* fix(agent): drive OpenCode over ACP so long tools and subagents are not force-stopped
`opencode run --format json` reports a tool only when it finishes, with the
call and its result in one event. The daemon never saw a tool in flight, so a
long silent tool or a delegated subagent was judged by the 10-minute OpenCode
idle window instead of the tool budget and force-stopped (GH #9614).
Built-in OpenCode 1.15+ and 2.x runtimes now run `opencode acp`, which on both
majors announces a tool when it starts executing. MessageToolUse is emitted
from that in_progress frame, so the daemon applies AgentToolWatchdog for the
whole tool. One code path covers both majors:
- model via session/set_config_option (2.x has no session/set_model); effort
via the advertised `effort` option, splitting 2.x `model#variant`
- resume via session/load; a missing session is reported as ResumeRejected
- cancel via session/cancel, which aborts the running tool on both majors
- MCP: 1.x keeps OPENCODE_CONFIG_CONTENT; 2.x gets servers in session/new,
which lifts the 2.x MCP refusal on this transport
- 2.x subagent text (`_meta["opencode/child-session"]`) stays out of the
parent's output; its tool calls are kept
- usage summed from the session export over this turn's assistant messages,
since the prompt result's usage is not per-turn on either major
The run transport stays as the fallback for custom profiles, older or
unknown versions, agents with custom_args, and MULTICA_OPENCODE_TRANSPORT=run.
Also fix the run transport on 2.x: tool_use parts carry the call id as `id`
(1.x: `callID`), so every 2.x tool message had an empty CallID.
Co-authored-by: multica-agent <[email protected]>
* fix(agent): bound OpenCode ACP teardown when the runtime stops reading
A runtime that stops reading stdin mid-prompt left the prompt write blocked
holding the client's write lock. The cancel path then waited on that lock to
send session/cancel, and the cancel grace and the process-group kill only ran
after the write, so the process could outlive the task indefinitely.
- session/cancel is sent only once the prompt is fully delivered, from its own
goroutine, so the grace starts immediately and a stalled pipe cannot hold it
- a turn OpenCode never answered (or a cancelled/timed-out setup) is torn
down forcibly: stdin closed first to release a blocked write, then SIGTERM to
the group and SIGKILL after the grace; an answered turn still exits on EOF
Regression tests cover a stalled prompt write under a 1s timeout and a runtime
that ignores session/cancel.
Co-authored-by: multica-agent <[email protected]>
* fix(agent): bound OpenCode ACP teardown when a setup write stalls
Setup requests (initialize, session/new, session/load, set_config_option)
are synchronous stdin writes that check their context only after the write
returns. A runtime that stops reading after initialize, with a large MCP
payload in session/new, left the run goroutine blocked in that write, so
the deferred teardown never ran and the process outlived the timeout.
A watcher now force-stops the process as soon as the run ends before the
prompt is fully delivered; closing stdin releases the blocked write. Once
the prompt is delivered the existing session/cancel path owns shutdown.
Regression test: 256 KiB MCP env, runtime stalls after initialize, 1s
timeout; returns in ~1s (60s before the fix).
Co-authored-by: multica-agent <[email protected]>
---------
Co-authored-by: Multica Eve <[email protected]>
Co-authored-by: multica-agent <[email protected]>
OpenCode 2.0 changed four things the daemon depends on. Two fail loudly —
`--dir` and `--variant` both exit 1 on an unrecognized flag, which is why every
task on a 2.x runtime died in ~130ms with empty output (GH #8586). Two are
silent: MCP config had no working channel left, and `opencode run` became a thin
client in front of a resident service, so signalling the client's process group
no longer stopped the run.
- `--dir` is dropped on 2.x; the workdir comes from cmd.Dir + PWD.
- The thinking level folds into the model string as `provider/model#variant`.
- Cancellation calls POST /api/session/<id>/interrupt in front of the existing
SIGTERM→SIGKILL, which is the only thing that stops a 2.x run.
- Runs carrying Multica-managed MCP servers are refused with an error naming
every source, rather than started without them: 2.x accepts MCP only through a
file in the task workdir, where the agent's own commits would capture the
credentials. Restoring MCP is tracked separately.
The 1.x path is unchanged, pinned by argv and MCP regression tests. Verified
against @opencode/cli 2.0.10 and opencode-ai 1.18.31 on macOS/arm64; Windows
cancellation is not yet verified end to end.
