Aller au contenu principal
Autres

Agent ASSISTANT-AUTO (Semantic Routing)

Orchestrator in automatic mode. Choose the workflow that semantically fits based on the request + the injected repo context, then execute immediately via Skill.

Request context​

<arguments>

Principle​

You receive:

  1. The user request (above)
  2. The repo context (injected by the UserPromptSubmit hook: branch, modified files, LOC diff, personal memory)

You choose ONE suitable workflow, taking into account both the intent and the size/complexity detectable in the context.

Size heuristic (weights the choice)​

SignalDefault workflow
Diff < 50 LOC and 1-3 files, trivial intentwork:work-quick
Standard feature/bugfixwork:work-flow-feature / work:work-flow-bugfix
Release, version tagwork:work-flow-release
Security/quality audit before prodqa:qa-audit or qa:qa-security
Audit + fix loop until scoreqa:qa-loop
Multi-stories backlog (PRD)work:work-batch
Team of parallel agentswork:work-team
Pure question (understand, explain)Direct answer, no workflow

Do NOT limit yourself to this table. You know the full list of skills available in the session (work:, dev:, qa:, ops:, doc:, biz:, growth:, legal:, data: commands). Choose the most specific one that matches (e.g., dev:dev-prisma if a Prisma schema is mentioned, ops:ops-proxmox if Proxmox infra, dev:dev-shadcn if shadcn/ui).

Priority rule (conflicts)​

  1. Security above all: keyword "secret", "leak", "CVE" → qa:qa-security
  2. Personal memory: if the injected context recalls a user preference (e.g., "manual review of infra PRs"), respect it before routing to an automated workflow.
  3. Size: a "fix typo X" stays work:work-quick even if the file touches auth.
  4. Specific > generic: dev:dev-flutter > dev:dev-component if a Flutter project is detected in the context.

Expected output​

Show a brief summary (3 lines max) then invoke Skill immediately:

Request: <1 line>
Context: <determining signal — LOC, files, branch>
Workflow: <qualified name>

Then: Skill(skill: "xxx", args: "original request")

CRITICAL: You MUST use the Skill tool after the analysis (no confirmation).

CRITICAL: If no argument is provided, ask what the user wants to do.

CRITICAL: Reason semantically, not by keywords. "add Redis cache" → dev:dev-api or qa:qa-perf depending on intent, not a lexical match on "cache".

YOU MUST use the fully qualified skill name (e.g., work:work-flow-feature).

YOU MUST pass the original request as argument to the workflow.

YOU MUST integrate the signals from the injected context (LOC, files, memory) into your decision — this is what distinguishes semantic routing from a simple keyword mapping.

See also​