Ports high-value agent engineering patterns from Claude Code's architecture
into the Qwen Code fork. Full details in the squashed commit.

Phase 5 added 'verify' as a third builtin agent. Update test assertions
that hardcoded the expected count (5→6) and name list to include 'verify'.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Phase 5 added 'verify' as a third builtin agent. Update compiled JS test
assertions that hardcoded expected count (5→6) and name list. Also suppress
pre-existing no-undef/console lint error in the compiled file.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- logger.test.ts/js: append process.pid to temp dir to prevent EACCES
  under parallel vitest workers sharing the same fixture path
- packages/vendor/tree-sitter/: add tree-sitter.wasm + tree-sitter-bash.wasm
  for compiled .js test paths (levelsUp=3 targets packages/vendor/, not core/vendor/)
- eslint.config.js: add Node globals block for packages/*/src/**/*.test.js
  with no-undef:off, no-console:off, and no-unused-vars caughtErrorsIgnorePattern

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…gnore

Root cause: tsc emits .js/.d.ts files alongside .ts sources in src/, and
vitest's default include picks up both. Every test ran twice, causing
temp-dir races between the .ts and .js versions of the same test.

- packages/core/vitest.config.ts: exclude *.test.js and *.spec.js so only
  TypeScript sources run (canonical)
- packages/core/.gitignore: ignore src/**/*.js, *.d.ts, *.js.map to
  prevent future compiled output from accumulating untracked
- Remove previously committed logger.test.js and subagent-manager.test.js
  (compiled duplicates; fixes belong in the .ts sources)

Result: 190/190 test files pass, 4968 tests, 0 failures.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…r test

- packages/cli/vitest.config.ts: restrict include to .ts/.tsx only;
  add *.test.js and *.spec.js to exclude (compiled outputs not tests)
- HistoryReplayer.test.ts: update TodoWriteTool test — isTodoWriteTool now
  always returns false (Phase 2 replacement), so todo_write results emit
  tool_call_update instead of plan; assert new behavior with objectContaining

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ectory linting

ESLint flat config resolves `files` patterns relative to the config root.
When the package runs `eslint src` from its own directory, paths arrive as
`src/...` rather than `packages/vscode-ide-companion/src/...`, so the root
config's scoped overrides (import/no-internal-modules, no-unused-vars,
react/prop-types) do not match.

- Add packages/vscode-ide-companion/eslint.config.js: extends root config,
  adds src/**/*.{ts,tsx,js} overrides with the correct rules for local invocation
- Remove now-redundant eslint-disable-next-line comments in webview/index.tsx
  (import/no-internal-modules is now off for the whole package)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…giene

- telemetry/sdk.ts: add buildLangfuseSpanProcessor() — activates from env
  vars (LANGFUSE_PUBLIC_KEY/SECRET_KEY/BASE_URL), sends traces to
  ${LANGFUSE_BASE_URL}/api/public/otel/v1/traces with Basic auth. Runs
  alongside primary OTLP exporter even when getTelemetryEnabled() is false.

- telemetry/sdk.test.ts: isolate Langfuse env vars in beforeEach/afterEach
  so host-env keys don't contaminate assertions about OTLPTraceExporterHttp
  call counts.

- openaiContentGenerator/pipeline.ts: OTel span per LLM call with GenAI
  semantic conventions (gen_ai.system, gen_ai.operation.name,
  gen_ai.request.model). Raw chunk delta logging via createDebugLogger
  ('OPENAI_PIPELINE') — reveals whether LiteLLM routes thinking tokens to
  reasoning_content or delta.content.

- .gitignore: suppress 741 co-located TypeScript compilation artifacts
  (docs/, integration-tests/, scripts/, packages/*/src/, vitest.config.*,
  __snapshots__/*.js.snap). Root tsconfig has no outDir so tsc emits
  alongside .ts sources. gitignore is the correct fix — noEmit on the base
  config breaks TS6310 composite project references.

- settings.schema.json: memoryConsolidation config block for Phase 3
  (Dream Memory Consolidation).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Enables qwen-code to read Claude Code project files without requiring
manual conversion or duplication:

- paths.ts, storage.ts: export CLAUDE_DIR = '.claude'; add '.claude' to
  SKILL_PROVIDER_CONFIG_DIRS so ~/.claude/skills/ and project/.claude/skills/
  are searched alongside .qwen and .agents variants.

- storage.ts: add getUserCommandsDirs() and getProjectCommandsDirs() returning
  both .qwen/commands and .claude/commands paths.

- FileCommandLoader.ts: use plural methods so ~/.claude/commands/ and
  project/.claude/commands/ are loaded every session.

- memoryTool.ts: add 'CLAUDE.md' to the context filename list. The existing
  upward-scan in memoryDiscovery already handles project-root CLAUDE.md files;
  this just enables the filename to be matched.

- memoryDiscovery.ts: route CLAUDE.md global lookup to ~/.claude/ instead of
  ~/.qwen/ (matching Claude Code's own location). All other filenames continue
  to use ~/.qwen/.

- skill-manager.test.ts: update base-dirs count assertions from 2 → 3.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Integrates obra/superpowers (https://github.com/obra/superpowers) as a
native, zero-config behavioral layer. No hook configuration required.

## What's included

14 skills bundled at packages/core/src/skills/bundled/superpowers/:
  brainstorming, test-driven-development, writing-plans, executing-plans,
  subagent-driven-development, systematic-debugging, dispatching-parallel-agents,
  using-git-worktrees, verification-before-completion, finishing-a-development-branch,
  requesting-code-review, receiving-code-review, writing-skills, using-superpowers

subagent-driven-development prompt templates adapted for quad-code:
  Task tool → Agent tool in implementer-prompt.md, spec-reviewer-prompt.md,
  code-quality-reviewer-prompt.md

## Bootstrap mechanism

prompts.ts dynamicSuffix now reads the bundled using-superpowers/SKILL.md
synchronously at session start and injects it wrapped in <EXTREMELY_IMPORTANT>
tags — matching the original superpowers SessionStart hook design. The model
receives the full behavioral scaffolding before the first user message with no
tool round-trip required.

The skill path resolves relative to prompts.js via import.meta.url, so it
works in both dev (dist/src/core/ → dist/src/skills/bundled/) and the CLI
bundle (copy_files.js carries .md files through the build).

## Effect

Every session opens with the superpowers pipeline active:
  brainstorm → plan → subagent-execute → verify → finish

Skills are also available by name via the Skill tool for on-demand invocation.
Project or user skills in .qwen/skills/ or .claude/skills/ take precedence.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…s them

SkillManager.loadSkillsFromDir scans one level deep, so bundled/superpowers/*/SKILL.md
was invisible. Move all superpowers skill dirs up to bundled/ directly and
update the prompts.ts using-superpowers path accordingly.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Reads installed_plugins.json and adds each plugin's commands/ dir to the
user command search path, so quad inherits all installed Claude Code
plugins (protolabs, rabbit-hole, etc.) with zero configuration.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…support

- Rename binary from quad/protocli to `proto`
- Replace ANSI Shadow ASCII art with larry3d font for proto
- Rename all user-facing "Qwen Code" strings to "proto" (91 files)
- Replace in-memory TaskStore with beads_rust (`br`) CLI backend
  - SQLite + JSONL persistence via `br` subprocess calls
  - Falls back to original in-memory store if `br` not installed
  - Status mapping: open->pending, in_progress, closed->completed, deferred->blocked
- Add protolabs MCP server config to ~/.qwen/settings.json
- Rewrite README with proto branding, feature comparison table, and usage docs
- Update remote to github.com/protoLabsAI/protoCLI

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…reaming

Tools that completed while the model was still streaming its response were
silently dropped by the isResponding guard in handleCompletedTools. This left
stale toolCalls with responseSubmittedToGemini=false, keeping streamingState
stuck on Responding indefinitely (the "I'm Feeling Lucky" spinner).

Fix: queue completed tools in a ref and drain the queue via useEffect when
isResponding transitions to false.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…eading newline

- Change QWEN_DIR from '.qwen' to '.proto' for all global and per-project paths
- SKILL_PROVIDER_CONFIG_DIRS reads from .proto, .qwen, .agents, .claude (backwards compat)
- Rename context file from QWEN.md to PROTO.md (reads QWEN.md + CLAUDE.md as fallbacks)
- Rewrite AGENTS.md with proto branding and updated project overview
- Trim leading newlines from first model content chunk (fixes blank line after diamond)
- Update all user-facing path references (~/.proto/settings.json, .proto/commands/, etc.)
- Add npm run ship script (build + link)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…points

Drop the unused sdk-java package (~22K lines) and rewrite the README
Quick Start section to walk through presetting an OpenAI-compatible
endpoint with modelProviders, API key via .env, and auth bypass.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Replace the append-only PROTO.md bullet list with Claude Code-style
file-per-memory architecture:

- Each memory is a separate .md file with name/description/type frontmatter
- 4 memory types: user, feedback, project, reference
- MEMORY.md auto-generated index (max 200 lines / 25KB), loaded into system prompt
- save_memory tool now creates individual files via memoryStore.createMemory()
- New type param on save_memory for categorizing memories
- Memory system prompt with taxonomy injected when MEMORY.md exists
- Old PROTO.md memories still load (backward compat), migration available
- Storage helpers: getGlobalMemoryDir(), getProjectMemoryDir()

New files: packages/core/src/memory/{types,frontmatterParser,memoryStore,
memoryPrompt,migration,index}.ts

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Phase 2 of the memory system overhaul:

- memoryExtractor.ts: spawns a restricted AgentHeadless after each
  conversation turn to extract memorable facts from recent messages.
  Fire-and-forget with cursor tracking (.cursor file) to avoid
  reprocessing. Max 5 turns, 2 minute timeout.
- memoryScan.ts: scan memory headers and format as text manifest
  for injection into extraction agent prompts
- memoryAge.ts: human-readable freshness (today/yesterday/N days ago)
  and staleness warnings for memories older than 30 days
- Extraction triggered from useGeminiStream.ts finally block (non-blocking)
- Restricted tool set: read_file, write_file, glob only

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- /memory list: shows all memory files across both scopes with type,
  age, staleness warnings, and description
- /memory forget <name>: deletes a memory by name from either scope
- Remove migration.ts (greenfield, no legacy memories to migrate)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Add Memory section to README with type table, file format example,
  slash commands, and auto-extraction description
- Update feature comparison table with memory row
- Fix clone URL to protoLabsAI/protoCLI
- Add memory commands to commands table
- Update AGENTS.md with memory system overview and commands

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…ages

Replace raw isSubmittingQueryRef boolean with generation-based QueryGuard
to prevent stale async finally blocks from releasing the lock prematurely.
Add popLast() to message queue for one-at-a-time cancel behavior instead
of dumping the entire queue into the input buffer. Improve queue display
with a labeled header showing message count.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Instead of holding queued messages until the turn finishes, drain them
in handleCompletedTools and inject alongside tool results so the model
sees user input immediately and can decide whether to act on it or
continue current work. Fallback drain on idle catches pure text turns.

Adds drain() to useMessageQueue with ref mirror for synchronous reads.
Uses ref bridge in AppContainer to cross the hook ordering boundary.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

useMessageQueue is now a pure external store (MessageQueueStore class)
subscribed via useSyncExternalStore. drain() and popLast() return values
synchronously — no refs, no stale closures, no hook ordering hacks.

Queue declared before useGeminiStream so drain is passed directly.
Idle-drain effect moved to AppContainer where both submitQuery and the
queue are in scope.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Run git status, branch, and recent commits at session init before the
agent touches anything. Surfaces dirty working trees with a warning so
the agent (and user) know the starting state. Supports an optional
configurable verification command.

Inspired by the "verify before building" pattern from Anthropic and
OpenAI harness engineering guides.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

After file-modifying tools (write_file, edit), run a configurable verify
command and inject failures into the model context alongside tool results.
The model sees build errors immediately and can self-correct without a
new turn.

Configure via settings: tools.verifyCommand (e.g. "npx tsc --noEmit").
Disabled by default — no overhead when unconfigured.

Implements the "separate evaluator" pattern from Anthropic/OpenAI harness
engineering guides: generation and evaluation are independent, and the
evaluator is skeptical by default.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…, verification

Four improvements informed by Anthropic/OpenAI harness engineering guides
and Claude Code prompt architecture audit:

1. Compression prompt: 9-section structured summary replacing 4-section
   XML snapshot. Preserves user messages verbatim, debugging history,
   reasoning chains, and precise work state. Adds "resume immediately"
   rule to prevent wasted turns acknowledging summaries.

2. Memory extraction: strict 2-turn budget (was 5). Turn 1: all reads
   in parallel. Turn 2: all writes in parallel. Prevents sprawl.

3. System prompt: failure recovery 6-step protocol, permission scope
   non-generalization ("approval for X is not approval for Y"),
   reversibility assessment before destructive actions.

4. Adversarial verification skill: rationalization trap list, spot-check
   warning, adversarial probes (concurrency, boundary, idempotency),
   structured PASS/FAIL/PARTIAL verdict format.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- Restore lsp field to CliArgs interface (removed by hooks cleanup cherry-pick)
- Add lsp: undefined to auth handler and gemini test fixtures
- Resolve client.ts conflict: keep both endTurnSpan and cache-safe param saving
- Fix speculationToolGate: ToolNames.TODO_WRITE → ToolNames.TASK_LIST
- mcp-tool: replace getStringifiedResultForDisplay with getDisplayFromParts,
  remove unused allowlist field and truncateTextParts method

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- constants.test: replace hardcoded count with Object.values(HookEventName).length
  so the test stays correct when the enum grows
- ideCommand.test: remove unused getIdeInstaller mock from vi.mock block

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- coreToolScheduler.test + nonInteractiveToolExecutor.test: add
  getDisableAllHooks mock to all mock configs (hooks cleanup change)
- mcp-tool.ts: restore output truncation via truncateParts helper using
  existing truncateToolOutput utility
- mcp-tool.ts: gate auto-reconnect on isConnectionError check so generic
  errors (Invalid parameters) don't trigger reconnect
- extension/npm.ts: fix CodeQL alerts — use replaceAll for '/' encoding
  in scoped package names

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Add file.on('error') handler to write stream in downloadNpmTarball so
  stream errors during res.pipe() properly reject the Promise instead of
  hanging forever
- Fix name.replace('/', '%2f') → replaceAll in downloadFromNpmRegistry
  (line 285 was missed in previous commit; line 374 already correct)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…up, follow-up suggestions (#19)

* fix: improve /compress split point selection for tool-heavy conversations

When conversation history is near the context window limit and dominated by
tool call/response cycles, findCompressSplitPoint would return a near-zero
split point because it only considered non-functionResponse user messages as
valid split points. This caused /compress to send almost no history to the
compression API (e.g. 29 tokens), producing a useless summary that inflated
token count instead of reducing it.

Changes:
- Track tool completion boundaries (positions after functionResponse) as
  fallback split points in findCompressSplitPoint
- Add user-with-functionResponse to the compress-everything safety check
- Use Math.max of primary and fallback split points for better coverage
- Add minimum content guard (5% threshold) to prevent futile API calls
- Add 4 new test cases covering tool-heavy conversation scenarios

Fixes QwenLM#2647

* fix: handle orphaned funcCall and improve compression logic for tool-heavy conversations

- Strip trailing orphaned funcCall (force=true) before split point calculation,
  so normal compression logic runs cleanly on the remaining history instead of
  requiring ad-hoc special-casing
- Remove redundant lastToolCompletionSplitPoint machinery: after fixing the
  i+2 index bug, lastSplitPoint already subsumes it, making Math.max redundant
- Add MIN_COMPRESSION_FRACTION constant (0.05) to guard against futile API
  calls when historyToCompress is too small relative to total history
- Add tests for orphaned funcCall handling (force=true compresses, force=false NOOP)
- Add test for MIN_COMPRESSION_FRACTION guard

Fixes QwenLM#2647

* fix: upgrade @lydell/node-pty to 1.2.0-beta.10 to fix PTY FD leak

The previous version (1.1.0) has a native-level bug on macOS where each
PTY spawn leaks one /dev/ptmx file descriptor that is never closed. Over
a long session with hundreds of shell commands, this exhausts the
system-wide PTY pool (kern.tty.ptmx_max = 511), breaking other programs
like tmux and new terminal windows.

Root cause: microsoft/node-pty#882

Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>

* fix(cli): add kitty sequence timeout management and buffer handling improvements in KeypressProvider

* remove hooks experimental and refactor hook Config

* fix(hooks): clean up abort listener in error handler

The error handler in hookRunner cleared the timeout but did not remove
the abort signal listener, unlike the close handler. When spawn fails
(e.g. executable not found), only the error event fires — the close
event is not guaranteed — so the abort listener leaked on the signal.

* feat(extension): add npm registry support for extension installation

- Add new npm extension installation channel via scoped packages (@scope/name)
- Implement npm.ts module with registry resolution, authentication, and download logic
- Support version pinning, dist-tags (latest, beta), and custom registries
- Handle private registry auth via NPM_TOKEN env var and .npmrc _authToken entries
- Update CLI install command with --registry flag for npm extensions
- Add comprehensive tests for npm package parsing and registry operations
- Update documentation for releasing and installing from npm registries
- Integrate npm updates into extension manager and update checking flow

This enables teams using npm for package distribution to publish Qwen Code extensions through their existing infrastructure, with full support for private registries and access control.

Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>

* feat(mcp): add reconnect command and implement auto-reconnect logic

* refactor(mcp): enhance reconnect logic and error handling

* test(mcp): update tests to include verification of reconnect command registration

* feat(cli): add follow-up suggestions feature

* fix: post-backport build fixes

- Restore lsp field to CliArgs interface (removed by hooks cleanup cherry-pick)
- Add lsp: undefined to auth handler and gemini test fixtures
- Resolve client.ts conflict: keep both endTurnSpan and cache-safe param saving
- Fix speculationToolGate: ToolNames.TODO_WRITE → ToolNames.TASK_LIST
- mcp-tool: replace getStringifiedResultForDisplay with getDisplayFromParts,
  remove unused allowlist field and truncateTextParts method

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(tests): update tests for upstream backport changes

- ideCommand.test: remove install subcommand tests (no longer in ideCommand)
- constants.test: update DISPLAY_HOOK_EVENTS length from 12→15 (3 new hook events)
- Footer.test: mock VoiceMicButton to avoid UIActionsProvider dep, add voice
  state fields to mock UIState, update snapshots
- client.test: add getDisableAllHooks mock for hooks cleanup change

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(tests): harden test assertions per code review

- constants.test: replace hardcoded count with Object.values(HookEventName).length
  so the test stays correct when the enum grows
- ideCommand.test: remove unused getIdeInstaller mock from vi.mock block

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(tests,core): resolve CI failures on upstream backport PR

- coreToolScheduler.test + nonInteractiveToolExecutor.test: add
  getDisableAllHooks mock to all mock configs (hooks cleanup change)
- mcp-tool.ts: restore output truncation via truncateParts helper using
  existing truncateToolOutput utility
- mcp-tool.ts: gate auto-reconnect on isConnectionError check so generic
  errors (Invalid parameters) don't trigger reconnect
- extension/npm.ts: fix CodeQL alerts — use replaceAll for '/' encoding
  in scoped package names

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(extension): harden npm download stream and fix slash encoding

- Add file.on('error') handler to write stream in downloadNpmTarball so
  stream errors during res.pipe() properly reject the Promise instead of
  hanging forever
- Fix name.replace('/', '%2f') → replaceAll in downloadFromNpmRegistry
  (line 285 was missed in previous commit; line 374 already correct)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

---------

Co-authored-by: LaZzyMan <zeusdream7@gmail.com>
Co-authored-by: tanzhenxin <tanzhenxing1987@gmail.com>
Co-authored-by: Qwen-Coder <qwen-coder@alibabacloud.com>
Co-authored-by: qqqys <qys177@gmail.com>
Co-authored-by: DennisYu07 <617072224@qq.com>
Co-authored-by: chinesepowered <nlai@rediffmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

IDE integration is no longer supported. Voice integration section covers:
- audio backend requirements (sox/arecord) per OS
- protoLabs hosted STT endpoint option
- self-hosted faster-whisper-server option
- settings.json config and /voice commands

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

No protoLabs-hosted voice service — self-hosted only.
Also removes the unused HOSTED_STT_ENDPOINT export.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Instead of loading all context files found (PROTO.md + AGENTS.md + QWEN.md +
CLAUDE.md) at every directory in the walk, now stops at the first filename
found per location. Global (~/.proto/, ~/.claude/) and each workspace directory
independently cascade through the priority list: PROTO.md > AGENTS.md > QWEN.md
> CLAUDE.md. Prevents context flooding when multiple compat aliases co-exist.

Also updates repository/homepage metadata to protoLabsAI/protoCLI and
protolabs.sh across root, cli, and sdk-typescript package.json files.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…inism

Recurring jobs get up to 15 min jitter. Ticking at 10:31 could land within
the jitter window of a 10:30 job depending on the random job ID. Pin jitterMs
to 0 so the test is deterministic — consistent with how other jitter tests work.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>