openclaw - 💡(How to fix) Fix [Bug]: opencode-go works via direct infer but fails in embedded agent runtime with session takeover

Bug type

Provider/runtime compatibility bug.

Summary

opencode-go/mimo-v2.5 works through direct inference, but fails consistently when used as the embedded agent runtime for the main Telegram/dashboard agent. The observed failure is EmbeddedAttemptSessionTakeoverError / session file changed while the embedded prompt lock was released.

This still reproduces after updating to OpenClaw 2026.5.28 stable, adding the provider-specific API key, restarting the gateway, creating fresh sessions, purging stale session files, and disabling the local session-snapshot stripping LaunchAgent.

Expected behavior

If openclaw infer model run --model opencode-go/mimo-v2.5 succeeds and the provider is configured/authenticated, the same model should be usable as an agent runtime model for normal Telegram/dashboard conversations, including follow-up turns and tool use.

Actual behavior

Direct inference succeeds, but the embedded agent runtime fails during normal conversation flow with a session takeover/lock error. Switching the same agent runtime to openai/gpt-5.5 works immediately as a control.

Reproduction / steps already tried

Environment: macOS install, OpenClaw updated from 2026.5.28-beta.4 to 2026.5.28 stable during the investigation.

1. Configure opencode-go/mimo-v2.5 as the agent runtime model for main.
2. Confirm direct inference works:

openclaw infer model run --model opencode-go/mimo-v2.5 --prompt "Responde exactamente: ok"

1. Use the same model in the embedded agent runtime through Telegram/dashboard.
2. First/direct call may work, but follow-up turns/tool sessions fail with EmbeddedAttemptSessionTakeoverError / session file changed while embedded prompt lock was released.

Troubleshooting steps that did not resolve it:

• Added OPENCODE_GO_API_KEY to the OpenClaw environment.
• Restarted the gateway.
• Started a clean Telegram session with /new.
• Purged stale files from agents/main/sessions/.
• Updated OpenClaw from 2026.5.28-beta.4 to 2026.5.28 stable.
• Ran openclaw doctor --fix.
• Performed full purge + restart + /new again.
• Disabled the local strip-session-snapshots LaunchAgent / compaction-adjacent file rewrite process.
• Retested after the above cleanup.

Control test:

• Switching the same agent runtime to openai/gpt-5.5 works immediately.
• Other working runtime models observed locally: github-copilot/gpt-5.4-mini, github-copilot/gpt-5.5, and anthropic/claude-opus-4-7.

Why this seems provider/runtime-specific

The important distinction is that direct model inference succeeds, but the embedded runtime path fails. That suggests this is not only a provider auth issue or a basic model availability issue. The failing path appears tied to the OpenClaw agent session lifecycle: session locks, follow-up turns, tool calling, compaction/session-file updates, or provider adapter behavior inside embedded agent execution.

Workaround

Use opencode-go/mimo-v2.5 only for direct inference for now:

openclaw infer model run --model opencode-go/mimo-v2.5 --prompt "ok"

For the main Telegram/dashboard agent, use a different runtime model such as openai/gpt-5.5 or github-copilot/gpt-5.4-mini.

Related issues that may be adjacent but do not seem to cover this exact failure

• #87790: separate Zen and Go API key env vars
• #79263: local-transport CLI env loading behavior
• #74056: missing OpenCode Go models in model list
• #43247: opencode-go provider missing from API key candidates

Suggested next step

Please verify whether the opencode-go provider adapter supports the same embedded agent runtime/session lifecycle as the other providers. A focused regression test could compare:

1. direct infer model run success,
2. first embedded agent turn success,
3. second embedded agent turn success,
4. embedded tool-call turn success,
5. behavior when session compaction/snapshot updates occur.