zcode-delegate

ZCode Delegate

You are the orchestrator. This skill lets you hand a bounded coding task to a separate implementer — the Z.AI ZCode CLI — then review what it produced and land it yourself. You write the brief and own the judgment; ZCode does the typing; you verify and commit.

Nothing here is specific to one orchestrating agent. The loop needs only the ability to run a shell command and read a file. (It is designed for and run on Claude Code; treat other orchestrators as designed-for, not yet proven.)

When NOT to use this

• The task is small enough to just do inline — delegation overhead is not worth it.
• ZCode is not installed, or its CLI has no model provider configured.
• You want to write the code yourself, or you only need a review.

Prerequisites (check once)

1. ZCode is installed. The CLI ships inside the desktop app — it is not on PATH and not on npm. The relay resolves it in this order: --zcode-path <file> or ZCODE_CLI first, then PATH, then the installed app bundle. On Linux the app is an AppImage with no fixed install path, so the flag or the environment variable is required there — the relay guesses nothing.
2. A model provider is configured for the CLI, with a key it can actually reach. Being signed into the desktop app is not enough — see below.
3. You are in (or will point --cd at) the target git repository.

The relay records the CLI version and how it was resolved into result.json, so a surprising install is visible after the fact.

Authenticating the headless CLI

Signing into the ZCode desktop app does not authenticate the CLI this relay drives. The CLI keeps its own config at ~/.zcode/cli/config.json, separate from the desktop app's, and nothing bridges the two. zcode login is the intended path, but where it fails with OAuth response is not valid JSON the way in is a Z.AI API key.

Two pieces are needed, and they are separate:

1. The provider block must exist in ~/.zcode/cli/config.json. It defines the provider, its endpoint and its models — the environment cannot supply this:

   {
     "provider": {
       "zai": {
         "kind": "anthropic",
         "options": { "apiKeyRequired": true, "baseURL": "https://api.z.ai/api/anthropic" },
         "models": { "glm-5.1": { "name": "GLM-5.1" } }
       }
     },
     "model": { "main": "zai/glm-5.1" }
   }
   

2. The key can live either in provider.zai.options.apiKey in that file, or in the environment as any one of ZAI_API_KEY, ZCODE_API_KEY, or ANTHROPIC_API_KEY. Prefer the environment — it keeps the secret off disk.

If a run fails with Model provider is missing an API key: <provider>, the provider block resolved but no key was found: set one of those variables and re-run.

Autonomy — read this before dispatching

ZCode's own term is mode. It has four values; only two are usable headlessly.

mode	Behaviour
yolo	Writes. ZCode's own default for --prompt, and this relay's write-capable default.
plan	Refuses edits. What --read-only selects.
build	Rejected by this relay. No permission client exists headlessly, so tools are blocked and the run exits 0 having done nothing.
edit	Rejected for the same reason.

Two limits stated plainly, because ZCode cannot enforce them:

• plan mode refused edits in testing, but the relay does not treat that as a guarantee. It takes a Git fingerprint before the run and reports a tri-state readOnlyViolation afterwards. Confirm touchedFiles came back empty rather than assuming no edits.
• ZCode has no --allowed-tools. Only the --disallowed-tools denylist exists, and it is genuinely enforced. An explicit allowlisted tool surface is therefore impossible here — do not assume one.

The loop

Run these five steps per task. Steps 1, 4, and 5 are your judgment; 2 and 3 are mechanical.

1. Write the brief

ZCode sees only what you send — no repo memory, no chat history. Everything the task needs goes in the brief: the goal, the current state, what to change, what to leave untouched, the project's actual gate commands (discover them from the repo's CLAUDE.md/AGENTS.md/Makefile — do not assume), and a report contract. Tell ZCode it will not commit. One task per brief. The relay delivers the brief as an attached file, so the command line no longer bounds its length — the model's context window still does. Full guidance and a template: references/writing-the-brief.md.

2. Dispatch

node "<skill-dir>/scripts/relay.mjs" --brief brief.txt --cd /path/to/repo
# read-only (review/diagnosis, no edits):   add --read-only
# continue a specific session:              add --session <sess_...>  (from result.json; send only the delta brief)
# continue the latest session for --cd:     add --resume-last
# withhold tools (denylist):                add --disallowed-tools "Write,Edit,Bash"
# point at the CLI explicitly:              add --zcode-path /path/to/zcode.cjs
# hard time limit (watchdog):               add --timeout 2h  (default: off)
# see all options:                          node .../relay.mjs --help

(<skill-dir> is this skill's installed directory — the folder containing this SKILL.md.)

The relay writes its artifacts to a temp dir, so the repo under review stays clean. It never commits — see step 5. Mechanics, flags, and the result.json shape: references/dispatch-and-poll.md.

3. Wait for completion

The relay blocks until ZCode finishes, so back it with whatever your orchestrator offers:

• Claude Code: run the Bash call with run_in_background: true; you are notified on completion.
• Plain shell / other agents: foreground for short tasks, or background it and poll the result file. The run is done when result.json exists with a status. A pre-run usage error exits 2 and writes no result file, so check the exit code too; a CLI that cannot be found exits 127 but does write a result.json with status zcode_unavailable.

Do not trust progress trackers over reality: read the working tree, not a status line.

4. Review — do not trust the self-report

• Re-run the project's gates yourself. Never take "gates passed" on faith.
• Read the diff against the brief: did ZCode do what was asked, nothing more and nothing less? touchedFiles is your starting point.
• On a --read-only run, check readOnlyViolation and confirm touchedFiles is empty.
• Run the relevant guard skills on the diff if you have them installed.

Full checklist: references/review-and-land.md.

5. Land it

The orchestrator commits. Only after the gates pass and the diff holds:

• Commit the verified work yourself, with a clear message.
• If it needs changes, send a delta brief with --session <sessionId> from the prior result.json, and review again.

Read-only second opinions

The relay doubles as a way to get an adversarial second opinion with no write risk: dispatch --read-only with a brief listing the agreed points, then each contested point with both positions, and ask ZCode to defend or concede each. Because plan mode's guarantee is measured rather than enforced here, verify touchedFiles came back empty instead of assuming no edits.

Authorization model

Delegation is something the human opts into. Once they have, committing verified, gate-passing work is the agreed contract. Two limits: surface, don't absorb (report ZCode's design decisions and defensible-but-unasked turns rather than silently keeping them) and stop for scope changes (if correct completion needs going beyond the brief, ask). The full treatment is in references/review-and-land.md.

References

• references/writing-the-brief.md — how to write a brief ZCode can execute blind: structure, the report contract, embedding the real gate commands.
• references/dispatch-and-poll.md — relay.mjs flags, the result.json contract, how the CLI is resolved, backgrounding, and recovery.
• references/review-and-land.md — the review checklist, the commit boundary, and the exact-session rework cycle.
• references/multi-task-queues.md — running a sequential queue: carrying constraints forward, progress tracking, and the end-of-run coherence check.