Dispatch a ticket — OpenCode runtime

Read dispatch-ticket for the shared protocol (binding identity, inputs/resolution,
worktree law, brief format, submission, watching, simplify, cleanup). This skill adds only
the OpenCode-specific launcher and verification.

Launcher matrix — OpenCode rows

builder  → opencode --agent builder -m <row: Model> --auto
shaper   → opencode --agent shaper -m <row: Model> --auto

Model comes from the role's orchestrator.md row, never from memory. The Model value
must include the provider prefix — opencode requires <provider>/<model> format (e.g.,
opencode-go/deepseek-v4-flash). Passing the bare model name without prefix throws
ProviderModelNotFoundError. If the model is already the opencode default, -m can be
omitted entirely.

No --effort flag — effort is unreachable on opencode (fact 2 below), so an opencode
Effort cell must be blank. A non-blank one is a misconfigured row: stop and ask the owner.

--auto auto-approves permissions that are not explicitly denied. With the agent's
permission: { "*": allow } frontmatter, --auto effectively means unrestricted.

Pre-dispatch verification

Verify the adapter exists in the worktree:

test -f <worktree-path>/.opencode/agents/<role>.md || echo "STOP: adapter missing"

Do NOT verify with opencode agent list --pure — it does not show project agents
(display-level bug, verified on v1.18.18). Use opencode debug agent <role> from the
worktree cwd instead to confirm resolution.

OpenCode walks CWD upward to find .opencode/agents/, so the pane's cwd must be the
worktree root or a child of it. The mandatory cwd gate in the shared protocol covers this.

Launch

herdr agent start "<role>-<ticket-id>" --kind opencode --pane <pane-id> --timeout 60000 \
  -- --agent <role> -m <row: Model> --auto

Measured runtime facts

Re-measured on herdr 0.8.0 and opencode 1.18.11:

1. herdr agent start --kind opencode works on 0.8.0 — exit 0, agent detected, process
   confirmed alive with pgrep rather than by trusting interactive_ready. agent start
   is the launch for all three runtimes.
2. Effort is unreachable. opencode run requires a positional message, so it cannot
   start a message-less persistent executor; the TUI form is the persistent one and has no
   --variant. The --variant form is invisible to herdr. Effort and orchestration
   visibility are mutually exclusive on opencode, and visibility wins — which is why opencode
   Effort stays blank.
3. opencode's idle is FABRICATED. herdr agent explain reports
   fallback_reason: default_known_agent_idle_fallback, and agent wait --until idle
   returned rc=0 in 8 ms on a pane nobody had touched. Its manifest carries 3 rules
   (claude 12, codex 7), covering only blocked and working. So on opencode working
   and blocked are OBSERVED while idle is ASSUMED: the watcher's start guard still works,
   terminal-state detection does not, and a verdict must come from an artifact (AST-032).
4. Runtime detection quality is a ladder. Codex's top rules are osc_title (1100, 1050)
   — the agent's own title, which beats scraping. Claude tops out at osc_title 1100 then
   falls to text regions. opencode establishes idle not at all. Verify by artifact on every
   runtime; on opencode it is the only thing that works.
5. Transcript reads return only the input box and footer. Reading an opencode TUI
   transcript via herdr agent read returns only the input box and footer. Tolerable under
   verify-by-artifact — review diffs, not pane narration.
6. Model requires provider prefix. -m deepseek-v4-flash fails with
   ProviderModelNotFoundError. The correct form is -m opencode-go/deepseek-v4-flash.
   The orchestrator.md Model column for opencode rows must carry the full
   <provider>/<model> string. Measured on opencode 1.18.18.

The two-source background check degrades here, and you must say so

dispatch-ticket requires BOTH sources before calling a pane finished: the process check and
the status line the transcript carries. On opencode the transcript read returns the input box
and the footer and nothing else, so the second source produces nothing — and the gate silently
becomes pgrep alone, which is the source that answered wrong in the field (AST-097).

Treat a single-source answer on opencode as a degradation and report it, the same way a
runtime fallback is reported. Where the decision that follows is a worktree removal, get the
second source another way — ask the pane directly and wait for a reply — or hand back rather
than remove.