On July 9, 2026, Qwen Code shipped its weekly update: three stable releases from v0.19.6 to v0.19.8, with more than 150 PRs merged. The surface spans model fallback, nested sub-agents, session management, a WeCom channel, and permission controls, but the through-line is a single one: keeping long-running agents alive in the real world. For teams that already run agents inside daily work, these changes cut closer to the pain than adding more parallel workers.
The engineering judgment of fallback: which errors to mask, which to surface
The interruption that hurts is not the model being wrong — it is a 429, 503, or 529 landing mid-task on a long job. Manually switching models and resending breaks rhythm and can lose accumulated context. The new model fallback is an opt-in chain: after the primary model exhausts its retry budget, the system switches to the next model on the chain.
The real design judgment sits in the trigger. Only capacity and rate-limiting errors trip the switch; authentication, quota, configuration, and client errors fail immediately. That boundary deserves a pause: swapping models rescues “no capacity right now,” but it never rescues a broken credential or a mis-written config. Mask the latter and you disguise a fixable configuration problem as a model problem, which only makes debugging harder. Fail-over belongs to transient faults; fail-fast belongs to deterministic errors. Where you draw that line decides whether a fallback system is trustworthy or self-deceiving.
Two details show the team reasoned through consequences. First, fallback only advances before visible output exists; if a fallback model has already streamed partial content and then fails, the system will not try the next one — no stitched-together answers assembled from multiple half-finished models. The next user turn always restarts from the primary, so work never gets stranded on the second choice. Second, configuration lands on clean scope flags: /model --project writes the repo’s .qwen/settings.json, --global writes ~/.qwen/settings.json, and the same flags apply to --fast, --voice, and --vision. Fallback chains can now be isolated per project — which repo deserves the premium model and which should bind the cheap one is a per-repo decision, not a machine-wide default. The operational reading is direct: configure the chain on a test repo first, and a production repo’s fallback behavior cannot be quietly changed by another project.
Nested sub-agents: depth with governance
Sub-agents can now spawn their own sub-agents, to a default depth of five, adjustable via model.maxSubagentDepth; set it back to 1 to restore the old behavior. Depth is not decoration: TUI and Web Shell render nesting as a tree with ↳ indentation, and when a parent exits, its children are promoted to root with a from <parent> annotation.
Runaway recursion is met with two layers of protection, and the layers think differently: at schema time, the deepest agents simply do not see the agent tool — they do not know they could spawn more; at runtime, spawn requests beyond the depth cap are rejected outright. One layer cannot see, one layer is refused; together they keep exponential growth inside the governance envelope.
The scope is drawn explicitly: nesting is on by default; Fork subprocesses, teammates, and workflow agents are excluded; Arena sessions can nest; background agents restore depth and cap on resume. Most coding tasks need one or two levels of delegation — if you find yourself needing a deeper tree, the thing to check first is whether the task definition is too vague, not whether the cap should go higher. That echoes a general rule of governance design: a cap’s value is not blocking reasonable demand but making unreasonable structure visible early.
Sessions and permissions: from feature list to working system
Web Shell session management absorbed seven PRs in one go: archive and delete, named groups, color labels, a scheduled-tasks page, and split-screen views. The rules carry protective design: active sessions cannot be archived or deleted; archived sessions cannot be reopened, returning 409 session_archived; deletion is a hard delete behind a confirmation. Overview cards sort by priority — approvals needed, then running, then idle — and split views assemble from a ?split=a,b URL parameter. For long-lived projects, sessions turning from a flat list into an organized resource changes “find last week’s context” from a search problem into a browsing problem.
Permissions add two missing pieces. The PreToolUse hook’s ask decision used to be equivalent to deny; it now raises a native confirmation dialog — allow once or suggest changes, with a deliberate absence of always-allow — and non-interactive or background contexts still fall back to deny. Permission rules also gained Tool(param:value) parameter-level syntax, so you can intercept only specific tool invocations, such as spawning against a particular model.
Other updates and fixes
The WeCom channel’s entry barrier is down to two fields: Bot ID and Secret, configured as {"type":"wecom","botId":...,"secret":...} and started with qwen channel start my-wecom. A long-lived WebSocket receives text, voice, and files; replies use WeCom markdown; outbound supports local image markers only. Auto-generated skills now get a review dialog: content is sanitized, reads are capped at 64 KiB, display is capped at 12 lines, and o opens the skill straight in your editor. Fixes include the context-window default corrected to 200k and 401 errors no longer hanging; the upgrade is one line: npm i @qwen-code/qwen-code@latest -g.
Three judgments to take away
What is worth remembering is not the feature list but the judgments. First, fallback should mask only transient faults while deterministic errors surface immediately — that is the balance between availability and debuggability, and it applies to any agent system you build yourself. Second, nested sub-agents govern delegation with “cannot see” plus “will be refused,” keeping flexibility inside a controllable envelope. Third, sessions and permissions becoming first-class features marks the shift of agents from experimental tools to systems you can operate long-term. Reliability was never the flashiest feature; it is the vote that decides whether an agent can go to production.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.