# External review-seat channel

The loaded `herdr` skill owns the command surface and lifecycle. The invocations below state the shape this channel uses; when the loaded skill's CLI differs, follow the loaded skill and report the drift rather than following this file.

## When the channel exists

- Require all three legs: this session can load the `herdr` skill; that skill's own `HERDR_ENV=1` probe passes; the detected mstar host maps to a supported agent kind below. Load the skill before running its probe or commands; use only its own guidance for command discovery and lifecycle operations.
- If any leg fails, state which leg failed and use the ordinary in-session review path unchanged. A missing skill is not an error; never discover or run Herdr commands outside the loaded skill's guidance.

## Kind mapping

This table maps the host detected by `mstar-host` to an agent kind; it is not an inventory of available kinds. Confirm kind availability through the loaded skill's agent surface; if unavailable, state the failed kind leg and use the ordinary path. Never guess a kind.

| Detected host | Mapped kind (subject to the loaded skill's agent surface) |
| --- | --- |
| `omp` | `omp` |
| `opencode` | `opencode` |
| `codex` | `codex` |
| `cursor` | `cursor` |
| `kimi` | `kimi` |
| `dsh`, `zcode`, unmapped | None — report "this host has no external review-seat channel" and use the ordinary path. |

## Open the channel

- Follow the loaded skill's layout rules: inspect the caller pane, honor a requested direction, otherwise split right for a wide caller or down for a narrow or tall caller. Default: `herdr pane split --current --direction right --cwd "$PWD" --no-focus`.
- Read the new pane id from the JSON response's `.result.pane.pane_id`; never predict it. Read that pane to confirm the shell is at its interactive prompt, or wait until it is, before starting one agent with a useful unique job name, e.g. `herdr agent start pr-reviewer --kind <mapped-kind> --pane <returned-pane-id>`. An `agent_pane_busy` refusal while the shell is becoming ready is not a channel failure: read the pane, wait for its interactive prompt and retry start once. Wait for agent readiness before prompting.
- Keep one channel session per PR and reuse it across rounds. Never create a workspace, tab, worktree or different cwd unless the user requested that topology. The review session's own review-worktree creation remains part of the ordinary review contract.

## Dispatch a review round

- Immediately before prompting the channel for each round, record the newest archived report path as that round's baseline; if none exists, record its absence.
- Send exactly one command entry as one prompt: `/amazing-pr-review pr:<n> [tier]`; append the tier only when the user named one. For example, `herdr agent prompt <agent-name> "/amazing-pr-review pr:<n>" --wait --timeout <timeout-ms>`; an explicit user tier is the command's trailing argument.
- Derive the wait timeout from the resolved tier's wall-clock budget printed by `mstar pr-review budget` (`PR_REVIEW_TIER_BUDGETS`), plus margin, converted to the channel's timeout unit. With no user tier, use the review's own inference ladder to size the wait, not a loop-specific default; do not add that inferred tier to the prompt. Never hand-write a minute count.
- The seat session is an ordinary independent host session: it resolves mstar skills, creates the review worktree, runs seats, synthesizes, posts and archives exactly per `commands/amazing-pr-review.md`. This channel adds no review rule.

## Read the result

- Wait for the channel agent to settle first; `blocked` on an approval/question UI is the recovery case below, not a review verdict. Use `mstar pr-review report-path` to resolve the review directory, then locate this round's **archived report file** there, asking the settled agent for its actual path if needed; the resolver's next available filename is not an existing result. Apply the same round-binding predicate under § Recovery / refusals (**Agent exits or pane is lost mid-round**), including its pre-dispatch baseline, before accepting verdict, tally, `comments:` state or review URL. If binding fails, follow that recovery rule to re-dispatch the round; never accept an older report as this round's result.
- Read pane output only as progress evidence, using the loaded skill's agent read surface. Expect alternate-screen truncation: if output is cut, ask the settled agent to restate the archived report path and read that file directly. Never re-run a review to recover text.

## Round loop obligations

- Per round: read the archived report → for `needs fixes` / `blocked`, apply the loop's existing dispatchability and decision boundaries → dispatch bounded fix Assignments, never implement fixes in the coordinator → implementers commit locally at the existing fix sites → wait for the push gate → push one batch → prompt the same channel agent again after the new head's waves, per the ordinary loop.
- A later round needs no extra prompt wording: the seat session's own § Re-review finds the prior report for the same target. Preserve prior reports and never rewrite history.

## Supervision boundary

- Do not nudge the pane session's seats; they belong to that session, which supervises them itself. Coordinator supervision is only waiting on the channel session and, on overrun, re-reading the report path and declaring elapsed time.
- A timeout does not prove non-delivery. Inspect the channel state and report path through the loaded skill's read surfaces, then continue waiting as appropriate; never re-send the round prompt merely because a wait timed out.

## Recovery / refusals

- **Skill absent:** name the missing skill leg; continue through the ordinary in-session path. Do not require installation to proceed.
- **Probe fails:** name the failed environment-probe leg; continue through the ordinary path without issuing channel commands.
- **No mapped kind:** report "this host has no external review-seat channel"; continue through the ordinary path, never substitute a kind.
- **Agent start fails or reports not-ready:** read the pane, repair its state using the loaded skill's guidance, then retry start once in an available shell. If a not-ready agent already occupies the pane, inspect it and wait for idle rather than starting over it; an approval/question UI takes the next decision boundary below. If startup still fails, report the actual cause and required pane-state repair before retrying this entry. If the pane state cannot be repaired, state why and use the ordinary in-session path.
- **Agent exits or pane is lost mid-round:** use the round baseline recorded under § Dispatch a review round. On loss, read the newest archived report and resume from it only when it is this round's output — its `head:` equals the round's resolved head, its verdict is not `incomplete` (a settled run per § Verdict synthesis), and it is the baseline's successor: a higher `report-rN.md` sequence in the same run directory, or a report whose `supersedes:` names the baseline. A baseline that never existed means any report carrying the resolved head and a non-`incomplete` verdict qualifies. Otherwise this round still owes its work: open a fresh channel session under the same layout/cwd rules and re-dispatch the same round. A rewritten history never resumes from a prior report. Loss, unlike a wait timeout, requires a replacement session.
- **Archived report with `comments: failed`:** keep the local report and its error as the recorded outcome under § Local report archive; that is not a review success. Do not advance the verdict transition from it — repair the posting cause and re-dispatch the round.
- **Approval or question dialog:** read the UI and stop for the user's decision; state the actual permission/product question and ask the user to decide it before resuming the channel. Never answer it as a review verdict or treat channel `blocked` as report `blocked`.

## Cleanup

- When the loop ends (`ship it`, user budget or `Blocked`), leave the pane and transcript in place as evidence. Never close a pane you did not create or move the user's focus.
- Delete the review worktree only through the existing review cleanup contract; channel cleanup adds no worktree deletion path.