Augment It
Search/

← Context/ loops

0.0.0.2Active

Loop through a spec — write the plan, implement, test, changelog, commit, repeat until Shipped

The phase loop that took Augment-From-DB-Flow from Signed-Off to Shipped in one day: plan, implement, prove, changelog, commit — per phase.

context-v/loops/ is an experimental folder (per the context-vigilance skill) and this is its first occupant here. Expect the shape to drift as more loops get codified.

This one was codified retrospectively — the run came first because there was no example to write against. Next time, invert it: author (or update) the loop doc before running, using this file as the template. The doc is the durable definition; the session is the execution.

What this loop is

A per-phase execution cadence for a spec that is already Signed-Off and decomposed into phases. The spec is the contract; each iteration turns one phase into a plan, lands it, proves it, records it, and pushes it — so the repo is never more than one phase away from a green, documented, pushed state. First proven on Augment From DB Flow (five phases in one session, 2026-07-22).

Preconditions: a Signed-Off spec with a phase decomposition whose Phase 1 is service/data-layer work (surfaces come after their capabilities — this ordering is what makes later phases UI-only and cheap); a proof-script convention; the changelog and git-conventions skills.

The loop at a glance

MERMAID
flowchart TD
  SPEC["Spec: Signed-Off,<br/>phases 1..N"] --> PICK["Take next phase"]
  PICK --> PLAN["1 · Author plan<br/>re-ground against LIVE code;<br/>correct spec drift explicitly"]
  PLAN --> IMPL["2 · Implement<br/>from named in-repo templates"]
  IMPL --> TEST["3 · Test, scriptable-first<br/>(the ladder, below)"]
  TEST --> GREEN{proof green?}
  GREEN -- "no" --> STOP1(["STOP — surface it;<br/>never build a phase on red"])
  GREEN -- "yes" --> LOG["4 · Changelog entry<br/>honest about what was NOT tested"]
  LOG --> FLIP["5 · Status flips<br/>plan → Shipped + post_ship_note;<br/>spec → Implementing / Shipped"]
  FLIP --> COMMIT["6 · Commit + push<br/>attempt(flow, capability, stepN)<br/>explicit paths only"]
  COMMIT --> MORE{phases left?}
  MORE -- "yes" --> PICK
  MORE -- "no" --> MILE["milestone(flow): verdict<br/>empty marker commit"]
  MILE --> DONE(["Loop ends.<br/>Usability iteration = a NEW loop"])
  PLAN -. "spec turns out WRONG<br/>(not just drifted)" .-> STOP2(["STOP — revise spec<br/>with the user, resume"])

The iteration (one phase per pass)

  1. Author the plan — context-v/plans/<Spec>-Phase-N-<Name>.md, spec_reference in frontmatter. Before writing steps, re-ground against the live code: read the exact files the phase touches. Where reality diverges from the spec's anticipated snippets, the plan corrects the spec and says so explicitly (Phase 1 found three such divergences; Phase 2 found a whole missing verb). The plan names its verification before implementation starts.

  2. Implement — smallest dependency-ordered steps, every new file copied from a named in-repo template (the plan lists which). Service verbs cross their three files (handler → capabilities map+timeout → typed client wrapper); remotes follow the scaffold of the newest shipped remote.

  3. Test, scriptable-first — climb the ladder, cheapest to dearest; stop climbing only where the next rung would pollute shared data:

    TEXT
    cost/risk ▲   ┌──────────────────────────────────────────────────────┐
          6   │ operator browser walk-through ── NAMED, not faked    │  humans only
          ────┼──────────────────────────────────────────────────────┤ ─────────────
          5   │ live end-to-end, side-effect-safe only               │
              │   ✓ stream-scan flip test   ✗ test persons in canon  │
          4   │ container rebuild + raw-NATS proof of new verbs      │
              │   (the running stack is OLD code until you rebuild)  │
          3   │ standing regression: prove-<spec>-capabilities.mjs   │  every phase,
              │   written in Phase 1, re-run every phase             │  scripted
          2   │ dev-server smoke: curl :PORT/remoteEntry.js          │
          1   │ builds — each remote + THE SHELL (catches            │
              │   federation-registration typos)                     │
          0   │ svelte-check + tsc --noEmit on everything touched    │
              └──────────────────────────────────────────────────────┘

  4. Changelog — one entry per phase, changelog-conventions shape, honest about what was NOT tested and why.

  5. Status flips — plan → Shipped + date_first_published + post_ship_note recording deviations and skipped checks; spec → Implementing on the first phase, Shipped (+ post_ship_note listing what remains human-only) when the last phase lands.

  6. Commit + push — one commit per phase: attempt(<flow-slug>, <capability>, stepN): <impact-first headline>, body per git-conventions (why before how, proof summary included). Stage explicit paths only — never sweep in unrelated dirty state (submodules like clients/* stay untouched for deliberate tidying). Push each phase; don't batch.

What one pass leaves behind (the artifact trail)

Every iteration deposits the same four artifacts plus one commit — this is the proving run's Phase 2, but every phase leaves the identical shape:

TEXT
augment-it/
├── context-v/
│   ├── specs/Augment-From-DB-Flow.md          ← status flip (+ post_ship_note at the end)
│   └── plans/
│       └── Augment-From-DB-Phase-2-….md       ← NEW: the plan, → Shipped + post_ship_note
├── changelog/
│   └── 2026-07-22_02_Org-Workbench-….md       ← NEW: one entry, honest about untested legs
├── apps/ | services/ | shell/                 ← the code, from named in-repo templates
└── (git) attempt(augment-from-db, org-workbench, step2): …   ← one pushed commit

Status lifecycles the loop drives

MERMAID
stateDiagram-v2
  direction LR
  state "Spec" as S {
    [*] --> SignedOff
    SignedOff --> Implementing: first phase starts
    Implementing --> Shipped: last phase lands
    note right of Shipped: post_ship_note lists what stays human-only
  }
  state "Each phase's Plan" as P {
    [*] --> Draft
    Draft --> Shipped_: proof green, same day
    note right of Shipped_: date_first_published + post_ship_note record deviations and skipped checks
  }

Exit conditions

  • All phases shipped → an empty milestone(<flow-slug>): <verdict> marker commit naming the step range and the known follow-ups. Then stop — usability iteration is a NEW loop with its own findings, not a tail on this one.

  • A phase's proof won't go green → stop the loop, surface it; don't proceed to a phase that builds on red.

  • The spec turns out wrong mid-phase (not just drifted — wrong) → stop, revise the spec with the user, resume.

Hard-won rules (from the first run)

  • Quote YAML revisions: entries. Any list item containing : breaks standard YAML parsers ("mapping values are not allowed in this context"). Double-quote every revision string at write time — this bit four files, two of them pre-existing.

  • The proof script is the loop's spine. Written once in Phase 1, re-run every phase; it converts "did we break the floor?" from a worry into a ten-second check.

  • svelte-check is not optional even when the build passes — it caught a prop declared in a type but missing from the destructure that rsbuild happily bundled.

  • Existing seams beat new machinery. Phase 5 was a half-day because curated_index_urls already existed; the plan-authoring step's job is to find that seam BEFORE writing code.

  • Splash content needs git add -f (the .gitignore content rule silently drops splash/src/content/*), and splash deploys only from main — a card authored on a work branch ships at reconciliation.

  • Cross-service needs go through NATS verbs, never shared DB access — the dedup read (content.urls.check) is the pattern, per the domains.ts precedent.