Coding agents

You call one skill per message. The skill does not start by itself.

Call /logic-grill when Intent slots are unbound. It emits a V3 chain. It does not run that chain.

The pipeline splits research from order-of-work, and one-hook design from implementation. Critics gate the next stage. They do not write the next artifact.

Staged pipeline

Work one stage at a time. `/hook-code` path-loads `coder.md`.

Stage	Call	Writes
Intent	/logic-grill	V3 chain. Does not run it
Facts	/logic-discovery	Conversation. Optional {slug}-discovery.md
Workspace order	/logic-plan	{slug}-plan.md
Plan audit	/logic-plan-critic	Verdict only
One hook	/hook-plan	{collection}-{slug}-hook-plan.md
Hook-plan audit	/hook-plan-critic	Verdict only
TypeScript	/hook-code	*.hook.ts
Hook audit	/hook-critic	Verdict only

question and idea stop after discovery. research writes remaining questions. It does not write implementation phases.

Pick a path. Each box is the first prompt for that skill. The next box grounds the file from the box above.

/logic-discovery then /logic-plan. Critics gate hook-plan, then /hook-code.

/logic-discovery

<task> warehouse can freeze a bin so pickers cannot allocate stock from it </task>

∞AgentExtra High

bin-freeze-discovery.md

Persist when `/logic-plan` needs a file. Type `feature`.

Open	Status
Document that stores freeze	UNKNOWN
Who may set freeze	UNKNOWN
Allocate block	UNKNOWN

Ground the discovery file. Do not name a job or a type in the first /logic-plan message.

/logic-plan

<problem> warehouse can freeze a bin so pickers cannot allocate stock from it </problem> <ground> @bin-freeze-discovery.md </ground>

∞AgentExtra High

tasks/bin-freeze-plan.md

Job `spawn`. Type `feature`. Status Draft.

new documents → document updates → new hooks → hook updates

Artifact	Path
Freeze flag	UNKNOWN. Ask.
Allocate block	UNKNOWN. Ask.

Audit the plan file. The critic does not rewrite it.

/logic-plan-critic

<plan> @tasks/bin-freeze-plan.md </plan>

∞AgentExtra High

Verdict

Live MCP `hook_graph`. Plan file unchanged.

Check	Result
Gradual order	Documents before hooks
Blast radius	Graph vs 🟢 / 🟡 / 🔴
Verdict	PASS or REVISE

After PASS and hook work, start /hook-plan. Do not write TypeScript. Do not ground a workspace plan file.

/hook-plan

∞AgentExtra High

warehouse-freeze-bin-hook-plan.md

Grain: one hook. No `*.hook.ts`.

Pattern, execute() phases, ok + data, replay, open questions.

Audit the hook-plan. After PASS, implement with /hook-code.

/hook-plan-critic

<ground> @warehouse-freeze-bin-hook-plan.md </ground>

∞AgentExtra High

/hook-code

<ground> @warehouse-freeze-bin-hook-plan.md </ground>

∞AgentExtra High

/hook-critic

Ground: @freeze-bin.warehouse.hook.ts

∞AgentExtra High

Grain

Reader question	Call
Bind outcome, stop, and the next skill.	/logic-grill
How does this work?	/logic-discovery
What do we build, and in what order?	/logic-plan
Is this workspace plan honest?	/logic-plan-critic
Design this one hook.	/hook-plan
Is this hook-plan honest?	/hook-plan-critic
Write the hook.	/hook-code
Review this hook .ts.	/hook-critic
What is the version-2 hook shape?	/hook-code-style

Discovery types: bug | feature | idea | requirement | question.

Plan types: bug | feature | fix | user-request | research.

Name discovery type in the first /logic-discovery prompt, or omit it so the skill asks.

Do not name a plan type or a job in the first /logic-plan message.

If the slash name does not start the skill, put the skill file on line 1:

@.agents/skills/logic-discovery/SKILL.md

∞AgentExtra High

Next steps

Complete Setup before the first MCP call.