feat(telemetry,ui): reasoning span attribute + collapsed thought summary (Phase 1 of #162)#165
Merged
Merged
Conversation
* feat(core): preserve task plan state in compaction summaries When context compaction fires, the agent loses awareness of its task plan (completed, in-progress, pending work) and may re-plan already-done tasks. Add extractTaskPlanSummary() that queries the TaskStore and produces a structured <task-plan> XML section with status markers ([x], [~], [ ], [-], [!]), priority labels, and parent-child indentation. Extend compactMessages() to accept an optional taskStore and append the plan to the compaction summary. Wire the TaskStore into agent-core at the compaction call site. Backward compatible: existing callers without taskStore remain unaffected. Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix: add error handling and recursive nesting to compaction task plan Address PR feedback from CodeRabbit: - Wrap extractTaskPlanSummary call in try/catch so TaskStore failures don't break compaction - Replace flat 2-level subtask rendering with recursive renderTask() that supports arbitrary nesting depth - Add tests for multi-level nesting and error fallback Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> --------- Co-authored-by: Automaker <automaker@localhost> Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
…houghts post-stream Phase 1 of the reasoning coordination tracked in #162. Captures delta.reasoning_content / delta.reasoning across stream chunks and surfaces it as gen_ai.response.thinking on the gen_ai chat span (gated on logPrompts, matching the completion event policy). Always emits gen_ai.usage.thinking_tokens when usage exposes it. Non-streaming responses get the same treatment by inspecting {thought:true} parts on the response — and the completion event no longer double-counts thoughts as content. Renders gemini_thought items as a compact "▸ thinking (N chars)" summary once the stream finalizes (live streaming render unchanged). Full text remains in ChatRecord, ACP agent_thought_chunk notifications, and Langfuse for downstream investigation. An in-TUI expand affordance is a follow-up. Once homelab-iac#31 (EMIT_REASONING_CONTENT) flips on, this also covers vLLM-served models that previously lost their <think> blocks at the gateway. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
WalkthroughThis pull request extends streaming support with explicit reasoning/thinking telemetry tracking and introduces task-plan state preservation during context compaction. Changes span pipeline telemetry (capturing and emitting Changes
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Possibly related issues
Possibly related PRs
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Review rate limit: 4/5 reviews remaining, refill in 12 minutes. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (1)
packages/core/src/core/openaiContentGenerator/pipeline.ts (1)
334-345: ⚡ Quick winDeduplicate telemetry truncation policy into a shared helper.
10_000and...[truncated]are repeated across streaming and non-streaming paths. Centralizing this reduces drift risk when policy changes.♻️ Suggested refactor
+const TELEMETRY_TEXT_LIMIT = 10_000; +const TELEMETRY_TRUNCATION_SUFFIX = '...[truncated]'; + +function truncateTelemetryText(text: string): string { + return text.length > TELEMETRY_TEXT_LIMIT + ? text.slice(0, TELEMETRY_TEXT_LIMIT) + TELEMETRY_TRUNCATION_SUFFIX + : text; +} ... - context.span.setAttribute( - 'gen_ai.response.thinking', - reasoningText.length > 10_000 - ? reasoningText.slice(0, 10_000) + '...[truncated]' - : reasoningText, - ); + context.span.setAttribute( + 'gen_ai.response.thinking', + truncateTelemetryText(reasoningText), + ); ... - 'gen_ai.completion': - responseText.length > 10_000 - ? responseText.slice(0, 10_000) + '...[truncated]' - : responseText, + 'gen_ai.completion': truncateTelemetryText(responseText), ... - span.setAttribute( - 'gen_ai.response.thinking', - reasoningText.length > 10_000 - ? reasoningText.slice(0, 10_000) + '...[truncated]' - : reasoningText, - ); + span.setAttribute( + 'gen_ai.response.thinking', + truncateTelemetryText(reasoningText), + );Also applies to: 738-770
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@packages/core/src/core/openaiContentGenerator/pipeline.ts` around lines 334 - 345, Centralize the telemetry truncation policy by adding a single helper (e.g., truncateForTelemetry or truncateTelemetryText) that takes a string, enforces the 10_000-char limit and appends '...[truncated]' when needed, then replace the inline logic in the non-streaming path that builds reasoningText and calls context.span.setAttribute('gen_ai.response.thinking', ...) (uses reasoningParts) and the equivalent streaming-path code (the block referenced around the other occurrence) to call this helper instead of duplicating 10_000 and '...[truncated]'; ensure the helper is exported/visible to both code paths and keep behavior identical.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@packages/core/src/agents/runtime/agent-core.ts`:
- Around line 474-481: The current early-return when estimateTokens(masked) <=
targetTokens skips the `<task-plan>` enrichment that compactMessages(...)
applies; instead of returning masked directly, always invoke
compactMessages(masked, targetTokens, { taskStore:
this.runtimeContext.getTaskStore?.() }) and assign compacted to its result
(await if it returns a Promise) so the task-plan summary enrichment runs even
when masking already meets the token target; use the existing variables masked,
targetTokens, taskStore and preserve the async handling used in the else branch.
In `@packages/core/src/agents/runtime/compaction.ts`:
- Around line 58-85: The orphan pass currently emits only a flat row and loses
orphaned subtrees; update extractTaskPlanSummary to treat orphan nodes as
additional roots and recursively render their children by calling the existing
renderTask for each orphan root (use the same indent and append " (orphan)" to
the root line), and modify renderTask to accept or use a visited Set to detect
and stop cycles (self-parented/cyclic tasks) to avoid infinite recursion;
reference renderTask, childrenMap, taskMap, rootTasks, and tasks when locating
where to add the orphan-root recursion and the visited-cycle guard.
---
Nitpick comments:
In `@packages/core/src/core/openaiContentGenerator/pipeline.ts`:
- Around line 334-345: Centralize the telemetry truncation policy by adding a
single helper (e.g., truncateForTelemetry or truncateTelemetryText) that takes a
string, enforces the 10_000-char limit and appends '...[truncated]' when needed,
then replace the inline logic in the non-streaming path that builds
reasoningText and calls context.span.setAttribute('gen_ai.response.thinking',
...) (uses reasoningParts) and the equivalent streaming-path code (the block
referenced around the other occurrence) to call this helper instead of
duplicating 10_000 and '...[truncated]'; ensure the helper is exported/visible
to both code paths and keep behavior identical.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: e3d1f9ba-54c8-4296-863d-e04c47a23387
📒 Files selected for processing (8)
docs/superpowers/plans/2026-04-30-p4-compaction-todo-preservation.mdpackages/cli/src/ui/components/messages/ConversationMessages.test.tsxpackages/cli/src/ui/components/messages/ConversationMessages.tsxpackages/core/src/agents/runtime/agent-core.tspackages/core/src/agents/runtime/compaction.test.tspackages/core/src/agents/runtime/compaction.tspackages/core/src/core/openaiContentGenerator/pipeline.thinking.test.tspackages/core/src/core/openaiContentGenerator/pipeline.ts
Comment on lines
+474
to
+481
| let compacted: Content[]; | ||
| if (estimateTokens(masked) <= targetTokens) { | ||
| compacted = masked; | ||
| } else { | ||
| const taskStore = this.runtimeContext.getTaskStore?.(); | ||
| const result = compactMessages(masked, targetTokens, { taskStore }); | ||
| compacted = result instanceof Promise ? await result : result; | ||
| } |
There was a problem hiding this comment.
Keep the task-plan summary on the masking-only path.
When observation masking already gets masked under targetTokens, this branch returns masked directly and never calls compactMessages(...), so the new <task-plan> enrichment is skipped on those compaction passes. That defeats the preservation behavior in a common case.
Possible fix
- let compacted: Content[];
- if (estimateTokens(masked) <= targetTokens) {
- compacted = masked;
- } else {
- const taskStore = this.runtimeContext.getTaskStore?.();
- const result = compactMessages(masked, targetTokens, { taskStore });
- compacted = result instanceof Promise ? await result : result;
- }
+ const taskStore = this.runtimeContext.getTaskStore?.();
+ const shouldCompact =
+ estimateTokens(masked) > targetTokens || !!taskStore;
+ const compacted = shouldCompact
+ ? await compactMessages(masked, targetTokens, { taskStore })
+ : masked;🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@packages/core/src/agents/runtime/agent-core.ts` around lines 474 - 481, The
current early-return when estimateTokens(masked) <= targetTokens skips the
`<task-plan>` enrichment that compactMessages(...) applies; instead of returning
masked directly, always invoke compactMessages(masked, targetTokens, {
taskStore: this.runtimeContext.getTaskStore?.() }) and assign compacted to its
result (await if it returns a Promise) so the task-plan summary enrichment runs
even when masking already meets the token target; use the existing variables
masked, targetTokens, taskStore and preserve the async handling used in the else
branch.
Comment on lines
+58
to
+85
| function renderTask(task: (typeof tasks)[0], indent: string) { | ||
| const marker = STATUS_MARKERS[task.status] ?? '[ ]'; | ||
| const priority = task.priority ? ` (${task.priority})` : ''; | ||
| lines.push(`${indent}${marker} ${task.title}${priority}`); | ||
|
|
||
| const children = childrenMap.get(task.id) ?? []; | ||
| for (const child of children) { | ||
| renderTask(child, indent + ' '); | ||
| } | ||
| } | ||
|
|
||
| // Render root tasks and their recursive children | ||
| for (const task of rootTasks) { | ||
| renderTask(task, ' '); | ||
| } | ||
|
|
||
| // Handle orphan subtasks (parent ID references a task not in the store) | ||
| const allIds = new Set(taskMap.keys()); | ||
| for (const task of tasks) { | ||
| if (task.parentTaskId && !allIds.has(task.parentTaskId)) { | ||
| const marker = STATUS_MARKERS[task.status] ?? '[ ]'; | ||
| const priority = task.priority ? ` (${task.priority})` : ''; | ||
| lines.push(` ${marker} ${task.title}${priority} (orphan)`); | ||
| } | ||
| } | ||
|
|
||
| lines.push('</task-plan>'); | ||
| return lines.join('\n'); |
There was a problem hiding this comment.
Don't drop orphaned task branches.
extractTaskPlanSummary() only recurses from root tasks, and the orphan pass emits a single flat row. That means self-parented/cyclic tasks or any orphan subtree with nested children disappear from the compacted summary, which undermines the state-preservation goal here.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed.
In `@packages/core/src/agents/runtime/compaction.ts` around lines 58 - 85, The
orphan pass currently emits only a flat row and loses orphaned subtrees; update
extractTaskPlanSummary to treat orphan nodes as additional roots and recursively
render their children by calling the existing renderTask for each orphan root
(use the same indent and append " (orphan)" to the root line), and modify
renderTask to accept or use a visited Set to detect and stop cycles
(self-parented/cyclic tasks) to avoid infinite recursion; reference renderTask,
childrenMap, taskMap, rootTasks, and tasks when locating where to add the
orphan-root recursion and the visited-cycle guard.
Code Coverage Summary
CLI Package - Full Text ReportCore Package - Full Text ReportFor detailed HTML reports, please see the 'coverage-reports-22.x-ubuntu-latest' artifact from the main CI run. |
This was referenced Apr 30, 2026
protoquinn Bot
pushed a commit
that referenced
this pull request
May 1, 2026
…r nuke) — bumps to v0.28.0 (#169) * fix(core): preserve tool history when building no-tools requests /recap (and any other caller without tools, e.g. /btw) was sending an empty conversation to the model. The no-tools branch in pipeline buildRequest dropped every assistant turn with tool_calls and every tool-role message wholesale, so in tool-heavy sessions the recap saw only bare user prompts and hallucinated context. - generateRecap now passes tools: [] so the strip path doesn't fire, matching cc-2.18's awaySummary pattern. - pipeline.ts no-tools branch now flattens instead of dropping: keeps assistant prose content and removes only the tool_calls field; tool results become [tool result] assistant notes truncated at 2000 chars. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(ci): release.yml now fires on auto-release/v* PRs (#160) auto-release.yml opens version-bump PRs from `auto-release/v*` branches into main, but release.yml's job gate only matched `head.ref == 'dev'`. Result: every auto-release PR was merging cleanly but skipping publish (v0.26.25 had to be dispatched manually). This adds the auto-release/* prefix to the gate and refreshes the stale top-of-file comment. Co-authored-by: Automaker <automaker@localhost> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: release v0.26.26 (#161) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * feat(telemetry,ui): reasoning span attribute + collapsed thought summary (Phase 1 of #162) (#165) * feat(core): preserve task plan state in compaction summaries (#163) * feat(core): preserve task plan state in compaction summaries When context compaction fires, the agent loses awareness of its task plan (completed, in-progress, pending work) and may re-plan already-done tasks. Add extractTaskPlanSummary() that queries the TaskStore and produces a structured <task-plan> XML section with status markers ([x], [~], [ ], [-], [!]), priority labels, and parent-child indentation. Extend compactMessages() to accept an optional taskStore and append the plan to the compaction summary. Wire the TaskStore into agent-core at the compaction call site. Backward compatible: existing callers without taskStore remain unaffected. Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * fix: add error handling and recursive nesting to compaction task plan Address PR feedback from CodeRabbit: - Wrap extractTaskPlanSummary call in try/catch so TaskStore failures don't break compaction - Replace flat 2-level subtask rendering with recursive renderTask() that supports arbitrary nesting depth - Add tests for multi-level nesting and error fallback Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> --------- Co-authored-by: Automaker <automaker@localhost> Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> * feat(telemetry,ui): capture reasoning on Langfuse span and collapse thoughts post-stream Phase 1 of the reasoning coordination tracked in #162. Captures delta.reasoning_content / delta.reasoning across stream chunks and surfaces it as gen_ai.response.thinking on the gen_ai chat span (gated on logPrompts, matching the completion event policy). Always emits gen_ai.usage.thinking_tokens when usage exposes it. Non-streaming responses get the same treatment by inspecting {thought:true} parts on the response — and the completion event no longer double-counts thoughts as content. Renders gemini_thought items as a compact "▸ thinking (N chars)" summary once the stream finalizes (live streaming render unchanged). Full text remains in ChatRecord, ACP agent_thought_chunk notifications, and Langfuse for downstream investigation. An in-TUI expand affordance is a follow-up. Once homelab-iac#31 (EMIT_REASONING_CONTENT) flips on, this also covers vLLM-served models that previously lost their <think> blocks at the gateway. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Automaker <automaker@localhost> Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: release v0.26.27 (#166) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> * chore(telemetry): rebrand to proto-cli, nuke qwen-logger Alibaba RUM ping (#167) * chore(telemetry): rebrand qwen-code identifiers to proto-cli Aligns telemetry / public-facing identifiers with the actual product name. Verified against the Langfuse instance: new spans land with service.name=proto-cli on scope=proto.openai-pipeline; existing proto.* tracers (proto.llm, proto.turn, proto.tools, proto.harness, etc.) were already correct. Changes: - SERVICE_NAME: qwen-code → proto-cli (resource attribute, the marquee label in Langfuse's service column) - All EVENT_* constants: qwen-code.* → proto.* (matches the existing proto.harness.* convention already in this file) - pipeline.ts tracer: qwen-code.openai-pipeline → proto.openai-pipeline (one straggler vs. the 9 other proto.* tracers in core/) - types.ts event.name literals (PromptSuggestion, Speculation): qwen-code.* → proto.* - acpAgent.ts agentInfo.name: qwen-code → proto-cli (visible to ACP clients like Zed when listing agents) - marketplace.ts User-Agent: qwen-code → proto-cli (extension fetch identifier sent to api.github.com / raw.githubusercontent.com) Out of scope (deliberately): - packages/core/src/telemetry/qwen-logger/* — separate analytics ping to gb4w8c3ygj-default-sea.rum.aliyuncs.com (Alibaba RUM, the upstream Qwen team's endpoint). Should be disabled rather than rebranded; tracking separately. - DEFAULT_SERVICE_NAME='qwen-code-oauth' in mcp/token-storage — renaming would orphan existing keychain entries. - Misc qwen-code-* file paths, tmp dir names, sandbox image tag, test fixtures — not telemetry / not user-visible labels. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore(telemetry): remove qwen-logger Alibaba RUM ping; keep useful events on Langfuse The qwen-logger system shipped usage telemetry to a fixed Alibaba RUM endpoint (gb4w8c3ygj-default-sea.rum.aliyuncs.com) — the upstream Qwen Code team's analytics pipeline. We don't operate that endpoint, the data isn't visible to us, and it labelled traffic as qwen-code-cli / qwen-code@${version}. Confirmed unused on our deployment; nuking. What's removed: - packages/core/src/telemetry/qwen-logger/ (entire directory: logger, event-types, tests) - packages/core/src/telemetry/integration.test.circular.ts (was a qwen-logger-specific circular-reference proxy-agent test, no longer applicable) - ~30 QwenLogger.getInstance(config)?.logXxxEvent(event) callsites in loggers.ts - QwenLogger exports from telemetry/index.ts and core/index.ts - QwenLogger spies and assertions in config.test.ts and the describe('logHookCall', ...) block in loggers.test.ts that was exclusively QwenLogger-shaped What's kept and rerouted to OTel/Langfuse: - HookCallEvent type and the logHookCall function — hook execution data is genuinely useful telemetry (which hook fired, success, duration, exit code, captured stdout/stderr, error). Now emits a proto.hook_call OTel log record via logs.getLogger(SERVICE_NAME) instead of the Alibaba ping. Existing call site in hookEventHandler.ts:619 still fires per hook execution. - LoopDetectionDisabledEvent likewise: was an empty no-op after the qwen-logger pull; rerouted to a proto.loop_detection_disabled OTel log record so the signal still reaches Langfuse. - New tests in loggers.test.ts assert OTel emission shape for logHookCall (success, error, sdk-not-initialized branches). Renamed (per "all not used" — no existing keychain entries to invalidate): - DEFAULT_SERVICE_NAME 'qwen-code-oauth' → 'proto-cli-oauth' - FORCE_ENCRYPTED_FILE_ENV_VAR 'QWEN_CODE_…' → 'PROTO_CLI_…' - file-token-storage encryption salt prefix and scrypt key seed switched to proto-cli; only invalidates non-existent tokens Verified live: kimi-k2.6 turn through the rebuilt CLI lands a Langfuse trace with service=proto-cli, scope=proto.openai-pipeline, gen_ai.response.thinking present. No outbound traffic to aliyuncs.com. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Automaker <automaker@localhost> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * chore: release v0.26.28 (#168) Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> --------- Co-authored-by: Automaker <automaker@localhost> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Phase 1 of the reasoning coordination tracked in #162. Surfaces the model's reasoning/thinking text on the
gen_ai chat ${model}Langfuse span and renders a compact summary post-stream so the chat UI stops dumping long thought blocks above each answer.gen_ai.response.thinking— accumulated fromdelta.reasoning_content/delta.reasoningacross stream chunks (and from{thought:true}parts on non-streaming responses), truncated to 10K chars to match the existinggen_ai.completionpolicy. Gated onlogPromptssince reasoning may carry user data references.gen_ai.usage.thinking_tokens— numeric, always emitted when usage exposesthoughtsTokenCount.gen_ai.content.completionevent no longer includes thought parts (they were silently mixed in before). Reasoning lives on its own attribute now.ThinkMessagerenders▸ thinking (N chars)intheme.text.secondaryonceisPending=false. Live streaming render is unchanged so users still see the model thinking. Continuation chunks (gemini_thought_content) render nothing post-stream — the parent ThinkMessage owns the summary.ACP delivery is untouched.
MessageEmitter.emitAgentThoughtcontinues to streamagent_thought_chunknotifications token-by-token, so Zed (and any other ACP client) keeps rendering thinking natively. 120 ACP tests pass.Phase 0 (gateway side) dependency
homelab-iac#31 adds the
EMIT_REASONING_CONTENTflag (default off) so the LiteLLM gateway re-emits stripped<think>text viadelta.reasoning_contentin-stream. Once that flag flips, vLLM-served models (QwQ etc.) that previously lost thinking entirely at the gateway start landing on this span and in this UI. Coordination thread: #162.Test plan
packages/core/.../pipeline.thinking.test.ts— 5 new tests: streaming reasoning_content, fallback toreasoning, logPrompts gating, 10K truncation, non-streaming{thought:true}extractionpackages/cli/.../ConversationMessages.test.tsx— 5 new tests: streaming render, post-stream summary, thousands separator, continuation visibility under both statespipeline.test.ts(24) and ACP suite (120) still passpackages/cli/src/uisuite (2108 tests) andpackages/core/src/core/openaiContentGeneratorsuite (302 tests) passnpm run typecheckcleannpm run lintcleanEMIT_REASONING_CONTENT=true— verify Langfuse trace showsgen_ai.response.thinkingfor a vLLM model, and TUI shows▸ thinking (N chars)after the answerKnown caveats (deferred to follow-up)
gemini_thought+gemini_thought_contentitems, the displayed char count counts only the first chunk. True total requires post-finalize coalescing inuseGeminiStream. Comment inConversationMessages.tsxflags this./thinkslash command or keyboard binding is a sensible next step.Closes none yet — keeps #162 open until Phase 2 (full Option B for gateway internal spans).
🤖 Generated with Claude Code
Summary by CodeRabbit
Release Notes