Skill v1.0.0
currentLLM-judged scan90/100version: "1.0.0" name: escalate-decision description: "Interrupt the developer only when a decision is truly theirs, and make the interruption answerable in one step: lead with the situation and the exact question, add the smallest evidence that makes the call possible, and list the real options. Use when a worker returns a question or fails, acceptance criteria are ambiguous, an action is risky or consequential (commit, merge, push, publish, deploy, delete, or exposing secrets), a review is inconclusive or disputed, or the rework iteration limit is reached — not for routine progress."
Escalate a decision
Automation fails when it either hides a decision or drowns the developer in them. Escalate when the answer genuinely belongs to the developer, and make each escalation cheap to answer: one situation, one question, the evidence that matters, and the options. The developer stays the final decision-maker; everything else you coordinate.
When to use
- A worker returned a question (is blocked) or has failed.
- Acceptance criteria are ambiguous or contradictory, and guessing would spend
real work on the wrong target.
- An action is risky or consequential — commit, merge, push, publish, deploy,
delete, or exposing secrets. These always need explicit approval.
- A review is
inconclusive, or a reviewer and implementer dispute a finding. - The rework loop reached
max_iterationswithout converging.
When not to use
- Routine progress or a completed task: that is
report-progress. - A question the task's own criteria or the project context already answers:
decide it and record the decision. Escalating what you can settle yourself trains the developer to ignore escalations.
Inputs
- The state that needs a decision, and what you tried.
- The smallest evidence that makes the decision possible — a worker's question,
review findings, a diff stat, command output.
- The options, with your recommendation.
Procedure
- Check that it is theirs. Confirm the call needs developer intent or a
risk appetite you do not have. If the criteria already answer it, decide and record the decision instead of asking.
- Write the decision brief, top-first:
- Situation — one line: what happened, and on which task or project.
- Question — the exact choice, phrased so a one-word answer resolves it.
- Evidence — the minimum that makes the call possible, quoted from its
source rather than summarized from memory.
- Options — the real alternatives, one line each, with your
recommendation.
- For a worker that returned a question, do not answer it. Read its report
and forward the question:
``bash python3 scripts/tm.py report "<name>" ``
The worker is waiting on the developer, not on you; answering on its behalf commits the project to a decision the developer never saw.
- Deliver it with
report-progress, which carries the shared
handoff-report shape. Keep the decision request at the top of that report; the developer should not have to read past it to find the question.
- Act only as approved. Record the decision on the ledger
(python3 scripts/tm.py task decide <id> approve|request-changes|reject|finalize) and resume or stop. Consequential work pauses until the answer arrives.
Output
A decision request the developer can answer in one step.
Failure
- If the question is buried, rewrite it — a reader should not have to infer the
choice from the report.
- Silence is not consent. Without an explicit answer, do not commit, merge,
push, publish, deploy, or delete on the developer's behalf.