NanoClaw patches OpenCode streaming delta recovery and fixes group reply routing

NanoClaw developers opened PR #3463 on September 20, 2026, resolving a critical token loss issue in its OpenCode streaming provider adapter. When streaming model completions through OpenCode-compatible proxies, certain proxy gateways emit token text inside 'message.part.delta' structures rather than standardized response blocks. In previous builds, the parser failed to recognize this alternative schema, causing the agent to silently drop generated tokens or prematurely terminate output generation. PR #3463 adds explicit fallback handling for delta text chunks, ensuring complete response reconstruction. Concurrently, PR #2328 corrects message destination routing in multi-party groups, defaulting reply targets strictly to the message origin rather than broadcasting across unrelated linked channels. In parallel, PR #2327 tackles post-compaction disorientation by injecting explicit destination reminders into prompt context following SDK auto-compaction routines, preventing agents from forgetting their active conversation target after context truncation. Meanwhile, issue #3858 highlights that native chat adapters omit sender display names, exposing only raw JIDs and complicating conversational personalization.

The facts

- PR #3463 introduces fallback parsing for message.part.delta text chunks in NanoClaw's OpenCode streaming adapter. - Prevents token dropouts and truncated completions when communicating through non-standard OpenCode proxy gateways. - PR #2328 ensures agent replies in multi-destination group chats default strictly to the originating message's channel. - PR #2327 automatically appends destination reminders to prompt context after context window compaction cycles. - Issue #3858 identifies an adapter limitation where sender display names are suppressed in favor of raw phone numbers or JIDs.

Why it matters

Edge-focused agent frameworks must contend with a fractured ecosystem of LLM gateways and chat protocols. Streaming parsers that strictly enforce a single response envelope will inevitably break when deployed behind reverse proxies, local model runners, or third-party router layers. Fallback delta parsing preserves compatibility across diverse model backends. Furthermore, maintaining accurate message destination state across context compaction cycles is vital for keeping agents coherent in multi-user group channels.

Current

Inspected on 2026-09-21. The NanoClaw stable-channel baseline is v2.3.0 published 2026-08-24T11:45:32Z. The primary source was open (not merged) when captured. Mainline merges, open proposals and packaged releases are distinct availability states; the release baseline does not establish that a proposal has shipped.

Evidence

Primary evidence: nanocoai/nanoclaw PR #3463 (https://github.com/nanocoai/nanoclaw/pull/3463). Supporting context: nanocoai/nanoclaw PR #2328 (https://github.com/nanocoai/nanoclaw/pull/2328); nanocoai/nanoclaw PR #2327 (https://github.com/nanocoai/nanoclaw/pull/2327); nanocoai/nanoclaw issue #3858 (https://github.com/nanocoai/nanoclaw/issues/3858). The source bodies and linked context were inspected. Test results quoted from contributors remain attributed reports, not independently reproduced experiments.

Operator take

Streaming consumers must be tolerant of slight envelope variations across API gateways. Robust parsers inspect multiple candidate paths for token payloads before discarding chunks. Engineering teams integrating NanoClaw with OpenCode or local model proxies should update their provider configurations to leverage delta fallback parsing and monitor group chat routing to ensure replies stay contained to origin channels.

Caveat

PR #3463 is an open pull request undergoing functional verification. Issue #3858 remains unresolved, meaning agents interacting over WhatsApp and Signal must parse raw phone numbers or handle missing user display names until adapter metadata extraction is enhanced.