Programmatic agent creation silently no-ops: prompt variable "{{model}}" has no value, loop swallows the error (dsh 0.1.1-rc.2)

[Kkkkkerwin]

Bug Description

Creating an agent programmatically (ctx.agents.create({agentOptions: {provider: "deepseek-official"}, ...}) then agent.followup(userMsg) + await agent.whenIdle()) silently produces a dead turn: the LLM is never called, stats are all zero, and the caller cannot distinguish "task done" from "task crashed". Two defects compound:

1. Prompt rendering fails without an explicit model - the host persona requires {{model}} (section deployment:persona), and the variable is bound to context.agent?.options.model (dsh-agent-loop/lib/index.js:1024-1026) with no fallback to agentDefaultModel. Programmatic callers that omit model in agentOptions (as our MCP bridge did, since the default is "") hit prompt variable "{{model}}" has no value for this assembly (section "deployment:persona") at the first renderPrompt (dsh-agent-loop/lib/index.js:611).

2. The loop swallows the error - kick() wraps turns in try { while (await this.turn()); } catch (_error) {} (dsh-agent-loop/lib/index.js:477-491), and whenIdle() resolves normally after an error turn (:460-465). The failure is only visible deep in the session log (turn/end with reason.kind === "error").

Steps to Reproduce

1. Create an agent without model in agentOptions:

   const handle = await ctx.agents.create({
     sessionId: SessionId(randomUUID()),
     meta: { cwd, agentPreset: "standard" },
     agentOptions: { provider: "deepseek-official" },  // no model
     setup: async (agentCtx) => { await ctx.agentPresets.mount(agentCtx, "standard"); }
   });
   handle.agent.followup(createUserMessage({ content: [{ type: "text", text: "hello" }], source: { kind: "plugin", plugin: "x" } }));
   await handle.agent.whenIdle();

[argszero]

Confirmed source-verified against the current tip cd5ef81481 (0.1.2-alpha.1) — your two-defect diagnosis is exact, and I'd add that this is a known pattern with an established precedent that agent-loop simply doesn't follow.

The decisive precedent is webhook. packages/webhook/webhook/src/session.ts:63-65 does exactly the fallback your bug is missing:

if (model === undefined) {
  const selected = ctx.agentDefaultModel.currentSelection()
  agentOptions = { provider: selected.provider, model: selected.model }
  modelSelection = { ...selected }
}

So agentDefaultModel is a real capability (ctx.agentDefaultModel.currentSelection(), defined at packages/core/agent-default-model/src/index.ts:88) and the webhook entry point already falls back to it when a caller omits model. But the agent-loop entry point does not — and the two affected sites are exactly the two sides of your bug:

Side 1 — the model system-prompt variable never sees agentDefaultModel. packages/core/agent-loop/src/index.ts:353:

ctx.systemPrompt.variable('model', context => context.agent?.options.model)

model?: string is optional (packages/core/agent/src/runtime-types.ts:26), so when a programmatic caller omits it, context.agent?.options.model is undefined, and interpolate throws prompt variable "{{model}}" has no value at system-prompt/src/index.ts:340. The route at agent.ts:458 uses this.options.model ?? '' (empty string, not undefined), but the variable uses the raw options.model — so it's undefined and throws. This is the asymmetry.

Side 2 — the render error masks the real route error, and kick() swallows both. agent.ts:344 calls renderPrompt(assembly) before buildRequest at :348. So the meaningful route error at agent.ts:482 (agent "…" has no provider/model) — which would actually be thrown — never gets a chance, because the prompt render fires first with the confusing {{model}} has no value message. Then kick() at :219-220 swallows it:

try {
  while (await this.turn()) {}
} catch (_error) {
  // Reported failures and cancellation are contained at the driver boundary.
}

and whenIdle() (:202-205) resolves normally. So the caller gets a silent dead turn with zero observable failure — worse than either bug alone.

The fix has a clean precedent to follow. Mirror the webhook fallback in the agent-loop entry: when building the model variable (or the route) and options.model is absent, consult agentDefaultModel.currentSelection() — the same way webhook does. That closes side 1. For side 2, the simplest high-value change is to let the route error surface: reorder so buildRequest's "has no provider/model" check (:482) runs before renderPrompt, or at minimum make kick()/whenIdle() report the swallowed error rather than resolving normally. The webhook pattern means this isn't a design question — it's "agent-loop entry never adopted the fallback webhook already has."

I'd be glad to help with a fork branch + tests if you want to take the webhook-fallback approach — it's a small, well-scoped change with an in-tree precedent to test against. Happy to hold for your call on the preferred fix shape.

中文摘要：已在 alpha.1 源码确认你的两缺陷诊断精确，并指出这是有既有先例但 agent-loop 未遵循的已知模式——webhook 入口（session.ts:63-65）在 model === undefined 时已回退到 ctx.agentDefaultModel.currentSelection()，而 agent-loop 入口没有。两个受影响点位：① agent-loop/src/index.ts:353 的 model 系统提示词变量只绑 context.agent?.options.model（可选字段，缺省为 undefined → {{model}} has no value 抛错；route 用 ?? '' 但不抛，变量却抛——不对称）；② agent.ts:344 renderPrompt 先于 :348 buildRequest 抛错，把 :482 的"has no provider/model"真错误淹没（:482 才是真实原因），随后 kick() :219-220 吞错 → whenIdle() 正常 resolve → 静默死轮。修复 = 镜像 webhook 回退（查 agentDefaultModel）+ 让真实 route 错误浮出。愿帮 fork 分支 + 测试。

[PerryLink]

Confirmed on master (c291e79). Both defects in your report are still present, and the in-tree precedent argument
(argszero's comment) is exactly right.

• The model system-prompt variable is bound to the raw optional field with no fallback:
  packages/core/agent-loop/src/index.ts:422
  ctx.systemPrompt.variable('model', context => context.agent?.options.model)
  So a programmatic caller that omits model hits prompt variable "{{model}}" has no value ... (section "deployment:persona").
• The loop still swallows it: kick() catches per-turn errors and whenIdle() resolves normally, so the caller cannot
  distinguish done from crashed — only the session log's turn/end shows the error.
• The precedent: webhook already falls back to the default selection when a caller omits the model —
  packages/webhook/webhook/src/session.ts:63-65 (if (model === undefined) { ctx.agentDefaultModel.currentSelection() ... }).
  agent-loop just never adopted the same fallback.

Workaround today (works, no fork needed): don't omit the model — resolve it yourself before create and pass it:

const selected = ctx.agentDefaultModel.currentSelection()
await ctx.agents.create({ ..., agentOptions: { provider: selected.provider, model: selected.model } })

and/or check agent.followup(...) result / the session log after whenIdle() rather than trusting silent resolution.

Fix direction for upstream: mirror the webhook fallback in the agent-loop entry (bind the {{model}} variable — or the
route — through agentDefaultModel.currentSelection() when options.model is absent), and let the "agent has no
provider/model" route error surface instead of the prompt-render error. Small, well-scoped, with a tested in-tree
precedent.