ltmoerdani · ltmoerdani · Jul 9, 2026 · Jul 9, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -8,6 +8,16 @@ All notable changes to the **OpenCode Go BYOK Provider** extension are documente
 
 - **`[Usage]` Monthly cost aggregation now respects the subscription anchor.** The monthly window was using calendar month after a regression, but OpenCode Go billing is anchor-based (subscription day/hour). The tracker now derives the window from three tiers: (1) user-configured anchor via "Set spent targets", (2) auto-anchor from the earliest SQLite row (matching actual billing start), (3) calendar month fallback. Also fixes `setManualSpentTargets` which previously computed the monthly cost for the old window before storing the anchor, causing tracked+baseline to mismatch the target. Now reads SQLite costs directly using the prospective window (with the new anchor).
 
+## [0.3.7] — 2026-07-09
+
+### Added
+
+- **`[Thinking]` Reasoning now surfaced to Copilot Chat as collapsible thinking blocks.** Previously, reasoning content from OpenCode models (DeepSeek, Kimi, GLM, Qwen, MiniMax, MiMo) was accumulated internally but never emitted to the VS Code Chat UI as a thinking part — so `chat.agent.thinkingStyle` (`collapsed` / `collapsedPreview` / `fixedScrolling`) had no effect, and reasoning either appeared as flat plain text or was silently dropped. The extension now streams reasoning per-chunk via the proposed `LanguageModelThinkingPart` API (available at runtime since VS Code ~1.102, well within our `^1.125.0` floor), across all four transports (chat-completions, Anthropic messages, OpenAI responses, Google Gemini) and both streaming + non-stream response paths. This makes `chat.agent.thinkingStyle` work for OpenCode BYOK models, matching the behavior of Copilot-hosted models. Falls back to the legacy accumulate-and-flush behavior on hypothetical older hosts via a runtime guard. Tool-call replication (`onReasoningContent`) and think-tag filtering (`opencodego.stripThinkTags`) remain intact and compose with the new surfacing. No `enabledApiProposals` declaration needed. Verified working with DeepSeek and Kimi in Copilot Chat. Fixes [#22](https://github.com/ltmoerdani/opencode-copilot-chat/issues/22) and [#71](https://github.com/ltmoerdani/opencode-copilot-chat/issues/71). See `docs/issues/33-20260709-thinking-part-byok-surfacing-research.md`.
+
+### Changed
+
+- **`[Streaming]` `[stream-summary]` log now reports total reasoning characters accurately.** Previously, `reasoningChars` in the log reflected only the remaining `reasoningContent` string, which is cleared by `flushToolCalls` (for tool-call replication) and `flushReasoningFallback` — so the metric showed `0` even when reasoning was streamed. Now tracks a monotonic `totalReasoningChars` counter that survives clears, giving accurate per-response reasoning metrics for debugging.
+
 ## [0.3.6] — 2026-07-08
 
 ### Fixed

diff --git a/docs/devlog.md b/docs/devlog.md
@@ -1,17 +1,58 @@
 # 🧠 OPENCODE COPILOT CHAT DEVLOG
-**Branch:** `main` | **Updated:** 2026-07-02 Asia/Jakarta | **Current Phase:** v0.3.5 — Session Cost + CopilotCredits ✅
+**Branch:** `fix/thinking-part-byok-surfacing-22-71` | **Updated:** 2026-07-09 Asia/Jakarta | **Current Phase:** v0.3.7 — Thinking Part Surfacing ✅ Solved
 
 ---
 
 ## ⚡ Session Handoff
 
 | Field | Value |
 |-------|-------|
-| **Last Session** | 2026-07-02 |
-| **Worked On** | Documented PR #60 (@Wallacy): SQLite-backed cost accuracy (fixes #59), DeepSeek context overflow fix, SSE log gating. Issue doc `30-20260630-pr60-*` written. Feature doc `03-20260605-go-usage-tracker.md` updated with SQLite integration section. Devlog entry added. Also documented PR #68 (@Wallacy): GLM thinking effort values, output channel popup removal, Authorization header on model fetch. Issue doc `31-20260702-pr68-*` written. Thinking controls feature doc `02-20260517-*` updated. |
-| **Stopped At** | `main` post-PR #68 merge. Issue docs + devlog complete. |
-| **Next Action** | → Any open issues or PR reviews. |
-| **Open Issues** | (1) VS Code API gap: `ProvideLanguageModelChatResponseOptions` lacks thread ID → `getCurrentSessionCost()` uses global most-recent (by design for now). (2) Qwen image requests can hit Alibaba quota. (3) `qwen3.6-plus-free` can loop tool calls. (4) Issue #57/#58 — Agent model visibility sync blocked by VS Code API limitation (workaround: `showAgentModelsInManagePanel: true`). |
+| **Last Session** | 2026-07-09 |
+| **Worked On** | Implemented + verified fix for issues #22 + #71 on branch `fix/thinking-part-byok-surfacing-22-71`. Reasoning from OpenCode models now streamed live via `LanguageModelThinkingPart` across all 4 transports (chat-completions, Anthropic, responses, Google) and both streaming + non-stream paths. `chat.agent.thinkingStyle` now respected. Manual test passed (DeepSeek + Kimi). Version bumped to 0.3.7. CHANGELOG, issue doc #33 (✅ Solved), and devlog updated. |
+| **Stopped At** | v0.3.7 ready. Pending commit + push + reply to #22/#71. |
+| **Next Action** | → Commit, push branch, open PR. Draft reply to #22/#71 (after merge or before). |
+| **Open Issues** | (1) VS Code API gap: thread ID → session cost. (2) Qwen image quota. (3) `qwen3.6-plus-free` tool-call loop. (4) #57/#58 agent model visibility. |
+
+---
+
+## 🔬 Issues #22 + #71 Deep-Dive — Thinking Part BYOK Surfacing — Session 2026-07-09 ✅ SOLVED
+
+**Action:** User asked to compare issues #22 (`chat.agent.thinkingStyle` not respected) and #71 (thinking tokens not displaying). Confirmed they are duplicates. Deep-dive research overturned the previous "upstream blocker" conclusion from doc `23-*`.
+
+**Key findings:**
+
+1. **Issues #22 and #71 are the same bug.** Both report reasoning from OpenCode models never rendered as a collapsible thinking block. #71 author explicitly frustrated ("negatively affecting results by a huge margin").
+
+2. **The "upstream blocker" narrative was wrong.** Doc `23-*` (2026-06-15) concluded this was blocked on `microsoft/vscode#318211` and unfixable extension-side. Research on 2026-07-09 found:
+   - `LanguageModelThinkingPart` API shipped to VS Code in **August 2025** (PR #259939). Our `engines.vscode: ^1.125.0` guarantees it is present at runtime.
+   - `Vizards/deepseek-v4-for-copilot` v0.6.2 (Marketplace, engine `^1.116.0`) **already solves this** for DeepSeek BYOK via `progress.report(new vscode.LanguageModelThinkingPart(text))` — with **no `enabledApiProposals`** in `package.json`.
+   - User @yinhx3 confirmed in #71: deepseek-v4-for-copilot "is able to display reasoning content."
+   - DeepSeek-v4 tracker has **zero open issues** about reasoning not displaying.
+
+3. **Root cause in our codebase:** `src/streaming.ts` `OpenAiResponseExtractor` accumulates `reasoningContent` but never emits it as a thinking part. `flushReasoningFallback` drops it silently when text/tool calls are present.
+
+**Verified implementation plan (6 steps):**
+1. Create `src/vscode.proposed.languageModelThinkingPart.d.ts` (type augmentation).
+2. Emit reasoning per-chunk via `LanguageModelThinkingPart` in the streaming extractor.
+3. Runtime guard `typeof vscode.LanguageModelThinkingPart === 'function'` (defensive only — our floor is 1.125.0).
+4. Refactor `flushReasoningFallback` to route through thinking part.
+5. **No `package.json` change** (no `enabledApiProposals` needed).
+6. Verify: `npm run compile` + manual test per model family + all three `thinkingStyle` values.
+
+**Documentation produced:**
+- `docs/issues/33-20260709-thinking-part-byok-surfacing-research.md` — comprehensive issue doc (status ✅ Solved).
+- `docs/issues/23-20260615-thinking-style-setting-not-respected.md` — marked ⚠️ Deprecated with redirect banner.
+- `/memories/repo/issue22-71-thinking-part-bypass.md` — repo memory snapshot with full research.
+
+**Implementation + verification:**
+- `src/vscode.proposed.languageModelThinkingPart.d.ts` — type augmentation (NEW).
+- `src/streaming.ts` — `emitThinkingPart()` helper + runtime guard; both extractors extended with `handleReasoning()`; all 4 transport call sites updated; `flushReasoningFallback` refactored; non-stream path (`extractChatCompletionParts`, `extractAnthropicParts`) also updated; `totalReasoningChars` monotonic counter for accurate log metrics.
+- `npm run compile` + `npx tsc --noEmit --strict` → both pass.
+- Manual test: DeepSeek + Kimi in Copilot Chat → reasoning rendered as collapsible thinking block. `chat.agent.thinkingStyle` respected. Tool-call replication intact.
+
+**Released in:** v0.3.7.
+
+**Next:** Commit, push branch, open PR, reply to #22/#71.
 
 ---
 

diff --git a/docs/issues/23-20260615-thinking-style-setting-not-respected.md b/docs/issues/23-20260615-thinking-style-setting-not-respected.md
@@ -1,4 +1,8 @@
-**Status:** 🟢 Active
+**Status:** ⚠️ Deprecated
+
+> **⚠️ Superseded by [`33-20260709-thinking-part-byok-surfacing-research.md`](./33-20260709-thinking-part-byok-surfacing-research.md).**
+>
+> The conclusion below ("blocked on upstream, no extension-side fix") was **overturned** on 2026-07-09 after deep-dive research found that `LanguageModelThinkingPart` is available in the VS Code runtime we target, and a shipping Marketplace extension (`Vizards/deepseek-v4-for-copilot`) already solves this exact problem. This document is retained for historical reference only — do not act on its "Recommended Action" section. See doc `33-*` for the verified, implementable fix plan.
 
 # `chat.agent.thinkingStyle` Not Respected — Reasoning Always Expanded