arcasilesgroup Advise
Discoverable wrapper around the ai-advise governance advisor: dispatches the agent via the Agent tool and renders a severity-tagged advisory (info | warn | concern). Never blocks, never modifies code.
Workflow
Principles: §10.6 SDD (every warning traces to an active decision or stack standard — no advice without a documented anchor); §10.4 DRY (the agent contract owns the analysis loop — the skill never paraphrases standards inline).
- Load stack contexts — read
.ai-engineering/manifest.ymlproviders.stacks; apply.ai-engineering/overrides/<stack>/conventions.mdso stack-specific standards are in scope. - Detect mode — first positional arg is
advise(default),gate, ordrift. Anything else is a path filter; mode defaults toadvise. - Dependency preflight — verify
.claude/agents/ai-advise.mdexists. STOP and report the exact missing path if absent; never paraphrase agent instructions inline. - Dispatch — invoke the
ai-adviseagent via the Agent tool with{mode, paths, severity_floor}. It runs in its own context window and returns the structured advisory. - Render — emit the advisory table grouped by severity (
concern→warn→info). Each row:File | Finding | Recommendation | Anchor, where Anchor is the standard or active decision the finding traces to. - Audit — emit
framework_eventkind=advisory_emittedwith{mode, file_count, warning_count, severity_distribution}. Never emitBLOCK/FAILoutcomes — those belong to/ai-verifyand git hooks.
Modes
| Mode | Trigger | Agent action |
|---|---|---|
advise (default) |
Post-edit in build | Scan changed files against stack standards + active decisions; emit severity-tagged warnings with recommendation. |
gate |
Pre-dispatch | Validate the proposed task respects governance boundaries (agent capabilities, expired risk acceptances, scope leakage). |
drift |
On-demand | Compare implementation against active architectural decisions; classify drift `none |
Output Contract
Grouped by severity, rendered as a markdown table. Severity scale: info < warn < concern (never error/critical/blocker).
# Guard Advisory: <mode>
## Summary
- Files checked: N
- Warnings: N (concern: N, warn: N, info: N)
## Warnings
| # | Severity | File | Finding | Recommendation | Anchor |
## Decision Context
[Active decisions that informed this advisory]
Differentiation
| Aspect | /ai-advise |
/ai-verify |
/ai-review |
|---|---|---|---|
| When | During development | Pre-release / pre-merge | Pre-merge narrative review |
| Blocking | Never (fail-open) | Can BLOCK on FAIL | Never (judgement only) |
| Scope | Changed files + active decisions | Full codebase or mode-specific | PR / branch / paths |
| Output | Severity-tagged warnings | Scored evidence-backed verdicts | Specialist-attributed findings |
| Engine | ai-advise agent (read-only) |
ai-verify agent + specialists |
ai-review agent + 9 specialists |
Shift-left lane: catch friction before it reaches the gates. /ai-verify is the gate; /ai-review asks "would a staff engineer approve this?".
Boundaries
- Never modifies code — advisory only.
- Never blocks execution — fail-open always.
- Never emits FAIL / BLOCK / CRITICAL — reserved for
/ai-verifyand git hooks. - Read-only for all files except
decision-store.json(drift annotations only, via the audit API) andstate/framework-events.ndjson(canonical outcomes). - Single agent dispatch — never reads the agent file inline; dispatches via the Agent tool so the agent runs in its own context window.
Examples
User: "advise on the changes I just made under src/auth/"
/ai-advise advise src/auth/
Dispatches the ai-advise agent in advise mode scoped to src/auth/: it loads cross-cutting + Python-stack standards, checks decision drift against decision-store.json rows intersecting src/auth/, and returns severity-tagged warn findings. The skill renders the advisory and emits a framework_event. No code modified.
Integration
Called by: operators via /ai-advise; /ai-build + /ai-autopilot (wave-end advisory pass). Calls: the ai-advise agent (.claude/agents/ai-advise.md) via the Agent tool — Agent-tool dispatch is the primary path. Inline fallback: on a harness WITHOUT a subagent/Agent-tool primitive, this skill is executed by reading the .claude/agents/ai-advise.md specialist file inline and running its steps in-context sequentially — inline-sequential is the floor, not a substitute for the isolated dispatch. See also: .claude/skills/ai-verify/SKILL.md (evidence-backed BLOCK lane), .claude/skills/ai-review/SKILL.md (narrative review), .ai-engineering/overrides/<stack>/conventions.md (stack overrides); D-134-06 (ai-guard agent → ai-advise rename), D-134-07 (cohesion test enforcement).
$ARGUMENTS