v0.1.32

Changelog

v0.1.32

✨ Features

• Tool System Completeness

  • Added full documentation for all registered tools: jobs, kill, question, memory, cron, and MCP dynamic tools
  • jobs tool: list and inspect background jobs started with bash async=true, with optional cleanup
  • kill tool: terminate a running background job by ID
  • question tool: AI can ask users multiple-choice questions during plan mode to clarify requirements
  • memory tool (Hermes): persistent memory via memory.md with read/add/update/delete actions across sessions
  • cron tool (Hermes/multi-agent): scheduled background tasks via sub-agents with @daily, @weekly, @every N schedules and one-shot support
  • MCP dynamic tools: tools/resources/prompts from MCP servers are auto-discovered and registered per session

• Plan Mode Question Tool

  • Added question tool, registered only in TUI + plan mode
  • AI can ask users multiple-choice questions; users select a preset option or type a custom answer
  • Helps clarify requirements before forming a plan, producing higher-quality proposals
  • Exposed via QuestionHandler optional interface (type assertion); does not pollute the public Agent interface

🐛 Bug Fixes

• Bash Tool Output Safety

  • Synchronous bash mode now enforces a 1 GB output limit using limitedBuffer, preventing OOM from unbounded bytes.Buffer growth

• Hermes /compact Command

  • Implemented the /compact slash command for Hermes messaging mode (previously a TODO stub)
  • Sets a ForceCompact flag on the session, consumed by the next agent run to trigger context compaction

• Session Durability

  • writeEntry now calls f.Sync() after writing, guaranteeing data survives crash or power loss
  • Corrupt session lines are now logged as warnings and skipped instead of blocking session load

• Hermes Approval Race Condition

  • ResolveApproval now uses select to avoid writing to an already-consumed channel when timeout and approval race

• Agent Sub-agent Panic Logging

  • sendParentEvent now logs the panic value before recovering, aiding diagnosis of closed-channel races

• Atomic File Write Cleanup

  • writeFileAtomic no longer uses defer os.Remove(tmpPath) which would attempt to delete an already-renamed file; cleanup is now explicit on each error path

• Agent Loop Detection Configurability

  • MaxConsecutiveNoText (stuck-detection threshold) is now configurable via AgentLoopConfig (default 95)
  • Fixed incorrect error message that added pre- and post-warning counters together

• Job Manager Auto-cleanup

  • AddJob now garbage-collects finished jobs older than 30 minutes (checked every 5 minutes)

• Cron Scheduler Error Logging

  • checkAndRun now logs store errors instead of silently swallowing them

• TUI Bash Output Display

  • Compressed bash tool output summary by removing blank lines to prevent excessive vertical height in the TUI collapsed view

• Vendored Search Tools

  • Added fallback to system grep / find when embedded rg / fd are unavailable for the current architecture

📦 Distribution

• Added Linux LoongArch64 (loong64) build and packaging targets, including tarball, Debian, and npm package metadata

✅ Tests

• Added unit tests for limitedBuffer truncation, JobManager GC, writeFileAtomic cleanup, sendParentEvent panic recovery, MaxConsecutiveNoText configurability, session fsync durability, corrupt-line tolerance, and QuestionTool metadata/mode-filtering/execution/error-handling

v0.1.31

Changelog

v0.1.31

🐛 Bug Fixes

• Terminal Input

  • Added Home/End cursor movement support in the TUI input box
  • Fixed the first submitted input being swallowed after canceling an approval prompt with Esc
  • Added command history navigation with Up/Down, including repeated selection through previous inputs

• A2A Security and Reliability

  • Changed the default A2A host from 0.0.0.0 to 127.0.0.1
  • Added Bearer token authentication for /a2a, REST A2A routes, and SSE events while keeping the Agent Card public
  • Replaced timestamp-based A2A task IDs with collision-resistant random IDs
  • Made A2A task store reads and writes use cloned task snapshots to avoid accidental shared mutation

• Path and Session Safety

  • Fixed path containment checks to use path-aware boundaries instead of string prefix checks
  • Prevented context extraFiles from escaping the working directory
  • Encoded unsafe Hermes session path components and enforced allowed_work_dirs during session creation
  • Restricted session deletion to .jsonl files under the configured session directory

• Auth, Approval, and Resource Limits

  • Switched Hermes HTTP/WebSocket token checks to constant-time comparison
  • Changed the Hermes WebSocket client to send auth via Authorization: Bearer ... instead of query strings
  • Cleaned up pending ACP permission requests on timeout and propagated ACP write errors
  • Added request/body size limits for ACP, read-tool image files, WeChat responses, and cron A2A responses
  • Added timeouts to cron A2A HTTP calls

• Memory, Context, and Concurrency

  • Added locking to memory store operations
  • Fixed memory.WriteAll() path handling and kept memory update/delete scoped to the requested section
  • Cloned gateway model settings before per-request temperature/top_p overrides
  • Passed agent callback context/message snapshots instead of shared references
  • Serialized cron job state transitions through the job store

• Configuration and Gateway Hardening

  • Gated !command API key resolution behind VIBECODING_ALLOW_SHELL_CONFIG=1
  • Fixed Gateway CORS to echo only the allowed request origin
  • Added a startup warning when Gateway listens beyond loopback in yolo mode without authentication
  • Hardened platform home/shell fallback behavior

🧪 Tests

• Added regression coverage for A2A auth, task ID uniqueness, task snapshot isolation, and persisted working task messages
• Added coverage for path traversal, unsafe session IDs, memory section operations, ACP cleanup, CORS behavior, UTF-8 truncation, and shell-config opt-in
• Ran focused package tests plus race tests for A2A, agent, gateway, and cron

📝 Docs

• Updated A2A, Hermes, Gateway, configuration, and security docs for the new authentication and hardening behavior

v0.1.29

Changelog

v0.1.29

🐛 Bug Fixes

• NPM Package Wrapper
  • Fixed npm/bin/vibecoding entry script to ensure installer packages ship the correct executable wrapper
  • Adjusted build-npm.sh and build-npm-packages.sh to include the wrapper consistently

v0.1.28

✨ Features

• Per-Model Temperature/Top-P Configuration

  • Added temperature and top_p fields to ModelConfig and Model for per-model parameter tuning
  • Wired through OpenAI and Anthropic providers with omitempty — nil means use API default
  • Wired through provider factory, agent loop, and ACP mode
  • Gateway supports per-request temperature/top_p override via ChatParams
  • When not configured, parameters are omitted entirely (no zero-value sent to API)

• OpenAI Responses API Support

  • Added a dedicated OpenAI Responses provider path under api: "openai-responses"
  • Supports Responses streaming, tool calls, reasoning summaries, and prompt cache parameters
  • Responses configuration is exposed under provider responses settings with default prompt cache enabled
  • Added model compat flags for supportsPromptCacheKey and supportsReasoningSummary

🧪 Tests

• Improved provider test coverage for OpenAI Responses API and Anthropic request parsing
• Reworked Anthropic tests to use in-memory HTTP mocks instead of port-binding test servers

📝 Docs

• Updated AGENTS.md version to v0.1.28

v0.1.27

✨ Features

• Hermes Mode (vibecoding hermes)

  • New messaging gateway mode for WeChat, Feishu, and WebSocket
  • Persistent per-user sessions with auto-archiving on /new
  • Default yolo mode for unattended operation
  • Smart approvals with tiered risk classification (low/medium/high)
  • User whitelist for platform access control
  • WebSocket streaming: real-time text_delta/think_delta/tool_call/tool_result/tool_diff/usage/done events

• A2A Protocol (vibecoding a2a)

  • New Agent-to-Agent protocol server (JSON-RPC 2.0 over HTTP + SSE streaming)
  • Standalone mode: vibecoding a2a start (port 8093)
  • Integration mode: hermes.json a2a.enabled: true shares hermes HTTP port
  • Agent Card at /.well-known/agent.json
  • Task lifecycle: submitted → working → completed/failed/canceled
  • REST endpoints: /a2a/send, /a2a/task, /a2a/task/cancel, /a2a/events
  • A2A Client: vibecoding a2a send <message> to send tasks to other A2A servers
  • A2A Discovery: vibecoding a2a discover <url> to fetch remote Agent Cards
  • A2A Scheduling: Cron jobs support --a2a-target to schedule tasks to A2A servers

• A2A Master Mode (--enable-a2a-master)

  • Configure multiple remote A2A agents via a2a-list.json
  • Registers a2a_dispatch tool for the LLM to automatically dispatch tasks to remote agents
  • Supports global (~/.vibecoding/a2a-list.json) and project-level (.vibe/a2a-list.json) config
  • --init-a2a-master-config generates a sample config file
  • Disabled by default, requires explicit opt-in

• A2A Config Initialization

  • vibecoding a2a --init-a2a-config generates a2a.json config template
  • vibecoding --init-gateway generates gateway.json config template (existing)
  • vibecoding --init-a2a-master-config generates a2a-list.json config template
  • All --init-* flags support --force to overwrite existing files

• Scenarios & Walkthroughs Documentation

  • New docs/scenarios.md (zh + en) covering 9 practical usage scenarios
  • Covers: daily coding, CI integration, multi-agent, VS Code ACP, A2A server,
    A2A Master cross-machine dispatch, Gateway HTTP, Hermes messaging, combined modes

• Documentation Overhaul

  • architecture.md: added all missing modules (a2a/acp/gateway/hermes/mcp/memory/messaging/vendored)
  • tools.md: added a2a_dispatch and skill_ref tool docs
  • cli-reference.md: added --enable-a2a-master, --init-a2a-master-config,
    --init-gateway, --force, a2a subcommand docs
  • README.md: updated architecture diagram, added running modes overview

• Pressure System

  • Context Pressure: EventContextPressure fired at 55% context usage (configurable via context_pressure_threshold)
  • Budget Pressure: EventBudgetPressure fired at 20% remaining iterations (configurable via budget_pressure_threshold)
  • One-shot events: fire once per threshold crossing, not every turn
  • Messaging platforms receive pressure warnings via progress callback

• Smart Approvals (Tiered Strategy)

  • Low risk: auto-approve
  • Medium risk: auto-approve + notify user
  • High risk (WebSocket): send approval_request, wait for user approval_response (5min timeout)
  • High risk (messaging): auto-reject + notify user
  • Command risk classification: low/medium/high based on bash command patterns

• Provider/Model Configuration

  • default_provider / default_model in hermes.json (overrides settings.json)
  • CLI flags -p/--provider and -m/--model for hermes start
  • Priority: CLI flags > hermes.json > settings.json

• Multi-Agent Mode (--multi-agent)

  • Enables sub-agent tools (spawn/status/send/destroy) in hermes sessions
  • Configurable via hermes.json multi_agent field or --multi-agent CLI flag

• Sandbox Mode (--sandbox)

  • Optional bwrap sandbox isolation (disabled by default)
  • Configurable via hermes.json sandbox field or --sandbox CLI flag

• MCP Integration

  • Hermes automatically loads MCP servers from global/project mcp.json
  • MCP tools registered per-session, connections auto-closed on session removal

• Progress Events for Messaging Platforms

  • Real-time tool execution progress sent to WeChat/Feishu during agent runs
  • Format: [tool]: args ✅/❌ for tools, 💭 ... for thinking process
  • Final summary sent after agent completes

• Memory Tool

  • memory tool with read/add/update/delete actions
  • Section-level operations (User Profile, Working Memory, Lessons Learned)
  • Defaults to .vibe/memory.md (project directory)
  • Lookup priority: memory.path config → .vibe/memory.md → <GLOBAL_DIR>/memory.md
  • /api/memory HTTP endpoint (GET/PUT) for memory access

• Hermes CLI Commands

  • hermes start — start daemon with all CLI flags
  • hermes stop — stop daemon via PID file + SIGTERM
  • hermes status — check daemon status via PID + HTTP health
  • hermes client — WebSocket client with streaming output and slash commands
  • hermes config init/show — configuration management
  • hermes wechat login/status — WeChat iLink management
  • hermes feishu setup/status — Feishu configuration
  • hermes webhook list — webhook route listing
  • hermes memory show/clear — memory management
  • hermes sessions list — active session listing (queries running instance)
  • hermes cron list/add/remove/enable/disable — cron job management
  • a2a start/stop/status/card — A2A server management

📝 Changes

• WeChat iLink implementation with zero external dependencies (5 files: types/protocol/auth/crypto/wechat)
• Feishu bot with official SDK and WebSocket long-connection
• Shell hooks for pre/post tool call external scripts (JSON stdin/stdout)
• Webhook inbound routing with HMAC-SHA256 signature verification
• WebSocket uses golang.org/x/net/websocket (stdlib compatible)
• PID file-based daemon management for hermes stop/status

🐛 Bug Fixes

• NPM Installer Packaging

  • Fixed release packaging flow so vibecoding-installer always ships executable entry bin/vibecoding.
  • Added scripts/npm-installer-wrapper.js as the single source of wrapper logic, reused by both
    scripts/build-npm.sh and scripts/build-npm-packages.sh to avoid drift.
  • Adjusted npm/.npmignore and npm/bin handling to avoid shipping accidental build artifacts and to keep
    package manifests (files) explicit.

• Hermes Webhook Delivery and Filtering

  • Webhook routes now treat unknown event types as non-matching unless the route explicitly allows *.
  • Added delivery_target to webhook routes so WeChat/Feishu delivery has a concrete recipient.
  • Updated webhook route listing and config templates to show the delivery target when present.

• OpenAI Responses Thinking Mapping

  • Mapped --thinking xhigh to reasoning.effort: "high" for the OpenAI Responses API.

🧪 Tests

• Reworked webhook router tests to wait on handler completion instead of sleeping, removing a race/flakiness source.
• Added coverage for webhook event rejection when the event type cannot be inferred.
• Added coverage for webhook delivery target handling.

v0.1.26

Changelog

v0.1.26

✨ Features

• Gateway Mode (vibecoding gateway)

  • New HTTP server exposing a standard OpenAI Chat Completions API (/v1/chat/completions, /v1/models, /health)
  • Any OpenAI-compatible client (Cursor, Continue, Open WebUI, Python SDK, etc.) can connect directly
  • Streaming (SSE) and non-streaming responses fully supported
  • Backend powered by VibeCoding agent loop with tool execution transparent to the caller

• Multi-Session Support

  • Built-in SessionPool for concurrent sessions, each with isolated agent, tools, and message history
  • Session association via x_session_id in request body; auto-created when absent
  • Configurable idle timeout (session.idleTimeoutSeconds) and max session limit (session.maxSessions)

• Sub-Agent Support in Gateway

  • Optional enableSubAgents config to enable multi-agent orchestration in gateway mode
  • Reuses existing AgentFactory / AgentManager / sub-agent tools with no core agent changes

• Bearer Token Authentication

  • Configurable via gateway.json with auth.enabled and auth.tokens list
  • Disabled by default; /health endpoint always unauthenticated

• Slash Commands via API

  • /clear, /mode, /model, /models, /sessions, /compact, /status, /skill, /skills, /help
  • Triggered when the last user message starts with /; processed at gateway layer without invoking LLM
  • Responses use standard OpenAI format with x_command extension field

• Tool Visibility Configuration (toolVisibility.mode)

  • "content" (default): tool status sent as text in content field during streaming
  • "sse_event": tool status sent as extended SSE events for custom clients
  • "none": fully transparent, client sees only final text

• System Prompt Handling (systemPromptMode)

  • "append" (default): client system messages appended to built-in system prompt
  • "ignore": client system messages discarded entirely

• Security: allowedWorkDirs

  • Directory whitelist for x_working_dir request-level overrides with path-separator-aware prefix matching
  • Three-layer security model: L1 auth + L2 directory control + L3 sandbox (bwrap)

• Sandbox Support in Gateway

  • Configurable via gateway.json sandbox.enabled / sandbox.level or --sandbox flag
  • Inherits detailed sandbox settings (allowedRead, deniedPaths, etc.) from settings.json

• Gateway Configuration (gateway.json)

  • Independent config file at ~/.config/vibecoding/gateway.json
  • Covers: listen address, auth, mode, sandbox, workingDir, allowedWorkDirs, session management, CORS, tool visibility, system prompt mode, request timeout, concurrency limit, logging
  • vibecoding --init-gateway to generate template; --force to overwrite

• Request Timeout & Concurrency

  • requestTimeoutSeconds (default 1800s); streaming keeps alive as long as data flows
  • maxConcurrentRequests (default 0 = unlimited)

📝 Docs

• Added docs/gateway-proposal.md with full architecture, API design, security model, and implementation plan
• Updated AGENTS.md version note

v0.1.25

v0.1.25

✨ Features

• Multi-Agent Mode

  • Added opt-in --multi-agent support across CLI, TUI, and ACP mode
  • Added AgentManager, EventRouter, and per-agent registries so agents have isolated tools, job managers, sessions, messages, and context
  • Added subagent_spawn, subagent_status, subagent_send, and subagent_destroy tools for delegated background work
  • Added multi-agent prompt guidance and safeguards that prevent nested sub-agent spawning

• Cron Task Support

  • Added internal/cron with persistent cron store and scheduler coverage
  • Added /cron command entry points in multi-agent TUI workflows

• Provider Vendor Adapter Layer

  • Added vendor adapter registration in internal/provider/vendor*.go
  • Centralized provider/model creation in internal/provider/factory
  • Added vendor detection for DeepSeek, Xiaomi, Kimi, MiniMax, Seed, Qianfan, Bailian, Gitee, OpenRouter, Together, Groq, Fireworks, OpenAI, and Anthropic
  • Preserved existing provider config format while allowing vendor-specific defaults and generic OpenAI/Anthropic-compatible fallback
  • Added model compat handling for thinking formats, reasoning effort support, max token field selection, adaptive Anthropic thinking, and DeepSeek/Xiaomi assistant reasoning_content

🐛 Bug Fixes

• Auto-initialized sessions on first append so sub-agents can write session entries without requiring explicit prior initialization
• Fixed sub-agent tests to wait for background runs and clean up spawned agents before temporary directory removal
• Preserved ACP Anthropic cache-control behavior while moving provider creation to the shared factory

📝 Docs

• Updated AGENTS.md with provider factory and vendor adapter guidance
• Replaced the multi-agent implementation checklist with a completed architecture/status document
• Removed the obsolete root todo.md

🧪 Testing

• Added coverage for provider vendor resolution, provider factory creation, OpenAI/Anthropic compat behavior, multi-agent manager/router/sub-agent flows, cron storage/scheduler behavior, and session auto-initialization
• Verified with make test (go test -v -race ./...)

v0.1.24

v0.1.24

✨ Features

• API Retry with Exponential Backoff
  • Automatic retry for transient errors (5xx, network failures, rate limits) on initial HTTP connection
  • Exponential backoff: baseDelay × 2^attempt, capped at 30 seconds
  • Does NOT retry on user abort (context.Canceled), 4xx client errors, or mid-stream failures
  • Configurable via retry settings (maxRetries, baseDelay, maxDelay)
  • Agent forwards retry events as status updates visible in TUI and print mode
  • ACP mode also receives retry configuration

🐛 Bug Fixes

• Anthropic cache_control Now Opt-In

  • Changed default cache_control behavior to off (was auto-enabled for official API base URL)
  • Require explicit cacheControl: true in provider config to enable prompt caching
  • ACP provider creation explicitly enables cache_control for Anthropic

• Anthropic Tool Result Grouping

  • Fixed consecutive toolResult messages to be grouped into a single user message
  • Anthropic API requires all tool_result blocks for preceding tool_use to appear together before other content
  • Image blocks from tool results are now appended after all result blocks in the same message

• Agent Tool-Only Loop Warning Ordering

  • Moved the no-text tool-loop warning to be injected after tool results are appended
  • Keeps assistant -> toolResult -> warning message ordering valid for provider and session transcripts
  • Warning messages are now also persisted to session storage

📝 Docs

• Comprehensive Configuration Documentation Rewrite
  • Added missing settings: cacheControl, idle compression, full sandbox fields (bwrapPath, allowedRead, allowedWrite, deniedPaths, passEnv, tmpSize), shellPath, shellCommandPrefix, sessionDir, skillsDir, theme, retry
  • Documented shell command apiKey format (!cmd) for password manager integration
  • Fixed key resolution order: config apiKey first, then derived env var
  • Fixed macOS config path: ~/Library/Application Support/vibecoding/
  • Added top-level fields reference table with all defaults
  • Added per-platform defaults for sandbox paths and env vars
  • Improved examples with Claude provider cacheControl, idle compression, project-level overrides, and custom sandbox paths

🧪 Testing

• Added retry tests covering IsRetryable, RetryDelay, and FormatRetryMessage
• Added Anthropic provider tests for consecutive tool result grouping
• Added a regression test covering tool-only warning placement after tool results

v0.1.22

Changelog

v0.1.22

✨ Features

• CLI/TUI MCP Auto-Loading
  • CLI/TUI startup now loads global and project mcp.json, connects configured MCP servers, and registers MCP tools before the agent tool list is frozen

🐛 Bug Fixes

• Markdown Rendering Style
  • Switched CLI print mode and TUI markdown rendering from Glamour auto-style detection to the fixed dark style for more consistent terminal output

🧪 Testing

• Added MCP config loader coverage for placeholder template filtering

🛠 Improvements

• Shared MCP Runtime
  • Moved MCP connection/tool registration out of ACP-only code into a shared runtime used by ACP and normal CLI/TUI sessions
  • Starter-template placeholder MCP servers are ignored during automatic startup loading

v0.1.21

Changelog

v0.1.21

✨ Features

• Plan/Apply Workflow

  • Added a built-in plan tool for structured task plans with pending, running, done, and failed step statuses
  • TUI now shows the current task plan and records plan updates in the transcript
  • Print mode and ACP now surface plan updates for non-interactive and editor-client flows

• Apply Confirmation

  • Added approval.confirmBeforeWrite to require approval before write and edit in agent mode
  • Enabled write/edit confirmation by default in generated settings
  • TUI approval prompts summarize write content by byte size instead of dumping full file content

• MCP Config Commands

  • Added /init_mcp to create project/global mcp.json with basic/full templates and optional --force
  • Added /mcps to list MCP servers from global and project mcp.json files
  • MCP config is now maintained in standalone mcp.json (separate from settings.json)

🧪 Testing

• Added coverage for the plan tool and write/edit approval gating
• Added HTTP-based MCP integration tests for tool/resource/prompt registration and callback paths
• Added SSE-based MCP integration tests for stream callbacks and message endpoint request/response flow

🛠 Improvements

• ACP MCP Hardening
  • Added MCP transport support for http and sse (alongside existing stdio)
  • Added MCP initialize/tool-discovery timeouts to avoid hanging ACP sessions
  • Added paginated tools/list fetching with upper page bounds
  • Added MCP resources/* and prompts/* discovery and tool registration
  • Added duplicate MCP server-name detection and MCP tool-name de-duplication
  • Added MCP inbound request/notification handling (ping, progress/logging/cancel notifications)
  • Added bridge for inbound sampling/createMessage to the active ACP provider/model
  • Added stricter close/error propagation

v0.1.20

v0.1.20

✨ Features

• Structured File Change Reporting
  • write and edit now attach structured file diff metadata to tool results
  • TUI tool details show full unified diffs while collapsed tool rows keep a compact +N -N summary
  • Print mode now emits clear file change summaries for non-interactive runs
  • ACP tool updates include diff metadata in raw output for compatible clients

🧪 Testing

• Added coverage for structured diff metadata from write and edit

---

v0.1.19

✨ Features

• TUI Tool Details Modal

  • Replaced Ctrl+O toggle-expand with a scrollable full-screen modal overlay showing all tool calls and results
  • Supports PgUp/PgDn, Up/Down, Home/End navigation; Esc/Ctrl+O/q to close
  • Tool headers now display file paths; removed content truncation in tool args display
  • Write tool results show diff summary in the one-line summary line
  • Key input is blocked while the modal is open to prevent accidental actions

• Write Tool Diff Summary

  • write tool now computes LCS-based line-level diff when overwriting files
  • Returns structured diff info (+N -N with line ranges) in the tool result
  • Skips diff computation for very large files (>200K line pairs) to avoid memory pressure

🛠 Improvements

• Unified Shell Args Across Sandbox Backends
  • All sandbox backends (none, mac, windows) now use platform.ShellArgs() for cmd.exe/PowerShell argument construction
  • Fixes Windows cmd.exe and PowerShell commands in sandboxed execution modes
  • ShellArgs now normalizes shell name to lowercase before matching

🧪 Testing

• Added TestNoneSandboxWrapCommandUsesPlatformShellArgs covering cmd.exe and PowerShell argument generation

v0.1.18

Changelog

v0.1.18

🐛 Bug Fixes

• TUI Nil Pointer Panic

  • Fixed a nil pointer panic in printMessageOnce when printedMessageIdx map was not initialized
  • Added nil check before accessing the map in the message printing logic

• Stream Commit Before Tool Execution

  • Added commitActiveStream() method to flush streaming content (thinking and assistant messages) to output before tool execution
  • Now properly commits active stream before EventToolCall and EventToolApprovalRequest handling
  • Ensures thinking and partial assistant responses are visible when tools run or approval is requested

🧪 Testing

• Added TestHandleAgentEventCommitsStreamBeforeApproval regression test for stream commit ordering