ls-tech-design

name	ls-tech-design
description	Transform a Feature Spec into implementable Tech Design with architecture, interfaces, and test mapping. Validates the spec as downstream consumer.

Tech Design

Purpose: Transform Epic into Tech Design with architecture, interfaces, and test mapping.

This phase is the downstream consumer of the Epic. If you can't design from it, the spec isn't ready. Validation is part of the quality gate.

When a Technical Architecture document exists, also read it before designing. The tech arch establishes the technical world — system shape, core stack, cross-cutting decisions, and the top-tier surfaces (primary domains or organizing surfaces) that structure the system. The epic defines what this feature does. The tech arch defines what technical world it lives in. Your design works within both.

Output Structure

The tech design always produces at least two documents:

Config A: 2 docs (default)

• tech-design.md — Index: decisions, context, system view, module architecture, work breakdown
• test-plan.md — TC→test mapping, mock strategy, fixtures, chunk breakdown with test counts

Everything lives in the index. Works when the design fits comfortably under ~1200-1500 lines. Typical for single-domain projects — a CLI, a backend service, a focused frontend feature.

Config B: 4 docs (when the index gets dense)

• tech-design.md — Index: decisions, context, system view, module architecture overview, work breakdown
• tech-design-[domain-a].md — Implementation depth for one domain (e.g., frontend, client, UI)
• tech-design-[domain-b].md — Implementation depth for the other domain (e.g., backend, server, API)
• test-plan.md — TC→test mapping, mock strategy, fixtures, chunk breakdown with test counts

The trigger is index density — when the index is approaching ~1200-1500 lines, split implementation depth into companion docs. Name companions by the project's actual domain boundaries (frontend/backend, client/server, renderer/engine — whatever fits the project). The index stays as the decision record and whole-system map. Companion docs carry the implementation detail.

Never go 3. You don't add just one companion doc. It's either everything in the index, or the index plus both companions. Two configurations, not a continuum.

The test plan is always its own document. If the work doesn't justify a separate test plan with TC traceability, this skill isn't the right tool — downshift to plan mode.

Companion docs maintain requirement traceability — they reference ACs and TCs so you can navigate from the companion back to the index and the epic.

Dual Role: Designer and Validator

As Validator

Before designing, validate the Epic:

• Can you map every AC to implementation?
• Are data contracts complete and realistic?
• Are there technical constraints the BA missed?
• Do the flows make sense from an implementation perspective?
• Answer every question in the Epic's "Tech Design Questions" section. These are legitimate technical questions raised during spec writing or validation that the spec intentionally deferred to this phase.

If issues found → return to BA for revision. Don't design from a broken spec.

When a tech arch exists, also validate against it:

• Does the epic's scope fit within the top-tier surfaces the tech arch defined?
• Are the cross-cutting decisions (auth, error handling, state management) compatible with what this epic needs?
• Are there tech arch assumptions that this epic's requirements challenge?
• Flag tensions early — they're cheaper to resolve before design than during implementation.

As Designer

Once validated, produce:

• Architecture decisions
• Module breakdown
• Interface definitions
• Test architecture
• Work plan (phases/stories)

---

The Altitude Model

Design from high to low. Don't skip levels. The template structures sections from system view down to interface definitions — use the altitude model to calibrate your depth at each level, but don't surface altitude labels in your output headings.

High Altitude (30,000 ft) — System Context

When a tech arch exists, the System Context inherits rather than re-derives. The tech arch already established system shape, boundaries, and communication patterns at 50k-20k ft. Your System Context narrows that to this epic's slice — which top-tier surfaces this epic touches, which external boundaries are relevant, what data flows through them for this epic's functionality. When no tech arch exists, derive from scratch.

## System Context

### External Systems
- **Backend API:** REST endpoints at `/api/v1/*`
- **Guidewire:** Embedded iframe, URL parameter communication
- **Auth:** JWT tokens from parent application

### Entry Points
- Route: `/locations/add`
- Triggered by: Guidewire "Add Location" button

### Data Flow Overview
Guidewire → Embed with params → Fetch locations → User selects → Return data → Guidewire

Medium Altitude (10,000 ft) — Module Architecture

When a tech arch exists, start from: which top-tier surfaces does this epic live in? The file tree and module responsibility matrix should nest within those inherited surfaces, not create a parallel organizing structure. If this epic needs a module that doesn't fit in any surface the tech arch defined, that's a deviation — document it in the Issues Found table and surface it upstream. The inherited surfaces are the shared vocabulary across all epics; respecting them keeps the system coherent.

When no tech arch exists, still ask: what are the primary organizing surfaces of this system? If you can infer them from the codebase, state them as locally derived context. If the surface map is materially unclear, flag it as a discussion point with the human — don't silently invent a full system architecture. The goal is coherent decomposition for this epic, not a substitute tech arch.

Human-first module design. The module structure should be designed for human navigability. If a human can't look at the responsibility matrix and immediately know where to go for any capability in this epic, it's over-segmented. If half the epic's functionality is jammed into one module, it's under-decomposed. Strong human abstractions are also the most model-navigable abstractions — models work better within clear responsibility boundaries than within structures optimized for technical purity.

Run a dimensional reasoning check before fixing module boundaries. Inherited surface fit, responsibility clarity, coupling, AC coverage, and testability do not always point toward the same decomposition; identify the tensions for this epic and weight which should win before you commit. See the Dimensional Reasoning Check reference.

## Module Architecture

src/features/add-location/
├── pages/
│   └── AddLocation.tsx        # Route entry
├── components/
│   ├── LocationList.tsx       # List display
│   └── LocationForm.tsx       # Create form
├── hooks/
│   ├── useLocations.ts        # Data fetching
│   └── useLocationSelection.ts # Selection state
├── api/
│   └── locationApi.ts         # API client
└── types/
    └── location.types.ts      # Shared types

### Module Responsibilities

| Module | Responsibility | ACs Covered |
|--------|----------------|-------------|
| AddLocation | Route, layout, flow control | AC-1 to AC-5 |
| LocationList | Display, filter, select | AC-10 to AC-20 |
| useLocations | Fetch, cache | AC-6 to AC-9 |

Low Altitude (Ground Level) — Interface Definitions

// types/location.types.ts
export interface Location {
  locRefId: string;
  locRefVerNbr: number;
  address: string;
  city: string;
  state: string;
  postalCode: string;
}

// hooks/useLocations.ts
interface UseLocationsReturn {
  locations: Location[] | undefined;
  isLoading: boolean;
  isError: boolean;
  error: Error | null;
}

// components/LocationList.tsx
interface LocationListProps {
  locations: Location[];
  selectedIds: Set<string>;
  onToggleSelection: (id: string) => void;
  onAddToPolicy: () => void;
}

---

Weaving Functional to Technical

At each altitude, connect back to ACs and TCs:

sequenceDiagram
    Note over User,Page: AC-10: User sees location list
    User->>Page: Navigate to route
    Page->>Hook: useLocations()
    Note over Hook,API: AC-6: API fetches locations
    Hook->>API: getLocations(sai)
    API-->>Hook: Location[]
    Hook-->>Page: { locations, isLoading }
    Note over Page,User: AC-11: Table displays locations
    Page-->>User: Render LocationList

---

Tech Design Writing Style: Rich, Layered Context

Tech designs are verbose and intentionally rich.

This is NOT about being minimal. Build a sophisticated web of context.

The Spiral Pattern

• Functional ↔ Technical — Repeatedly connect requirements to implementation. Don't just list interfaces — show how they fulfill ACs.
• High level ↔ Low level — Spiral through abstraction layers. Go high → low → back to high → lower. Not a linear descent.
• Back and forth — Revisit topics from different angles. Mention the same concept in system context, in module breakdown, in interface definitions.
• Redundant connections — Rich, layered, multiple paths to the same information.

Why This Works

The goal is redundant connections — multiple paths through the material so the model (and humans) can navigate complexity.

A web of weights around the material, not a thin thread.

If someone enters the design at the interface section, they should still understand why this interface exists (AC reference). If they enter at the module section, they should still see how data flows (sequence connection).

Anti-Pattern: Thin Linear Design

## Bad: Linear descent, minimal context
- System: calls API
- Module: LocationList
- Interface: LocationListProps

Pattern: Woven Context

## Better: Rich connections

At 30,000 ft: "The system needs to display account locations (AC-10).
This requires a fetch from the XAPI..."

At 10,000 ft: "LocationList handles AC-10 through AC-20. It receives
locations from useLocations (established in system context above) and
displays them with selection capability (supporting the return flow
we'll detail at ground level)..."

At ground level: "LocationListProps includes selectedIds (supporting
AC-20 selection requirement) and onAddToPolicy (the return trigger
from our sequence diagram)..."

---

Test Architecture

Mock Strategy: API Boundary

Mock at the API layer, not hooks.

// ✅ CORRECT
jest.mock('@/features/add-location/api/locationApi');

// ❌ WRONG
jest.mock('@/features/add-location/hooks/useLocations');

Why API boundary: Tests the real integration (Component → Hook → React Query → mock). Catches hook wiring bugs.

When inherited top-tier surfaces exist, they inform where to enter for high-leverage testing, not where to mock. Top-tier surfaces are internal responsibility zones — they're the natural entry points from which a single test exercises broad functionality with consistent input patterns. But mocking stays at external boundaries per the service mock philosophy: mock where your code ends and external systems begin (network, database, filesystem), never between your own internal domains. A test that enters at a top-tier surface boundary should exercise all internal modules within it and mock only what's truly external.

TC to Test Mapping (Critical)

The test plan must explicitly map every TC from the Epic to a test. This is the Confidence Chain in action: AC → TC → Test → Implementation.

Test plan table format:

TC	Test File	Test Description	Status
TC-6a	AddLocation.test.tsx	shows loading during fetch	Planned
TC-6b	AddLocation.test.tsx	hides loading after fetch	Planned
TC-10a	LocationList.test.tsx	renders location rows	Planned

Rules:

• Every TC from Epic must appear in this table
• TC ID must be visible in table OR in the test name/comment (for traceability)
• If you can't map a TC to a test, either the TC is untestable (return to spec) or you're missing a test boundary
• Group by test file for clarity; the TC column is what matters for traceability

→ See the Testing reference section in this skill for test code organization patterns.

---

Work Plan: Chunking for Stories

Break work into manageable pieces. Each chunk becomes a story or set of stories. The chunk is the Tech Lead's unit of decomposition; chunks inform how stories are organized when the epic is published (usually 1:1, sometimes a chunk splits into multiple stories or merges with another).

Dimensional reasoning check: chunk boundaries are compositional. Functional coherence, dependency sequencing, and scope manageability do not always agree; weigh the tension before partitioning the work.

Chunks vs. Phases

Chunks are vertical slices (by functionality):

• Chunk 0: Infrastructure (types, fixtures, error classes)
• Chunk 1: Initial Load flow (ACs X-Y)
• Chunk 2: Selection flow (ACs Z-W)

Phases are horizontal stages (by workflow):

• Skeleton: Stubs that compile but throw
• TDD Red: Tests that run but fail
• TDD Green: Implementation that passes

The relationship: each chunk goes through all phases.

Chunk 0 (Skeleton) → Chunk 0 (Red) → Chunk 0 (Green) →
Chunk 1 (Skeleton) → Chunk 1 (Red) → Chunk 1 (Green) → ...

Some teams prefer completing all skeletons first (all Chunk Skeletons, then all Chunk Reds). Choose based on team preference and dependency structure. The default is vertical: complete each chunk fully before starting the next.

Chunk 0: Infrastructure (Always First)

• Types and interfaces
• Test fixtures
• Test utilities
• Error classes (NotImplementedError)

Subsequent Chunks

## Chunk 1: Initial Load

**Scope:** Page component, data fetching, loading/error states
**ACs:** AC-1 to AC-9
**TCs:** TC-1a through TC-9b

**Files:**
- src/features/add-location/pages/AddLocation.tsx
- src/features/add-location/hooks/useLocations.ts
- src/features/add-location/api/locationApi.ts

**Relevant Tech Design Sections:** §System Context — Data Flow,
§Module Architecture — AddLocation Page, §Low Altitude — useLocations Hook,
§Flow 1: Initial Load Sequence, §Testing Strategy — Initial Load Tests

**Non-TC Decided Tests:** Empty state render (no locations), loading
skeleton timing assertion (Tech Design §Testing Strategy)

**Test Count:** 12 tests + 2 non-TC
**Running Total:** 14 tests

The "Relevant Tech Design Sections" field lists which headings from this tech design are relevant to the chunk. This directly supports story creation: when publishing the epic, these references help select which tech design content is relevant to each story's Technical Design section.

The "Non-TC Decided Tests" field lists tests this chunk needs that aren't 1:1 with a TC -- edge cases, collision tests, defensive tests. These must be carried forward into stories during technical enrichment so they aren't lost.

Chunk Dependencies

Chunk 0 → Chunk 1 → Chunk 2
              ↘      ↗
               Chunk 3

---

Dependency and Version Grounding

When a tech arch exists, distinguish between inherited and epic-scoped decisions. Core stack choices (framework, runtime, data layer, auth) are already settled — don't re-research them unless something specific to this epic challenges them. Epic-scoped dependencies (feature-specific libraries, adapters, connectors) get fresh research. If deeper research reveals that an inherited decision needs revision, treat it as a deviation: proceed with the better approach, document the rationale in the Issues Found table, and surface it upstream for backfill.

Dependency and version choices must be grounded by current web research, not training data. This is especially important for fast-moving ecosystems (build tools, frameworks, runtimes, packaging tools). Training data goes stale; the npm registry and GitHub releases don't.

Before pinning any epic-scoped version, research the current ecosystem status: latest stable version, known breaking changes, compatibility with the project's existing stack, and whether the package is actively maintained. Document your findings in a Stack Additions table:

Package	Version	Purpose	Research Confirmed
[package]	[version]	[why this package]	Yes — [key finding from research]

Also document packages considered and rejected, with rationale. This prevents future designers from re-evaluating the same alternatives.

---

Upstream Document Evolution

Downstream work regularly surfaces new facts that reveal the need to realign upstream decisions. When the tech design discovers that a tech arch decision or an epic assumption needs revision — whether through deeper dependency research, implementation reality, or evolved understanding:

• Proceed with the better approach. Don't block progress waiting for upstream approval.
• Document the deviation and rationale in the Issues Found table. Include what the upstream document says, what the design does instead, and why.
• Surface it upstream to the orchestrator or human as a suggested update. The orchestrator tracks the backfill task.
• The upstream documents are the starting position, not inviolable sources of truth. Common sense and fresh appraisal of the problem space based on new knowledge, not artificial adherence to upstream documents.

This applies to both the epic and the tech arch. The Issues Found table in the template already supports this — use the "Resolved — deviated" status for design-time deviations.

---

Verification Scripts

The tech design defines the project's verification gates before implementation begins. These become the quality gates that story technical sections reference — getting them right here prevents ad-hoc discovery during implementation.

Every project needs at least four verification tiers: red-verify (everything except tests — for TDD Red exit when stubs throw), verify (standard development gate), green-verify (verify + test immutability guard — for TDD Green exit), and verify-all (deep verification including integration and e2e suites). Define the specific commands for each tier in the tech design so stories can reference them consistently.

---

Test Count Reconciliation

Test counts drift between documents. This is a known trap — the index work breakdown says one number, the test plan per-chunk totals say another, the per-file totals say a third. Each fix round can introduce new inconsistencies.

After completing the test plan, do a single mechanical reconciliation pass: per-file test counts sum to per-chunk totals, per-chunk totals sum to the index work breakdown summary. One pass, three cross-checks. If anything doesn't add up, fix it before self-review. This catches arithmetic drift in one pass instead of burning multiple verification rounds on it.

---

Validation Before Handoff

Before handing off:

• Every TC mapped to test file
• All interfaces defined
• Module boundaries clear
• Chunk breakdown complete with relevant tech design section references per chunk
• Non-TC decided tests identified and assigned to chunks
• Test counts estimated (TC tests + non-TC tests)
• No circular dependencies
• Top-tier surfaces stated (inherited from tech arch, or locally derived with rationale)
• Modules nest within inherited surfaces (or deviations documented in Issues Found)
• Inherited stack decisions respected (or deviations documented with rationale)
• Epic-scoped dependency additions grounded in fresh research

Self-review (CRITICAL):

• Read your own design critically
• Is the spiral pattern present?
• Are there redundant connections or just thin threads?

The BA/SM validates by confirming they can derive stories from the design. The Tech Lead validates by confirming they can add story-level technical sections from the design. If either can't, the design isn't ready.

→ Verification prompt: examples/tech-design-verification-prompt.md — Ready-to-use prompt for external validation before handoff

---

Output: Tech Design

The tech design expands significantly from the epic — typically 6-7x. The design includes:

• System context (external systems, data flow)
• Module architecture (files, responsibilities, AC mapping)
• Sequence diagrams (per flow)
• Interface definitions (types, props, signatures)
• TC-to-test mapping
• Chunk breakdown with test counts

The verbose, spiral style is intentional. It creates the redundant connections that help both humans and models navigate the complexity.

---

Reference: confidence-chain

The Confidence Chain

Every line of code traces back through a chain:

AC (requirement) → TC (test condition) → Test (code) → Implementation

Validation rule: Can't write a TC? The AC is too vague. Can't write a test? The TC is too vague.

This chain is what makes the methodology traceable. When something breaks, you can trace from the failing test back to the TC, back to the AC, back to the requirement.

---

Reference: dimensional-reasoning

Dimensional Reasoning Check

Some decisions in this methodology are compositional: choosing feature boundaries, top-tier surfaces, module boundaries, or story partitions. These decisions have multiple valid organizing axes. Models have the metacognitive capacity to identify and weigh competing considerations but will not do so spontaneously. The failure mode is single-axis reasoning: one compelling consideration absorbs the decision before competing dimensions are surfaced.

Before committing to a structure:

1. Enumerate the competing considerations for this specific system or feature.
2. Identify opposition — where do those considerations pull toward different structures or partitions?
3. Weight and resolve — which tensions matter most here, and why? Commit with rationale.

The factors vary by decision point. The process stays the same.

---

Reference: Verification: The Scrutiny Gradient

Verification: The Scrutiny Gradient

Upstream = more scrutiny. Errors compound downward.

The epic gets the most attention because if it's on track, everything else follows. If it's off, everything downstream is off.

The Gradient

Epic:  #################### Every line
Tech Design:   #############....... Detailed review
Stories:       ########............ Key things + shape
Implementation:####................ Spot checks + tests

Epic Verification (MOST SCRUTINY)

This is the linchpin. Read and verify EVERY LINE.

Verification Steps

1. BA self-review -- Critical review of own work. Fresh eyes on what was just written.

2. Tech Lead validation -- Fresh context. The Tech Lead validates the spec is properly laid out for tech design work:

   • Can I map every AC to implementation?
   • Are data contracts complete and realistic?
   • Are there technical constraints the BA missed?
   • Do flows make sense from implementation perspective?

3. Additional model validation -- Another perspective (different model, different strengths):

   • Different model, different strengths
   • Adversarial/diverse perspectives catch different issues

4. Fix all issues, not just blockers -- Severity tiers (Critical/Major/Minor) set fix priority order, not skip criteria. Address all issues before handoff. Minors at the spec level compound downstream -- zero debt before code exists.

5. Validation rounds -- Run validation until no substantive changes are introduced, typically 1-3 rounds. The Tech Lead also validates before designing -- a built-in final gate. Number of rounds is at the user's discretion.

6. Human review (CRITICAL) -- Read and parse EVERY LINE:

   • Can you explain why each AC matters?
   • No "AI wrote this and I didn't read it" items
   • This is the document that matters most

Tech Design Verification

Still detailed review, but less line-by-line than epic.

What to Check

• Structure matches methodology expectations
• TC-to-test mapping is complete
• Interface definitions are clear
• Phase breakdown makes sense
• No circular dependencies

Who Validates

• Tech Lead self-review -- Critical review of own work
• BA/SM validation -- Can I shard stories from this? Can I identify coherent AC groupings?
• Tech Lead re-validation -- Can I add story-level technical sections from this?

Story Verification

Stories go through a two-phase validation reflecting their two-phase authoring.

Functional Stories (after BA/SM sharding)

Less line-by-line, more shape and completeness:

• Coverage gate: every AC/TC assigned to a story
• Integration path trace: no cross-story seam gaps
• Each story coherent and independently acceptable
• Tech Lead confirms they can add technical sections

Technically Enriched Stories (after Tech Lead enrichment)

Story contract compliance check:

1. Tech design shard present -- substantial, story-scoped tech design content in Architecture Context and Interfaces
2. TC-to-test mapping present -- every TC mapped to a test approach with file names and approaches from the tech design
3. Non-TC decided tests present -- edge/integration tests from tech design carried forward or explicitly noted as absent
4. Technical DoD present -- specific verification commands
5. Spec deviation field present with citations -- checked tech design sections listed, even when no deviations
6. Targets, not steps -- technical sections describe what, not how

Consumer gate: could an engineer implement from this story alone, without reading the full tech design?

Implementation Verification

Spot checks + automated tests.

What to Check

• Tests pass (full suite)
• Types check clean
• Lint passes
• Spot check implementation against tech design
• Gorilla testing catches "feels wrong" moments

---

Multi-Agent Validation Pattern

Liminal Spec uses this pattern throughout:

Artifact	Author Reviews	Consumer Reviews
Epic	BA self-review	Tech Lead (needs it for design)
Tech Design	Tech Lead self-review	BA/SM (needs it for story derivation) + Tech Lead (needs it for technical sections)
Published Stories	BA/SM self-review	Engineer (needs them for implementation)

Why This Works

1. Author review -- Catches obvious issues, forces author to re-read
2. Consumer review -- Downstream consumer knows what they need from the artifact
3. Different model -- Different strengths catch different issues. Use adversarial/diverse perspectives for complementary coverage.
4. Fresh context -- No negotiation baggage, reads artifact cold

The Key Pattern: Author + Downstream Consumer

If the Tech Lead can't build a design from the epic -> spec isn't ready. If the BA/SM can't derive stories from the epic -> epic isn't ready. If the Engineer can't implement from published stories + tech design -> artifacts aren't ready.

The downstream consumer is the ultimate validator.

---

Orchestration

How to run validation passes is left to the practitioner. This skill describes:

• WHAT to validate -- Which artifacts, which aspects
• WHEN to validate -- Checkpoints in the flow

Leaves flexible:

• HOW to validate -- Which models, how many passes
• Specific orchestration -- Based on your setup and preferences

---

Checkpoints

Before Tech Design

• Epic complete
• BA self-review done
• Model validation complete
• All issues addressed (Critical, Major, and Minor)
• Validation rounds complete
• Tech Lead validated: can design from this
• Human reviewed every line

Before Publishing Epic

• Tech Design complete (all altitudes: system context, modules, interfaces)
• Tech Lead self-review done (completeness, richness, writing quality, readiness)
• Model validation complete (different model for diverse perspective)
• All issues addressed (Critical, Major, and Minor)
• Validation rounds complete (no substantive changes remaining)
• TC -> Test mapping complete (every TC from epic maps to a test)
• Human reviewed structure and coverage

Before Implementation

• Functional stories complete (all ACs/TCs assigned, integration path traced)
• Technical enrichment complete (all six story contract requirements met)
• Consumer gate passed: engineer can implement from stories
• Different model reviewed stories (if high-stakes)

Before Ship

• All tests pass
• Gorilla testing complete
• Verification checklist passes
• Human has seen it work

---

Reference: Writing Style Reference

Writing Style Reference

Deep-dive guide for documentation that serves both humans and AI agents. Load when writing Epics and Tech Designs.

---

Quick Navigation (If You're Stuck Mid-Doc)

Problem	Jump to
Everything feels flat / equal weight	Branches, Leaves, Landmarks • Before/After #1
Altitude jump / reader lost	The Smooth Descent • Extended Example
Functional ↔ technical drift	The Weave • Before/After #3
Unsure how deep to go	Bespoke Depth • Scoring Rubric
Complex diagram overwhelming	Progressive Construction
Fast final-pass check	Quick Diagnostic

---

The Core Insight: Documentation Has Three Dimensions

Most technical documentation uses one dimension: hierarchy. Categories, subcategories, sections. This works for reference lookup but fails for understanding. Good documentation uses three dimensions—and the third is what makes it work.

Dimension 1: Hierarchy — Parent/child relationships. What contains what.

Dimension 2: Network — Cross-references. What connects to what.

Dimension 3: Narrative — Temporal and causal flow. What leads to what, and why.

The third dimension is where context lives. Not just what things are, but how they connect, why they exist, and what happens when you use them.

This matters for AI agents because LLMs trained on internet text—50TB of messy, narrative, temporal human writing—encode knowledge using narrative substrate. When you conform to that structure, you get efficient encoding and better retrieval almost for free. The relationships you'd otherwise need to enumerate explicitly are encoded in the temporal flow.

Why Narrative Is Compression

Consider two ways to express the same information:

Enumerated (explicit relationships):

- ConversationManager exists
- Session exists  
- ConversationManager creates Session
- Session handles messages
- Relationship: ConversationManager owns Session

Narrative (implicit relationships):

When a user starts chatting, ConversationManager creates a Session to handle
the conversation. The Session manages message history and coordinates with
external services. The manager holds the session reference throughout the
conversation lifecycle.

Both encode the same facts. The narrative version is shorter because relationships emerge from temporal flow ("when... creates... manages... throughout"). You don't enumerate "Relationship: A owns B"—the ownership is implicit in "manager holds the session reference."

For LLMs, narrative structure activates learned patterns from training data. For humans, narrative matches how we naturally think. We remember journeys better than lists.

---

Branches, Leaves, and Landmarks

Think of your document as a tree. Prose paragraphs establish branches—the structural limbs that hold everything together. Bullet lists hang leaves—specific details attached to their branch. Diagrams create landmarks—spatial anchors that help readers navigate.

This creates attentional hierarchy. When all information has equal visual weight (flat bullets, uniform paragraphs), readers and models spread attention evenly. Key insights get lost in noise.

The Branch: Prose Paragraphs

Prose establishes context, importance, and relationships. It tells the reader why something matters before presenting what specifically exists. Two to five sentences. One clear point per paragraph.

The Leaves: Bullet Lists

Lists enumerate specifics after context is established. They hang from the branch that prose creates. Never more than two levels deep in any single section.

Good list usage:

OAuth tokens are retrieved from keyring storage where other CLI tools have 
already obtained and stored them. We're not implementing OAuth flows—just 
reading tokens.

Token locations:
- ChatGPT: ~/.codex/auth/chatgpt-token
- Claude: ~/.claude/config

Poor list usage:

Authentication:
- API keys supported
- OAuth supported
- ChatGPT tokens
- Claude tokens
- Token refresh

The second example forces readers to infer all relationships. The cognitive load is on the reader instead of the writer.

The Landmarks: Diagrams

Diagrams encode spatial relationships that prose can't express efficiently. They create memory anchors through visual variation.

User Command → AuthManager → Check method
                  ↓
            API Key ──→ Config → Headers
                  ↓
            OAuth ──→ Keyring → Token → Headers

The same information in prose would take 3-4 sentences and lose the spatial relationship. Diagrams are compression.

The Ratio: A Compass, Not a Rule

Most well-structured sections land around:

• ~70% prose — Paragraphs establishing context, relationships, purpose
• ~25% lists — Specifics, enumerations, steps
• ~5% diagrams — Spatial relationships, architecture, flow

This isn't prescription. Some sections need more diagrams (architecture overviews). Some need more lists (API references). The ratio is a compass: if you're at 90% bullets, you're probably missing branches. If you're at 100% prose, you're probably missing scannable specifics.

When a section feels off, check the ratio. Monotonous structure often explains the problem.

Putting It Together

## Session Management

Conversations persist through the Session abstraction. When a user starts 
chatting, ConversationManager creates a Session to hold conversation state—
message history, active provider, pending tool calls.

Sessions coordinate between three subsystems:

    Session
    ├── MessageHistory (stores conversation)
    ├── ModelClient (sends to LLM)
    └── ToolRouter (handles tool calls)

Key methods:
- `sendMessage(content)` — Format, send, process response, update history
- `getHistory()` — Return full message history for context
- `processToolCall(call)` — Route to executor, await result, append

The separation between Session and ConversationManager matters for testing.
Sessions test with mocked clients. Managers test lifecycle without message flow.

Five elements: branch paragraph, diagram landmark, detail leaves, closing branch. Complete concept.

---

The Altitude Metaphor

Documentation exists at different altitudes. Higher = broader view, less detail. Lower = narrower focus, more specifics.

25,000 ft   PRD: "The system enables collaborative AI conversations"
    ↓
15,000 ft   Tech Approach: "ConversationManager orchestrates Sessions"
    ↓
10,000 ft   Phase README: "Session.sendMessage() formats per provider spec"
    ↓
5,000 ft    Checklist: "Task 3: Wire Session to ModelClient with retry"
    ↓
1,000 ft    Code: const response = await client.send(formatted, { retries: 3 })

The failure mode isn't being at the wrong altitude—it's jumping altitudes without bridges.

The Smooth Descent

Each document should bridge levels, not exist at one. Start higher than you'll finish. Descend gradually. Each level answers questions raised by the level above.

PRD says: "User can authenticate with ChatGPT OAuth" Reader asks: How does that work technically?

Tech Design says: "Read token from ~/.codex keyring" Reader asks: What's the implementation approach?

Phase Doc says: "Use keyring-store module, mock filesystem in tests" Reader asks: What are the specific tasks?

Checklist says: "1. Import keyring-store 2. Add getToken() wrapper 3. Create mock"

No gaps. Each level makes sense in context of the previous one.

The Consistency Ladder

The same capability should be visible at every altitude level:

Altitude	Example
25K (PRD)	User can start a conversation and receive a response
15K (Approach)	ConversationManager wires CLI → Codex → ModelClient
10K (Phase)	Implement createConversation(), wire CLI command, mock ModelClient
5K (Checklist)	1) Add CLI command 2) Implement createConversation 3) Write mocked test

If a capability appears at one level but not another, something is missing. The ladder is the alignment test.

Extended Example: Tracing Through Altitudes

Epic (25K feet):

Users can execute tools (read files, run commands) through the AI assistant. The assistant requests permission before executing.

Reader understands: what capability exists, who controls it. Reader wonders: how does this work technically?

Tech Design - Overview (15K feet):

Tool execution flows through three components. Session detects when the model requests a tool call. ToolRouter matches the request to an executor. The CLI presents approval before execution proceeds.

Reader understands: which components, how they connect. Reader wonders: what are the specific interfaces?

Tech Design - Details (10K feet):

Session.processResponse() checks for tool_calls in model output. When found, it extracts the tool name and arguments, then calls ToolRouter.route(toolCall). Before execution, Session emits a 'tool_request' event that the CLI handler intercepts.

Reader understands: specific methods, data flow, event mechanism. Reader wonders: what are my implementation tasks?

Implementation Checklist (5K feet):

1. Add tool_calls detection to Session.processResponse()
2. Implement ToolRouter.route() with executor registry
3. Wire CLI approval handler to 'tool_request' event

Each level answered the question raised by the previous level. That's the smooth descent.

---

Functional and Technical: The Weave

Traditional process separates functional requirements ("user can chat") from technical design ("WebSocket with JSON-RPC"). Product writes the PRD, throws it over the wall, engineering writes the tech spec. The gap that opens between them is where projects fail.

Better: weave functional and technical together at every altitude level.

In high-level docs (mostly functional, touch technical):

• User outcome: "User can resume saved conversations"
• Technical grounding: "Verify by loading JSONL, continuing from last message"

In technical docs (mostly technical, ground in functional):

• Architecture: "Session coordinates ModelClient and ToolRouter"
• Functional anchor: "This enables users to review tool actions before execution"

In implementation docs (deep technical, verify via functional):

• Task: "Implement resumeConversation() method"
• Functional test: "User can continue conversation where they left off"

The weave prevents drift. When functional and technical stay interlocked, you can't over-engineer (functional bounds what's needed) and you can't under-deliver (technical serves functional outcomes).

Functional Verification Grounds Testing

Technical tests without functional grounding: "Test that ConversationManager.createConversation() returns Conversation object"

This can pass while the user still can't chat. The test verified mechanism, not outcome.

Functional test criteria: "User can start conversation, send message, receive response"

The test name describes the user capability. The test implementation exercises the technical path. The assertion verifies functional success. This is the weave in action.

---

Bespoke Depth: Signal Optimization

The anti-pattern: uniform depth across all topics. Everything documented to the same depth means nothing stands out.

Better: go deep where it matters, stay shallow where it doesn't.

The Four Questions

Before diving into any topic, ask:

Question	Deep if...	Shallow if...
Is this complex or simple?	Complex	Simple
Is this new or already done?	Novel	Existing
Is this critical or optional?	Critical	Optional
Will implementers struggle here?	High risk	Low risk

Scoring Rubric

Score each topic 1-5 on four dimensions, then sum:

• Complexity: How many moving parts?
• Novelty: Is this new or already implemented?
• Criticality: Does the project fail if this is wrong?
• Risk: How likely is confusion or error?

Score	Depth
4-8	Shallow (one paragraph, maybe a link)
9-13	Medium (2-3 paragraphs, small list)
14-20	Deep (multi-paragraph, diagram, examples)

Example:

• OAuth token retrieval: 4+5+5+4 = 18 → Deep
• Config file parsing: 2+1+2+2 = 7 → Shallow

Token Budget Thinking

Instead of 10 topics × 500 tokens = 5,000 tokens of uniform depth:

• 2 critical topics × 1,500 tokens = 3,000 tokens of real depth
• 8 simple topics × 100 tokens = 800 tokens of appropriate brevity
• Total: 3,800 tokens, higher signal-to-noise

The critical topics got the depth they need. The simple topics didn't waste tokens.

---

Progressive Construction: The Whiteboard Phenomenon

There's a difference between seeing a complex diagram and building one. People who build understand deeply. People who receive are overwhelmed.

When you whiteboard a system, you add one box at a time, connect it, then add the next. Each step scaffolds the next. Documentation should mimic that process.

Build Diagrams in Stages

Instead of dropping a 20-component diagram:

Step 1: System has three layers.
    CLI → Library → External

Step 2: Library entry point is ConversationManager.
    CLI → ConversationManager → External

Step 3: Manager coordinates Codex and Session.
    CLI → ConversationManager → Codex → Session → External

Step 4: Session routes to ModelClient and ToolRouter.
    CLI → ConversationManager → Codex → Session → ModelClient
                                       ↘ ToolRouter

By the final diagram, readers have constructed the understanding.

Conceptual Spiral

Progressive construction applies to concepts, not just diagrams. Revisit from multiple angles, each pass adding detail:

• Pass 1: "Phase 2 adds tool execution to the conversation flow."
• Pass 2: "Session detects tool calls and routes them to ToolRouter."
• Pass 3: "ToolRouter executes the tool, returns results, Session resumes the model call."
• Pass 4: "Here is the sequence diagram of that cycle."

Each pass deepens without forcing a leap. The spiral guides into complexity without drowning.

---

Before/After Transformations

Example 1: Flat List → Rich Context

Before:

CLI Features:
- Interactive REPL
- One-shot command mode  
- JSON output flag
- Provider switching

After:

The CLI supports three interaction modes for different audiences. Interactive 
REPL serves humans who want conversational flow. One-shot commands serve 
automation and testing. JSON output serves programmatic consumption.

Modes available:
- Interactive: `codex` → enters REPL, `quit` to exit
- One-shot: `codex chat "message"` → execute and exit
- JSON output: Add `--json` flag for structured response

The prose establishes why (the branch). The bullets enumerate what (the leaves).

Example 2: Altitude Jump → Smooth Descent

Before:

## Tool Execution
The system enables AI-assisted tool execution.

const result = await executor.run(tool, args, { timeout: 30000 });

After:

## Tool Execution
The system enables AI-assisted tool execution, where the model can request 
actions like reading files, running commands, or making API calls.

Tool execution follows a request-approve-execute cycle. The model requests 
a tool call, the system presents it for approval, execution runs sandboxed, 
and results return to the model.

    Model Request → Approval Gate → Executor → Result → Model

const result = await executor.run(tool, args, { timeout: 30000 });

Three altitudes (concept → mechanism → implementation), smooth descent between each.

Example 3: Separated → Woven

Before:

## Authentication Implementation
AuthManager reads from config.toml or keyring. API keys use ConfigReader.
OAuth tokens use KeyringStore.

Implementation:
- ConfigReader.get('api_key')
- KeyringStore.retrieve(provider)

After:

## Authentication Implementation
Users authenticate through two paths: API keys (for personal accounts) and
OAuth tokens (for reusing existing ChatGPT or Claude subscriptions).

AuthManager abstracts this choice. When a user starts a conversation, the 
manager checks the configured auth method. For API keys, it reads from config.
For OAuth, it retrieves tokens from keyring.

This abstraction enables provider switching without re-authentication—users
configure once, the system handles the rest.

Technical components:
- ConfigReader: Load from config.toml or environment
- KeyringStore: Retrieve OAuth tokens (path varies by provider)

Opens with functional context. Shows mechanism. Grounds in benefit. Then enumerates components.

---

Common Failure Modes

Bullet Soup

Symptom: Every section is bullets. No paragraphs. Lists all the way down. Fix: Add prose branches. Explain why the list matters before presenting it.

Wall of Text

Symptom: Dense paragraphs. No lists. No diagrams. No variation. Fix: Break up with lists for enumerations, diagrams for spatial relationships.

Altitude Yoyo

Symptom: Document bounces between vision and implementation randomly. Fix: Pick an altitude and stay there, or descend smoothly. Don't yoyo.

All Technical, No Functional

Symptom: Describes mechanisms without purpose. Fix: Ground in functional outcome. "When a user sends a message, ConversationManager... This enables users to..."

Premature Depth

Symptom: Edge cases before establishing normal path. Fix: Normal path first, edge cases second.

Orphaned Diagrams

Symptom: Diagram appears without prose context. Fix: Introduce diagrams with prose, then reference them. The diagram confirms; prose explains.

---

Writing for Agents

This reference will be loaded by AI agents writing Epics and Tech Designs. Understanding how agents read—and fail to read—makes the difference between documentation that works and documentation that wastes context.

How Agents Read Differently

Humans skim, backtrack, ask questions, fill gaps with intuition. Agents process sequentially, can't ask clarifying questions, and treat ambiguity as noise rather than invitation.

Human reading: Scan headings → jump to relevant section → skim for keywords → read closely when relevant → ask if confused.

Agent reading: Load document into context → process sequentially → attempt task → fail or succeed based on what was explicit.

This means:

• Gaps that humans bridge intuitively become blockers for agents
• Implied relationships that humans infer need to be stated
• "Obviously" is never obvious—if it matters, write it

Context Window Economics

Every token of documentation is a token not available for reasoning or code generation. Agents operate under hard context limits. This creates pressure for compression—but compression mustn't sacrifice clarity.

The solution is signal density: every token earns its place.

Low signal density:

The configuration system is designed to be flexible and extensible. 
It supports multiple configuration sources and can be extended by 
implementing the IConfigSource interface.

High signal density:

Configuration loads from config.toml. Extend via IConfigSource interface.
See /src/config/ for implementation.

Same information, half the tokens. Strip:

• Obvious filler ("designed to be flexible")
• Vague claims without specifics ("can be extended"—how?)
• Redundant phrasing

What Agents Need Explicitly

Implicit for Humans	Explicit for Agents
"Handle errors appropriately"	"Catch ConfigError, log message, return null"
"Test this thoroughly"	"Write tests for: valid input, empty input, malformed input"
"Wire up the components"	"Import X from Y, instantiate with config, pass to Z constructor"
"Follow the pattern from Phase 1"	"Copy the approach from Session.sendMessage(): validate → transform → execute → handle result"

The Explicit Stack

For any task, agents need these layers explicit:

1. Scope — What to do, what NOT to do
2. Inputs — What exists, where to find it
3. Outputs — What to produce, where to put it
4. Sequence — What order, what depends on what
5. Verification — How to confirm success

Before (human-readable):

Implement the auth flow. Make sure it works with both API keys and OAuth.

After (agent-executable):

Implement AuthManager.authenticate():

Scope:
- Implement API key and OAuth paths
- Do NOT implement token refresh (out of scope for Phase 1)

Inputs:
- AuthConfig from config.toml (method: 'api_key' | 'oauth', credentials)
- Existing KeyringStore for OAuth token retrieval

Outputs:
- Returns AuthToken { token: string, expiresAt: Date }
- Throws AuthError on failure

Sequence:
1. Read config.method
2. Branch: API key → read from config, OAuth → read from keyring
3. Validate token format
4. Return AuthToken

Verification:
- Test: API key path returns token from config
- Test: OAuth path retrieves from keyring mock
- Test: Invalid config throws AuthError

The second version is longer but executable. An agent can complete it without asking questions.

The Isolation Test

Agents often read documents in isolation—loaded into fresh context without the conversation history that produced them. The document must stand alone.

Test: Cover the rest of the document. Read only this section. Could you complete the work described?

If yes: section is self-contained. If no: identify what's missing and add it.

Progressive Loading Strategy

For large specifications, don't load everything into every context. Instead:

1. Core context (always loaded): Architecture overview, key patterns, current phase
2. Task context (loaded for specific work): Relevant phase spec, related component docs
3. Reference context (linked, not loaded): API docs, style guides, examples

This mirrors bespoke depth at the document level. Load what matters for this task; link to the rest. Mention what exists even if you don't include it—agents can request additional context if they know it exists.

---

Quick Diagnostic

When a section feels wrong but you can't identify why:

1. Altitude check: Am I jumping levels without bridges?
2. Branch check: Do lists have prose context, or are they orphaned?
3. Depth check: Am I going deep everywhere, or allocating based on importance?
4. Weave check: Is functional purpose present, or is this pure mechanism?
5. Variation check: Is the structure monotonous?
6. Narrative check: Does this read as a journey, or as disconnected facts?

The Writing Loop (Mid-Doc Recalibration)

When you drift—and you will—run through:

1. Check altitude: Where am I? 25k, 15k, 10k, 5k?
2. Check bridge: Did I explain why the next level exists?
3. Check weave: Does each technical detail map to a functional outcome?
4. Check depth: Did I go deep where complex and shallow where simple?
5. Check structure: Did I vary form to distribute attention?

If any answer is "no," rewrite that paragraph or section. This loop is the fast path to high-signal documentation.

---

The Meta-Test

This document should demonstrate what it teaches:

• ✓ Opens with core insight, builds from there
• ✓ Spirals through ideas with increasing depth
• ✓ Mixes structural forms (prose, lists, diagrams, tables, code)
• ✓ Creates attentional weight through variation
• ✓ Includes before/after transformations
• ✓ Weaves functional with technical
• ✓ Uses progressive disclosure

If this document is comprehensible and useful, the principles work.

---

Remember: You're not following rules. You're thinking about how information encodes and transmits. The patterns are heuristics that usually work. When they don't fit, understand why and adapt.

---

Reference: Testing Reference

Testing Reference

Philosophy: Service Mocks

Service mocks are in-process tests at public entry points. They test as close to where external calls enter your code as possible, exercise all internal pathways, and mock only at external boundaries. Not unit tests (too fine-grained, mock internal modules). Not end-to-end tests (too slow, require deployed systems). Service mocks hit the sweet spot.

Core Principle

Test at the entry point. Exercise the full component. Mock only what you must.

Your Code
┌─────────────────────────────────────────────────────┐
│  Entry Point (API handler, exported function, etc.) │ ← Test here
│         ↓                                           │
│  Internal logic, state, transformations             │ ← Exercised, not mocked
│         ↓                                           │
│  External boundary (network, DB, filesystem)        │ ← Mock here
└─────────────────────────────────────────────────────┘

Why Service Mocks Work

Traditional unit tests mock at module/class boundaries — testing UserService by mocking UserRepository. This hides integration bugs between your own components.

Service mocks push the mock boundary outward to where your code ends and external systems begin. You test real integration between your modules while keeping tests fast and deterministic.

The insight: Your code is one unit. External systems are the boundary.

Mock Strategy

Boundary	Mock?	Why
Off-machine (network, external APIs, services)	Always	Speed, reliability, no external dependencies
On-machine, out-of-process (local database, Redis)	Usually	Speed; judgment call based on setup complexity
In-process (your code, your modules)	Never	That's what you're testing

The Two Test Layers

Coverage comes from two complementary layers:

Layer 1: Service mocks (primary)

• Many tests, fast, in-process
• This is where TDD lives
• Coverage goals met here
• Run on every save, every CI build

Layer 2: Wide integration tests (secondary)

• Few tests, slower, require deployed environment
• Verify deployed pieces work together
• Catch configuration and wiring issues
• Run locally before merge, post-CD as verification — NOT on CI

┌──────────────────────────────────────────────────┐
│  Service Mocks (many, fast, in-process)         │  ← TDD lives here
│  Coverage goals met here                         │
└──────────────────────────────────────────────────┘
                        +
┌──────────────────────────────────────────────────┐
│  Wide Integration Tests (few, slower, deployed)  │  ← Smoke tests, critical paths
│  Run locally + post-CD, not CI                   │
└──────────────────────────────────────────────────┘

Confidence Distribution

Service mocks provide high confidence for logic and behavior. Wide integration tests provide confidence for deployment and wiring. Together they cover most failure modes.

What they can't cover: Visual correctness, UX feel, edge cases you didn't anticipate. That's what gorilla testing is for.

---

API Testing (Deepest Section)

API testing is the cleanest application of service mocks. The entry point is obvious (the HTTP handler), the boundaries are clear (external services), and the response is easily asserted. This is the pattern to internalize — UI testing adapts it with more friction.

Pattern: Test the Route Handler

Get as close to the HTTP handler as possible. Use your framework's test injection (Fastify's inject(), Express's supertest, etc.) to send requests without network overhead.

// Service mock test for POST /api/prompts
describe("POST /api/prompts", () => {
  let app: FastifyInstance;

  beforeEach(async () => {
    app = buildApp();  // Your app factory
    await app.ready();
  });

  afterEach(async () => {
    await app.close();
  });

  describe("authentication", () => {
    // TC-1: requires authentication
    test("returns 401 without auth token", async () => {
      const response = await app.inject({
        method: "POST",
        url: "/api/prompts",
        payload: { prompts: [] },
      });

      expect(response.statusCode).toBe(401);
    });
  });

  describe("validation", () => {
    // TC-2: validates input
    test("returns 400 with invalid slug format", async () => {
      const response = await app.inject({
        method: "POST",
        url: "/api/prompts",
        headers: { authorization: `Bearer ${testToken()}` },
        payload: {
          prompts: [{ slug: "Invalid:Slug", name: "Test", content: "Test" }],
        },
      });

      expect(response.statusCode).toBe(400);
      expect(response.json().error).toMatch(/slug/i);
    });
  });

  describe("success paths", () => {
    // TC-3: creates prompt and returns ID
    test("persists to database and returns created ID", async () => {
      mockDb.insert.mockResolvedValue({ id: "prompt_123" });

      const response = await app.inject({
        method: "POST",
        url: "/api/prompts",
        headers: { authorization: `Bearer ${testToken({ sub: "user_1" })}` },
        payload: {
          prompts: [{ slug: "my-prompt", name: "My Prompt", content: "Content" }],
        },
      });

      expect(response.statusCode).toBe(201);
      expect(response.json().ids).toContain("prompt_123");
      expect(mockDb.insert).toHaveBeenCalledWith(
        expect.objectContaining({ slug: "my-prompt", userId: "user_1" })
      );
    });
  });

  describe("error handling", () => {
    // TC-4: handles database errors gracefully
    test("returns 500 when database fails", async () => {
      mockDb.insert.mockRejectedValue(new Error("Connection lost"));

      const response = await app.inject({
        method: "POST",
        url: "/api/prompts",
        headers: { authorization: `Bearer ${testToken()}` },
        payload: { prompts: [{ slug: "test", name: "Test", content: "Test" }] },
      });

      expect(response.statusCode).toBe(500);
      expect(response.json().error).toMatch(/internal/i);
    });
  });
});

Setting Up Mocks

Mock external dependencies before importing the code under test. The pattern is framework-agnostic:

// Mock external boundaries — database, auth service, config
const mockDb = {
  insert: vi.fn(),
  query: vi.fn(),
  delete: vi.fn(),
};
vi.mock("../lib/database", () => ({ db: mockDb }));

vi.mock("../lib/auth", () => ({
  validateToken: vi.fn(async (token) => {
    if (token === "valid") return { valid: true, userId: "user_1" };
    return { valid: false };
  }),
}));

// Reset between tests
beforeEach(() => {
  vi.clearAllMocks();
});

What Makes a Good API Service Mock Test

1. Tests one behavior — authentication, validation, success path, or error handling
2. Uses real request/response — not calling internal functions directly
3. Mocks only external boundaries — database, auth service, external APIs
4. Asserts on observable behavior — status code, response body, side effects
5. Traces to a TC — comment links back to spec

Wide Integration Tests for APIs

After service mocks verify logic, wide integration tests verify the deployed system works:

// Integration test — runs against deployed staging
describe("Prompts API Integration", () => {
  const baseUrl = process.env.TEST_API_URL;
  let authToken: string;

  beforeAll(async () => {
    authToken = await getTestAuth();
  });

  test("create and retrieve prompt round trip", async () => {
    const slug = `test-${Date.now()}`;

    // Create
    const createRes = await fetch(`${baseUrl}/api/prompts`, {
      method: "POST",
      headers: { Authorization: `Bearer ${authToken}`, "Content-Type": "application/json" },
      body: JSON.stringify({ prompts: [{ slug, name: "Test", content: "Test" }] }),
    });
    expect(createRes.status).toBe(201);

    // Retrieve
    const getRes = await fetch(`${baseUrl}/api/prompts/${slug}`, {
      headers: { Authorization: `Bearer ${authToken}` },
    });
    expect(getRes.status).toBe(200);
    expect((await getRes.json()).slug).toBe(slug);

    // Cleanup
    await fetch(`${baseUrl}/api/prompts/${slug}`, {
      method: "DELETE",
      headers: { Authorization: `Bearer ${authToken}` },
    });
  });
});

When to run:

• Locally before merge
• Post-CD as deployment verification
• NOT on CI (too slow, requires deployed environment)

---

UI Testing (Lighter Section)

UI testing follows the same service mock philosophy but with more friction. The "entry point" is less clear, browser APIs complicate mocking, and visual/UX correctness can't be verified programmatically.

Same ideals, messier execution. UI tests can't match API test confidence. Aim for behavioral coverage, then rely on gorilla testing for visual/UX verification.

The Principle Applied to UI

Mock at the API layer (fetch calls, API client). Let UI framework internals (state, hooks, DOM updates) run for real. Test user interactions and their effects.

UI Code
┌─────────────────────────────────────────────────────┐
│  User Interaction (click, type, submit)             │ ← Simulate here
│         ↓                                           │
│  Component logic, state, framework internals        │ ← Runs for real
│         ↓                                           │
│  API calls (fetch, client library)                  │ ← Mock here
└─────────────────────────────────────────────────────┘

HTML/JS (No Framework)

For plain HTML with JavaScript, use jsdom to load templates and test behavior:

import { JSDOM } from "jsdom";

describe("Prompt Editor", () => {
  let dom: JSDOM;
  let fetchMock: vi.Mock;

  beforeEach(async () => {
    dom = await JSDOM.fromFile("src/prompt-editor.html", { runScripts: "dangerously" });
    fetchMock = vi.fn(() => Promise.resolve({ ok: true, json: () => ({ id: "new_id" }) }));
    dom.window.fetch = fetchMock;
  });

  // TC-3: Submit valid form creates prompt
  test("submitting form calls POST /api/prompts", async () => {
    const doc = dom.window.document;
    doc.getElementById("slug").value = "new-prompt";
    doc.getElementById("name").value = "New Prompt";
    doc.getElementById("prompt-form").dispatchEvent(new dom.window.Event("submit"));

    await new Promise((r) => setTimeout(r, 50));

    expect(fetchMock).toHaveBeenCalledWith("/api/prompts", expect.objectContaining({ method: "POST" }));
  });
});

React / Component Frameworks

Same principle — mock API layer, let framework run for real:

import { render, screen, waitFor } from "@testing-library/react";
import userEvent from "@testing-library/user-event";

// Mock API layer, NOT hooks or components
vi.mock("@/api/promptApi");

describe("PromptList", () => {
  // TC-7: displays prompts when loaded
  test("renders prompt list from API", async () => {
    mockPromptApi.getAll.mockResolvedValue([{ id: "1", name: "Prompt 1" }]);

    render(<PromptList />);

    await waitFor(() => {
      expect(screen.getByText("Prompt 1")).toBeInTheDocument();
    });
  });

  // TC-8: shows error on failure
  test("displays error when API fails", async () => {
    mockPromptApi.getAll.mockRejectedValue(new Error("Failed"));

    render(<PromptList />);

    await waitFor(() => {
      expect(screen.getByText(/error/i)).toBeInTheDocument();
    });
  });
});

E2E for Critical Paths (Playwright)

E2E tests serve as wide integration for UI — verify the full deployed stack works:

test("user can create and view prompt", async ({ page }) => {
  await page.goto("/prompts");
  await page.click('[data-testid="new-prompt-button"]');
  await page.fill('[data-testid="slug-input"]', "e2e-test");
  await page.fill('[data-testid="name-input"]', "E2E Test");
  await page.click('[data-testid="submit-button"]');

  await expect(page).toHaveURL(/\/prompts\/e2e-test/);
  await expect(page.getByText("E2E Test")).toBeVisible();
});

Run locally and post-CD, not on CI.

UI Testing Limitations

Acknowledge the gap: UI testing cannot match API testing confidence. Visual correctness, UX polish, interaction feel — not verifiable programmatically.

Plan for more gorilla testing. Plan for iterative polish. The GORILLA phase exists partly for this.

---

CLI Testing

For CLI tools, the entry point is the command handler. The same service mock principle applies: test at the entry point, exercise internal modules through it, mock only at external boundaries (filesystem, network, child processes).

The Principle Applied to CLI

CLI Code
┌─────────────────────────────────────────────────────┐
│  Command handler (yargs, commander, etc.)           │ ← Test here
│         ↓                                           │
│  Internal orchestration (executors, managers)        │ ← Exercised, not mocked
│         ↓                                           │
│  Pure algorithms (parsing, transforming)             │ ← Can test directly (no mocks needed)
│         ↓                                           │
│  Filesystem / network / child processes             │ ← Mock here
└─────────────────────────────────────────────────────┘

Layer	Mock?	Why
Command handler	Test here	Entry point
Internal orchestration (executors, managers)	Don't mock	Exercise through command
Pure algorithms (no IO)	Can test directly	No mocking needed, supplemental coverage
Filesystem / network / child processes	Mock	External boundary

Correct Structure

tests/
├── commands/              # Entry point tests (primary coverage)
│   ├── edit-command.test.ts    # Full edit flow, mocks filesystem
│   ├── clone-command.test.ts   # Full clone flow, mocks filesystem
│   └── list-command.test.ts    # Full list flow, mocks filesystem
└── algorithms/            # Pure function tests (supplemental)
    └── tool-call-remover.test.ts  # No mocks, edge case coverage

Anti-Pattern

tests/
├── edit-operation-executor.test.ts  # ❌ Internal module with mocked fs
├── backup-manager.test.ts           # ❌ Internal module with mocked fs
├── tool-call-remover.test.ts        # ✓ Pure algorithm, ok
└── edit-command.test.ts             # ✓ Entry point, ok

The anti-pattern tests internal modules in isolation with mocked dependencies. This hides integration bugs between your own components — exactly what service mocks avoid. An agent seeing API and UI examples ("test the route handler," "test the component") will pattern-match to "test the executor, test the manager" unless given explicit CLI guidance.

---

Convex Testing

Convex functions are serverless handlers. Same service mock principle — mock external boundaries, test the function directly:

describe("withApiKeyAuth wrapper", () => {
  beforeEach(() => {
    process.env.CONVEX_API_KEY = "test_key";
  });

  test("validates API key and calls handler", async () => {
    const handler = vi.fn(async (ctx, args) => ({ userId: args.userId }));
    const wrapped = withApiKeyAuth(handler);

    const result = await wrapped({}, { apiKey: "test_key", userId: "user_1" });

    expect(result.userId).toBe("user_1");
    expect(handler).toHaveBeenCalled();
  });

  test("rejects invalid API key", async () => {
    const wrapped = withApiKeyAuth(vi.fn());

    await expect(wrapped({}, { apiKey: "wrong", userId: "user_1" })).rejects.toThrow("Invalid");
  });
});

---

TC Traceability

Every test must trace to a Test Condition from the Epic. This is the Confidence Chain in action.

In Test Code

describe("POST /api/prompts", () => {
  // TC-1: requires authentication
  test("returns 401 without auth token", async () => { ... });

  // TC-2: validates slug format
  test("returns 400 with invalid slug", async () => { ... });
});

In Test Plan

TC ID	Test File	Test Name	Status
TC-1	createPrompts.test.ts	returns 401 without auth token	Passing
TC-2	createPrompts.test.ts	returns 400 with invalid slug	Passing

Rules:

• TC ID in comment or test name
• Every TC from spec must have at least one test
• Can't write a test? TC is too vague — return to spec

---

Anti-Patterns

Asserting on NotImplementedError

// ❌ Passes before AND after implementation
it("throws not implemented", () => {
  expect(() => createPrompt(data)).toThrow(NotImplementedError);
});

// ✅ Tests actual behavior
it("creates prompt and returns ID", async () => {
  const result = await createPrompt(data);
  expect(result.id).toBeDefined();
});

Over-Mocking

// ❌ Mocking your own code hides bugs
vi.mock("../hooks/useFeature");
vi.mock("../components/FeatureList");

// ✅ Mock only external boundaries
vi.mock("../api/featureApi");

Testing Implementation Details

// ❌ Internal state
expect(component.state.isLoading).toBe(true);

// ✅ Observable behavior
expect(screen.getByTestId("loading")).toBeInTheDocument();

---

Test Organization

tests/
├── service/           # Service mock tests (primary)
│   ├── api/
│   │   └── prompts.test.ts
│   └── ui/
│       └── prompt-editor.test.ts
├── integration/       # Wide integration tests
│   ├── api.test.ts
│   └── ui.test.ts
└── fixtures/
    └── prompts.ts

Track running totals across stories. Previous tests must keep passing — regression = stop and fix.

---

Template

Technical Design: [Feature Name]

Purpose

This document translates feature requirements into implementable architecture. It serves three audiences:

Audience	Value
Reviewers	Validate design before code is written
Developers	Clear blueprint for implementation
Story Tech Sections	Source of implementation targets, interfaces, and test mappings

Prerequisite: The epic must be complete (all ACs have TCs) before starting this document.

Expected Length: A complete tech design expands significantly from the epic — typically 6-7×. The richness comes from redundant connections: the same concepts appearing at multiple altitudes, woven through functional and technical perspectives. Shorter usually means insufficient depth. Longer usually means scope creep.

Output structure — choose one:

Config	Documents	When
A: 2 docs	tech-design.md (index) + test-plan.md	Index fits comfortably under ~1200-1500 lines. Single-domain projects (CLI, backend service, focused frontend feature).
B: 4 docs	tech-design.md (index) + tech-design-[domain-a].md + tech-design-[domain-b].md + test-plan.md	Index approaching ~1200-1500 lines. Multi-domain projects (frontend + backend, client + server).

Never go 3 — don't add just one companion. Either everything lives in the index, or split both domain companions out at the same time. Name companions by the project's actual domain boundaries (frontend/backend, client/server, API/UI — whatever fits).

The index remains the decision record and whole-system map regardless of configuration. In Config B, the index retains: purpose, spec validation, context, tech design Q answers, system view, module architecture overview, dependency map, work breakdown summary, deferred items. Companion docs carry implementation depth and maintain requirement traceability (AC/TC references, cross-links back to the index).

The test plan is always its own document. If the work doesn't justify a separate test plan with TC traceability, this skill isn't the right tool.

---

⚠️ About this template — guidance vs. output. This template contains inline methodology commentary (marked with ✏️ or explaining why the template is structured a certain way). That commentary is instruction to you about how to write — it is not content for the output document. Do not reproduce methodology labels, altitude references, spiral pattern mentions, or "this repetition is intentional" explanations in your tech design. Your headings should be descriptive ("System Context", "Module Architecture"), not methodology-labeled ("High Altitude: System Context"). Write the weave — don't announce it.

---

Spec Validation

Before designing, validate the Epic is implementation-ready. You are the downstream consumer—if you can't design from it, the spec isn't ready.

Validation Checklist:

• Every AC maps to clear implementation work
• Data contracts are complete and realistic
• Edge cases have TCs, not just happy path
• No technical constraints the BA missed
• Flows make sense from implementation perspective

Issues Found:

This table serves two purposes. First, it captures pre-design validation issues — problems with the spec that need resolution before design begins. Second, it captures design-time deviations — places where the tech design intentionally diverges from the epic because implementation reality demands a different approach. Both are expected. Deviations without documented rationale look like bugs to verifiers; deviations with rationale are accepted as design decisions and reduce verification churn significantly.

Issue	Spec Location	Resolution	Status
[Pre-design issue — spec problem]	AC-X	[Fix needed or applied]	Resolved
[Design-time deviation — different approach chosen]	AC-Y, Data Contracts	[What the epic says, what the design does instead, and why]	Resolved — deviated
[Clarification — spec is ambiguous, design makes it concrete]	AC-Z, A3	[What was ambiguous, how the design interprets it]	Resolved — clarified

If blocking pre-design issues exist, return to BA for revision. Don't design from a broken spec. Document what you found — even minor issues — so there's a record of spec evolution.

Design-time deviations are different: the spec may be correct but the implementation reality demands a different approach. Document the deviation, explain the rationale, and keep designing. The epic is the requirements source of truth; the tech design is the implementation source of truth. When they diverge, the deviation table is where that divergence is made explicit.

---

Context

This section establishes the "why" behind architectural choices. Write 3-5 paragraphs covering the landscape that shaped this design. Someone entering this document without prior conversation should understand:

• What triggered this feature? Business driver, user need, or technical debt that created the work.
• What constraints shaped the design? Existing systems that can't change, timeline pressure, risk tolerance, team capabilities.
• What major decisions were made early? Technology choices, boundary decisions, trade-offs accepted.
• How does this connect to adjacent work? Related features, future phases, dependencies on other teams.

The goal is rich context that survives isolated reading. Don't summarize—immerse the reader in the problem space so architectural choices feel inevitable rather than arbitrary.

Example (note the paragraph depth):

This feature addresses Guidewire users who need to add multiple locations to commercial policies during the quoting process. Currently, users exit Guidewire entirely, navigate to the legacy location system, and manually re-enter policy context. The round-trip takes 8-12 minutes and is the #2 complaint in broker feedback surveys.

The primary constraint is iframe embedding—Guidewire's extension framework prohibits navigation or popups. All data must flow through URL parameters (in) and redirect URLs (out). This limits payload size to roughly 2KB encoded, which influenced our decision to return location IDs rather than full location objects. The parent application will re-fetch details as needed.

We chose to isolate this feature under the /locations namespace, creating a parallel implementation rather than extending v1 flows. This increases some code duplication (shared components will be extracted in Phase 2) but eliminates integration risk during the policy renewal window in Q4. The v1 system handles 40% of premium volume; we cannot risk destabilization.

The design assumes the XAPI team delivers their location search endpoint by Sprint 23. If delayed, Chunk 2 (search functionality) slides but Chunk 1 (browse existing locations) can proceed independently.

---

System View

✏️ Highest altitude — broadest view, external boundaries. Establish the full picture before descending into modules. Config B: this section stays in the index.

Start at the highest level. How does this feature fit into the broader system? What crosses the application boundary?

System Context Diagram

Show external actors and systems. For complex integrations, consider building the diagram progressively—start with core actors, then add external systems, then show the full picture. This helps readers construct understanding rather than absorb a complete diagram all at once.

flowchart LR
    subgraph External
        GW[Guidewire Policy Center]
    end

    subgraph "Location App"
        UI[React UI]
        XAPI[Experience API]
    end

    subgraph "Backend Services"
        API[Domain API]
        DB[(Database)]
    end

    GW -->|"query params"| UI
    UI -->|"API calls"| XAPI
    XAPI -->|"domain calls"| API
    API -->|"read/write"| DB
    UI -->|"redirect + data"| GW

External Contracts

What crosses the boundary? This section connects to the epic's Data Contracts and establishes what the implementation must honor. These contracts become the fixed points around which internal architecture flexes.

Incoming (from Guidewire):

Describe what arrives and why. The table enumerates; the prose contextualizes.

Parameter	Required	Source	Purpose
param1	Yes	Query string	Description

Outgoing (to Guidewire):

Describe what returns and the format constraints. Note any size limits, encoding requirements, or ordering expectations.

Data	Format	Destination	Purpose
Location data	Base64 JSON	Redirect URL	Return selected/created locations

Error Responses:

Errors are part of the contract. Define shapes so tests can mock realistic failures and UI can handle them gracefully. These error shapes should appear again in the testing section—that redundancy is intentional, creating multiple paths to the same information.

Source	Status	Code	Shape	Client Handling
XAPI	400	VALIDATION_FAILED	{ status: 'ERROR', code: string, messages: [...] }	Show validation message
XAPI	500	INTERNAL_ERROR	{ status: 'ERROR', code: string, messages: [...] }	Show generic error