moonlight-lupin/agent-skills / plugins/dev-pipeline/skills/orchestrator
orchestrator
VerifiedUse when the user asks how to approach building a feature or says 'build X', 'implement X', 'add feature X' without naming a phase ('SDD it', 'spec-driven development', 'run the dev pipeline', 'what phase comes next') — the dev-pipeline stage router: complexity S0-S3 and risk R0-R3 first, then grill → spec (+clarify, plan, tickets) → test-evals → build → review-commit (two-axis pass, budgeted confirm rounds, commit/deploy gate). Owns risk-based review routing, complexity-aware planning and the model lanes.
| File | Kind | Size |
|---|---|---|
| SKILL.mdexecutable | skill md | 6.6 KB |
| references/coding-principles.mdexecutable | script | 5.3 KB |
| references/flow-map.mdexecutable | script | 1.4 KB |
| references/lanes.mdexecutable | script | 3.2 KB |
| references/risk-routing.md | markdown | 5.2 KB |
| references/size-tiers.mdexecutable | script | 2.1 KB |
references/risk-routing.md
View on GitHub# Routing — complexity and risk (single home)
This replaces size-driven review depth with **two independent classifications**.
Use this file for routing decisions. Older stage documents may say "tier";
that means the *effective review tier* below, not the complexity band.
## Step 0 — classify both axes, before building
**Complexity S0–S3** selects planning and task-decomposition work.
**Risk R0–R3** selects independent review depth. State both and the derived
review tier. Do not infer safety from a small diff. Reclassify when the actual
diff or exposed boundary changes; never change either axis silently. Lowering a
classification requires the user's consent.
| Band | Complexity signal | Planning path |
|---|---|---|
| S0 | Trivial, one obvious change, ~20 lines or less | Intent/assumptions, edit, verify |
| S1 | Small, one behavior/seam, ~300 lines or less | Short acceptance-criteria plan, RED test if behavior changes |
| S2 | Multiple behaviors/seams, >300 lines, or >5 files | Spec, plan, named test cases, task decomposition |
| S3 | Large, cross-component, multi-team, or difficult migration plan | Full spec, dependencies, milestones, task slices and integration checks |
Take the highest applicable complexity signal. Complexity does **not** by
itself increase the number of reviewer lanes or confirmation rounds.
| Band | Exposure / consequence (highest matching risk wins) | Review tier |
|---|---|---|
| R0 | Documentation, metadata, mechanical change; no observable behavior change and no sensitive boundary | T0 for S0; otherwise T1 |
| R1 | Bounded internal behavior, trusted inputs, reversible changes, no outside trust boundary | T1 |
| R2 | Untrusted/externally supplied data, user-controlled parsing, resource exhaustion, nontrivial concurrency or filesystem access without established security isolation | T2 |
| R3 | Authentication, authorization, permissions, secrets, money, irreversible data loss/migration, trust-boundary escape (e.g. symlink/path traversal), public release or live deployment | T3 |
For ambiguous external exposure, choose the higher risk until evidence rules
it out. A five-line auth edit is R3. A 1,000-line internal refactor may be
S3/R1. A CSV importer consuming user-provided files is at least R2 even if
its diff is only ten lines.
## Derive effective review tier (compatibility contract)
Downstream skills continue to read T0–T3 *review tiers*; their existing
reviewer lanes, confirm-round budgets, and live-deploy three-reviewer gate
remain unchanged. Do **not** take max(complexity, risk): complexity selects
the planning path, risk selects review depth.
| Risk | Effective review tier | Step A | Step B reviewer lanes | Confirm budget |
|---|---|---|---|---|
| R0 / S0 | T0 | none | 0–1 | 0 |
| R0 / S1–S3 or R1 | T1 | two-axis pass | 1 (A) | 1 |
| R2 | T2 | two-axis pass | 2 (A+B) | 2 |
| R3 | T3 | two-axis pass | 3 (A+B+C) | 3 |
The full acceptance suite MUST run again after each fix before a review
finding is closed. Add focused regression tests for reproduced failure modes;
check the net diff after fixes, not only the specific repaired lines.
An exhausted budget with open BLOCKING/MAJOR findings is **NOT SHIP**;
stop and ask the user whether to redesign, spend another round, or explicitly
accept the residual risk. A test failure introduced by a fix cannot be
waived merely because the original finding was resolved.
## Stage routing: complexity first, risk reviews afterward
- S0: edit + verification, with risk-appropriate review (S0/R3 is still T3).
- S1: short plan + test-evals (behavior changes) + build + review-commit.
- S2/S3: grill if fuzzy, spec, plan, test-evals, build, review-commit.
- Goldfish spec review: offer for S2/S3 after the approved plan; it is a
planning aid, NOT a reviewer lane or a substitute for security checks.
- Per-task review and confirm budgets come from the effective T-tier, NOT
from S-size. For large S2/R1 work, separate planning tasks but use T1
review depth.
- User-requested review counts or models may add scrutiny beyond the default
depth; they may not reduce required security review or remove the
three-reviewer gate. Lowering a risk classification always needs the user's
consent. Any live deploy requires the three-reviewer gate whatever the
requested depth.
- A new or discovered risk boundary upgrades R immediately. Do not silently
downgrade R after a review passes.
- Converge loop eligibility and caps follow complexity, not the T-tier:
S0/S1 skip the loop; S2 caps at 2 rounds, S3 at 3
(see `build/references/converge-loop.md`).
## Empirical basis and caution
Four paired pilots found costly T1/T2 review did not help in ordinary tasks,
while the two attacker-facing tasks benefited from deeper review. This is a
directional sample, NOT proof of a general effect. Preserve the risk-based
heuristic as an explicit, revisable policy rather than promising that review
can catch every bug.
## Legacy note
T0–T3 are retained as *review depth aliases* for compatibility with
`build`, `review-commit`, and review-lane reference documents. Older wording
that says size selects T-tier is superseded by this document. Do not change
their model pins or require new CLI capabilities merely to adopt routing.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the work | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixes | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation code | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |