opencode-auto-resume-v2

OpenCode v2 plugin adapter for opencode-auto-resume — the stall-detection / auto-continue plugin.

OpenCode 2.x loads plugins through the new v2 plugin API ({ id, setup(context) }) and rejects the v1 Plugin(ctx, options) entrypoint that opencode-auto-resume ships (see upstream issue Mte90/opencode-auto-resume#33). This adapter bridges the two so the v1 recovery engine runs unchanged on OpenCode 2.x:

1. Builds a v1-shaped client bridge on top of the v2 plugin context (context.session.prompt, context.session.interrupt, context.session.context, context.session.command).
2. Invokes the upstream v1 plugin with that legacy context.
3. Re-registers the resulting v1 hooks into the v2 domain API: the event loop (context.event.subscribe), tools (context.tool.transform), and prompt/tool/command hooks.
4. Translates v2 bus events into the v1 event shape the engine expects.

The recovery logic itself is untouched — it comes entirely from the upstream package, so this adapter automatically stays in sync with upstream v1 changes.

What 0.2.0 fixes

0.1.1 had three v2/v1 mismatches that produced a "Tool execution interrupted" storm: the engine aborted live tool calls and then stood down instead of resuming.

| Bug | Effect | Fix | |---|---|---| | session.status() returned {}, and v2 messages use content: [{type:"tool"}] while the engine parses v1 parts[]/tool_calls | The engine could not see in-flight tools, so its 15s "parent stuck" check aborted mid-tool runs | status() now reports live sessions/tools from the event bus, and normalizeMessages() converts v2 messages (only running tools become tool_use/pending) | | session.execution.interrupted was translated to session.interrupted even for aborts the adapter itself issued | The engine read its own abort as a user cancel (userCancelled = true) and skipped the resume prompt | Aborts issued by the adapter are remembered for 15s; their interrupted event is swallowed | | Every v2 event was forwarded as message.updated with no role/tokens | Activity tracking was blind, and the stall detector worked from stale state | Events are mapped explicitly (session.step.ended → finish/tokens, session.inbox.delivered → user message, tool/step/text/reasoning events refresh activity); unknown events are dropped |

Also fixed: tool.ids is now answered from context.tool.list() instead of throwing on every check, and the default resume prompt is a guarded instruction rather than the bare string continue (see upstream #45).

Behaviour that is intentionally preserved: a genuinely silent stream still recovers at the engine's 180s chunk timeout, because status() stops reporting "busy" after 90s without events.

Upstream PR Mte90/opencode-auto-resume#44 proposes merging this adapter into the upstream package itself (as its ./server entrypoint). Once that lands and is published, you can install opencode-auto-resume directly on 2.x and drop this adapter.

Requirements

• OpenCode 2.x (opencode --version ≥ 2.0). The adapter relies on the v2 plugin context exposing tool, session, event and command domains.

Install

opencode plugin add opencode-auto-resume-v2

or add it to opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["opencode-auto-resume-v2"]
}

Then remove the legacy v1 entry if it is still configured:

opencode plugin remove opencode-auto-resume 2>/dev/null || true

Configuration options pass through to the upstream engine, e.g.:

{
  "plugin": [
    ["opencode-auto-resume-v2", {
      "chunkTimeoutMs": 180000,
      "gracePeriodMs": 3000,
      "maxRetries": 3
    }]
  ]
}

See the upstream README for the full option reference.

Verification

Plugin loading and plugin stderr are written to:

~/.local/share/opencode/log/opencode.log

Look for:

msg="loading plugin" id=opencode-auto-resume-v2

and any [auto-resume-v2] ... / failed to load plugin lines. A successful load emits the upstream engine's own [auto-resume] info: ... config OK / ... ready lines once events are received.

License

GPL-3.0-or-later, matching the upstream plugin this adapter wraps.