Panda Suite Docs
v1.9.2GitHub ↗
Panda Suite / Architecture

Panda Suite architecture

Panda Suite is twenty-four independent instruction skills. Panda Lead is a small router; the host model and its existing tools perform work. There is no server, database, scheduler, agent loop, or programmatic skill invocation API.

The supplied briefs are the implementation specification. The initial catalog and the five essential expansion skills are in scope.

Decisions

  1. Use Markdown skills with progressive references. Each specialist can be copied alone and carries its own workflow, outputs, limits, and essential authorization rules. No mandatory shared file outside the skill directory.
  2. Route by requested artifact and authorization. Exploration, execution, and delivery describe user expectations; they are not runtime services. A focused explanation is answered directly. Explicit specialist invocation wins.
  3. Keep upstream skills unchanged. Vendored files retain exact bytes and original identity. Additional provenance and evaluation notes live outside their instructions. See research and integration.
  4. Use native host capabilities. Load selected instructions through documented discovery; delegate only when available, authorized, and useful. Reading a file is described as instruction loading, not a simulated tool invocation.
  5. Use one portable manifest and a Claude compatibility manifest. Both point to the same twenty-four folders. A Codex fallback is added only if local testing demonstrates that this installed CLI needs it. No MCP or hooks are needed.
  6. Evaluate claims at the correct level. Validators prove package invariants; real fixture tasks test behavior. No performance or token-saving claim follows from scenario presence, word counts, or a single successful model response.

Boundaries

Brainstorm explores problems and options. Plan defines ordered engineering work. Spec defines technical contracts and current/proposed behavior. Execute changes authorized source. Review identifies defects. Simplify assesses unnecessary complexity. Write documents engineering evidence. Manual documents observed user journeys. Debugger owns causal investigation. Buddhist Method is consulted at decision points; it is never a compulsory pipeline stage.

Skill dependencies are optional and flow outward from Lead to specialists. Specialists do not route back to Lead or require each other. A compact handoff preserves evidence, artifacts, decisions, unknowns, next action, and authorization. Only a material scope, requirement, or risk change needs a new decision from the user when implementation is already authorized.

Refactor owns bounded structural improvement through baseline, incremental edits and compatibility verification. Simplify retains general complexity assessment; Plan, Execute, Test and Review retain their focused deliverables. Direct Refactor requests can stop at analysis or planning. Peer guidance is optional, with no mandatory chain or new runtime. See the Refactor design and evaluation.

Learn owns understanding unfamiliar existing code through source tracing, adaptive teaching and optional practice. Research owns bounded evidence questions; Write owns requested engineering documents; Debugger owns causal failures. A tiny explanation stays direct unless a specialist is explicitly selected. Learning does not authorize application edits or execution. See Learn design and trials.

Setup owns project initialization, existing-repository preparation and setup readiness audits: current-version selection, minimal structure, agent instructions, ignore rules and verified commands. It does not start a lifecycle for later feature work. See Setup design and evaluation.

Craft is an optional explicitly requested composition workflow for selected Panda or external specialists. Lead retains general workflow selection and coordination; Craft focuses on their shared outcome contract, handoffs and acceptance evidence. Neither is a prerequisite for the other. This overlap is intentional; added behavioral benefit remains an evaluation question. See Craft design and checks.