_DIR"
+export PI_SKIP_VERSION_CHECK=1
+
+pi --no-extensions \
+  -e npm:pi-commandcode-provider@next \
+  --provider commandcode \
+  --model deepseek/deepseek-v4-flash
+```
+
+Inside pi:
+
+```txt
+/login
+```
+
+Then:
+
+1. choose **Use a subscription**
+2. choose **Command Code**
+3. complete the browser auth flow
+4. if automatic transfer fails, paste the copied Command Code API key into pi
+5. send this message:
+
+```txt
+Reply exactly: manual-npm-ok
+```
+
+Expected:
+
+- login succeeds
+- a Command Code credential is saved under the temporary `PI_CODING_AGENT_DIR`
+- the model replies exactly `manual-npm-ok`
+
+### 3. Post-login print-mode test
+
+Using the same exported temp variables from above:
+
+```sh
+pi --no-extensions \
+  -e npm:pi-commandcode-provider@next \
+  --no-session \
+  -p \
+  --provider commandcode \
+  --model deepseek/deepseek-v4-flash \
+  "Reply exactly: manual-npm-ok"
+```
+
+Expected:
+
+```txt
+manual-npm-ok
+```
+
+### 4. Cleanup isolated pi config
+
+Only run this if these variables were created by the test above:
+
+```sh
+rm -rf "$PI_CC_TEST_AGENT_DIR" "$PI_CC_TEST_SESSION_DIR"
+unset PI_CC_TEST_AGENT_DIR PI_CC_TEST_SESSION_DIR
+unset PI_CODING_AGENT_DIR PI_CODING_AGENT_SESSION_DIR PI_SKIP_VERSION_CHECK
+```
+
+## Stable release flow
+
+After the `next` package is verified, set the intended stable version:
+
+```sh
+npm version 0.1.1 --no-git-tag-version
+```
+
+Replace `0.1.1` with the intended stable version.
+
+Update `CHANGELOG.md`, then run checks:
+
+```sh
+npm test
+npm run format:check
+npm pack --dry-run
+git diff --check
+```
+
+Commit and tag:
+
+```sh
+git add .
+git commit -m "Release 0.1.1"
+git tag -a v0.1.1 -m "Release 0.1.1"
+```
+
+Publish stable:
+
+```sh
+npm publish --tag latest --access public
+```
+
+If npm asks for browser or OTP auth, run the publish command manually and complete the npm prompt.
+
+Verify npm:
+
+```sh
+npm view pi-commandcode-provider version dist-tags --json
+npm view pi-commandcode-provider@0.1.1 version --json
+```
+
+Expected:
+
+- `latest` points to the stable version
+- the stable version exists on npm
+
+Push commit and tag:
+
+```sh
+git push origin main
+git push origin v0.1.1
+```
+
+## GitHub follow-up
+
+Comment on the related PR and issue after publishing and pushing:
+
+```sh
+gh pr comment <number> --body "Shipped in \`pi-commandcode-provider@0.1.1\` / tag \`v0.1.1\`."
+
+gh issue comment <number> --body "Shipped in \`pi-commandcode-provider@0.1.1\` / tag \`v0.1.1\`."
+```
+
+Only comment on PRs or issues actually included in the release.

package/index.ts

@@ -12,8 +12,8 @@
  * Models are fetched from Command Code's Provider API at startup.
  */
 
-import { calculateCost, createAssistantMessageEventStream } from "@mariozechner/pi-ai"
-import type { ExtensionAPI } from "@mariozechner/pi-coding-agent"
+import { AssistantMessageEventStream, calculateCost } from "@earendil-works/pi-ai"
+import type { ExtensionAPI } from "@earendil-works/pi-coding-agent"
 
 import { COMMAND_CODE_CLI_VERSION, createStreamCommandCode, DEFAULT_API_BASE } from "./src/core.ts"
 import { DEFAULT_MODELS_URL, fetchCommandCodeModels } from "./src/models.ts"
@@ -22,8 +22,53 @@ import { getApiKey, login, refreshToken } from "./src/oauth.ts"
 const API_BASE = process.env.COMMANDCODE_API_BASE ?? DEFAULT_API_BASE
 const MODELS_URL = process.env.COMMANDCODE_MODELS_URL ?? DEFAULT_MODELS_URL
 
+type CommandCodeModelCost = {
+  input: number
+  output: number
+  cacheRead: number
+  cacheWrite: number
+}
+
+const ZERO_MODEL_COST: CommandCodeModelCost = {
+  input: 0,
+  output: 0,
+  cacheRead: 0,
+  cacheWrite: 0,
+}
+
+// The Provider API supplies the current model list. Keep known display pricing
+// here until the Provider API exposes prices directly.
+const MODEL_COSTS: Record<string, CommandCodeModelCost> = {
+  "claude-opus-4-7": { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
+  "claude-opus-4-6": { input: 5, output: 25, cacheRead: 0.5, cacheWrite: 6.25 },
+  "claude-sonnet-4-6": { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3.75 },
+  "claude-haiku-4-5-20251001": { input: 1, output: 5, cacheRead: 0.1, cacheWrite: 1.25 },
+  "gpt-5.5": { input: 5, output: 30, cacheRead: 0.5, cacheWrite: 0 },
+  "gpt-5.4": { input: 2.5, output: 15, cacheRead: 0.25, cacheWrite: 0 },
+  "gpt-5.3-codex": { input: 2, output: 8, cacheRead: 0.5, cacheWrite: 0 },
+  "gpt-5.4-mini": { input: 0.75, output: 4.5, cacheRead: 0.075, cacheWrite: 0 },
+  "google/gemini-3.5-flash": { input: 1.5, output: 9, cacheRead: 0.15, cacheWrite: 0 },
+  "google/gemini-3.1-flash-lite": { input: 0.25, output: 1.5, cacheRead: 0.03, cacheWrite: 0 },
+  // 4× usage deal: 75% off (permanent, no expiry)
+  "deepseek/deepseek-v4-pro": { input: 0.435, output: 0.87, cacheRead: 0.003625, cacheWrite: 0 },
+  "deepseek/deepseek-v4-flash": { input: 0.14, output: 0.28, cacheRead: 0.028, cacheWrite: 0 },
+  "moonshotai/Kimi-K2.6": { input: 0.95, output: 4, cacheRead: 0.16, cacheWrite: 0 },
+  "moonshotai/Kimi-K2.5": { input: 0.6, output: 3, cacheRead: 0.1, cacheWrite: 0 },
+  "zai-org/GLM-5.1": { input: 1.4, output: 4.4, cacheRead: 0.26, cacheWrite: 0 },
+  "zai-org/GLM-5": { input: 1, output: 3.2, cacheRead: 0.2, cacheWrite: 0 },
+  "MiniMaxAI/MiniMax-M2.7": { input: 0.3, output: 1.2, cacheRead: 0.06, cacheWrite: 0 },
+  "MiniMaxAI/MiniMax-M2.5": { input: 0.27, output: 0.95, cacheRead: 0.03, cacheWrite: 0 },
+  "Qwen/Qwen3.6-Max-Preview": { input: 1.3, output: 7.8, cacheRead: 0.26, cacheWrite: 1.63 },
+  "Qwen/Qwen3.6-Plus": { input: 0.5, output: 3, cacheRead: 0.1, cacheWrite: 0 },
+  // 2× usage deal: 50% off through June 22, 2026
+  "Qwen/Qwen3.7-Max": { input: 1.25, output: 3.75, cacheRead: 0.25, cacheWrite: 1.56 },
+  "stepfun/Step-3.5-Flash": { input: 0.1, output: 0.3, cacheRead: 0.02, cacheWrite: 0 },
+  "xiaomi/mimo-v2.5-pro": { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+  "xiaomi/mimo-v2.5": { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+}
+
 const streamCommandCode = createStreamCommandCode({
-  createStream: createAssistantMessageEventStream,
+  createStream: () => new AssistantMessageEventStream(),
   calculateCost,
   apiBase: API_BASE,
 })
@@ -57,7 +102,7 @@ export default async function (pi: ExtensionAPI) {
       name: model.name,
       reasoning: model.reasoning,
       input: ["text"] as const,
-      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+      cost: MODEL_COSTS[model.id] ?? ZERO_MODEL_COST,
       contextWindow: model.contextWindow,
       maxTokens: model.maxTokens,
     })),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-commandcode-provider",
-  "version": "0.1.1-next.0",
+  "version": "0.3.0",
   "description": "pi custom provider for Command Code API (commandcode.ai)",
   "type": "module",
   "keywords": [
@@ -22,15 +22,19 @@
     "index.ts",
     "src/",
     "README.md",
+    "CHANGELOG.md",
+    "CONTRIBUTING.md",
+    "RELEASE.md",
     "LICENSE"
   ],
   "scripts": {
-    "test": "npm run typecheck && tsx tests/test-pure-functions.ts && tsx tests/test-models.ts && tsx tests/test-oauth.ts && tsx tests/test-abort.ts && tsx tests/test-stream.ts && node tests/test-pi-local.mjs",
+    "test": "npm run typecheck && tsx tests/test-pure-functions.ts && tsx tests/test-models.ts && tsx tests/test-pricing.ts && tsx tests/test-oauth.ts && tsx tests/test-abort.ts && tsx tests/test-stream.ts && node tests/test-pi-local.mjs && node tests/test-omp-compat.mjs",
     "typecheck": "tsc --noEmit",
     "format:check": "prettier --check '**/*.{ts,mjs,json,md}'",
     "format": "prettier --write '**/*.{ts,mjs,json,md}'",
     "test:unit": "tsx tests/test-pure-functions.ts",
     "test:models": "tsx tests/test-models.ts",
+    "test:pricing": "tsx tests/test-pricing.ts",
     "test:oauth": "tsx tests/test-oauth.ts",
     "test:abort": "tsx tests/test-abort.ts",
     "test:stream": "tsx tests/test-stream.ts",
@@ -43,13 +47,20 @@
     ]
   },
   "devDependencies": {
-    "@mariozechner/pi-coding-agent": "0.72.0",
     "@types/node": "25.6.0",
     "prettier": "^3.5.0",
     "tsx": "4.21.0",
     "typescript": "6.0.3"
   },
   "dependencies": {
-    "@mariozechner/pi-ai": "0.72.0"
+    "@earendil-works/pi-ai": "0.75.5"
+  },
+  "peerDependencies": {
+    "@earendil-works/pi-coding-agent": "^0.75.5"
+  },
+  "peerDependenciesMeta": {
+    "@earendil-works/pi-coding-agent": {
+      "optional": true
+    }
   }
 }

package/src/converters.ts

@@ -39,7 +39,11 @@ export function numberValue(value: unknown): number | undefined {
 }
 
 function defaultAuthPaths(home: string): string[] {
-  return [join(home, ".commandcode", "auth.json"), join(home, ".pi", "agent", "auth.json")]
+  return [
+    join(home, ".commandcode", "auth.json"),
+    join(home, ".omp", "agent", "auth.json"),
+    join(home, ".pi", "agent", "auth.json"),
+  ]
 }
 
 function apiKeyFromCredentialRecord(value: unknown): string | undefined {
@@ -284,3 +288,30 @@ export function mapFinishReason(reason: unknown): StopReason {
   }
   return "stop"
 }
+
+function promptPartToText(value: unknown, depth = 0): string {
+  if (depth > 10) return ""
+  if (typeof value === "string") return value
+  if (Array.isArray(value))
+    return value
+      .map((v) => promptPartToText(v, depth + 1))
+      .filter(Boolean)
+      .join("\n")
+  if (!isRecord(value)) return ""
+  const text = stringValue(value.text)
+  if (text) return text
+  const content = promptPartToText(value.content, depth + 1)
+  if (content) return content
+  return ""
+}
+
+export function systemPromptToText(value: unknown): string {
+  if (value === undefined || value === null) return ""
+  if (typeof value === "string") return value
+  if (Array.isArray(value))
+    return value
+      .map((v) => promptPartToText(v, 0))
+      .filter(Boolean)
+      .join("\n\n")
+  return promptPartToText(value, 0)
+}

package/src/core.ts

@@ -18,6 +18,7 @@ import {
   recordOrEmpty,
   stringValue,
   toolsToJson,
+  systemPromptToText,
 } from "./converters.ts"
 import type {
   AssistantMessageEventStreamLike,
@@ -131,8 +132,13 @@ export function createStreamCommandCode(deps: CoreDependencies) {
     const stream = deps.createStream()
 
     async function run() {
+      // OMP may pass the env-var name "COMMANDCODE_API_KEY" as the apiKey
+      // value instead of resolving it. Filter out this specific string.
+      const hostKey =
+        options?.apiKey && options.apiKey !== "COMMANDCODE_API_KEY" ? options.apiKey : undefined
+
       const apiKey =
-        options?.apiKey ??
+        hostKey ??
         getApiKey({
           env: deps.env,
           authPaths: deps.authPaths,
@@ -149,7 +155,7 @@ export function createStreamCommandCode(deps: CoreDependencies) {
           usage: defaultUsage(),
           stopReason: "error",
           errorMessage:
-            "No Command Code API key. Run /login and select Command Code, set COMMANDCODE_API_KEY env var, or configure ~/.commandcode/auth.json or ~/.pi/agent/auth.json.",
+            "No Command Code API key. Run /login and select Command Code, set the COMMANDCODE_API_KEY env var, or configure ~/.commandcode/auth.json, ~/.pi/agent/auth.json or ~/.omp/agent/auth.json",
           timestamp: now(),
         }
         stream.push({ type: "error", reason: "error", error: msg })
@@ -172,7 +178,7 @@ export function createStreamCommandCode(deps: CoreDependencies) {
       let reader: ReadableStreamDefaultReader<Uint8Array> | undefined
       let textBlock: TextContent | undefined
       let currentTextIdx = -1
-      let thinkingBlock: string[] = []
+      let thinkingIdx = -1
       let finished = false
 
       const abortUpstream = () => {
@@ -204,29 +210,18 @@ export function createStreamCommandCode(deps: CoreDependencies) {
         currentTextIdx = -1
       }
 
-      const flushThinkingBlock = () => {
-        if (thinkingBlock.length === 0) return
-        const thinkingText = thinkingBlock.join("")
-        thinkingBlock = []
-        output.content.push({ type: "thinking", thinking: thinkingText })
-        const idx = output.content.length - 1
-        stream.push({
-          type: "thinking_start",
-          contentIndex: idx,
-          partial: output,
-        })
-        stream.push({
-          type: "thinking_delta",
-          contentIndex: idx,
-          delta: thinkingText,
-          partial: output,
-        })
-        stream.push({
-          type: "thinking_end",
-          contentIndex: idx,
-          content: thinkingText,
-          partial: output,
-        })
+      const endThinking = () => {
+        if (thinkingIdx < 0) return
+        const tc = output.content[thinkingIdx]
+        if (tc && tc.type === "thinking") {
+          stream.push({
+            type: "thinking_end",
+            contentIndex: thinkingIdx,
+            content: (tc as { thinking: string }).thinking,
+            partial: output,
+          })
+        }
+        thinkingIdx = -1
       }
 
       const handleEvent = (event: unknown) => {
@@ -234,6 +229,7 @@ export function createStreamCommandCode(deps: CoreDependencies) {
 
         switch (event.type) {
           case "text-delta": {
+            endThinking()
             if (!textBlock) {
               textBlock = { type: "text", text: "" }
               output.content.push(textBlock)
@@ -262,12 +258,32 @@ export function createStreamCommandCode(deps: CoreDependencies) {
 
           case "reasoning-delta": {
             endTextBlock()
-            thinkingBlock.push(stringValue(event.text) ?? "")
+            const delta = stringValue(event.text) ?? ""
+            if (thinkingIdx < 0) {
+              output.content.push({ type: "thinking", thinking: delta })
+              thinkingIdx = output.content.length - 1
+              stream.push({
+                type: "thinking_start",
+                contentIndex: thinkingIdx,
+                partial: output,
+              })
+            } else {
+              const tc = output.content[thinkingIdx]
+              if (tc && tc.type === "thinking") {
+                ;(tc as { thinking: string }).thinking += delta
+              }
+            }
+            stream.push({
+              type: "thinking_delta",
+              contentIndex: thinkingIdx,
+              delta,
+              partial: output,
+            })
             break
           }
 
           case "reasoning-end": {
-            flushThinkingBlock()
+            endThinking()
             break
           }
 
@@ -277,6 +293,7 @@ export function createStreamCommandCode(deps: CoreDependencies) {
 
           case "tool-call": {
             endTextBlock()
+            endThinking()
             const toolCall: ToolCallContent = {
               type: "toolCall",
               id: stringValue(event.toolCallId) ?? "",
@@ -355,7 +372,7 @@ export function createStreamCommandCode(deps: CoreDependencies) {
             model: model.id,
             messages: messagesToCC(context.messages),
             tools: toolsToJson(context.tools),
-            system: context.systemPrompt ?? "",
+            system: systemPromptToText(context.systemPrompt),
             max_tokens: generateMaxTokens(model, options),
             temperature: 0.3,
             stream: true,
@@ -436,7 +453,7 @@ export function createStreamCommandCode(deps: CoreDependencies) {
         }
 
         endTextBlock()
-        flushThinkingBlock()
+        endThinking()
 
         stream.push({
           type: "done",