fix(workflow): pass the agent model and effort to opencode run (#1193)

[kittimzhe]

What

A workflow step with agent: on OpenCode launched opencode run --format json <prompt> --agent <name> and nothing else, while pull deploys every OpenCode agent as mode: subagent. The report's request-body forensics (#1193) show:

• OpenCode 1 rejects a subagent from run --agent — ! agent "reviewer" is a subagent, not a primary agent. Falling back to default agent — so the definition's prompt and model never run.
• OpenCode 2 applies the definition's prompt but ignores the file's model: and variant:, running on the configured default.

Change (the model half)

• The driver passes the step agent's mapped --model explicitly — the codex family in the same invocation already did. Per the report's table, both majors honor an explicit --model even in fallback.
• The mapped effort rides with it: --variant on OpenCode 1, inside the model id (prov#effort) on OpenCode 2, where run has no separate variant flag. The major is probed once per executable (--version) and cached; a failed probe reads as 1, the pinned devDependency's major.
• local-driver reads the effort off the deployed file's variant field (reverseFromOpencode already collects it into tool_extras.opencode) and passes it through to the adapter.

The mode half — whether pull keeps writing subagent and documents tool_extras.opencode.mode as the escape, or writes a primary-mode copy for workflow-destined agents — is a design call I've left on the issue rather than picking unilaterally.

Verification

Red tests first, both shapes pinned (workflow-local-tool-adapter.test.ts):

• OpenCode 1: run args contain --model anthropic/claude-sonnet-4-5 and --variant high, no #high
• OpenCode 2: run args contain --model anthropic/claude-sonnet-4-5#high, no --variant

Real CLI, after npm run build, against the pinned devDependency (OpenCode 1.18.23), agent file exactly as pull deploys it:

$ opencode run --format json --agent reviewer "Reply with just: ok"
! agent "reviewer" is a subagent, not a primary agent. Falling back to default agent

$ opencode run --format json --agent reviewer --model anthropic/claude-sonnet-4-5 --variant high "Reply with just: ok"
(no usage error — the flags parse, the session starts; the run's own error is upstream provider auth in this sandbox)

The second command shows the emitted shape is accepted by OpenCode 1's CLI; whether the model is honored is the report's table (row 2: explicit --model → model sent, even in fallback). I did not re-run their request-body capture.

Suites: 132/132 (adapter + driver + agent-format), e2e 2/2, tsc and lint clean.

Closes the model half of #1193.

[github-actions]

No confident findings in the three-file head-commit diff inspected. The PR description includes a post-build real-CLI verification record; no missing-testing finding is warranted.

Review limitation: The checkout is shallow, and the requested three-dot diff fails with no merge base. I inspected cf03ed3...8c0b26f instead, so I cannot certify that the entire requested PR diff was reviewed.

No PR code was run, built, or installed.

[kittimzhe]

Closing — superseded by #1192, which landed both halves. I missed the pointer posted 23 minutes after my claim and opened this before re-reading the thread; sorry for the duplicate.

The independent run did converge: the model half here is the same shape #1192 shipped — --model M --variant V on OpenCode 1, --model M#V on OpenCode 2, the major probed once per executable and cached. I'd left the mode half open as a design call, and #1192's answer (keep subagent; on OpenCode 1 a subagent step inlines the instructions without --agent, warning about the file settings that then don't apply) is the cleaner one.

Thanks for the careful-repro note on the issue — glad the forensics held up.