| name | lakebase-sftdd-workflows |
| description | Spec-First Test-Driven Development (SFTDD) with evolutionary design, against paired Lakebase branches: the design lane is Spec Driven Development (SDD) and the build lane is canonical Beck-style Test Driven Development (RED-GREEN-REFACTOR), composed with paired-branch primitives (cheap experiments, parent-aware schema diff, real per-branch databases). Spec-first means the spec is drafted, reviewed, and frozen at gates before any build cycle runs; evolutionary means the spec, the architecture, and the database all evolve increment over increment. Use when planning a new feature, running design-spec gates, running TDD cycles, comparing parallel experiments, or detecting workflow bad smells. Imports software-design-principles canon. Builds on lakebase-scm-workflows + lakebase-release-workflows. |
| user-invocable | true |
lakebase-sftdd-workflows – agent contract
Agent-facing contract: hard rules, phase flow, agent prompt index, and concrete code patterns for the substrate primitives.
For the human-facing overview (lexicon, roles narrative, how-to-use prompts, project entry points) see README.md.
Hard rules
The contract every agent (Navigator, Driver, Orchestrator) and every human collaborator must honor.
- Tests are immutable until the test list itself is renegotiated through the PO. Never delete or weaken a test to make it pass.
- "Minimal code" means minimal honest code that satisfies the current test list, not just the current test. Use the test list as your horizon.
- After every GREEN, ask: "would a fresh reader infer the right concept from this API now?" If no, request REFACTOR before the next test.
- Test at the outermost public boundary that maps to the AC. Inner-loop unit tests are reserved for pure logic that can't be exercised through the outer boundary.
- A correct refactor should not change the outer-boundary tests. A refactor that requires editing tests is suspect.
- Never make a private method public to test it.
- Test count is a lagging indicator. The leading indicator is "how cheap is the next test?" Rising cost = design problem.
- Spike code is throwaway. Promote nothing from a spike branch into a TDD branch except notes.
- Experiments are scoped to a story (the branch forks from feature HEAD). N=1 is one experiment per story (iterative refinement, no promote/synthesize ceremony); N>=2 races competing strategies for that story. The PO accepts a story's experiment by merge (git-merge + run its migrations against the feature branch DB), or discards/revises it.
See agents/navigator.md and agents/driver.md for per-role specializations.
Phases and gates
The kit is Spec-First Test-Driven Development (SFTDD): it composes two disciplines back to back. The design lane is Spec Driven Development (SDD) – spec drafting, architectural review, test-list construction, all driven by /design – and the build lane is Test Driven Development (TDD) – the RED → GREEN → REVIEW → REFACTOR cycles driven by /build. Spec-first means SDD freezes the spec at the spec + test_list gates; TDD only starts once they are approved, so the build always runs against a reviewed spec.
Iterative and evolutionary. The freeze is per increment, not forever. Across increments, all three dimensions evolve and are never treated as static: the spec and features are living , each sprint the PO re-plans, folding in what the last working software revealed; the architecture evolves under fitness functions (@architectural-design-principles/evolutionary-architecture.md); and the database evolves by migration on the paired branch, diffed against its parent (@architectural-design-principles/evolutionary-database-design.md). The rule is: spec-first within an increment, evolutionary across increments.
Gates are keyed (not numbered): spec / plan / test_list / promote / deploy. The SDD lane runs PER STORY (it streams), and experiments + plans are per-story.
| Step (per story unless noted) | Output | HITL gate |
|---|
| Spec drafting | feature-spec.{md,json} (feature) + per-story story.{md,json} + acs/<AC>.{md,json} | spec gate |
| Architectural review | layer + architectural_notes on each AC; architecture.{md,json} (holds the NFRs) | folds into the spec gate |
| Test-list construction | per-story test-list-per-story.json (scoped from the feature test-list.{md,json}) | test_list gate |
| Build (per AC) | cut a per-story experiment; RED -> GREEN -> REVIEW -> REFACTOR cycles producing tests + code | Continuous: smells; per story: accept (merge) / discard / revise |
| Deploy | working software on the experiment branch (reachable + verify) | deploy gate |
Sprint planning has its own plan gate. For N>=2 experiments within a story, the menu-pick promote/synthesize decision is recorded under synthesis/<F>/. Refuse to transition if prior-step artifacts are missing or invalid.
Enhancement review (placeholder, TODO)
Enhancements are decided iteratively. After a feature ships, new work is proposed and prioritized with the Product Owner during sprint planning (/plan) - the same gate that commits any sprint - so the loop keeps returning to the PO for the next increment.
TODO (not yet implemented): add a review role that inspects the running, deployed software (post-deploy) and recommends concrete enhancements back to the PO as candidate features for the next /plan. This mirrors the existing review lens (the Architect Reviewer / Navigator review the spec and code) but aimed at observed runtime behavior rather than source. To flesh out: the trigger point (post-deploy), the inputs (the reachable endpoint + feature-status), the output (enhancement candidates feeding planning/feature-proposals.md), and the PO hand-off. Until this lands, enhancements are surfaced manually at sprint planning.
The phases + gates above are the PER-FEATURE pipeline that /design (phases 0 to 2 + gates 1 to 4) and /build (phase 4) run. They sit inside a larger orchestrated loop:
Orchestrated commands (the dev loop)
Every command is a thin invocation of the deterministic orchestrator driver (lakebase-sftdd-drive), scoped to a phase range. The driver sequences the work and spawns each role as a subagent (claude -p --agent <role>); routing is code, not an LLM. Gates always pause for a human decision: --gates interactive (the default for the live slash commands) stops at each gate so the human answers; --gates proxy (headless / CI) has the Human Proxy answer.
Tier 1: /sprint = plan ─> [PLAN GATE] ─> per feature: /design ─> /build ─> /deploy
▲ │
└──────────────────── working software feeds back ─────────────────┘
Tier 2: /plan /design /build /deploy (run ONE phase, then stop + suggest next)
/spike (throwaway exploration, outside the loop)
/sprint [name] (Tier 1, the top-level orchestrator): runs the whole sprint as one continuous flow, plan to the plan gate, then claim + drive each backlog feature design -> build -> deploy. Re-invoked per cycle; resumable (halts at the next HITL gate for the human, continues on re-run). lakebase-sftdd-drive --sprint <name>./plan [name] (Tier 2, sprint planning, ABOVE the per-feature loop): the Spec Author proposes the candidate breakdown (.sftdd/planning/feature-proposals.md); the Architect t-shirt-sizes the candidates (.sftdd/planning/estimates.json, XS/S/M/L/XL); the Product Owner commits the backlog by authoring a feature-request.md per feature that fits sprint capacity; the deterministic sync-backlog step projects .sftdd/sprints/<name>/backlog.json (committed ids + sizes); the sprint plan gate is the HITL checkpoint. Stops there (does not flow into design). Requires project intake (product-overview.md + nfrs.md, +design-brief.md for UI) as a precondition. --sprint <name> --plan-only./design <feature-id>: the SDD (Spec Driven Development) lane. Claims the paired branch (Step 0), enforces the feature's feature-request.md + project intake (Step 0.5, a precondition, NOT a gate), then drives the per-story design lane (Spec Author -> Architect Reviewer -> Test Strategist) to the spec + test_list gates, producing the executable spec. --only design./build <feature-id>: the TDD (Test Driven Development) lane. RED -> GREEN -> REVIEW -> REFACTOR cycles + per-story acceptance against the frozen spec, to ready-for-review (requires the SDD lane done). --only build./deploy <feature-id> [--target local] [--story <s>]: deploys the merged feature (or one story's branch) + verifies reachable + feature-verify; the deploy gate is the working-software review the PO signs off (the local target is the only one implemented; remote release is the scaffolded merge.yml). --only deploy. For a hands-on review the human can run ./scripts/run-dev.sh to serve the app locally (migrates + hot-reload) and open it in a browser./spike <slug> [--for <feature>]: throwaway exploration on its own paired branch, OUTSIDE the workflow (no gates). Notes carry forward into a feature's design-spec gate; code is never promoted. lakebase-sftdd-spike.
The same orchestrated path runs for real and headless; headless, the Human Proxy stands in for the human at every supply + gate (below).
Headless / Human Proxy mode
By default every gate is HITL (the workflow halts for the Product Owner). When LAKEBASE_SFTDD_HUMAN_PROXY=1 (set by CI and the smoke), the approver role is performed by the human-proxy identity, a diligent stand-in, not a rubber stamp. For the artifact gates (spec/plan/test_list/promote), it approves a gates.json gate and emits gate.approved only when both hold (the deploy gate is certified differently, see the /deploy bullet below):
- Given the artifacts: the gate's expected artifacts EXIST (a missing one is refused).
- Format-conformant: each validates against its declared format (JSON against its schema; narrative MD against its required sections, see
references/spec-format.md + lakebase-sftdd-gate-conformance). A malformed artifact, or one missing a required section, is refused.
So the producing role's job here is to HAND the approver complete, conformant artifacts, recording its recommended resolutions (decisions, NFR acceptances, orderings) INSIDE them rather than leaving open questions for a human reply. A gate advances because real well-formed work was verified, never because it was skipped; a missing/malformed artifact hard-blocks in CI exactly as for a human.
Beyond the gates, the Human Proxy stands in wherever the path needs human input (lakebase-sftdd-human-proxy has two subcommands, supply and approve; both validate-then-place, neither fabricates or skips):
- Project intake (precondition of
/plan + /design): supplys product-overview.md / nfrs.md / design-brief.md from $LAKEBASE_SFTDD_RECORDED_INTAKE_DIR; lakebase-sftdd-intake then passes because they're present + conformant.
/plan backlog: the Architect sizes the candidates live; the Proxy supplys the recorded feature-request.md files (the PO's groomed sprint). sync-backlog projects backlog.json (committed ids + sizes)./deploy gate: confirms the app came up reachable AND the verify passed, then records gate.approved; never approves a non-reachable or failed-verify deploy.
Check the mode with [ "$LAKEBASE_SFTDD_HUMAN_PROXY" = "1" ]. Absent/unset = normal HITL.
Configuration (one source of truth per setting)
Every knob has exactly ONE home; see CONFIG.md for the full table + writers. In brief:
- Project settings (what the project IS:
uiTrack, gates, deployTarget, the per-role model matrix, build cadence) live only in .lakebase/sftdd-config.json, resolved file -> code default by resolveSftddSettings. There is no env or flag override at read time. The writers are create-project (create-time, e.g. --ui-track) and the drive's write-through flags (--gates / --deploy-target / --no-sizing), which persist INTO the file before it is read.
uiTrack is the single door for the UX lane. It drives BOTH the UX Designer (design-guide / ia.md / adherence gate) AND the e2e harness (create-project derives e2e from it, and refuses a UI project without it). A UI project can never run with the UX lane off.- Run-mode knobs (record/replay, headless, debug, e.g.
LAKEBASE_SFTDD_HUMAN_PROXY, _AUTO_CONTINUE, _RECORD_DIR) are per-invocation LAKEBASE_SFTDD_* env vars, read via sftddEnv (one door each). They are NOT project settings and never belong in sftdd-config.json.
- Capture-time conditions live in a scenario's
scenario.json and are funneled into create-project flags by capture-scenario.sh; they never reach the drive directly.
Agent roles (the per-role agent runtime)
Each role is a separate agent definition under agents/ with frontmatter (name, a description that is the auto-selection criteria, least-privilege tools, a strongly-recommended model, memory: project, color) and a body that is its system prompt. The roles communicate only through the artifacts on disk, the artifact is the inter-agent API.
agents/product-owner.md – the PO's facilitator: runs the intake interviews + drafts product-overview.md / nfrs.md / design-brief.md, authors the sprint's feature-request.md files at /plan, and is the approver at every HITL gate. Headless, the Human Proxy plays the PO.agents/spec-author.md – the Spec Author (BA). At /plan, proposes the feature breakdown from product-overview.md + nfrs.md (feature-proposals.md, the PO's input). At /design phase 0, turns one feature's feature-request.md into the structured draft spec (feature-spec.{md,json} + stories + ACs).agents/ux-designer.md – between phase 0 and 1, UI projects only: owns design-guide.{md,json} + ia.md and the UX adherence gate. Skipped for API/CLI/Infra-only features.agents/architect-reviewer.md – phase 1, populates layer and architectural_notes, covers every nfrs.md Required item via architecture.json brief_ref, imports software-design-principles.agents/test-strategist.md – phase 2, builds the Beck-style ordered test list.agents/navigator.md – phase 4 PLAN + RED + REVIEW.agents/driver.md – phase 4 GREEN + REFACTOR.agents/release-engineer.md – /deploy: deploys the built increment to its target, polls reachable, runs the feature verify, hands the evidence to the PO for the deploy gate. Composes on lakebase-release-workflows for remote/release-on-merge.
The orchestrator is the deterministic driver (lakebase-sftdd-drive), not an LLM agent: it routes over workflow-state.json, hands each phase to the right role agent above, carries artifacts forward, and surfaces every gate to the PO. It writes no spec/code/test/deploy.
How the orchestrator runs them. The role defs are scaffolded into the project's .claude/agents/ (so Claude Code can discover + spawn them; the skill copy is the source). The driver computes the next action as a pure function of the recorded state, then spawns the role for that phase via claude -p --agent <role>. Routing is code, not a model: there is no LLM orchestrator session. Before spawning a role, the driver resolves the model from the project's .lakebase/sftdd-config.json via resolveSftddSettings (roles.<role>.model, the single source of truth for project settings). Resolution is sftdd-config.json role model ?? legacy .lakebase/agent-config.json (override ?? recommended) ?? built-in recommended ?? inherit: agent-config.json is honored only as a fallback for projects scaffolded before sftdd-config.json. The HIL sets per-project models at lakebase-create-project; each role's recommended model lives in its definition's model: (mirrored in RECOMMENDED_MODELS).
References
references/spec-format.md – full .sftdd/ directory layout + markdown ↔ JSON contract.references/orchestrator-contract.md – how the agent DRIVING /sprint /design /build /deploy must behave: drive to completion via lakebase-sftdd-next, surface only HITL gates + blockers, report outcomes (not process), verbose/eval narration opt-in. The orchestrator's counterpart to agent-operating-rules.md; the command templates load it.references/next-schema.md – the lakebase-sftdd-next / .sftdd/next.json "what next" surface the orchestrator contract drives on.references/agent-logging.md – structured agent log format + per-role emit points. Every role emits what it is doing via lakebase-sftdd-log (debug = reasoning, info = outputs) to the centralized .sftdd/agent-log.jsonl.scripts/sftdd/schemas/ – JSON Schemas validated by spec-sync.ts.../software-design-principles/SKILL.md – engineering canon (SOLID, DRY, clean code, layered architecture, cross-cutting concerns, NFRs). Required reading for Architect Reviewer and Navigator.
tag → runner map
Every AC declares a layer ("API" / "E2E" / "Infra" in ac.schema.json). The Driver dispatches to the runner that matches the current cycle's layer, not a single uniform npm test. The substrate enforces this: markGreen refuses to advance a layer-tagged cycle until recordRunnerOutcome has logged at least one run for the matching tag.
A test-list item's kind refines this: a kind:"client" item (a UI-presentation AC the architecture routes to the SPA's own harness) runs through the CLIENT toolchain regardless of layer , npm --prefix client test (Vitest component test) or npm --prefix client run test:e2e (Playwright spec), authored under client/tests/. behavior/fitness items run through the layer runner below.
| AC.layer | tag | Default runner | Notes |
|---|
API | api | npm test (Node), ./mvnw test (Java/Kotlin), uv run pytest (Python) | The project's primary test runner. Driver runs it as-is. |
E2E | e2e | npm run test:e2e (Node, alias for playwright test) or uv run --extra dev pytest tests/e2e (Python, pytest-playwright) | Wired by lakebase-create-project --enable-e2e (ships the Node playwright.config.ts + smoke for Node, the tests/e2e/conftest.py live_server fixture for Python). Driver exports BASE_URL at the paired-branch app endpoint before invoking; scripts/run-tests.sh runs whichever matches the project. A missing tests/e2e/conftest.py is a scaffold-defect to surface, never author it in the build. |
Infra | infra | npm run test:infra (alias for lakebase-infra-runner) | Wired by lakebase-create-project --enable-infra. Ships three substrate-side checks: migrations-clean, schema-diff-computable, connection-reachable. JUnit XML output via --junit-output matches vitest's reporter shape. When no runner is wired, the Driver flags the cycle and surfaces to PO; do not silent-skip. |
Each cycle records its runner outcome via recordRunnerOutcome({ scope, cycleId, experimentSlug, layer, passed }). The substrate uses these counts for outcomes.by_tag, the e2e-row-perma-red smell detector, and the design-spec gate guard.
import { recordRunnerOutcome, markGreen } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/run-cycle";
recordRunnerOutcome({ scope, cycleId: c1.cycle_id, experimentSlug: "checkout", passed: true });
markGreen(scope, c1.cycle_id, "added POST handler + repository write");
Operations
Concrete invocations of the substrate primitives.
Design-spec gate (phase 3)
import { analyzeForGate, recordPlan, writePlan } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/design-spec-gate";
import { canCutAnotherExperiment } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/budget";
import { attachSpikeInputs } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/spike-carryforward";
const analysis = analyzeForGate(".sftdd", "F1", "S1");
recordPlan(".sftdd", analysis.proposed_plan, "kevin@example.com");
writePlan(".sftdd", analysis.proposed_plan);
attachSpikeInputs({ tddDir: ".sftdd", featureId: "F1", slugs: ["explore-cart-storage"] });
const ok = canCutAnotherExperiment(".sftdd", "F1");
if (!ok.ok) throw new Error(`budget: ${ok.reason}`);
Experiment (phase 4)
import { cutExperiment, listExperiments, readOutcomes, writeOutcomes, deleteExperiment }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/experiment";
const exp = await cutExperiment({
instance: "proj-checkout",
tddDir: ".sftdd",
featureId: "F1",
storyId: "S1-submit",
experimentSlug: "s1-submit",
branch: "exp/F1/S1-submit",
parentBranch: "feature/F1",
});
writeOutcomes(".sftdd", "F1", "S1-submit", "s1-submit", { status: "succeeded", tests_passed: 2 });
await deleteExperiment({
instance: "proj-checkout",
tddDir: ".sftdd",
featureId: "F1",
storyId: "S1-submit",
experimentSlug: "s1-submit",
deleteBranchToo: false,
});
Spike (side-mode)
import { cutSpike, listSpikes, deleteSpike }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/spike";
import { collectSpikeInputs, attachSpikeInputs }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/spike-carryforward";
await cutSpike({
instance: "proj-checkout",
tddDir: ".sftdd",
spikeSlug: "explore-cart-storage",
branch: "spike-explore-cart-storage",
parentBranch: "staging",
notes: "---\nfor_feature: F1-checkout\n---\n\n# explore-cart-storage\n\nTried postgres arrays. Worked.\n",
});
await deleteSpike({
instance: "proj-checkout",
tddDir: ".sftdd",
spikeSlug: "explore-cart-storage",
});
const inputs = collectSpikeInputs({ tddDir: ".sftdd", featureId: "F1-checkout" });
attachSpikeInputs({ tddDir: ".sftdd", featureId: "F1-checkout", slugs: ["explore-cart-storage"] });
Cycle (RED → GREEN → REFACTOR)
import { beginCycle, markGreen, markRefactored, flagSmells, listCycles, openBranchDsn }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/run-cycle";
const scope = {
tddDir: ".sftdd",
feature_id: "F1",
story_id: "S1",
ac_id: "AC1",
experiment_slug: "checkout",
branch_id: "checkout",
};
const dsn = await openBranchDsn({ instance: "proj-checkout", branch_id: "checkout" });
const c1 = beginCycle({
...scope,
test_id: "T1",
test_description: "POST /orders returns 201 on valid cart",
navigator_plan: "force the public boundary to accept a Cart payload",
});
markGreen(scope, c1.cycle_id, "added POST handler + repository write");
markRefactored(scope, c1.cycle_id, "extracted CartRepository");
flagSmells(scope, c1.cycle_id, ["test-deletion-attempt"]);
Smells (after every cycle)
import { runDetectorsForScope, writeSmellsLog, readSmellsLog, SMELL_CATALOG }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/smells";
const hits = runDetectorsForScope(".sftdd", scope);
if (hits.length) {
writeSmellsLog(".sftdd", hits);
}
Per-experiment cap (every cycle, N≥2)
import { checkPerExperimentCap, recordExperimentCap, clearExperimentCap }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/experiment-cap";
import { readPlan } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/design-spec-gate";
const plan = readPlan(".sftdd", "F1", "S1");
const check = checkPerExperimentCap({
tddDir: ".sftdd",
featureId: "F1",
experimentSlug: "exp-postgres-arrays",
cap: plan?.budget.per_experiment,
cycleCount: listCycles(scope).length,
});
if (check.capped) {
recordExperimentCap({
tddDir: ".sftdd",
featureId: "F1",
experimentSlug: "exp-postgres-arrays",
hit: check.hit!,
});
}
clearExperimentCap({
tddDir: ".sftdd",
featureId: "F1",
experimentSlug: "exp-postgres-arrays",
});
The default cap (30 cycles, 60 minutes wall-clock) is seeded by analyzeForGate. Cycle cap is evaluated before wall-clock cap so a cycle-bound experiment that is also slow returns the max_cycles reason deterministically.
Compare (N≥2 convergence)
import { compareExperiments }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/compare-experiments";
import { renderComparisonReport, writeComparisonReport }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/comparison-report";
const report = compareExperiments(".sftdd", "F1", "S1");
const md = renderComparisonReport(report);
writeComparisonReport({ tddDir: ".sftdd", featureId: "F1", report });
Promote (N≥2, single winner)
import { promoteExperiment }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/promote-experiment";
const result = promoteExperiment({
tddDir: ".sftdd",
featureId: "F1",
storyId: "S1",
winnerSlug: "exp-postgres-arrays",
hitlApproved: true,
approverEmail: "kevin@example.com",
});
Synthesize (N≥2, menu-pick)
import { synthesizeExperiments }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/synthesis";
const result = await synthesizeExperiments({
instance: "proj-checkout",
tddDir: ".sftdd",
featureId: "F1",
storyId: "S1",
picks: [
{ source_slug: "exp-postgres-arrays", capability: "storage schema" },
{ source_slug: "exp-json-blob", capability: "API surface" },
],
synthesizedSlug: "exp-synth",
branch: "checkout-synth",
parentBranch: "staging",
hitlApproved: true,
approverEmail: "kevin@example.com",
});
Spec sync (drift detection)
import { validateSpec, readFeature, writeFeature, readWorkflowState, writeWorkflowState }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/spec-sync";
const drift = validateSpec(".sftdd");
Test list transformation
import { readMasterTestList, writeMasterTestList, viewByAc, viewsForAllAcs, writePerAcViews }
from "@databricks-solutions/lakebase-app-dev-kit/sftdd/test-list";
const list = readMasterTestList(".sftdd", "F1");
writePerAcViews(".sftdd", "F1", list);
Gates state machine (structured HITL approvals)
Design: ADR-0004.
.sftdd/features/<F>/gates.json is the substrate's authoritative gate state. selection-log.md stays as the human-readable narrative-of-record; the substrate dual-writes it at every state change. Agents read gates.json; humans read the log. Never regex-scan the log for state.
Named gates: spec / plan / test_list / promote / deploy. Statuses: open / approved / superseded / withdrawn. Each gate's record carries the approver, timestamp, captured artifact hashes (sha256 of normalized content), and a history[] log of every action.
import {
readGates,
writeGates,
defaultGatesState,
} from "@databricks-solutions/lakebase-app-dev-kit/sftdd/gates";
import { approveGate } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/approve-gate";
import { verifyGateIntegrity } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/verify-gate-integrity";
import { withdrawGate } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/withdraw-gate";
const state = readGates("F1", { tddDir: ".sftdd" });
approveGate({
featureId: "F1",
gate: "spec",
approver: "po@example.com",
hitlApproved: true,
artifactInputs: { "feature-spec.md": specContent, "feature-spec.json": featureJsonContent },
tddDir: ".sftdd",
});
const v = verifyGateIntegrity({
featureId: "F1",
gate: "spec",
currentInputs: { "feature-spec.md": currentSpec, "feature-spec.json": currentFeatureJson },
tddDir: ".sftdd",
});
withdrawGate({
featureId: "F1",
gate: "spec",
approver: "po@example.com",
reason: "scope rewrite",
tddDir: ".sftdd",
});
Per-gate artifact scope (convention enforced at the orchestrator call site, not in the substrate):
| Gate | artifactInputs keys |
|---|
spec | feature-spec.md, feature-spec.json |
plan | plan.json |
test_list | test-list.json |
promote | promote_ref (synthesized string: <winner-slug>:<branch_id>) |
deploy | certified from deploy-evidence (reachable + verify), not a hashed artifactInputs set |
Concurrent-safe: approveGate and withdrawGate serialize through a .gates.lock file (fs.openSync with wx flag). Two callers cannot lose either approval. Override: HUSKY=0 not applicable here; the lock is mandatory.
Brownfield adoption
For features that pre-date the gates state machine (approvals only in selection-log.md):
import { migrateGatesFromSelectionLog } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/migrate-gates";
migrateGatesFromSelectionLog({
featureId: "F1",
tddDir: ".sftdd",
currentInputsByGate: {
spec: { "feature-spec.md": readFileSync("...feature-spec.md", "utf8"), "feature-spec.json": readFileSync("...feature-spec.json", "utf8") },
plan: { "plan.json": readFileSync("...plan.json", "utf8") },
test_list: { "test-list.json": readFileSync("...test-list.json", "utf8") },
},
});
Feature-status integration
feature-status surfaces the compact gates view via the gates field of FeatureStatusSnapshot. See references/feature-status-schema.md. For full state including history + artifact_hashes, call readGates() directly.
Flow patterns
End-to-end orchestrator patterns the deterministic driver runs in response to /design and /build. Each flow is what the driver assembles from the primitives above; the human-facing prompts that trigger each flow live in README.md.
Author + validate spec (response to /design)
import { writeFileSync, mkdirSync } from "fs";
import { join } from "path";
import { validateSpec } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/spec-sync";
const tdd = ".sftdd";
const fdir = join(tdd, "features", "F1-checkout");
const sdir = join(fdir, "stories", "S1-place-order");
mkdirSync(join(sdir, "acs"), { recursive: true });
writeFileSync(join(fdir, "feature-spec.json"), JSON.stringify({
id: "F1", name: "Checkout flow", status: "draft", tdd_mode: "N=1", stories: ["S1"],
}));
writeFileSync(join(fdir, "feature-spec.md"), "# Checkout flow\n\nDesign intent.\n");
writeFileSync(join(sdir, "story.json"), JSON.stringify({
id: "S1", asA: "shopper", iWantTo: "place an order",
soThat: "I receive the goods", feature_id: "F1",
}));
writeFileSync(join(sdir, "story.md"), "# Place order\n\nNarrative.\n");
writeFileSync(join(sdir, "acs", "AC1.json"), JSON.stringify({
id: "AC1", layer: "API", given: "valid cart", when: "POST /orders",
then: "201 with order id", status: "draft", story_id: "S1",
}));
writeFileSync(join(sdir, "acs", "AC1.md"), "# AC1\n\nAC narrative.\n");
const drift = validateSpec(tdd);
if (drift.length) console.warn(drift);
N=1 cycle end-to-end (response to /build, simple case)
import { writeMasterTestList } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/test-list";
import { analyzeForGate, writePlan, recordPlan } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/design-spec-gate";
import { cutExperiment, writeOutcomes } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/experiment";
import { beginCycle, markGreen } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/run-cycle";
import { runDetectorsForScope, writeSmellsLog } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/smells";
const tdd = ".sftdd"; const featureId = "F1"; const storyId = "S1-submit";
writeMasterTestList(tdd, { feature_id: featureId, ordered_for: "design-momentum", items: [
{ id: "T1", description: "POST /orders returns 201 on valid cart", ac_id: "AC1", status: "pending" },
{ id: "T2", description: "POST /orders rejects empty cart with 400", ac_id: "AC1", status: "pending" },
]});
const analysis = analyzeForGate(tdd, featureId, storyId);
recordPlan(tdd, analysis.proposed_plan, "kevin@example.com");
writePlan(tdd, analysis.proposed_plan);
const exp = await cutExperiment({
instance: "proj-checkout", tddDir: tdd,
featureId, storyId, experimentSlug: "s1-submit", branch: "exp/F1/S1-submit", parentBranch: "feature/F1",
});
const scope = { tddDir: tdd, feature_id: featureId, story_id: storyId, ac_id: "AC1",
experiment_slug: exp.experiment_slug, branch_id: exp.branch_id };
const c1 = beginCycle({ ...scope, test_id: "T1",
test_description: "POST /orders returns 201 on valid cart",
navigator_plan: "force the public boundary to accept a Cart payload" });
markGreen(scope, c1.cycle_id, "added POST handler + repository write");
const hits = runDetectorsForScope(tdd, scope);
if (hits.length) writeSmellsLog(tdd, hits);
writeOutcomes(tdd, featureId, storyId, exp.experiment_slug, { status: "succeeded", tests_passed: 2 });
N≥2 race + promote/synthesize
import { cutExperiment, writeOutcomes } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/experiment";
import { compareExperiments } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/compare-experiments";
import { promoteExperiment } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/promote-experiment";
import { synthesizeExperiments } from "@databricks-solutions/lakebase-app-dev-kit/sftdd/synthesis";
const tdd = ".sftdd"; const featureId = "F1"; const storyId = "S1";
await cutExperiment({ instance: "proj-checkout", tddDir: tdd, featureId, storyId,
experimentSlug: "exp-postgres-arrays", branch: "checkout-pg-arrays", parentBranch: "staging" });
await cutExperiment({ instance: "proj-checkout", tddDir: tdd, featureId, storyId,
experimentSlug: "exp-json-blob", branch: "checkout-json-blob", parentBranch: "staging" });
writeOutcomes(tdd, featureId, storyId, "exp-postgres-arrays", { status: "succeeded", tests_passed: 5 });
writeOutcomes(tdd, featureId, storyId, "exp-json-blob", { status: "succeeded", tests_passed: 5 });
const report = compareExperiments(tdd, featureId, storyId);
if (report.recommendation === "promote") {
const winner = report.rows.find((r) => r.signal === "winning")!;
promoteExperiment({ tddDir: tdd, featureId, storyId, winnerSlug: winner.experiment_slug,
hitlApproved: true, approverEmail: "kevin@example.com" });
} else if (report.recommendation === "synthesize") {
await synthesizeExperiments({
instance: "proj-checkout", tddDir: tdd, featureId, storyId,
picks: [
{ source_slug: "exp-postgres-arrays", capability: "storage schema" },
{ source_slug: "exp-json-blob", capability: "API surface" },
],
synthesizedSlug: "exp-synth", branch: "checkout-synth",
hitlApproved: true, approverEmail: "kevin@example.com",
});
}
Adapters
Bundled: markdown.ts (no-op default – the spec IS the tracking), jira.ts (stub). Project skills wire in the adapter they want via .sftdd/adapters/<name>.json config.
The SpecAdapter interface extends SyncEventHooks – implementations may opt into onPhaseTransition, onCycleComplete, onSmellDetected hooks for status-mirroring to external trackers. Adapter failures must degrade gracefully – the on-disk spec is the source of truth.
CLI bins
For non-agent invocation (debugging, CI introspection):
| Command | Purpose |
|---|
lakebase-feature-status <featureId> [--tdd <dir>] [--json] | One-screen snapshot of a feature's TDD workflow state. Use --json for machine-readable payload. |
lakebase-sftdd-next (--feature <F> | --sprint <S>) [--json] | The authoritative, strictly read-only "what do I do next?" surface, computed from the same engine the drive runs on: the reconciled state, the decision menu (the real HIL choices, each with its correct enact command + a prompt to pose), and any blockers. The drive also auto-emits it to .sftdd/next.json on every stop, so an orchestrating agent's contract is "on any stop, read next.json and present its options" instead of improvising. See references/next-schema.md. |
node dist/scripts/sftdd/spec-sync.cli.js <tddDir> | Walk the .sftdd/ tree and print drift reports. Exits 0 even when reports exist (warn-only). |
node dist/scripts/sftdd/test-list.cli.js <tddDir> <featureId> [storyId] | Regenerate per-AC views from the feature-level master test list. With a storyId, instead write that story's scoped per-story test list (stories/<story>/test-list-per-story.json), the streaming build lane's per-story input. |
