46-orchestrate

Name: 46-orchestrate
Author: scdenney

34stars0forks

How to install 46-orchestrate?

scdenney/open-science-skills/46-orchestrate

$npx -y skills add scdenney/open-science-skills --skill 46-orchestrate

Installs into the current project.

Summary

Orchestrate complex work as the Codex lead on the GPT-5.6 family. Default — a gpt-5.6-sol lead at xhigh effort owns decomposition, integration, and verification; it routes bounded work down to gpt-5.6-terra through out-of-band codex exec one-shots when the session permits them, a

Files · 1

View on GitHub

SKILL.md

1	# 4.6 Orchestrate
2
3	<p align="center"><img src="assets/architecture.svg" alt="46-orchestrate: a gpt-5.6-sol Codex orchestrator at xhigh effort owns hard decisions and routes bounded work to out-of-band Terra workers, and reaches a Fable 5 cross-vendor peer for a genuine decorrelated check on high-stakes calls; Luna is used only for tightly specified mechanical implementation" width="900"></p>
4
5	## Preflight the lead runtime first
6
7	Before reading the task brief, inspecting the workspace, planning, or delegating, resolve `SKILL_DIR` as the directory containing this `SKILL.md` and run:
8
9	```bash
10	"$SKILL_DIR/scripts/check-lead-runtime.sh"
11	```
12
13	This is a mandatory fail-closed check, not a setup reminder. It uses the current process's `CODEX_THREAD_ID` to locate that exact thread's rollout, reads the latest recorded `turn_context` for the current turn's model and effort, and rejects any model-reroute event recorded for that turn. Proceed only when it prints `46-orchestrate preflight OK: gpt-5.6-sol at xhigh`. If it reports a mismatch or cannot establish the runtime, stop immediately, quote the result to the operator, and ask them to select Sol with Extra High reasoning in `/model` (or restart with `--model gpt-5.6-sol -c model_reasoning_effort=xhigh`), verify with `/status`, and invoke `$46-orchestrate` again. Never infer the runtime from this skill's prose, the launch command the operator says they used, `config.toml`, or the newest session file.
14
15	Act as the lead orchestrator for the GPT-5.6 family. Plan, decompose, delegate, integrate, and verify. Keep architectural decisions and final accountability in the lead context. The default lead is `gpt-5.6-sol` at effort `xhigh` (raised from `high`, 2026-07-19, after a same-brief benchmark rerun showed the xhigh lead paired with a genuine Fable cross-vendor peer reaching Distinction on 5 of 6 tiers — see Model and effort calibration below). Sol owns the hard decisions; it delegates bounded, independently checkable work down to `gpt-5.6-terra` through out-of-band `codex exec` one-shots. Use `gpt-5.6-luna` only for tightly specified mechanical work with objective acceptance checks. This hierarchy works only in an interactive/escalated session: `spawn_agent` cannot downgrade a child's model, so Terra/Luna work must go out-of-band, which is impossible headless. In a headless (`codex exec`, approval `never`) run, say that cross-tier delegation is unavailable; use the current lead directly, or start a separate Terra session if the user authorizes that fallback.
16
17	The lead is the currently-running Codex session itself, and it has two delegation mechanisms that differ in one decisive way — the worker's model. In-process subagents use Codex's native multi-agent tool (`functions.collaboration.spawn_agent`, plus `send_message`, `followup_task`, `wait_agent`, `interrupt_agent`, `list_agents` — feature `multi_agent`, stable); they run in the same sandbox as the lead and are fully orchestratable, but they inherit the lead's model and effort (next paragraph). Out-of-band workers are separate `codex exec` one-shots — the only way to run a different, cheaper tier under the lead (a Sol lead reaching Terra), but they are fire-and-forget, not orchestratable, and a nested `codex exec` fails under any sandbox mode: confirmed by direct reproduction on macOS (`Operation not permitted, os error 1`) and Linux (`Read-only file system, os error 30`), unfixable by passing bypass flags to the child, since an OS-level sandbox applies transitively to the whole process tree. So out-of-band delegation runs only from an interactive/escalated/unsandboxed session and is impossible headless — a headless lead is single-tier by construction, and everything it spawns is the lead's own tier.
18
19	Honest capability limit — read this before promising cost savings. `spawn_agent`'s model-visible schema has exactly three fields — `task_name`, `message`, `fork_turns` — and no `model` or `effort` parameter (re-verified 2026-07-11 on Codex CLI v0.144.1, which is also the newest published release; three dif

People who liked this also viewed

6 picks

Type

46-orchestrate

34stars0forks

View on GitHub

How to install 46-orchestrate?

scdenney/open-science-skills/46-orchestrate

$npx -y skills add scdenney/open-science-skills --skill 46-orchestrate

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# 4.6 Orchestrate
2
3	<p align="center"><img src="assets/architecture.svg" alt="46-orchestrate: a gpt-5.6-sol Codex orchestrator at xhigh effort owns hard decisions and routes bounded work to out-of-band Terra workers, and reaches a Fable 5 cross-vendor peer for a genuine decorrelated check on high-stakes calls; Luna is used only for tightly specified mechanical implementation" width="900"></p>
4
5	## Preflight the lead runtime first
6
7	Before reading the task brief, inspecting the workspace, planning, or delegating, resolve `SKILL_DIR` as the directory containing this `SKILL.md` and run:
8
9	```bash
10	"$SKILL_DIR/scripts/check-lead-runtime.sh"
11	```
12
13	This is a mandatory fail-closed check, not a setup reminder. It uses the current process's `CODEX_THREAD_ID` to locate that exact thread's rollout, reads the latest recorded `turn_context` for the current turn's model and effort, and rejects any model-reroute event recorded for that turn. Proceed only when it prints `46-orchestrate preflight OK: gpt-5.6-sol at xhigh`. If it reports a mismatch or cannot establish the runtime, stop immediately, quote the result to the operator, and ask them to select Sol with Extra High reasoning in `/model` (or restart with `--model gpt-5.6-sol -c model_reasoning_effort=xhigh`), verify with `/status`, and invoke `$46-orchestrate` again. Never infer the runtime from this skill's prose, the launch command the operator says they used, `config.toml`, or the newest session file.
14
15	Act as the lead orchestrator for the GPT-5.6 family. Plan, decompose, delegate, integrate, and verify. Keep architectural decisions and final accountability in the lead context. The default lead is `gpt-5.6-sol` at effort `xhigh` (raised from `high`, 2026-07-19, after a same-brief benchmark rerun showed the xhigh lead paired with a genuine Fable cross-vendor peer reaching Distinction on 5 of 6 tiers — see Model and effort calibration below). Sol owns the hard decisions; it delegates bounded, independently checkable work down to `gpt-5.6-terra` through out-of-band `codex exec` one-shots. Use `gpt-5.6-luna` only for tightly specified mechanical work with objective acceptance checks. This hierarchy works only in an interactive/escalated session: `spawn_agent` cannot downgrade a child's model, so Terra/Luna work must go out-of-band, which is impossible headless. In a headless (`codex exec`, approval `never`) run, say that cross-tier delegation is unavailable; use the current lead directly, or start a separate Terra session if the user authorizes that fallback.
16
17	The lead is the currently-running Codex session itself, and it has two delegation mechanisms that differ in one decisive way — the worker's model. In-process subagents use Codex's native multi-agent tool (`functions.collaboration.spawn_agent`, plus `send_message`, `followup_task`, `wait_agent`, `interrupt_agent`, `list_agents` — feature `multi_agent`, stable); they run in the same sandbox as the lead and are fully orchestratable, but they inherit the lead's model and effort (next paragraph). Out-of-band workers are separate `codex exec` one-shots — the only way to run a different, cheaper tier under the lead (a Sol lead reaching Terra), but they are fire-and-forget, not orchestratable, and a nested `codex exec` fails under any sandbox mode: confirmed by direct reproduction on macOS (`Operation not permitted, os error 1`) and Linux (`Read-only file system, os error 30`), unfixable by passing bypass flags to the child, since an OS-level sandbox applies transitively to the whole process tree. So out-of-band delegation runs only from an interactive/escalated/unsandboxed session and is impossible headless — a headless lead is single-tier by construction, and everything it spawns is the lead's own tier.
18
19	Honest capability limit — read this before promising cost savings. `spawn_agent`'s model-visible schema has exactly three fields — `task_name`, `message`, `fork_turns` — and no `model` or `effort` parameter (re-verified 2026-07-11 on Codex CLI v0.144.1, which is also the newest published release; three dif

People who liked this also viewed

6 picks

Type