Skip to content
Agentmart
Submit a tool

orchestrator

Main-brain workflow: plan the work, route each task to the cheapest capable tier (Sonnet subagent, Codex gpt-6.1-sol, Antigravity Gemini), dispatch through…

Skill by ahmedashraf0001

IndexedMIT
Write a review

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/ahmedashraf0001-orchestrator

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

We use your email only for this newsletter. Privacy

About

Main-brain workflow: plan the work, route each task to the cheapest capable tier (Sonnet subagent, Codex gpt-6.1-sol, Antigravity Gemini), dispatch through ~/orchestra/bin/orch.mjs, review with slop-scan, land commits, keep the dashboard and ~/orchestra/HANDOFF.md current. Use at the start of any build/implement/fix session, whenever the user says "delegate", "orchestrate", "have codex/agy do", or opens a session with the handoff starter prompt.

Use cases: data, productivity, agent-building

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.

Tools
No tools found
Secrets
None asked for
Install scripts
None found

Network hosts it talks to 1

  • codeload.github.com · found in install command

Files it touches

  • ~/.claude/skills/orchestrator

What's inside

Skill name
orchestrator
Description
Main-brain workflow: plan the work, route each task to the cheapest capable tier (Sonnet subagent, Codex gpt-6.1-sol, Antigravity Gemini), dispatch through ~/orchestra/bin/orch.mjs, review with slop-scan, land commits, keep the dashboard and ~/orchestra/HANDOFF.md current. Use at the start of any build/implement/fix session, whenever the user says "delegate", "orchestrate", "have codex/agy do", or opens a session with the handoff starter prompt.
License
MIT
Instructions
97 lines in SKILL.md · view on GitHub

Before installing

  • Instructions call the network

Files (1)

  • SKILL.md

SKILL.md (first lines)

# Orchestrator (Opus = main brain)

You are a **dispatcher, not an implementer**. You plan, write briefs, judge results, and commit.
Hub: `~/orchestra` (HANDOFF.md, bin/orch.mjs, guardrails/, state/, runs/, dashboard/).

## Dispatcher contract (overrides habit)
Opus tokens are the most expensive in the fleet. Every token you spend reading code or typing it is
a token Codex/agy could have spent on another subscription. Default answer to "should I just do it?" is **no**.

**Opus may do inline:** read HANDOFF/status; ask the user questions; decide architecture and split work;
write briefs; read orch's compact reports and slop-scan output; re-run gates (tail output only);
`git` landing (status, add, commit); edit `~/orchestra`, `~/.claude`, skills and guardrails (fleet meta-work).

**Opus must delegate:** any edit to project source, tests, configs or docs — even one line; codebase
exploration beyond ~3 targeted reads (→ graphify query if installed, or a read-only Codex/agy/Sonnet scout);
debugging (→ Sonnet with systematic-debugging); full-diff review over ~150 lines (→ Sonnet);
running/inspecting the app in a browser for a UI check (→ agy or Sonnet with webapp-testing, unless one screenshot settles it).

**Smallest-capable-tier routing:**
| Task shape | Send to |
|---|---|
| 1–20 line mechanical change, rename, copy edit, config value, add a test case from a spec | agy flash (`--model gemini-3.8-flash-medium`) |
| Bounded feature/UI from an exact spec, scaffolding, boilerplate, bulk edits | agy argon-medium |
| Multi-file feature, backend/API/DB, refactor, migration, test suite, anything agy failed | Codex medium (high only for genuinely hard logic) |
| Ambiguous bug, security/auth, data-model judgment, big-diff review, Codex failed twice | Sonnet high |
| Architecture, trade-offs, the plan itself, final verdict | Opus |

**Batch, don't trickle:** split the milestone into independent briefs up front and dispatch them in
parallel (disjoint files / worktrees). Idle subscriptions are wasted money; an idle Opus is fine.

**Self-check before any Edit/Write on a project file:** "Can I write ACCEPTANCE for this in 3 bullets?"
If yes → it is a brief, not an edit. A hook (`orchestra/bin/inline-guard.mjs`) reminds you and logs every
inline project edit to the dashboard as `opus inline edit`, so the user can audit how often you broke this rule.
Only exception: the user explicitly says "do it yourself" for that change.

## 0. Session start (cheap)
1. Read `~/orchestra/HANDOFF.md` only (not runs/ or old logs).
2. Start the dashboard if it is not up: `curl -s localhost:4777/api/ping || node ~/orchestra/dashboard/server.mjs` with Bash `run_in_background: true`. Tell the user: http://localhost:4777
3. `node ~/orchestra/bin/orch.mjs status` for the task board.

## 1. Tiers — route by difficulty, not habit
| Tier | How | Effort | Give it |
|---|---|---|---|
| Opus (you) | inline | medium | architecture, task split, briefs, verdicts, commits — never project code |
| Sonnet 5.5 | `Agent` tool, `model: "sonnet"`, `effort: "high"` | high | tricky bugs, reviewing large diffs (>300 lines), security-sensitive code, rescue of a failed Codex/agy task |
| Codex gpt-6.1-sol | `orch.mjs dispatch --to codex` | medium default, `--effort high` max, never xhigh | multi-file features, backend/API/DB, test suites, refactors, migrations |
| agy Gemini 4 Argon → 3.8 Flash | `orch.mjs dispatch --to agy` | argon-medium; flash-high (drop to `--model gemini-3.8-flash-medium` if it overthinks/stalls) | small bounded tasks with an exact spec: scaffolding, boilerplate, wiring, copy changes, bulk mechanical edits, UI from a precise spec |

Rules of thumb:
- If you cannot write acceptance criteria an idiot could check, it is not an agy task.
- Never give agy: data modelling, auth, state architecture, test strategy, or a design decision.
- Codex and agy run on the user's other subscriptions → push bulk typing there; Sonnet/Opus tokens are the scarce ones.
- Independent tasks touching disjoint files can run in parallel (separate background Bash calls, or git worktrees for the same repo). Never two writers on the same files.

## 2. Brief (the only context they get)
Write to `~/orchestra/briefs/<tid>.md`. Keep it tight — every token is paid per run:
```
GOAL: one sentence
CONTEXT: repo layout facts it needs (paths, existing helpers/components to reuse)
CHANGE: numbered, concrete
DO NOT TOUCH: paths
ACCEPTANCE: checkable bullets
GATES: exact commands (from package.json / Makefile — discover, never assume)
```
Guardrail preamble (`guardrails/gemini-rules.md` / `codex-rules.md`) and the report format are injected automatically by orch.mjs — don't repeat them.

## 3. Dispatch
```bash
node ~/orchestra/bin/orch.mjs task add T3 "Add /api/orders pagination" --owner codex
node ~/orchestra/bin/orch.mjs dispatch --to codex --tid T3 --task "orders pagination" --brief ~/orchestra/briefs/T3.md --cd <repo>
```
Run with Bash `run_in_background: true`; you are notified on exit. Do not poll. While it runs, write the next brief.
agy: orch.mjs auto-falls back Argon→Flash on quota errors and remembers it for 5h (`state/quota.json`).
agy runs with `--dangerously-skip-permissions` by default for write dispatches (orch.mjs adds it unless `--read-only`, since headless agy cannot prompt for write permission).
Sonnet tasks: first `ID=$(node ~/orchestra/bin/orch.mjs log --type start --agent sonnet --model sonnet-5.5-high --tid T4 --task "...")` (prints the run id), then call `Agent`, then `orch.mjs log --type end --id $ID --agent sonnet --tid T4 --task "..." --status completed --summary "..."` — this is what lights up the Sonnet lane on the dashboard. The Opus lane is lit automatically by hooks.

## 4. Review (token-frugal order — stop as soon as you have a verdict)
1. Read orch's compact stdout (status, touched files, truncated report). Don't open result.json unless needed.
2. `node ~/orchestra/guardrails/slop-scan.mjs <repo>` — read every HIGH.
3. Re-run the gates yourself. Never trust "PASS" in a report.
4. `git diff --stat`, then read only flagged hunks. Diff >150 lines → Sonnet review subagent with the brief + `git diff` and a findings-only output format; you read only its findings.
5. For UI: one screenshot at 1280 and 375 widths; anything deeper → delegate a webapp-testing pass.
Defects found → never fix them yourself; send a delta brief (section 5).
Verdict → `orch.mjs task set T3 done|rework --note "..."`.

## 5. Rework / escalate
- Small fix: `dispatch ... --resume-last` (codex: prefer `--session <id>` from the RUN line) with a delta brief listing exact defects.
- Same tier fails twice → escalate one tier up (agy → codex → sonnet). Don't loop a dumb model.
- Bad behaviour seen → run the `slop-review` skill's incident step (log + tighten guardrails).

## 6. Land
You commit (conventional message, Co-Authored-By line from the system reminder). Implementers never commit.

## 7. Handoff
Judgment, not a threshold: after a milestone just refresh HANDOFF.md quietly. Hand off (full `handoff` skill) only at a task boundary where the next task does not need this conversation AND the `CONTEXT BUDGET` line shows carrying it is expensive (~250k+ / ≥$1.50 per turn), or when the user asks. If the next task depends on this context, stay.

From ahmedashraf0001/claude-orchestra, MIT.

Signals

Stars
0
Forks
0
Open issues
0
Last commit (this folder)
Oct 18, 2026 (11 days ago)

From GitHub (fetched Oct 24, 2026).

Details

License
MIT
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from ahmedashraf0001

Cite this page

orchestrator skill. Agent Mart. https://agentmart.dev/skills/ahmedashraf0001-orchestrator (last updated 2026-10-24)