claude-code - 💡(How to fix) Fix Background sub-agents intermittently return context-decoupled hallucinations (ignore prompt entirely, reported as 'completed')

Summary

Sub-agents spawned via the Agent / Task tool (run_in_background: true) intermittently return completions that are completely decoupled from their input context — a fully-formed answer to a question that was never asked, on a consistent and bizarre off-topic theme (obscure Portuguese river/island geography), bearing no relation to the prompt, the repository, or any tool the agent had available. The orchestrator is told the agent "completed" with a confident, plausible-length message that is 100% hallucinated.

Environment

• Claude Code: 2.1.143 (cli entrypoint)
• Model: claude-opus-4-7[1m] (Opus 4.7, 1M context). Sub-agents inherited the parent model; the affected agent type (generic-dev:feature-developer) has no model: frontmatter override.
• Platform: macOS (Darwin 25.3.0)
• Affected feature: background sub-agents launched through the Agent/Task tool

What was asked

In a normal orchestration session I delegated well-scoped software tasks to several background sub-agents (code audit, a TypeScript refactor, a Swift dead-code removal, and a review+test pass). The prompts were ordinary multi-file engineering instructions about a local Safari-extension codebase. No prompt contained anything about geography, Portugal, or the hallucinated content.

What happened (controlled within one session)

Worked correctly (controls):

• a08d89fe489c874c9 — feature-developer, Swift task — correct, verified output.
• a336f548ae621889a, a31c4960fdd261b4a — Explore audit agents — correct output.

Failed with context-decoupled hallucination:

• af220485a3df5cf96 — feature-developer. Zero tool uses. User-prompt timestamp 2026-05-17T05:46:12.458Z → assistant 2026-05-17T05:46:18.474Z (~6 s). Its first and only completion was an unprompted essay about "Ilha da Morosa / Moura, Portugal" (a tidal islet, Moorish history, the Guadiana/Ardila rivers). I inspected the full transcript — only 4 records: the input user message is the exact, correct task prompt; the two attachments are the standard deferred_tools_delta and skill_listing. The input was clean; the model's very first generation ignored it entirely.
• a0ddb68f220163958 — feature-developer. Ran ~21 min / 143 tool calls and actually did correct, independently-verifiable work, but its final assistant message was a hallucination about "Teixeira do Minho / Cávado River basin, Braga, Portugal". Notably that message's text content contains literal scaffolding tokens — [TRACE], assistant: [THINKING], [/THINKING] — emitted as plain prose, i.e. the model produced something resembling a serialized/foreign transcript rather than a normal response.

Key diagnostic facts

1. The zero-tool-use failure (af22…) proves this is not prompt-injection from file content (the agent read nothing) and not a poisoned agent/plugin/prompt: grep -ri across ~/.claude (agent defs, plugins, marketplaces) finds no occurrence of any hallucinated term (morosa, cávado, teixeira do minho, moura, moorish). The transcript shows a clean input prompt.
2. Two independent failures in the same session share a highly specific signature: obscure Portuguese hydrography/geography, fully decoupled from input.
3. The [TRACE] / [THINKING] literal-token leak in a0dd…'s output suggests context corruption or a foreign-context bleed at the inference/serving layer, not ordinary sampling drift.
4. Non-deterministic: the same agent type (feature-developer) and the same session both succeeded and failed; one Portuguese failure occurred on turn 1 (0 tools), the other only on the final message after 143 correct tool calls.

Impact

• Silent corruption of delegated work. A background agent reports completed with a confident, well-formed message that is entirely unrelated to the task. There is no error, no crash — only a content mismatch.
• A 21-minute, 143-tool-call run's real, correct work was nearly discarded because only the final summary message was inspected. Orchestration workflows that trust sub-agent final messages will silently propagate garbage.
• Erodes trust in all delegated output unless every sub-agent result is independently re-verified against the repository.

Reproduction

Non-deterministic. Within a single Claude Code session, spawn multiple background sub-agents (subagent_type such as generic-dev:feature-developer, inheriting Opus 4.7 1M context) with substantial multi-file engineering prompts. A subset return Portuguese-geography hallucinations decoupled from input — either immediately (0 tool uses) or as the final message after otherwise-correct work. No project content is required to trigger it (the 0-tool case never read a file).

Artifacts

Sub-agent JSONL transcripts (local, ephemeral under /private/tmp/claude-501/.../tasks/):

• af220485a3df5cf96.output — 4 records; the cleanest reproduction (clean input, immediate decoupled hallucination, 0 tools). Most diagnostic.
• a0ddb68f220163958.output — ~848 KB; correct work for 143 tool calls, then a decoupled hallucination as the final message with leaked [TRACE]/[THINKING] scaffolding tokens.

Happy to provide full transcripts to Anthropic on request.