astra-advisor

# Astra Advisor Buy a small amount of GPT-6 Astra judgment at a decision point. The parent agent gathers evidence, implements, and verifies; the advisor returns a verdict. Default advisor: `gpt-6-astra`, reasoning `high`. Honor an explicit user model or effort choice; do not silently substitute a different model if Astra is unavailable. ## Decide whether to consult Consult when both conditions hold: 1. A wrong decision would cause substantial rework or consequences, or two genuinely different fix attempts have failed. 2. There is a real choice, unresolved failure, or concrete residual concern that an independent review could resolve. Useful moments: before committing to architecture, schema or API contracts; before finalizing a production migration; before declaring consequential work complete; before starting unattended automation that changes external state or accumulates material cost. For automation, include interval, state/deduplication, verification, stop conditions, action limits, and expected per-run cost. Gather enough evidence to frame the question before consulting. Skip mechanical changes, settled choices, a first failed attempt, and easily cancelled read-only polling. An explicit request for an Astra second opinion overrides this relevance gate. If the parent already runs on Astra, decide locally by default; an explicitly requested independent review can still justify a separate Astra context. ## Bound the exchange - Default to one consult per task, maximum three executed advisor interactions, counting follow-ups. Use at most one reconciliation follow-up per question. - Announce the topic, model, trigger, and count before each interaction: `Consulting GPT-6 Astra on <decision> (trigger: <reason>, interaction 1/3).` - Aim for an 800-word brief; stay below 1,200 words including excerpts. Provide at most five precise file pointers. Request an answer under 300 words. - Batch related questions. Follow up only for new evidence or a specific conflict; do not retry until the advisor agrees. Stop at the cap and explain unresolved gaps. - These are interaction and prompt limits, not guaranteed billing caps: reasoning, tool calls, and runtime-provided context also consume resources. Respect any user spend or time limit; do not invent dollar savings from word counts. ## Prepare the brief ```text DECISION: <one focused question> CONTEXT: <what is being built and for whom> OPTIONS: <A/B with real tradeoffs; include your leaning and why> CONSTRAINTS: <hard requirements, existing infrastructure, team, deadline> EVIDENCE: <observed facts, exact errors, relevant excerpts> FILES: <up to five paths, line ranges, and why they matter> ANSWER: Verdict / Why / Risks / Would change my mind; under 300 words. ``` For debugging, replace OPTIONS with ATTEMPTS and what each result ruled out. For completion review, use WHAT CHANGED, HOW VERIFIED, and SPECIFIC WORRIES. Do not send the full conversation, whole files, or secrets when a focused excerpt suffices. ## Dispatch This skill instructs a bounded advisor delegation when the gate above is met and the host permits delegation. Read [the advisor prompt](references/advisor.md), prepend it to the brief, and use [the runtime instructions](references/runtime.md). Use a fresh context with the model explicitly pinned; a skill name alone does not select a model. The advisor may read only the named evidence and must not edit, publish, call mutating connectors, or delegate further. A read-only shell sandbox does not itself constrain every connector, so use tool restrictions when available. Respect the user's instruction to skip delegation or require approval. If the host cannot run the chosen model within the allowed permissions, report the limitation and continue the parent task where possible without claiming an advisor review. ## Use the answer Treat the verdict as advice. Check decisive claims against the evidence, resolve a material contradiction with one narrow follow-up or source read, and make the final call yourself. Explain any reasoned disagreement. A review does not authorize a deployment, publication, or other action outside the user's scope. Tell the user the model consulted, the verdict, and what changed because of it. Never present an attempted or failed invocation as a completed review. The sibling `astra-orchestrator` handles the reverse direction: an Astra parent delegates substantial fact-finding to explicitly selected lower-cost GPT workers.