Skip to content
Agentmart
Submit a tool

jules-orchestrator-discipline

Six hard-won orchestrator rules for dispatching, watching, and harvesting Google Jules sessions — derived from the 2026-05-17 silent-fail incidents.

Skill by netzkontrast

IndexedMIT
Write a review

Chalk Labs

Need help deploying jules-orchestrator-discipline?

Chalk Labs sets up agents like this inside real teams: connected to your tools, tested on your data, with a human in the loop where it matters.

Book a call

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/netzkontrast-jules-orchestrator-discipline

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

Six hard-won orchestrator rules for dispatching, watching, and harvesting Google Jules sessions — derived from the 2026-05-17 silent-fail incidents. Use when the user says "dispatch Jules", "Jules session", "fan out specs", "orchestrate Jules", "Jules COMPLETED", "verify branch on remote", or "Codex bot review", or whenever a fresh Claude Code session begins coordinating any Jules-side work for `the-agency-system`.

Use cases: research, 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/jules-orchestrator-discipline

What's inside

Skill name
jules-orchestrator-discipline
Description
Six hard-won orchestrator rules for dispatching, watching, and harvesting Google Jules sessions — derived from the 2026-05-17 silent-fail incidents. Use when the user says "dispatch Jules", "Jules session", "fan out specs", "orchestrate Jules", "Jules COMPLETED", "verify branch on remote", or "Codex bot review", or whenever a fresh Claude Code session begins coordinating any Jules-side work for `the-agency-system`.
License
MIT
Instructions
123 lines in SKILL.md · view on GitHub

Files (1)

  • SKILL.md

SKILL.md (first lines)

## The rule (one sentence)

When you orchestrate Jules, verify every COMPLETED state on the remote
*before* trusting it, and never burn a third session on a spec that has
silently failed twice.

## Why this exists

On 2026-05-17 a single orchestration session burned ~15 Jules slots
because the model trusted the `COMPLETED` lifecycle event without
checking whether a branch actually landed on `origin`. The Jules backend
has a real silent-fail mode (see `Plan/JULES_PROTOCOL.md §8`) — the
session reports COMPLETED, the sandbox `git status` is clean, but no
branch ever materialises on the remote. Re-dispatching a fresh session
for the same spec wastes quota; the work already exists on the original
session's API record and only needs extraction.

This skill encodes the six rules so any future session inherits them
without having to re-learn the lesson.

## The six rules

1. **Verify before trust.** After every Jules `COMPLETED`, confirm the
   branch exists on `origin` BEFORE acting on the state:
   ```bash
   git ls-remote origin | grep <sid-or-branch-fragment>
   ```
   If the branch is missing, the session silently failed. Recover via
   `tools/jules-patch-extract.py <sid>` (see the
   `silent-fail-recovery` skill). NEVER re-dispatch a fresh Jules
   session on the same spec for the same work.

2. **2 silent-fails → switch.** If two Jules sessions on the same spec
   end in silent-fail, do NOT dispatch a third. Spawn a local subagent
   (Agent tool) and have it apply the API-extracted patch instead. The
   third Jules slot will almost certainly fail the same way and the
   quota is more valuable elsewhere.

3. **Patches > 2 KB never `head`/`grep`/`cat`-ed.** Always go through
   `tools/jules-patch-extract.py` — it writes the patch to
   `/tmp/jules-patches/<sid>-out{i}.patch` and emits only
   `{bytes, files, first_files[]}` stats. See the
   `context-safe-patch-handling` skill for the rationale.

4. **Wrap every `mcp__github__pull_request_read` in a focused subagent.**
   The main thread must never page through PR bodies — they routinely
   exceed 20 KB and pollute context. Dispatch a subagent whose only job
   is to read the PR and return a structured summary. Spec 106
   documents the permanent fix; until it ships, wrap manually.

5. **Wait for Codex bot review before merging non-trivial PRs.** On
   2026-05-17 the Codex bot caught 7 P1s that the spec-writing subagent
   missed. Non-trivial = touches >3 files OR changes any plugin
   surface (skills, MCP, hooks, manifests). Trivial doc/config tweaks
   may merge without waiting.

6. **`git rebase origin/Master` before publishing a PR.** Stale
   merge-base produces phantom-files in the diff — files the PR did
   not actually change appear as deletions. Rebase first, then
   `git push --force-with-lease`. The default branch is `Master`
   (capitalised); PRs targeting `main` are rejected with HTTP 422.

## Red flags

| Symptom | What it means | Action |
|---|---|---|
| `jules_get` returns `COMPLETED` but `git ls-remote` shows no branch | Silent-fail (Rule 1) | Run `tools/jules-patch-extract.py <sid>` |
| Two sessions on the same spec both silent-failed | Spec or backend is hostile (Rule 2) | Stop. Dispatch a local subagent. |
| About to `cat` a `.patch` file from `/tmp/jules-patches/` | Context poisoning risk (Rule 3) | Stop. Use the extractor's stats output instead. |
| Main thread is calling `mcp__github__pull_request_read` directly | Context bloat (Rule 4) | Wrap the call in a subagent via Agent tool. |
| PR has no `@codex` review yet, planning to merge | Quality regression risk (Rule 5) | Wait for the review unless the diff is trivial. |
| PR diff shows files you didn't touch | Stale base (Rule 6) | `git rebase origin/Master` + `git push --force-with-lease` |
| Creating a PR with `base: main` | Wrong default branch | Use `base: Master`. |

## Decision flow

```
Jules session reports COMPLETED
        │
        ▼
git ls-remote origin | grep <sid>
        │
   ┌────┴────┐
   │         │
 branch    branch
 present   missing
   │         │
   │         ▼
   │   tools/jules-patch-extract.py <sid>
   │         │
   │         ▼
   │   Was this the 2nd silent-fail on this spec?
   │         │
   │     ┌───┴───┐
   │     │       │
   │    yes      no
   │     │       │
   │     ▼       ▼
   │  local   apply patch,
   │ subagent push, open PR
   │     │       │
   ▼     ▼       ▼
       rebase origin/Master
              │
              ▼
       wait for Codex review (if non-trivial)
              │
              ▼
              merge
```

## References

- `Plan/JULES_PROTOCOL.md §8` — canonical recovery flow for silent-fail
- `Plan/_session-state/2026-05-18-next-session-goal.md` — origin of these
  six rules (the lessons-learned section)
- `skills/agentic/silent-fail-recovery/SKILL.md` — the recovery runbook
- `skills/agentic/context-safe-patch-handling/SKILL.md` — extractor usage
- `tools/jules-patch-extract.py` — the extractor script itself

From netzkontrast/agency, MIT.

More from netzkontrast/agency

Signals

Stars
0
Forks
0
Open issues
0
Last commit (this folder)
Jun 8, 2026 (4 months ago)

From GitHub (fetched Oct 24, 2026).

Details

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

More from netzkontrast

Done for you

Rather not set it up yourself?

Chalk Labs scopes, builds and runs agents for your team: the task audit, the eval harness, the integrations and the weekly ops report. Fixed scope, priced before we start.

See how we build AI agents →Read the case studies

Chalk Labs is a two-founder AI and growth studio. We build AI agents, CRMs and MVPs for Web3 and AI companies, and have shipped 50+ product and token launches.

launches shipped
50+
launches shipped
media placements
500+
media placements
founders on every account
2
founders on every account

Press we've placed: Forbes · CoinDesk · Decrypt · Cointelegraph · The Block

Rahil Jain, Co-founder, Chalk Labs. Ex-Goldman Sachs engineer. You talk to the person who builds it.

Get jules-orchestrator-discipline running in your business

Tell us what you want to automate. A founder replies within one working day with a plan, not a pitch deck.

Goes to the Chalk Labs founders, not the newsletter. Privacy

Need help deploying jules-orchestrator-discipline?

CallGet help
Cite this page

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