instructa/agent-skills/delegate-sol

Delegate to Sol

Delegate through Herdr with one explicit mode:

• review: Keep Sol and all Terra gatherers read-only. Require Sol to launch two or three distinct

gpt-5.6-terra subagents at medium reasoning, collect their evidence, and own the final review.

• implement: Make Sol the sole writer. Stop editing until Sol finishes, then inspect and verify its work.

Default to review when authorization is ambiguous. Never infer permission to edit from phrases such as "ask Sol", "take a look", "analyze", or "check this". Start Sol with gpt-5.6-sol and xhigh reasoning unless the user explicitly requests another Sol reasoning level.

Prepare the handoff

Create a self-contained task containing:

• the objective and relevant project context;
• the exact scope and files when known;
• current evidence, errors, or competing hypotheses;
• constraints and explicit non-goals;
• required verification and return format.

Do not forward the entire conversation. Preserve only information needed for the delegated task.

Before implement, inspect the workspace state and preserve unrelated or pre-existing changes. Tell Sol not to commit, check in, shelve, push, switch branches, or alter workspace configuration unless the user explicitly requested it.

Run the delegation

Resolve this skill's directory and invoke its script. Pass the task with --task, --task-file, or stdin.

bash <skill-dir>/scripts/delegate-sol.sh review \
  --target-title "Sol Reviewer" \
  --task "Review the current change for correctness, regressions, and missing tests."

bash <skill-dir>/scripts/delegate-sol.sh implement \
  --task "Implement the agreed fix within the stated scope and run the relevant verification."

The script must:

1. Honor --target first, then --target-title. If a requested target is missing, is the current

caller pane, or is busy, stop instead of falling back.

1. Without an explicit target, always create a fresh Sol pane. Reuse only a user-designated

--target or the single exact, settled --target-title match when that pane was created by this skill. Never reuse the calling Codex pane or an arbitrary Codex pane.

1. Start new panes with gpt-5.6-sol and xhigh reasoning, submit the task, wait for settlement,

and return visible output. For a fresh pane, place the full task in a mode-600 temporary packet outside the project and pass only a short packet-reading instruction as Codex's initial CLI prompt. Keep the packet until Sol settles, then delete it through an exit trap.

1. In review, require Sol to spawn two or three parallel, read-only context gatherers with

model: "gpt-5.6-terra", reasoning_effort: "medium", and fork_turns: "none". Give each a distinct evidence-gathering scope, wait for all of them, and synthesize their results. Terra gatherers must not decide approval or edit files.

1. Correlate the captured terminal tail to Sol's final response with a unique per-invocation handoff

ID. Verify that it reports two or three distinct structured Terra records, each with the required model, effort, fork setting, distinct scope, and evidence disposition.

1. Leave the Sol pane open for follow-up work.

Do not manually duplicate these steps unless the script reports a concrete failure.

Default to one complete Sol turn per hypothesis. If the response lacks the required handoff sections, report an incomplete handoff; do not automatically send an open-ended continuation. On timeout in a fresh pane owned by this invocation, send Escape, collect its last output, and stop. For an explicitly reused pane, leave it untouched on timeout to avoid cancelling unrelated work.

Handle the result

For review, report Sol's findings and Terra-gathered context alongside your own assessment. Do not implement findings unless the user separately authorizes changes.

For implement:

1. Resume ownership only after Sol settles.
2. Inspect the actual changed files and version-control state.
3. Run proportionate verification independently.
4. Report what Sol changed, what you verified, and any remaining risk.

If Sol is blocked, show the blocking output. Never auto-approve permissions, destructive actions, commits, check-ins, shelvesets, pushes, or branch changes.

If no Herdr environment is available, stop and explain that the skill must run from an agent pane inside Herdr.