al-refine

name	al-refine
description	One `status: ready` task to a fresh Test Specification or Verification Plan for AL/Business Central. Technical task -> ready-for-implementation. Verify task -> ready-for-verification.

Style: Concise — cut filler, keep grammar. Opinionated — pick a side. Arrows (→) for causality. Technical terms exact, code and errors quoted verbatim.

/al-refine, Task to Test Specification / Verification Plan

Fill one named status: ready task in the tasks/ folder. Existing proof content is untrusted; regenerate from current app/tests, ground symbols, sharpen intent, write the task body. One task per run.

Branch by kind: in the task file's frontmatter:

• technical → fresh Test Specification, flip ready → ready-for-implementation.
• verify → fresh Verification Plan, flip ready → ready-for-verification.
• provision / breaking-change → not a refine target: no proof artifact, leave status: ready, notify "ops task → run /al-provision" or "ops task → run /al-validate-breaking-changes".

Layer. Authors the Test Specification / Verification Plan each pyramid layer verifies (see test-strategy.md) using the grammar in test-specification.md.

Preconditions

• Branch matches ^\d{3}-. If not: Stop. Run /al-event-model (or /al-design for backend-only).
• Spec folder holds architecture.md. Missing → Stop, run /al-design.
• User/API-facing features: event-model.md also present.
• Target task is named and has status: ready (context exists for /al-refine only). One nuance, kind: technical only: a technical task blocked with all depends_on: done, no replan flag in the body, and its own slice already open (at least one same-slice: sibling is not blocked) → the upstream close just didn't flip it; open it ready here and proceed, no /al-steer bounce. But if every task in its slice: is still blocked, the slice itself has not been opened — that open is the prior slice's cross-slice gate still to be earned (a clean /al-code-review, or a user/API-facing prior slice's verify pass), not a forgotten flip: Stop, Next: /al-steer to name the pending gate; do not self-open, or the review gate is skipped. A kind: verify task gets no self-open — it reaches ready only when /al-code-review opens it on a clean per-slice review (carrying review: clean). A blocked verify task with all technical deps done and no review: clean is slice-done owing its first code review → Stop, Next: /al-code-review T-NNN; do not self-open and write the plan, or the review gate is skipped. blocked on an unsatisfied edge or a replan flag → /al-steer. A done task carries downstream evidence; reopen only through /al-steer.
• Verify task (kind: verify) but no event-model.md → contract violation, Stop, route to /al-steer. Verify tasks only exist for user/API-facing features.
• Read test-specification.md, test-strategy.md, test-layout.md (the Unit-vs-Integration scope call is the placement rule there — a case whose codepath needs real BaseApp behaviour cannot be scoped Unit), voice-contract.md, and markdown-spec-discipline.md before writing.
• Scan .not-yet-specified/*.md at repo root when present — the deferred-question ledger /al-steer grooms. A question the task's behaviour touches means the spec would encode a guess on a question the user already flagged as undecided: Stop, name the question file, route to /al-steer (or /al-grill-adr for a domain rule) to graduate it first. No question touches the task → proceed; the scan costs one folder read.

Regenerate proof

Preserve scope-time context: title, description, the depends_on: / refactors: / fixes: frontmatter edges, slice:, constraints, risks, source context, acceptance intent. Rewrite the proof section whole. Do not preserve stale Test Specification, Verification Plan, AAA cases, coverage tables, examples, or charters because they existed in the task body.

Technical task: Test Specification

Spec how one technical task's behaviour is proved so /al-implement can drive red → green without re-deriving intent. Answer before writing:

• What is the task delivering? Resolve from task description, slice context in architecture.md, event-model.md when present, CONTEXT.md, and codebase.
• Is there meaningful branching? No branching → Expected Behaviors with B# IDs. Branching rule, policy, calculation, status combination → Decision Matrix with R# IDs. Multiple unrelated groups → split or route to /al-steer.
• What is each AAA case's scope? Unit for AL-Runner decision proof; Integration for container, database, event, TestPage, posting, install, permission, or wiring proof. /al-refine proposes scope. /al-implement may change it and must reconcile the task file. Every Integration case is a push-up from Unit — see Surface push-ups.
• What procedure names should exist? Propose short PascalCase AL test procedure names. These populate Covered By, AAA headers, and Procedure:.
• What order? AAA cases list Unit first, then Integration; within each scope, coverage ID order.
• What does codebase actually expose? Real codeunits, tables, fields, pages, procedures, events, and APIs on the boundary. Every exact name rests on current evidence.
• Which objects and signatures does the task land? Write New and Modified Objects per the grammar in test-specification.md; /al-implement consumes these signatures instead of minting names mid-TDD. Seed New: vs Modified: from architecture.md's new / extends markers, then override by workspace state at refine time — an object an earlier task already landed is Modified: regardless of its design-time marker. Architecture silent on a needed object → mint it when it serves a listed slice slot; a missing slot is trigger #6, route /al-steer.
• What vocabulary names the coverage? Project terms from CONTEXT.md first, then BC display labels, then exact AL names only when traceability or ambiguity requires them.

Unanswerable → cannot write the spec yet. Flip or keep status: blocked and resolve via /al-research (BC behaviour), /al-grill-adr (domain rule), /grill-me (intent the user must adjudicate), or /al-steer (replan).

Verify task: Verification Plan

Spec how the slice is checked through its user/API-facing surface. Write only subsections that apply:

• Journey Examples for Scope: E2E, each marked Record: yes or Record: no (see below). /al-page-script records the Record: yes set; /al-user-verification walks the Record: no set.
• Contract Examples for Scope: Contract. Name the client or harness.
• Exploration Charters for Scope: Exploration. Charter plus 2-4 prompts; no exact click script.

Answer before writing:

• Which slice does this verify? Read slice: in the task file's frontmatter, resolve to event-model.md timeline step. Title + description quote that step's Role, Action, Business Event, View, Status vocabulary.
• Which surface is exercised? BC Web Client, API endpoint, Postman collection, curl, integration harness, or another named client. Name the surface inline so downstream skills do not guess.
• Which examples cover checkable outcomes? UI workflow → at least one E2E journey. API/client slice → at least one Contract example. Each has action bullets and observable-check bullets.
• Which E2E journeys need a recording? Mark each E2E example Record: yes only when no AL test layer can automate the behaviour (control add-in, canvas, web-client-only behaviour) — the generation-time push-down call (see test-strategy.md). Behaviour a Unit or Integration test already pins (or that a technical task in this slice should pin — push it down via /al-steer rather than record it) is Record: no: walked for acceptance, never recorded, because a recording that doubles a lower test is pure cost. Most examples are Record: no; a slice with zero Record: yes skips /al-page-script.
• Which exploration is useful? Add charters for new workflows, major workflow changes, and error/user-guidance changes. Exploration findings become tasks unless a functional failure is observed.
• Which event-model.md slots are cited? Every Role / Action / Business Event / View / Status name in a verify example is backed by grep against event-model.md or workspace lookup on the underlying BC surface.

Unanswerable → cannot write Verification Plan yet. Flip or keep status: blocked and resolve via /al-research (BC surface), /grill-me (intent), or /al-steer (wrong slice boundary or missing prerequisite).

Surface push-ups, commit nothing

A test scoped above the deepest layer that could check the behaviour is a push-up (see test-strategy.md): every Integration case, every Record: yes E2E, every Contract example. For each, record a Contract notes line — why the layer below cannot hold it + the named seam from testability.md that would reach it, or the wall — and emit the Push-up report (voice-contract.md) as its own chat section — each line passes pre-send check 3: why the layer below cannot hold it, in named objects, actionable without this session. /al-refine only proposes: nothing is written but the task file, so this surfaces the tax without gating — the handoff stop is the user's review point, and /al-implement is where a push-up is actually committed. Record: no E2E (already pushed down) and Exploration (no checkable floor) are not push-ups and carry no justification.

Ground exact names

Every exact BC-specific symbol in a Test Specification or Verification Plan meets the evidence bar in voice-contract.md: workspace hit this session or quoted fetch; conflicts and design-artifact facts escalate to /al-research. Names already cited by /al-design or /al-event-model count only when grep against those upstream files returns the name this session.

Minted names in New and Modified Objects meet the bar's minted-name clause in voice-contract.md.

Artifacts stay clean; chat carries the citation. Researched: bullets land in Contract notes — refine's at write, /al-implement's at reconcile.

Sharpen vague language inline

When a domain rule is implicit, when edge discovery surfaces a case the user must adjudicate, when an upper bound is missing, when a boundary contradicts another rule, or when intent splits (validate as schema check vs. business rule check) → run /grill-me. Fuzzy language shipped to /al-implement becomes fuzzy code; fuzzy verification becomes weak sign-off. Replace vague phrasing with exact object + procedure + status names inline.

Second opinion on non-trivial plans

Cross-check via /al-second-opinion. Prompt body for technical tasks: task title + description + proposed Test Specification + CONTEXT.md language excerpt if resolved + "what behaviours, decision rows, negatives, boundaries, scopes, procedure mappings, or object/signature landings are missing or wrong? AND does this surface any of the eight replan triggers? AND does wording use project vocabulary where applicable? Return a bulleted list." Prompt body for verify tasks: task title + slice context from event-model.md + proposed Verification Plan + "what user-facing journeys, contract checks, exploration prompts, boundaries, or exception paths are missing or wrong? AND do examples name real surfaces? AND does this surface any of the eight replan triggers? Return a bulleted list."

Reconcile each returned bullet; accept by updating or reject with session rationale. If a rejection encodes a durable principle, escalate via /al-steer to /al-grill-adr or /al-design.

Numbering and handles

Technical coverage IDs are B# for Expected Behaviors and R# for Decision Matrix. Verify IDs are V# for E2E, C# for Contract, and X# for Exploration.

Stable handles:

• Technical: AL test procedure name in Covered By / Procedure.
• Verify: example ID + title, e.g. V1 BlocksReleaseFromSalesOrderPage.

Document verification

After writing the fresh Test Specification or Verification Plan, run the document-integrity check yourself, inline (no subagent), before flipping status: — verify the task file against doc-integrity.md, scoped to this T-NNN: the proof-section check (a populated Test Specification needs a New and Modified Objects block) and, for a verify task, the E2E Record: flag check.

A fail (structural or boundary blocker) blocks the ready-for-implementation / ready-for-verification flip; fix it or route to /al-steer. A warn does not block; include it in the handoff. This gate checks document integrity only, not whether the planned proof is sufficient.

Status flip

After writing and reconciling the fresh proof section, locate the task file by its T-MMM filename and flip its status: frontmatter line only:

technical: status: ready → status: ready-for-implementation
verify:    status: ready → status: ready-for-verification

No in-progress state.

Advisor checkpoint. Before writing the first Test Specification or Verification Plan into the task file, do a final shape check — coverage, the New/Modified split, the push-up set. For non-trivial proof, /al-second-opinion gives an independent cross-family read; the shape is hard to retract once downstream skills consume it.

Next step

Read off the flipped task — the handoff passes the pre-send checks in voice-contract.md:

• Technical task → ready-for-implementation: Next: /al-implement T-NNN — drive the Test Specification red→green.
• Verify task → ready-for-verification: the verify task arrived carrying review: clean (/al-code-review ran at slice-done and opened it); the status:-only flip preserves it. Next: state-conditional on the slice's recordings — /al-page-script T-NNN (a Record: yes Journey Example's recording missing) or /al-user-verification T-NNN (all present, or no Record: yes example). The review already ran — do not route back through /al-code-review.
• Stayed/flipped blocked: Next: /al-research (BC fact), /al-grill-adr (domain term), or /al-steer (replan) per why it would not ground.

If state can't be read, fall back to /al-implement for a technical task; for a verify task (which reached /al-refine only because /al-code-review already ran) → /al-user-verification, or /al-steer when unsure — never back through /al-code-review.

Composition

Runs after	/al-scope or dependency restoration opened one named task to status: ready
Hands off to	/al-implement for ready-for-implementation technical tasks; for ready-for-verification verify tasks, state-conditional on the slice's recordings — /al-page-script (a Record: yes recording missing) else /al-user-verification (/al-code-review already ran at slice-done; the verify task arrives carrying review: clean)
Calls directly	/al-research (BC facts), /al-second-opinion (non-trivial Test Specification / Verification Plan) — the only skills it invokes
Replan venue	/al-steer
Sidebands	/al-grill-adr (fuzzy domain term), /grill-me (fuzzy intent)