gate-walk

name	gate:walk
description	Walk Blake through pending gate items (decisions + scripts) one at a time, with AI assistance. Resolves the human-only inbox without context-switching.
requires_stores	["gate"]
default_invoker	human
user_invocable	true

Walk gate items

Blake has a queue of human-only items in the gate store: decisions only Blake can answer (business / scope / design) and scripts only Blake can run (sudo deploys, secret rotations). This skill walks them sequentially, helping with context but never answering on Blake's behalf.

Why default_invoker is human

The gate's answer field has actor: human in its schema. The CLI auto- detects $CLAUDECODE and would otherwise infer ai_autonomous, which the schema rejects. Pass --invoker human explicitly on every answer call. Without it, the answer write fails with a clear actor-mismatch error.

Other gate verbs have different actor requirements — match the flag to the verb, don't blanket --invoker human:

Verb	Required actor	What to pass
answer	human	--invoker human
defer, resume	ai_with_human	--invoker human (accepted) or --invoker ai_with_human
cancel	ai_autonomous	omit --invoker (auto-detect wins in CLAUDECODE shells)

Discover

stores gate --help
stores gate schema --json
stores gate list --status pending --json

The walk

For each pending item:

1. Print it — the question, options (if decision) or command (if script), any task_ref link.
2. Add context — read the linked observation/task if task_ref is set: stores observations show <ref-id> or equivalent. Surface the relevant bits of context Blake needs to decide. Don't editorialize.
3. Pause for Blake. This is the load-bearing moment: you do not answer. Wait for Blake's choice (if decision) or for Blake to confirm they ran the script (if script).
4. Record the answer:

# Decision Blake answered:
stores gate answer <gate-id> --answer "<choice>" --invoker human

# Decision Blake wants to defer (resurfaces when defer_until arrives):
stores gate defer <gate-id> --defer-until <YYYY-MM-DD>

# Decision Blake wants to cancel (no further action; record-keeping):
stores gate cancel <gate-id>

5. Move to the next pending item.

When NOT to use this skill

• Single gate item: just answer it directly via stores gate answer .... The walk overhead isn't worth it.
• Item needs deeper investigation: stop the walk, hand context to a separate research session, come back later.

Rules

• Never call stores gate answer without --invoker human. The CLI rejects it; you'd be working around schema enforcement.
• Surface context, do not decide.
• One item at a time. No batching.
• If you spot follow-up work for Blake while walking (e.g. a decision implies a new task), file it as a fresh gate item or task entry — don't cram it into the current item's resolution.