Release v0.21.14

Highlights

• Adds qwen sessions ps command and live-session registry to list and manage running interactive sessions with JSON output. (#8969, #9261, #9366)
• Improves Web Shell reliability by decoupling input from SSE streams, fixing image format crashes, and preventing duplicate messages. (#9323, #9295, #9407)
• Minimizes spam comment visibility gaps from over an hour to near zero by checking new comments against the blocklist immediately. (#9266)
• Adds /advisor slash command for independent read-only second opinions and enables workflow agents to pin specific working directories. (#7567, #8972)
• Enhances review skills to analyze scripts against actual execution lanes and supports GitLab repositories via Aone Code provider. (#9263, #9226)
• Fixes autofix convergence to write handoff decisions to failure.md and ensures bilingual content in failure handoff comments. (#9371, #9386)

Breaking Changes

No known breaking changes.

Complete Change List

Features

• Adds a live-session registry and the qwen sessions ps command to list running interactive sessions with optional JSON output. (#8969) by @qqqys
• feat(daemon): attach skill-toggle mutation metadata to settings_changed (#9051) by @samuelhsin
• Chains Terminal-Bench release evaluation by submitting SWE-bench runs first and dispatching Terminal-Bench tasks only after SWE results are published. (#9120) by @DennisYu07
• Improves Web Shell sidebar session management with consistent hover details, compact status indicators, and persistent workspace expansion behavior. (#9311) by @ytahdn
• Decouples the Web Shell composer from SSE catch-up to prevent input locking and allows prompt submission even when the stream is disconnected. (#9323) by @ytahdn
• Added end-to-end support for session-scoped media references, ensuring image previews persist across refreshes and reconcile consistently. (#9310) by @ytahdn
• Enabled workflow agents to pin to a specific working directory using agent({workingDir}), allowing them to operate outside default bounds. (#8972) by @qqqys
• Enhanced the review skill to analyze shell and CI scripts against the specific lanes and environments where they actually execute. (#9263) by @wenshao
• Added the /advisor slash command to request an independent, read-only second opinion on the current conversation without mutating history. (#7567) by @yiliang114
• Introduced privacy-safe diagnostics to correlate oversized or mutated tool results using size measurements and HMAC hashes without exposing content. (#9039) by @doudouOUC
• Adds a trusted-only GET /workspaces/:workspace/sessions/live-state endpoint returning volatile session snapshots and a catalog version token to reduce polling. (#9261) by @doudouOUC
• Consolidates Local Control into a single daemon-owned implementation with a secondary listener, unified security model, and revocable pairing credentials. (#9106) by @yiliang114
• feat(autofix): seed the takeover round counter with /takeover from N (#9321) by @wenshao
• feat(daemon): make serve new-file mode configurable (QWEN_SERVE_NEW_FILE_MODE) (#9364) by @yiliang114
• Web Shell now fully disables file drag-and-drop when fileUploadEnabled is false and supports configuring a default upload directory via fileUploadDirectory. (#9382) by @ytahdn
• Added pollable HTTP endpoints GET /session/:id/turns/current and GET /session/:id/turns/:promptId to check daemon turn status without SSE. (#9080) by @BenGuanRan
• Updates WebShell to use workspace-scoped session live-state for efficient sidebar updates, reducing full catalog refreshes. (#9366) by @doudouOUC
• Adds Aone Code as a second review provider, enabling meta, issue-context, and fetch-pr commands for GitLab-based repositories. (#9226) by @wenshao
• The exported HTML viewer now includes a global Expand all / Collapse all toolbar to toggle all collapsible sections like thinking blocks and tool outputs at once. (#9367) by @yiliang114
• When the severity floor resolves to Critical-only, the CLI automatically moves inline Suggestion comments to the review body deferral list instead of posting them to GitHub. (#9279) by @wenshao
• Review comments posted via --comment now use plain reviewer prose instead of template scaffolding, while retaining concrete evidence triggers and outcomes. (#9027) by @wenshao
• The sandboxed verification lane now includes a deterministic flakiness gate that re-runs modified unit tests multiple times to detect and report non-deterministic failures. (#9130) by @wenshao
• Added --resume flag to fetch-pr to resume interrupted reviews by validating on-disk state and reusing the existing worktree. (#9092) by @wenshao
• Autofix failure handoff comments now include bilingual content with a collapsed Chinese details block containing translated failure analysis. (#9386) by @wenshao

Bug Fixes

• Minimizes spam comments immediately upon creation by checking new comments against the blocklist, reducing visibility gaps from over an hour to near zero. (#9266) by @yiliang114
• fix(ci): drop pull_request_review events on closed PRs at the route gate (#9299) by @wenshao
• Repairs seven review pipeline defects found in live runs, including fixing incremental anchor logic and enabling multi-call build-and-test dimensions. (#9175) by @wenshao
• Fixed an issue where foreground agents were incorrectly marked as failed due to missing routing fields in initial SSE events. (#9330) by @ytahdn
• Fixed context usage details to exclude skills listed in skills.disabled, ensuring accurate context cost analysis. (#9346) by @callmeYe
• Updated agent-team prompts and TeamCreate description to accurately reflect automatic final answer delivery when teammates go idle. (#9284) by @yiliang114
• Makes transient resource-exhaustion and read I/O errors retryable while keeping symlink loops and malformed records as terminal failures. (#9362) by @doudouOUC
• Wraps the agent-tab view in a non-fatal ErrorBoundary to prevent render errors in one tab from exiting the entire session. (#9292) by @yiliang114
• Bounds string width and code point caches to 500 entries with oldest-entry eviction to prevent unbounded memory growth in long sessions. (#9185) by @yiliang114
• Clamps compression output budget to the remaining context window size instead of using a fixed 20K-token limit. (#9109) by @ZijianZhang989
• The Weixin channel now refreshes the typing indicator every 4 seconds during long turns to prevent it from expiring prematurely. (#9358) by @yiliang114
• The autofix convergence brake now explicitly instructs the agent to write its handoff decision to failure.md instead of a wrapper file. (#9371) by @wenshao
• The autofix fleet scan now fails closed on API enumeration errors to prevent dispatching busy PRs and marks dispatched PRs pending leg materialization. (#9329) by @wenshao
• Fixed a bug where enableCacheSharing defaulted to false instead of true, ensuring cache-aware suggestions work without explicit user configuration. (#9233) by @yiliang114
• fix(core): dispatch manually assigned team tasks to their owner (#9289) by @yiliang114
• Adds workspace wipe guards to triage and Serve A/B workflows to prevent unsafe rm -rf operations on non-canonical paths. (#9277) by @yiliang114
• Ensures incremental review anchors are only reused when the certifying model matches the current run to prevent incorrect skips. (#9184) by @wenshao
• fix(artifacts): verify and canonicalize record_artifact workspace paths (#9142) by @zjgzx1988
• Prevents session crashes by omitting unsupported image formats like HEIC and TIFF, replacing them with a text notice for the model. (#9295) by @yiliang114
• Improves memory recall reliability by injecting deterministic candidates if live recall does not settle within a 100ms ceiling. (#8716) by @yiliang114
• The compose-review command now enforces GitHub's 65,536-character review limit by trimming Chinese translations and deferred content first, with truncation as a last resort. (#9247) by @wenshao
• The web shell now uses backend-authoritative queue states to prevent duplicate messages and ensure draft payloads are only restored after proven delivery failures. (#9407) by @ytahdn
• Certification bars in the reverse-audit path now report specific failure names like 'receipt lead contradicts the phrase' to improve diagnostic clarity. (#9272) by @wenshao
• Unit tests in packages/cli now fail immediately with an actionable error message and the exact npm run build command if required build prerequisites are missing. (#9171) by @yiliang114
• Cleared the deferred-suggestion backlog from previous review rounds and fixed behavior issues in persistRecoveredLedger and rendering. (#9342) by @wenshao
• Enabled pagination for review thread fetching to ensure all threads are resolved instead of only the oldest 100. (#9390) by @qqqys

Internal Changes

• revert(web-shell): restore pre-#8098 composer animations at 50% opacity (#9349) by @ytahdn
• Updated review skill documentation to reflect the settled 3-round huge tier cap and pinned test coverage for chunk-based round gates. (#9258) by @yiliang114
• Removed thirteen unused internal helpers from the settings dialog utility module to clean up legacy code. (#9379) by @qqqys
• Simplified the review checkout self-heal logic by removing complex path guards and reverting to a core wipe-and-retry mechanism. (#9327) by @wenshao

Full Changelog: v0.21.13...v0.21.14