shared/workflows/phase-1-advise
Source:
shared/workflows/phase-1-advise.md
workflow: Phase 1 — advise
The plan/clarify/present phase. Wraps
shared/advisor.md+shared/question-first.md. Mandatory.
Steps
- Summarize the context in 3–5 bullets (drawn from
task-slug/context.md). - Identify ambiguity: list the open questions the agent has after reading context. Cap at 3 user-facing; rest become
auto-defaultedchoices. - Apply
question-first.md: ask up to 3 questions. Wait for answers (or proceed silently under--auto+defaults.question_first.silent: true). - Present approaches: 2–4 options, one-line trade-off each, one marked recommended (with one-line evidence: “you picked X on the last 3 similar tasks”).
- Wait for user choice.
- Write the chosen plan to
<task-slug>/plan.md— 1-pager: goal, approach, scope, success criteria, risks, rollback. - Log every fork to decisions.jsonl.
Plan.md template
# plan: <task-slug>## goal<one sentence, the user's confirmed restatement>## approach<chosen option>: <one-line summary>## scopein-scope: - <item> - <item>out-of-scope: - <item>## success criteria1. <testable criterion>2. <testable criterion>## risks- <risk>: <mitigation>## rollback<one sentence: how do we undo this if it goes wrong>## decisions log entries- scope=full (user-answered)- test_framework=jest (override-applied from core.yaml)- pr_strategy=single (auto-defaulted, based on 5 prior tasks in this repo)Recommendation logic
- Read
$ADK_DATA_HOME/improve/learning/decisions.jsonl+summary.mdfor prior matchingfork_ids. - Read
$ADK_CONFIG_HOME/core.yaml.defaults.<skill>.*. - Read
<repo>/.adk/core.yaml.defaults.<skill>.*(override layer). - Recommendation = highest-priority match.
- If no prior data: recommendation = the most common cross-user-base default, marked
[no personal history].
Honesty
- If the agent has too little context to recommend, say so: “no recommendation — I don’t know your team’s preference on PR strategy. Pick one and I’ll learn.”
- If the recommendation is uncertain (< 3 prior matches), mark it
[weak signal].