Use only when the user explicitly invokes this entry to start a clearly stated task under ai-collaboration in this project. Recommended general entry: reuses existing Task Intake shortcuts, the same conditional Requirements Analyst contract as ai-collab-requirements when needed, prepares the minimum coherent task input, and routes by risk. Discovery alone does not select the workflow.
Install
npx skillscat add calvinfungcn/engrelay/ai-collab-start Install via the SkillsCat registry.
AI Collab Start
Activation
Apply this entry only when the current request explicitly selects ai-collaboration
under docs/ai-collaboration/README.md#optional-activation-and-coexistence.
- If the project WorkflowMode is
disabled, explain that explicit re-enabling is
required and stop. Do not change the mode or start a lifecycle. - If the user selected another controller for this request, follow that controller.
Write no collaboration runtime file and dispatch no role. - Discovery, a
.ai/folder, role names or an installed Skill do not activate it.
Read First
Use project-relative paths only:
docs/ai-collaboration/workflows.md#1-task-intake, and the templates below when
preparing the task inputdocs/ai-collaboration/templates/current-task.template.mddocs/ai-collaboration/templates/handoff.template.md.ai/current-task.mdand.ai/handoff.md, when they exist- the named
openspec/changes/<change-id>/artifacts, when the user names a change - only when Task Intake cannot safely summarize material business ambiguity (step 5):
docs/ai-collaboration/workflows.md#requirements-analyst-conditional-role-actionanddocs/ai-collaboration/agent-prompts.md#requirements-analyst
Reuse a reference already read in this session only while its content is known
unchanged; reread it for a new agent, missing context, a changed input or a rule
question (docs/ai-collaboration/agent-prompts.md#common-preflight). Mandatory
module instructions and triggered Skills still apply.
Procedure
- Same-work check: only when the current request belongs to the existing batch
or the same unfinished acceptance, and that batch is executable, waiting, or
has a named pending transfer, reuse its identity, acceptance and cumulative review count.
Do not reset the count or create a duplicate batch. Continue it throughai-collab-continue, which first applies its source-matched recovery to a
same-batch completed review or response handoff that the live file predates
(docs/ai-collaboration/handoff-schema.md#source-matched-recovery). - Unrelated unfinished batch: when the request is unrelated, do not continue the
existing batch under step 1; keep its live.aiinputs untouched. Explain the
conflict and request the smallest owner choice. Replacement needs explicit owner
authorization. Response-only discussion may proceed without writes. - Closed previous batch: use
README.md#closed-batch-guidanceunderdocs/ai-collaboration/. An active parent change alone does not send start back
to closed continue. Check the sole tasks.md and dependencies for the clearly
identified, authorized follow-up, then perform Task Intake here without asking
for the same request again. Preserve counts/finite limits for the same unfinished
acceptance; distinct scope gets its own mapping without rewriting old history.
A single post-closure review follows the existing bounded-review rule, including
independence and no new repair allowance. For ambiguous requests or
stale notices report actual closure, remaining/unknown scope and one applicable
next step; do not create work or archive state just for that answer.
Before preparing authorized new input, archive the closed current-task, handoff
and any baseline under.ai/archive/<batch id>/by the existing closure rule,
without overwriting records. Referenced evidence stays where it is. - Prepare the minimum coherent
.ai/current-task.md: task ID, change ID if any,
Execution Batch ID mapped to OpenSpec checklist items, riskS,MorHwith a
reason, the observable outcome, must-have acceptance, out-of-scope work, required
checks, Constraints (recording the default manual relay for formal agent dispatch,
or an explicit automation scope when explicitly authorized by the user) and the Initial role.
Keepopenspec/changes/<change-id>/tasks.mdas the only implementation checklist;
do not create a second task list. - Route by the existing rules:
- material business ambiguity that Task Intake cannot safely summarize: set
Initial role to Requirements Analyst and run the conditional role action in
this same responsible session perworkflows.md#requirements-analyst-conditional-role-action. Do not ask the
owner to invokeai-collab-requirementsseparately; technical ambiguity
stays with Chief Designer; - a new contract, a complexity exception, risk
Hor missing design basis goes to
Chief Designer with a complete Operator Command; - a clear, bounded
SorMscope goes to Implementation Engineer; - an existing implementation that needs review goes to Code Reviewer;
- a design-only mapping whose authorized target also needs implementation keeps
that implementation mapped unless an independent design deliverable is evidenced
(docs/ai-collaboration/README.md#design-to-implementation-relay).
Reuse an already confirmed requirement baseline; do not add a repeated final
confirmation.
- material business ambiguity that Task Intake cannot safely summarize: set
- If the first executable action needs an owner decision, keep this role with
same-sessioncontinueandawaiting-owner-decisionplus an Owner Decision
Brief. Do not dispatch an implementer.
Boundaries
- Do not implement the work in this entry. It only prepares the input and route.
- Do not reopen closed evidence or write over unfinished unrelated work.
- Owner acceptance and deployment remain outside this entry.
Output
Write .ai/current-task.md and .ai/handoff.md at each real boundary. The handoff
must carry complete Routing and a copy-ready Operator Command for an agent dispatch,
or an Owner Decision Brief for a wait, as defined bydocs/ai-collaboration/handoff-schema.md.
For an agent dispatch, the first Current Status line is the advisoryRecommended capability tier: Cn (docs/ai-collaboration/handoff-schema.md#capability-tier-guidance).
Confirm it at the same readback and echo the same tier in the final reply; a
response-only fallback carries the same line. Under manual relay, include the top human
transfer notice in handoff, state where the Operator Command is located, and stop the turn
for operator transfer without autonomously launching the next role.
EngRelay material and modifications: Copyright (c) 2026 calvinfungcn and EngRelay
contributors, MIT. The full notice accompanies the installed kit at the project-root
path docs/ai-collaboration/README.md#engrelay-license. Preserve it and any upstream
notice when redistributing a standalone Skill.