Diagnosis: "build agent empty input" after orchestrator output

Status: Resolved for interview injections. Date: 2026-08-10 Related PR: #818 (fix/preset-tui-slash-command) — same root class as the original /preset fix. Suspected sibling bug reported by user: During superpowers / brainstorm skill conversations, when the orchestrator asks for confirmation or work is interrupted (subagent completes, background task finishes), a build agent turn sometimes appears with an empty user input.

TL;DR

The build agent turn with empty input is the same class of bug as the original /preset issue fixed in #818: a plugin hook calls sessionSdk.promptAsync({ body: { parts: [createInternalAgentTextPart(...)] } }) without specifying an agent field. opencode then resolves the agent via agents.defaultInfo(), which falls back to the built-in build agent whenever default_agent is unset, user-overridden, or not effectively applied. The synthetic: true flag hides the injected text from the TUI, so the user perceives the build turn as having "empty input."

Update: the incomplete-todo continuation path passes agent: 'orchestrator', and interview injections now use the interview-only runtime boundary, which always targets the orchestrator. Smartfetch remains a separate call site.

Root cause (causal chain, cross-validated)

Orchestrator enters input-wait. After emitting a confirmation question (skill flow), the assistant turn finishes. opencode's per-session Runner transitions to Idle (packages/opencode/src/effect/runner.ts:115-138, packages/opencode/src/session/run-state.ts:60-63). The session is no longer "busy" from the Runner's perspective.
Plugin hook fires promptAsync with a synthetic part and historically no agent field. The interview service now delegates these calls through InterviewSessionRuntime, whose v1 implementation supplies agent: 'orchestrator'; the v2 bridge uses the v2 orchestrator session API directly.
- Continuation nudge (continuation-evaluator.ts) sets agent: 'orchestrator' and only runs when continueOnIdle is enabled.
opencode does not guard promptAsync against busy/input-wait state. The HTTP handler at packages/opencode/src/server/routes/instance/httpapi/handlers/session.ts:311-329 does not call assertNotBusy and does not consult the Question service. It proceeds straight to promptSvc.prompt. The Runner, being Idle, immediately startRuns the new turn (runner.ts:131-134). No queue, no reject, no cancellation of the pending question.
Agent resolves to build. packages/opencode/src/session/prompt.ts:636-637:
```
const agentName = input.agent
const ag = agentName ? yield* agents.get(agentName) : yield* agents.defaultInfo()
```
When input.agent is omitted, opencode uses agents.defaultInfo() (packages/opencode/src/agent/agent.ts:328-340), which returns the first visible mode: "primary" agent — build (declared first in the agent registry, agent.ts:141-155). omos attempts to set default_agent = "orchestrator" via its config hook (src/index.ts:546-551), but only when default_agent is absent. build is selected whenever:
- config.setDefaultAgent === false (plugin config disables it)
- The user's opencode.json already sets a different default_agent
- The config hook didn't run or didn't apply (SDK/runtime version skew: plugin built against @opencode-ai/sdk v1.4.3, installed runtime v1.18.3)
- The orchestrator agent isn't registered at config-load time
"Empty input" is the synthetic flag's visual effect. synthetic: true only controls TUI visibility. packages/opencode/src/session/message-v2.ts:206 still includes synthetic text parts in model messages (the only filters are !part.ignored and part.text !== ""; there is no synthetic filter when building model messages). The user sees the build agent respond to a turn with no visible user message — perceived as "empty input." (createInternalAgentTextPart appends a \n marker, so the text is non-empty from the LLM's perspective.)
Session agent is durably corrupted. packages/opencode/src/session/prompt.ts:672-689 compares current.agent !== info.agent and, if they differ, calls sessions.setAgentModel({ agent: info.agent, ... }). A single agent-less promptAsync that resolves to build permanently rewrites the session's agent to build — every subsequent turn also routes to build until explicitly reset.

Confirmed-affected call sites

File:line	Trigger	Body omits `agent`?	Gate
`src/hooks/task-session-manager/continuation-evaluator.ts` (`promptAsync`)	`session.idle` / `session.status(idle)` on orchestrator session with incomplete todos when the opt-in beta `backgroundJobs.continueOnIdle` is `true`	No (`agent: 'orchestrator'`)	`continueOnIdle`, process-local one-attempt gate (reserve→commit), `hasInputWait`, `isCurrentContinuation`, `isFallbackInProgress`, `backgroundJobBoard.hasTerminalUnreconciled`, malformed/active SDK short-circuits
`src/interview/service.ts`	User submits interview input	No — uses `InterviewSessionRuntime.continue()`	`sessionBusy` lock, interview active state
`src/interview/service.ts`	Interview URL notification	No routing ambiguity — uses `InterviewSessionRuntime.notify()`	none
`src/tools/smartfetch/secondary-model.ts:252`	Smartfetch secondary model query	Yes	none

Correct pattern (for comparison)

src/hooks/foreground-fallback/index.ts:635-639 explicitly includes the agent:

const promptBody = {
  parts: lastUser.parts,
  model: ref,
  ...(agentName ? { agent: agentName } : {}),
};

This is the pattern every promptAsync caller in omos should follow.

Why the `hasInputWait` gate in task-session-manager is not sufficient

The gate exists and works in the common case (see continuation-evaluator.ts and task-session-manager/index.test.ts continuation cases). Notes:

Continuation is opt-in beta. backgroundJobs.continueOnIdle defaults to false; set it to true to enable continuation SDK calls. Idle reconciliation remains active either way. When enabled, a process-local reserve/commit gate allows at most one promptAsync per session epoch between real user messages.
Documented race window (when enabled). IDLE_RECONCILE_DELAY_MS = 2_000. The idle-reconciliation comment admits late completions can still race the window. If session.idle fires and the timer expires before question.asked is delivered, and the SDK liveness reads resolve before the wait is tracked, a nudge can still fire.
Input-wait is not the only trigger. The interview/skill path (src/interview/service.ts) does not consult hasInputWait at all — it injects on user dashboard actions, which can happen while the orchestrator is mid-question. Its continuation routing is nevertheless explicit now.
Continuation path sets agent: 'orchestrator'. The historical missing- agent misroute on the continuation nudge is fixed in continuation-evaluator.ts; interview uses its runtime boundary as well. Smartfetch remains outside this interview fix.

Why this is the same class as the #818 `/preset` fix

#818's original bug: /preset used createInternalAgentTextPart() to trigger an LLM turn that was invisible in the TUI (synthetic: true). The fix moved /preset to pure TUI dialogs (src/tui-preset.ts uses only api.ui.dialog / DialogSelect / DialogPrompt / DialogConfirm — no promptAsync).

The historical bug used the same createInternalAgentTextPart + promptAsync pattern and omitted agent, so the invisible turn could route to build instead of the orchestrator. The interview path no longer makes those raw calls.

Fix implemented

The interview service no longer owns raw session prompt calls. Its narrow InterviewSessionRuntime boundary routes all continuation prompts to the orchestrator. v1 uses the nested SDK client; v2 uses a marker/context bridge and the v2 session methods without expanding the global client shim.

Hardening (optional, larger scope)

Input-wait guard on the interview/skill path. Consult hasInputWait (or an equivalent signal) before injecting in src/interview/service.ts. Do not inject while the orchestrator is waiting for user input.
Post-injection agent assertion. After each promptAsync, assert current.agent was not changed out from under the orchestrator; if it was, restore it via setAgentModel.
Investigate the default_agent application reliability on opencode v1.18.x. The plugin was built against @opencode-ai/sdk v1.4.3; the installed runtime is v1.18.3. The config hook's default_agent = 'orchestrator' mutation may not be applied reliably under this skew. (Note: #799 tracks the package upgrade.)
Shrink or eliminate the IDLE_RECONCILE_DELAY_MS race for sessions that have a pending question.asked / permission.asked.

Evidence index

omos source

Continuation nudge (fixed agent + opt-in beta + one-attempt gate): src/hooks/task-session-manager/continuation-evaluator.ts, continuation-attempt-gate.ts, backgroundJobs.continueOnIdle in src/config/schema.ts
Missing agent field (skill flow): src/interview/service.ts:622, 871, 933, 1007
Correct pattern for comparison: src/hooks/foreground-fallback/index.ts:635-639
omos sets default_agent only when absent: src/index.ts:546-551
createInternalAgentTextPart produces synthetic: true: src/utils/internal-initiator.ts:9-21
CONTINUATION_NUDGE is non-empty: src/hooks/task-session-manager/continuation-evaluator.ts
hasInputWait / continuation gates: continuation-evaluator.ts, input-wait-tracker.ts
IDLE_RECONCILE_DELAY_MS race window: src/hooks/task-session-manager/index.ts (IDLE_RECONCILE_DELAY_MS)
disableDefaultAgents preserves build and plan: src/cli/config-io.ts:564-600

opencode source (`anomalyco/opencode` @ `dev`)

promptAsync HTTP handler (no busy/input-wait guard): packages/opencode/src/server/routes/instance/httpapi/handlers/session.ts:311-329
prompt HTTP handler (no guard): same file, :295-309
Internal prompt always starts a new turn: packages/opencode/src/session/prompt.ts:1052-1071
Agent selection: input.agent ?? defaultInfo(): packages/opencode/src/session/prompt.ts:636-637
defaultInfo() → default_agent or first visible primary: packages/opencode/src/agent/agent.ts:328-340
build agent definition (default primary, first in registry): packages/opencode/src/agent/agent.ts:141-155
Destructive setAgentModel overwrite on agent change: packages/opencode/src/session/prompt.ts:672-689
Synthetic parts included in model messages: packages/opencode/src/session/message-v2.ts:206
Runner ensureRunning (Idle = run now, no queue): packages/opencode/src/effect/runner.ts:115-138
Runner Idle transition on turn end: packages/opencode/src/session/run-state.ts:60-63
assertNotBusy (NOT used by prompt/promptAsync): packages/opencode/src/session/run-state.ts:71-75

SDK types

default_agent doc: "Falls back to 'build' if not set or invalid": node_modules/@opencode-ai/sdk/dist/v2/gen/types.gen.d.ts:1255-1257
SessionPromptAsyncData.body.agent? is optional: node_modules/@opencode-ai/sdk/dist/v2/gen/types.gen.d.ts:3241-3269
build is a built-in agent: node_modules/@opencode-ai/sdk/dist/v2/gen/types.gen.d.ts:1273-1279

Follow-up

Smartfetch still has its own secondary-model prompt path and is outside the interview runtime boundary.
Continue keeping cache-sensitive v2 interview rewriting restricted to the trailing marker message.

Regression coverage lives in src/interview/runtime.test.ts, src/interview/finalization.test.ts, and src/v2/interview-bridge.test.ts.

build-agent-empty-input-diagnosis.md 12 KB パーマリンク 履歴 Raw