Phase 4 — Implementation

The story cycle. Sprint plan → prepared story → code + tests → layered review → repeat. This is where most of your time is actually spent — and where you choose how much of it to spend supervising.

Owner: Dev (e11-agent-dev). The sprint/story skills are invoked directly — the SM and QA agents are retired (where their duties went).

Workflows in this phase

Code	Workflow	Purpose	Output
`[SP]`	`e11-sprint-planning`	Generate/update the sprint board (owns the `sprint-status.yaml` schema)	`{implementation_artifacts}/sprint-status.yaml`
`[SS]`	`e11-sprint-status`	Summarize the board, route to the next workflow	(in-session)
`[CS]`	`e11-create-story` (Create)	Prepare a story with full implementation context	`{implementation_artifacts}/stories/…`
`[VS]`	`e11-create-story` (Validate)	Validate story readiness before Dev starts	validation report
`[DS]`	`e11-dev-story`	Execute story tasks + tests, capability-driven alloy pre-load	code + tests, updated story file
`[CR]`	`e11-code-review`	Five-layer adversarial review + structured triage — spec optional	findings, triaged; deferred items → `deferred-work.md`
`[CP]`	`e11-checkpoint-preview`	Human-in-the-loop change walkthrough (deep page)	guided review + trail file
`[QA]`	`e11-qa-generate-e2e-tests`	Generate API + E2E coverage with the project's own framework	tests in the project tree
`[ER]`	`e11-retrospective`	Epic-end review, run on the multi-perspective roster	`{implementation_artifacts}/retrospective-epic-<n>.md`
`[CC]`	`e11-correct-course`	Continue / update PRD / redo architecture / restart decision	`{planning_artifacts}/sprint-change-proposal-<date>.md`
`[DF]`	`e11-dev-flow`	Quick build/fix/refactor without story ceremony (Quick flow)	spec + code
`[DA]`	`e11-dev-auto`	One fully unattended story iteration (deep page)	code + spec + `e11-dev-auto-result.json`

Standard story cycle

text

[SP]      → Sprint board (once per sprint)
   │
   ▼
[CS]/[VS] → Story prepared: tasks, ACs, SDK stamp from Phase 3
   │
   ▼
[DS]      → Capability-driven alloy pre-load, tasks IN ORDER,
            tests pass before any [x]. "Alloy pre-load complete"
            recorded in the Dev Agent Record.
   │
   ▼
[CR]      → Fresh context. Five layers in parallel → orchestrator
            triage → decision-needed / patch / defer / dismiss.
   │        Patches applied → re-verify; defers → deferred-work.md.
   ▼
[CP]/[QA] → Optional: guided human walkthrough / extra E2E coverage.
   │
   ▼
[CS]      → Next story…    [ER] at epic end.

Prefer zero supervision? [DA] runs clarify → plan → implement → self-review unattended and HALTs done or blocked — see Unattended runs.

Code review: layers + triage

e11-code-review works with or without a spec: a five-tier target cascade (explicit ref → active story → recent artifacts → git state → ask) finds what to review; with no spec it runs in no-spec mode and reclassifies would-be "decision-needed" findings. The layers are data ([[workflow.review_layers]] in customize.toml), six by default:

blind-hunter — e11-review-adversarial-general, no context, pure hostility.
edge-case-hunter — e11-review-edge-case-hunter, boundary walk, E11 also_consider conditions passed in.
verification-gap — e11-review-verification-gap, "what could regress without a test catching it?"
reachability — e11-review-reachability, "deployed today, what would not work — and is each gap owned?" Activation nobody performs, a runtime read path nobody asserts, a capability nobody registers. Screens itself out on changes with no deployed surface.
acceptance-auditor — spec mode only: diff vs ACs.
alloy-auditor — e11-review-alloy-auditor: convention compliance per @engineering11/* import (checked against skill:e11-knowledge), pre-load/citation verification, skip-honesty audit, alloy:// cite resolution.

Finder skills never assign severity — the orchestrator rates low/medium/high only after reading the surrounding code, dedups same-claim+same-fix findings, and routes each into one of four buckets: decision-needed / patch / defer / dismiss. If any layer fails, a zero-finding run warns instead of reporting clean (failed_layers tracking). Disable or add layers via override — see Customization.

When to detour

Dev halts on an alloy gap. [R]ead the package source / [S]kip the convention check (visibly) / [A]scalate to the Architect. No "press on and guess" option — by design.
Tests fail mid-[DS]. The task stays [ ]. Fix the test or the code; don't disable the test.
Something is materially wrong. [CC] early — three stories in is cheaper than ten.
Review in the authoring context. Don't. Fresh session, ideally a stronger model.
Story too small for the cycle. That's [DF] — see Quick flow.

Hand-off

No Phase 5 hand-off. At epic end [ER] runs the roster over what shipped — including a platform-debt pass (alloy gaps and Tier C follow-ups accumulated during the epic) — and routes to the next epic or [CC].

Source

Workflow SKILL files:

Phase 4 — Implementation ​

Workflows in this phase ​

Standard story cycle ​

Code review: layers + triage ​

When to detour ​

Hand-off ​

Source ​