calvinfungcn

ai-collab-continue

Use only when the user explicitly invokes this entry to resume the current ai-collaboration task in this project. Reads the latest task, batch, handoff, evidence and review allowance, then follows only the authorized next step for the current agent. It never self-reviews, resets a count or reopens a closed batch. Discovery alone does not select the workflow.

calvinfungcn 2 Updated 20h ago
GitHub

Install

npx skillscat add calvinfungcn/engrelay/ai-collab-continue

Install via the SkillsCat registry.

SKILL.md

AI Collab Continue

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 another controller owns the task, 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

Start with applicable project instructions and project-relative paths:

  • .ai/current-task.md and .ai/handoff.md, the latest authority for Next action
  • the named authority needed to confirm the actual route, outcome and remaining
    scope: mapped OpenSpec tasks, applicable result, review index and linked round

Then read the sections triggered by the actual action, not every execution branch:

  • route interpretation or close: docs/ai-collaboration/README.md#routing-contract
    (including #closed-batch-guidance); use
    handoff-schema.md#routing-section-schema when interpreting its fields
  • executable continue: handoff-schema.md#autonomous-continuation; agent or human
    dispatch: handoff-schema.md#dispatch-commands and the assigned role sections
  • execution, verdict/count or allowance questions, including post-closure review:
    README.md#review-round-circuit-breaker
  • a named recovery candidate from step 1: handoff-schema.md#source-matched-recovery
  • an actual write boundary: handoff-schema.md#expected-state-validation;
    read-only or unpersisted return: handoff-schema.md#response-handoff-fallback

These reference paths are under docs/ai-collaboration/. Project instructions and
mandatory reading from triggered Skills still apply; this is no prohibition on
additional reading needed to resolve uncertainty.

Reuse a reference already read in this session only while its content is known
unchanged. Reread the live .ai/ files, the named authority and the anchored rule
sections for a new agent, missing context, a changed input or a rule question
(docs/ai-collaboration/agent-prompts.md#common-preflight).

Procedure

  1. Compare the request with the latest state: change ID, Execution Batch ID, State
    revision, Review round, current role, current agent and status. A delayed or
    duplicate message is checked against that state before any action.
    If the live handoff routes to a step already superseded by a completed
    same-batch transfer: a review record linked by the review index, or a result or
    response file whose path the live handoff (including Next output files) or the
    current task explicitly names (for example, a verdict is counted while the live
    file still dispatches that review).
    First apply handoff-schema.md#source-matched-recovery. Read each named path
    once; never scan a directory. A new container must carry a real prior SHA256 and
    matching revision; only legacy index-linked records and legacy response handoffs
    follow the schema's legacy compatibility rule. Restore only the single pending
    transfer of the batch, stop on any conflict or missing evidence, persist only if
    this Agent may write the handoff, read back, and then follow the restored route.
    An equal Next action, Next role and Review round alone do not show the transfer
    was applied. Never recount, replay or reopen a batch from a candidate alone.
  2. Resolve actual waits and limits before executing any command:
    • close: follow README.md#closed-batch-guidance. Briefly state the evidenced
      outcome and mapped result, known remaining scope or unknowns, no next role or
      Operator Command for that closed batch, and one applicable next step. Distinguish
      ordinary accepted, owner-limited acceptance, stopped and superseded. Do not write
      a new handoff just to answer status or replay the closed batch.
      If this current request clearly identifies and authorizes distinguishable
      follow-up work, use workflows.md#1-task-intake in this session after checking
      scope, dependencies, authority, real waits and execution ownership. Preserve the
      old closure via the existing archive rule; retain cumulative verdicts and finite
      limits for the same unfinished acceptance. A genuinely different scope gets its
      own mapping and count without rewriting history. This uses the existing workflow
      action, not an implicit invocation of the explicit-only ai-collab-start Skill;
      do not ask the owner to resend equivalent authorization. Explicit single reviews
      after closure use the existing bounded-review rule and independent reviewer,
      never a fresh repair allowance. Ambiguous requests and stale notices only
      get state-aware guidance; do not infer work or authorization from them. A bare
      “continue” is covered only by a recorded standing authorization for the parent
      proposal with exactly one unambiguous ready item (README.md#closed-batch-guidance);
      otherwise it is ambiguous and gets the closed-batch guidance. For a closed
      independent design batch, the first Current Status paragraph and reply name the
      completed design, remaining parent work and one copyable request
      (README.md#design-to-implementation-relay), not an Operator Command.
    • awaiting-owner-decision (including human dispatch): present the recorded
      Owner Decision Brief; continue only after an applicable reply. Silence is
      not approval. Reuse an already answered decision without asking again. An
      older wait caused only by the former default count of three may resume without
      a new reply only when its recorded Owner Decision Brief proves the migration
      basis in README.md#review-round-circuit-breaker (no non-convergence, material
      limitation or scope trade-off) and no other wait or limit remains; cite that
      basis and keep the unknowns.
    • continue + blocked: check the Resume condition and current evidence;
      perform a cheap safe check if available. Resume only if the actual condition
      is resolved; record that evidence and update the wait/summary before executing
      the authorized step. Otherwise retain the wait and report its resume condition.
      A no-action forward notice alone is not a blocker.
    • reached review ceiling: follow the existing circuit breaker and any applicable
      finite owner extension. Never start an autonomous cycle past a reached ceiling, or
      reset the count. Use the evidenced owner-decision route when required.
    • confirmed same-finding stall: allow only its one technical analysis inside the
      original scope (README.md#review-round-circuit-breaker). An executable disposition it
      proposes proceeds within existing authorization; use the evidenced owner-decision
      route only when the owner conditions there apply. Never start another repair/review
      loop for that finding without an applicable authorization.
  3. Otherwise follow the latest executable Next action:
    • continue: immediately perform the stated authorized next step if capable
      and permitted. Read the Continuation Command under Current Status when
      present; on legacy records recover the step from Current Status and linked
      authority, without requiring new fields or owner approval. A replacement
      session may assume the same responsibility after checking latest state and
      other same-batch execution ownership; record actual Current agent in Routing and increment State revision, keep the
      responsible role, batch, scope, risk and Review round. Formal role transfers,
      synthesis-agent dispatch and independent reviews still use dispatch.
      Local CLI/IDE host choice within authorized offline work is your execution
      detail: never ask the owner where to run or to approve continue again.
      A prior runner exiting is not environment blockage; do not launch competing
      writers or kill another agent's process. Report a concrete capability or
      permission limit through the existing recovery route when needed. Ending the
      active session on purpose is a boundary even for the same role: save the
      Continuation Command first. A foreseeable long check may leave one short
      checkpoint with its external log, stated as not yet passed
      (handoff-schema.md#planned-session-end-and-long-check-checkpoint).
    • dispatch to an agent: follow the complete Operator Command only if the current
      agent is the named role and is independent of the producer of the work under
      review. A reviewer never reviews its own design or implementation. If not
      eligible, do not perform the review; present the original Operator Command for
      the correct recipient and the reason. When this agent completes its role and
      dispatches the successor under manual relay (default for new Full tasks), save,
      read back, and validate the handoff with expected state, place the top human transfer
      notice, explain the transfer in chat, and stop the turn without autonomously
      launching the next role. Explicit automation applies only when authorized in
      Constraints and not overridden by a narrower user instruction.
  4. A delayed request targeting a closed batch or an earlier revision reports the
    latest actual state; it does not replay dispatch or change prior acceptance.
    Distinguish it from a clear current follow-up request under step 2.
  5. If the authoritative inputs are missing or conflicting, name the exact file and
    conflict, and give the smallest recovery direction. Never invent a role, verdict,
    count or acceptance from chat history.

Boundaries

  • Keep full Operator Commands and Owner Decision Briefs intact when you copy or
    update them.
  • Update .ai/handoff.md only at a real boundary using the canonical schema, then
    read back its Routing and validate it with the expected batch, action, next role
    and Review round decided for that boundary (handoff-schema.md#expected-state-validation).
    Include Human Action Required in new/updated handoffs; for executable continue
    include its Continuation Command, and omit it on actual waits/dispatch/close.
    A forward notice needs no current human action/reply. This entry does not
    schedule or wake processes, change host configuration, or override disabled mode.
  • A read-only request uses the response handoff fallback described in
    docs/ai-collaboration/handoff-schema.md.
  • At a real boundary, reconcile active narrative with the live task, handoff, cited
    review and the sole tasks.md: correct only stale active facts, never stable
    acceptance or history.
  • Owner acceptance, Git operations and deployment remain outside this entry.

Output

State the Next action you followed, the evidence you checked, and any unfinished
mapped work. For an agent dispatch, confirm where the Operator Command is written and
that its first Current Status line carries the advisory capability tier, echoed in
the final reply; a response-only fallback carries the same line. Under manual relay,
state that the turn stops for operator transfer. For a human decision,
present the brief and omit the line. For close, state that no next role exists.

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.