George930502 Prompt Review and Dispatch
Orchestrate $prompt-master-gpt5 and $grill-with-docs. Default to
current-conversation execution. After approval, continue the exact prompt in
the current task so the discussion and execution remain visible in one
conversation; this mode does not require a foreground message-injection tool.
Use
background-task only when the user explicitly asks for another Codex task or
thread. Read
references/approval-protocol.md before the
first user decision; it is the single source for state, approval, and
execution.
Every user decision in every stage follows
../grilling/NATIVE-INPUT.md.
1. Intake
Capture the rough prompt and GPT-5.6 target surface. Set
execution_mode: current-conversation unless the user explicitly requests
another Codex task or thread.
For current-conversation, label the destination as the current conversation;
do not call list_threads, read_thread, or wait_threads, and do not invent a
destination ID.
For background-task, resolve the destination with the host's thread-list/read
capabilities and show its title and ID before approval.
Completion criterion: the execution mode and every working-state field required by the approval protocol exist; background mode additionally requires one verified destination task.
2. Audit
Invoke $prompt-master-gpt5 in audit-only mode. Consume its gate as
authoritative; this orchestrator does not redefine the audit dimensions.
Completion criterion: the audit returns exactly one gate:
ready-for-polish or needs-purpose-clarification.
3. Align
For an incomplete prompt, invoke $grill-with-docs. Resolve facts from files
and docs; ask only the highest-impact unresolved decision.
Re-run audit-only after each answer. Repeat until the audit is ready and the
user passes the native alignment gate. A complete user-supplied contract may
skip grilling, but still requires that gate.
Completion criterion: the audit is ready and alignment is explicitly selected.
4. Polish
Invoke $prompt-master-gpt5 in polish mode with the aligned brief and audit.
Preserve settled decisions.
Completion criterion: one final copyable prompt passes the prompt-master audit.
5. Approve and execute
Show the exact prompt and selected execution mode, then execute the approval protocol. Rejection returns to alignment with the selected reason as new evidence; a new draft requires a new approval.
After approval in current-conversation mode, treat the exact draft as the
next instruction for this same running Codex task and execute it here. Keep
progress in the current conversation; do not call list_threads, read_thread,
wait_threads, send_message_to_thread, create_thread, fork_thread, or
handoff_thread. Use the background dispatch section of the approval protocol
only for an explicitly requested
background-task. Current-conversation execution is not dispatch and does not
weaken the verified-dispatch gate: background dispatch still requires its
verified send evidence. Inline completion requires verified evidence of the
same-task continuation and the approved prompt's success criteria.
Completion criterion: current-conversation mode has verified inline-continuation
evidence showing executed_draft_sha256 equals draft_sha256, the approved
prompt's success criteria have actual result/artifact/test evidence, and
execution is complete;
or the explicitly requested background dispatch reaches complete with
verified send evidence and exact-byte hash equality.