AMQ collaborative specification
Invoke the amq-spec skill through your host's skill interface. It defines
a structured two-agent specification flow; invocation syntax depends on the host.
Use canonical phases in order:
Research -> Discuss -> Draft -> Review -> Present -> Execute
Detailed step-by-step protocol lives in references/spec-workflow.md.
This file is the concise operational entrypoint.
Parse Input
From the user prompt, extract:
- topic: short kebab-case spec name (e.g.,
auth-token-rotation)
- partner: partner agent handle (default:
codex)
- problem: the full design problem statement
If topic/problem are unclear, ask for clarification.
Pre-flight
- Verify AMQ is available:
which amq
- Verify the AMQ root is discoverable (
.amqrc, AMQ env vars, or the default .agent-mail layout); otherwise run: amq coop init
- Use thread name:
spec/<topic>
First Action: Send problem to partner IMMEDIATELY
The entire point of the spec workflow is parallel research — both agents
exploring the problem independently, then comparing notes. Every second you
spend researching before sending is a second your partner sits idle waiting
for the problem statement. That's why the send comes first, even though your
instinct might be to "research first to give better context."
amq send --to "<partner>" --kind question \
--labels workflow:spec,phase:request \
--thread "spec/<topic>" --subject "Spec: <topic>" --body "<problem>"
Send the user's problem description verbatim — your own analysis goes in the
research phase, not the kickoff. If you pre-analyze, you bias the partner's
independent research, which defeats the purpose of having two perspectives.
Label Convention
Labels are how both agents and the receiver-side protocol table know which phase the conversation is in. Use existing AMQ kinds plus labels to express spec workflow semantics:
| Phase | Kind | Labels |
|---|
| Problem statement | question | workflow:spec,phase:request |
| Research findings | brainstorm | workflow:spec,phase:research |
| Discussion | brainstorm | workflow:spec,phase:discuss |
| Plan draft | review_request | workflow:spec,phase:draft |
| Plan feedback | review_response | workflow:spec,phase:review |
| Final decision | decision | workflow:spec,phase:decision |
| Progress/ETA | status | workflow:spec |
Quick Command Skeleton
# Initiate spec with problem statement
amq send --to "<partner>" --kind question \
--labels workflow:spec,phase:request \
--thread "spec/<topic>" --subject "Spec: <topic>" --body "<problem>"
# Submit independent research
amq send --to "<partner>" --kind brainstorm \
--labels workflow:spec,phase:research \
--thread "spec/<topic>" --subject "Research: <topic>" --body "<findings>"
# Discuss and align
amq send --to "<partner>" --kind brainstorm \
--labels workflow:spec,phase:discuss \
--thread "spec/<topic>" --subject "Discussion: <topic>" --body "<analysis>"
# Draft plan
amq send --to "<partner>" --kind review_request \
--labels workflow:spec,phase:draft \
--thread "spec/<topic>" --subject "Plan: <topic>" --body "<plan>"
# Review plan
amq send --to "<partner>" --kind review_response \
--labels workflow:spec,phase:review \
--thread "spec/<topic>" --subject "Review: <topic>" --body "<feedback>"
# Optional final decision message
amq send --to "<partner>" --kind decision \
--labels workflow:spec,phase:decision \
--thread "spec/<topic>" --subject "Final: <topic>" --body "<final plan>"
When You RECEIVE a Spec Message
If you receive a message labeled workflow:spec, your action depends on the phase:
| Label | Your action |
|---|
phase:request | Read the problem statement, do your own independent research first, then submit findings as brainstorm + phase:research |
phase:research | Before reading: check if you've already submitted your own research on this thread. If not, do your own research and submit it first. This preserves research independence — reading the partner's findings before forming your own view contaminates your perspective. Once your research is submitted, read the thread and start discussion as brainstorm + phase:discuss. |
phase:discuss | Reply with your analysis, continue discussion until aligned |
phase:draft | Review the plan and send feedback as review_response + phase:review. Your job here is review, not implementation — the plan needs to survive scrutiny before anyone builds it. |
phase:review | Revise plan if needed, or confirm alignment |
phase:decision | Stop. A phase:decision message is agent-to-agent alignment, not user approval, so do not implement from a spec decision alone. Only the human authorizes implementation, recorded as a structural gate to the initialized human handle (conventionally user; see the Operator Gates section in the amq-cli operations guide). Wait until the initiator confirms the human approved on the gate thread and assigns you work. |
Why the partner doesn't implement: The spec workflow is a design process.
The initiator owns the relationship with the user and presents the final plan.
If the partner implements without approval, the user loses control over what
gets built. The agent-to-agent phase:decision message is alignment, not
authorization: human approval is a structural gate to the initialized human
handle, and partner agents must not implement from a spec decision alone.
Implementation starts only after the initiator explicitly tells you the human
approved and assigns work.
Protocol Discipline
- Use the existing wake while waiting — if a live injecting wake notifies
this terminal, finish independent work and yield. On its doorbell, run
amq drain --include-body. Do not start amq watch, amq monitor, or a
background polling loop for peer replies. Check amq wake check --me <handle> --json when the capability is unknown. A bounded watch is a fallback only
without an injecting wake; notify-only supervisor consumers remain valid.
These rules exist because violations silently break the workflow's value proposition:
- Send before researching — parallel research is the whole point. Pre-researching wastes your partner's time and biases the outcome toward your initial framing.
- Submit your own research before reading partner's — reading first contaminates your independent perspective. Two agents who read the same code and reach the same conclusion is less valuable than two agents who explore independently and then compare notes.
- Don't skip phases — each phase builds on the previous. Collapsing directly to a finished spec skips the discussion where misunderstandings surface.
- Use
spec/<topic> threads and the label convention — this is how both agents (and the tooling) know which phase the conversation is in. Without consistent labels, the receiver-side protocol table above breaks.
- Don't enter plan mode during research if it blocks tool usage — you need tools to explore the codebase.
- Present the final plan to the user before executing, and raise a structural gate. The initiator owns the user relationship. After the decision phase, present the plan in chat AND raise a structural human gate using the initialized human handle (conventionally
user) on a stable gate/<topic> thread, then wait for explicit approval on that thread. The agent-to-agent phase:decision message is alignment only; partner agents must not implement from it. See the Operator Gates section in the amq-cli operations guide for canonical mechanics, seeding, and guardrails.
Reference
For full protocol details, templates, and phase gates, see: