nhahan / webgpt

WebGPT

Delegate bounded work to save Codex usage. User instructions about division of work, model,
monitoring, permissions, follow-ups and retention override defaults. Never silently take over
WebGPT's assignment; Codex may perform separately assigned work concurrently.

Keep Codex ↔ WebGPT coordination in English, preserving source quotations and the requested
deliverable language. Reply in the user’s language unless they request another; leave browser UI language unchanged.
Every conversation must use Chat, never Work, including setup tests, browser-only work and
open. Verify Chat before composing or selecting a connector; return to Chat if a shortcut opens Work.

Route

• webgpt open [path]: follow [open.md](references/open.md); skip delegation below.
• Worker delegation: read [workspace.md](references/workspace.md).
• Explicit browser-only work: use [browser-reply.md](references/browser-reply.md); skip Worker commands.

• Read [setup.md](references/setup.md) only for installation or missing capability, and
  [recovery.md](references/recovery.md) only for actionable exceptions or same-chat follow-ups.
  Client tunnel/connection errors require setup's origin recovery before dispatch.

For Worker tasks, register first and use the returned connection.name and connection.registrationId.
With connection.status: ready, skip configuration/setup reads, directory searches and Plugins
rediscovery. Missing/invalid metadata routes to recovery.md; keep the registered task, do not register again.

Delegate

Assign one bounded deliverable with its relevant tests per new chat. Investigate only enough to
specify the objective, context, constraints and success criteria; leave routine commands to WebGPT.
Avoid open-ended projects, arbitrary duration limits and separate chats for tiny actions. Assign
file ownership/worktrees for concurrent changes. No duplicate task documents or tracking files.

Prepare one concise English prompt before opening the tab: task, necessary context/constraints and
success criteria. Reference accessible project briefs by path instead of restating them.
For Worker tasks, add the named connector/task token and “Save the final outcome
through the Worker.” Tool schemas supply the completion protocol; do not repeat it. Never send the
controller key or connection URL. Retain task and owned chat/tab identifiers in context.

Use documented browser initialization; reuse known browser identity without scanning all surfaces.
For Worker tasks, select the existing connector in Chat by verified registration ID, or
use a verified Chat launch URL. Mentioning a connector in prose does not select it.
Import helpers once per runtime (again after reset/resume); allow 60 seconds per helper call:

var {dispatch, deleteAndClose} = await import('<skill>/scripts/browser.mjs');
var sent = await dispatch(tab, prompt, 'xh', {
  startup: {connectorName, authorizeConversation: true}
});

Modes: m|medium = Medium (중간; also Standard/표준), h|high = High,
xh|xhigh = Extra High (default), p|pro = Pro. Pass any of these aliases to dispatch.
Omit startup for browser-only work. The helper verifies mode, sends once and checks startup.
Choose Always allow / 항상 허용 for authorized WebGPT permission prompts whenever offered;
this is the helper default. Only an explicit narrower user request uses
permissionScope:'conversation'. Preserve unrelated permissions and runtime gates.

On submitted, reuse the returned URL and wait. Permission selections or not_observed do not
prove terminal readiness; do not poll startup. On submission_unconfirmed, inspect the existing
user turn without resending. Resolve other actionable states via recovery.md.
Emit only relevant evidence, not full AX trees/exports or unnecessary screenshots. If helper
imports are unavailable, use documented CUA; do not read helper source during normal operation.

Wait, review, close

Use client.mjs finish <ids> for quiet waiting, verified result text and collection in one call;
do not repeat status/read/collect. Do separately assigned Codex work after verified submission.
Without requested monitoring, inspect progress only for a due backup or explicit help/failure.
On backupDue, read recovery.md: one scoped chat check, then resume <id>, not another finish.
Waiting requires a live parent runtime; it cannot wake exited Codex.

Review once against the outcome and relevant diff/evidence. Hash integrity is not correctness.
Do not reread the returned artifact or rerun reported passing tests on the delivered version
unless evidence is missing/conflicting, failures or integration changes need checks, or the user
requests verification. Distinguish PASS/FAIL/NOT_RUN. A partial result
needs continuation; a cleanupError needs recovery. Do not duplicate WebGPT's investigation.

After preserving/reviewing results, call deleteAndClose(tab, cua, browserId, ownedChatUrl).
Permanently delete the exact owned task chat, verify deletion, close its tab and verify absence.
This cleanup is already authorized: do not ask again due to permanence, generic confirmation
notices or a new CLI session. Honor explicit retention; never delete open chats, uncollected work
or unrelated chats. Keep the connector, Worker and tunnel for reuse. Actual tool rejection or a
user-requested cleanup pause routes to recovery.md; tab closure alone never proves deletion.