Support OpenCode `--agent` passthrough for sub-agent routing per workflow

[cameronmpalmer]

Problem

OpenCode supports named agents via opencode run --agent <name>, allowing users to define agents with different system prompts and permission profiles. For example:

• Planner agent: read-only analysis (edit: ask, bash: ask), system prompt instructs "never edit files"
• Implementer agent: full editing (edit: allow, bash: allow), system prompt instructs writing production code

This maps perfectly to roborev's review/fix workflow split — use a conservative read-only agent for reviews, and a capable editing agent for fixes.

However, roborev's OpenCode integration invokes:

opencode run --format json [--model MODEL] [prompt]

There is no mechanism to pass --agent <name> through to the OpenCode CLI. This means roborev always gets whichever agent OpenCode defaults to, regardless of whether the job is a review or a fix.

Why this matters

In non-interactive mode (opencode run --format json), OpenCode auto-approves permission prompts — so edit: ask becomes effectively allow. But the system prompt still governs behavior. A Planner agent whose instructions say "never edit files" will refuse to make changes even when permissions allow it.

This creates a deadlock for the fix workflow:

• The default agent (Planner) won't edit files due to its system prompt
• The agent that can edit (Implementer) can't be selected by roborev

The workaround today is to use a different top-level agent for fixes (e.g., fix_agent = "claude-code"), which defeats the purpose of having OpenCode's agent system.

Proposed solutions

Any of these would resolve the issue (in rough order of preference):

1. Per-workflow OpenCode agent config keys

# .roborev.toml or global config
opencode_agent = "implementer"              # default for all opencode workflows
review_opencode_agent = "planner"           # override for reviews
fix_opencode_agent = "implementer"          # override for fixes

roborev would then invoke opencode run --agent <name> --format json ....

2. Generic per-agent extra args

A more general mechanism that works for any agent:

# Extra CLI args appended when invoking a specific agent
[agent_args]
opencode = "--agent implementer"

# Or per-workflow
[review_agent_args]
opencode = "--agent planner"

[fix_agent_args]
opencode = "--agent implementer"

3. opencode_cmd config key

Similar to the existing claude_code_cmd, codex_cmd, and cursor_cmd keys — allow overriding the command used to invoke OpenCode:

opencode_cmd = "/path/to/wrapper-script"

This is the least ergonomic option but would unblock the use case immediately.

Environment

• roborev v0.37.0
• OpenCode (latest, installed via opencode-ai npm package)
• OpenCode agents defined in ~/.config/opencode/agent/ with distinct system prompts and permissions

[danielaborg]

I like option 2 the best, and regardless of whether it's implemented or not I'd still like to have option 3 for flexibility and configurability.

Option 1 is fine (easy to understand and configure - good, fit-to-purpose but inflexible configuration = less good but easy to use). But: option 2 is almost as simple to understand and allows infinitely much more workflow and configuration creativity, big win in the long term.