@hegemonart/get-design-done 1.57.1 → 1.57.3

package/dist/claude-code/.claude/skills/bandit-status/SKILL.md

@@ -1,94 +0,0 @@
----
-name: gdd-bandit-status
-description: "Surface read-only per-(agent, bin, delegate) bandit posterior snapshot - alpha/beta/mean/stddev/count/last-used per arm. Phase 27.5 (v1.27.5) diagnostic. Use when investigating 'why did the bandit pick tier X for agent Y?' or when verifying posterior convergence after enabling adaptive_mode: full."
-argument-hint: ""
-tools: Read, Bash
----
-
-# gdd-bandit-status
-
-## Role
-
-You are a deterministic, read-only diagnostic skill. You do not spawn agents and do not modify the posterior. You read `.design/telemetry/posterior.json` (path declared by `scripts/lib/bandit-router.cjs`'s `DEFAULT_POSTERIOR_PATH`), aggregate per-`(agent, bin, delegate, tier)` arm state, and emit a single Markdown table. Read-only per Phase 27.5 D-11 - to reset, use `/gdd:bandit-reset` (Phase 23.5). See `./reference/bandit-integration.md` for setup, interpretation, and convergence guidance.
-
-## Invocation Contract
-
-- **Input**: none.
-- **Output**: a Markdown bandit-status table to stdout. The table is the entire output.
-
-## Procedure
-
-### 1. Locate the posterior file
-
-Read `.design/telemetry/posterior.json`. Missing → emit empty-state message:
-
-```
-## Bandit Posterior Snapshot
-
-No posterior data yet — run a few pipeline cycles with `adaptive_mode: full` first.
-
-No posterior data found at `.design/telemetry/posterior.json`.
-
-Possible reasons:
-- `adaptive_mode` is `static` or `hedge` (bandit silent — see `.design/budget.json`).
-- No spawns have fired since Phase 27.5 wiring landed.
-- Posterior was cleared via `/gdd:bandit-reset`.
-
-See `reference/bandit-integration.md` for setup guidance.
-```
-
-Skip to Section 4 (Record). Parse failure (truncated/corrupted) → emit `Posterior file exists but is unparseable. Run /gdd:bandit-reset to start fresh, or restore from a backup.`
-
-### 2. Parse the posterior
-
-Schema:
-
-```json
-{
-  "schema_version": "1.0.0",
-  "generated_at": "<ISO>",
-  "arms": [{ "agent": "...", "bin": "...", "tier": "...", "delegate": "...", "alpha": N, "beta": N, "last_used": "...", "count": N }]
-}
-```
-
-The `delegate` field is optional - absent = Phase 23.5 legacy slice (rendered as `-` in the table).
-
-### 3. Render the table
-
-Compute per arm: `mean = alpha / (alpha + beta)` (3 decimals), `stddev = sqrt(alpha*beta / ((alpha+beta)^2 * (alpha+beta+1)))` (3 decimals).
-
-Sort by `(agent ASC, bin ASC, delegate ASC where '-' first, tier ASC opus<sonnet<haiku, last_used DESC)`. Group by agent for readability.
-
-Emit:
-
-```
-## Bandit Posterior Snapshot
-
-Per-(agent, bin, delegate, tier) posterior state. Read-only — to reset use `/gdd:bandit-reset` (Phase 23.5).
-
-Posterior file: `.design/telemetry/posterior.json` (last updated: <generated_at>)
-Total arms: <count>
-
-| Agent | Bin | Delegate | Tier | Alpha | Beta | Mean | Stddev | Count | Last Used |
-|-------|-----|----------|------|-------|------|------|--------|-------|-----------|
-| ...   | ... | ...      | ...  | ...   | ...  | ...  | ...    | ...   | ...       |
-
-> Mean = alpha / (alpha + beta). Stddev = sqrt(alpha*beta / ((alpha+beta)^2 * (alpha+beta+1))).
-> Delegate '-' = Phase 23.5 legacy slice (equivalent to 'none').
-> See `reference/bandit-integration.md` for interpretation.
-> Read-only — use `/gdd:bandit-reset` to clear posterior state.
-```
-
-Precision: alpha/beta 2 decimals; mean/stddev 3 decimals; count integer; `last_used` truncated to minute (`YYYY-MM-DDTHH:MM`); null `last_used` renders `-`.
-
-After the table, surface a per-`(agent, bin)` best-arm summary: for each unique pair, identify highest-mean arm (tie-broken by `count` DESC) - answers "why did the bandit pick tier X?" at a glance.
-
-### 4. Record
-
-Append one JSONL line to `.design/skill-records.jsonl`: `{"skill":"gdd-bandit-status","ts":"<ISO>","arms_seen":<count>,"posterior_present":<bool>}`. Skill writes ONLY to skill-records.jsonl (telemetry); never touches the posterior.
-
-## Cross-references
-
-- `./reference/bandit-integration.md` - operator guide; interpretation patterns.
-- `scripts/lib/bandit-router.cjs` (Phase 23.5) - posterior shape, `DEFAULT_POSTERIOR_PATH`, `loadPosterior()`.
-- `scripts/lib/bandit-router/integration.cjs` (27.5-01), `hooks/budget-enforcer.ts` (27.5-02), `scripts/lib/session-runner/index.ts` (27.5-03), `scripts/lib/bandit-arbitrage.cjs` (27.5-04), `/gdd:bandit-reset` (Phase 23.5) - only surface that mutates the posterior.

package/dist/claude-code/.claude/skills/benchmark/SKILL.md

@@ -1,65 +0,0 @@
----
-name: gdd-benchmark
-description: "Harvest and synthesize per-component design benchmarks from 18 design systems and produce canonical component specs at `reference/components/<name>.md`. Use when adding a new component spec, running a benchmark wave, listing corpus coverage, or refreshing a spec after a design-system version bump."
-argument-hint: "<component> | --wave <N> | --list | --refresh <component>"
-tools: Read, Write, Bash, Grep, Glob, Task, WebFetch
----
-
-# /gdd:benchmark
-
-Harvest per-component design knowledge from 18 design systems and synthesize canonical
-specs at `reference/components/<name>.md`. The 18-source corpus + fallback chain lives in
-`../../connections/design-corpora.md`. Per-skill output discipline + completion-marker
-conventions are at `../../reference/shared-preamble.md#output-contract-reminders`.
-
-## Invocation Modes
-
-| Invocation | Action |
-|------------|--------|
-| `/gdd:benchmark <component>` | Harvest + synthesize a single component |
-| `/gdd:benchmark --wave <N>` | Run a full wave (1 = Inputs, 2 = Containers, etc.) |
-| `/gdd:benchmark --list` | Show corpus coverage - which specs exist, which are pending |
-| `/gdd:benchmark --refresh <component>` | Re-harvest a spec (for design-system version bumps) |
-
-## Single-Component Flow (`/gdd:benchmark <component>`)
-
-1. **Check if spec exists** - `Glob("reference/components/<component>.md")`. If found and `--refresh` was not passed, confirm before overwriting.
-
-2. **Harvest** - spawn `component-benchmark-harvester` with required-reading `@connections/design-corpora.md`, target component, raw-output path `.planning/benchmarks/raw/<component>.md`, and a salvage hint to `.planning/research/impeccable-salvage/`. Acceptance: raw harvest exists with ≥4 source sections.
-
-3. **Synthesize** - spawn `component-benchmark-synthesizer` with required-reading `@.planning/benchmarks/raw/<component>.md`, `@reference/components/TEMPLATE.md`, `@reference/anti-patterns.md`. Output: `reference/components/<component>.md`. Also update `reference/components/README.md` (add entry in correct category). Acceptance: spec ≤350 lines, cites ≥4 systems, follows `TEMPLATE.md` sections, has a WAI-ARIA keyboard contract + a failing-example block.
-
-4. **Report** - print spec path, line count, systems cited.
-
-## Wave Mode (`/gdd:benchmark --wave <N>`)
-
-Read the wave definition from `reference/components/README.md` and run each component sequentially (not parallel - each harvest is network-bound and the raw files are large).
-
-| Wave | Components |
-|------|-----------|
-| 1 | button, input, select-combobox, checkbox, radio, switch, link, label |
-| 2 | card, modal-dialog, drawer, popover, tooltip, accordion, tabs |
-
-Print progress per component: `[N/total] harvesting <component>…`
-
-## List Mode (`/gdd:benchmark --list`)
-
-Read `reference/components/README.md` and diff against `reference/components/*.md` files. Print a table with columns: Component, Status, Wave, Lines.
-
-## Refresh Mode (`/gdd:benchmark --refresh <component>`)
-
-Same as single-component flow but skips the "already exists" guard. Use when a design system ships a breaking update to a component's spec.
-
-## Source List
-
-`../../connections/design-corpora.md` - 18 design systems with canonical URLs, licensing, and fallback chain (canonical → archive.org → Refero MCP → Pinterest MCP).
-
-## Output Artifacts
-
-| Artifact | Purpose |
-|----------|---------|
-| `.planning/benchmarks/raw/<component>.md` | Raw multi-source harvest (input only) |
-| `reference/components/<component>.md` | Canonical spec (distributed with plugin) |
-| `reference/components/README.md` | Corpus index (updated by synthesizer) |
-
-## BENCHMARK COMPLETE

package/dist/claude-code/.claude/skills/bootstrap-ds/SKILL.md

@@ -1,43 +0,0 @@
----
-name: gdd-bootstrap-ds
-description: "Bootstraps a design system for a GREENFIELD project that has none - no Figma, no tokens, no component library. Takes a brand input (primary color + optional secondary + tone tags + target framework) and emits a coherent OKLCH token system (color tints, modular type scale, 4pt/8pt spacing, radius + motion defaults) in 3 variants to pick from, then scaffolds proof components (button/input/card). Use at the start of a brand-new project, or when /gdd:discover finds no existing design system. Never invents a brand; never overwrites an existing DS."
-argument-hint: "[--primary <color>] [--secondary <color>] [--tone <tags>] [--framework web|native-ios|native-android|flutter]"
-user-invocable: true
-tools: Read, Write, Bash, Glob, Grep, AskUserQuestion, Task
----
-
-# /gdd:bootstrap-ds
-
-Greenfield design-system bootstrap. Closes the gap that GDD's `design-context-builder` assumes a design system already exists (in code or Figma) - a brand-new project has none. This skill is the **front door**: it collects a brand input and hands it to `agents/ds-generator.md`, which emits the token system + proof components per `reference/ds-bootstrap-rubric.md` (deterministic math in `scripts/lib/ds/token-scale.cjs`).
-
-## When to use
-
-- At the start of a brand-new project (no `tailwind.config`, no token file, no DS).
-- When `/gdd:discover` / `design-context-builder` reports **no existing design system** and the user opts to bootstrap one.
-
-If a design system **already exists**, do NOT run this - defer to `design-context-builder` (it maps the existing one). State that and stop.
-
-## Steps
-
-1. **Collect the brand input.** From flags, or via `AskUserQuestion` for anything missing:
-   - **primary** (required) - the brand color (hex / rgb / `oklch()`).
-   - **secondary** (optional) - a second brand color (emitted only if supplied - the rubric's ≤2-brand-colors rule).
-   - **tone tags** (optional) - `calm` / `corporate` / `editorial` / `playful` / `bold` (maps to the type ratio + chroma treatment).
-   - **target framework** (optional) - `web` (default) / `native-ios` / `native-android` / `flutter`. Detect from the project if absent (Phase 34 routing).
-2. **Delegate to `ds-generator`** (via `Task`): it resolves the primary to OKLCH, runs `token-scale.cjs` for **3 variants** (conservative / balanced / bold), and presents them.
-3. **Pick a variant.** Surface the 3 variants (primary `500`, type ratio, spacing baseline, radius, feel); the user picks ONE.
-4. **Emit + scaffold (proposal → confirm).** `ds-generator` emits the chosen token set (role-named CSS custom properties + the framework mapping) and scaffolds **button / input / card** as a coherence proof. Nothing is written to `src/` without confirmation.
-
-## Do Not
-
-- Do not invent a brand (no logomark, no typeface choice, no third brand color) - emit a token *system*, not an identity.
-- Do not overwrite an existing design system - defer to `design-context-builder`.
-- Do not add a color-conversion dependency - `token-scale.cjs` emits native CSS `oklch()`.
-
-## Output
-
-End with:
-
-```
-## BOOTSTRAP-DS COMPLETE
-```

package/dist/claude-code/.claude/skills/brief/SKILL.md

@@ -1,145 +0,0 @@
----
-name: gdd-brief
-description: "Stage 1 of 5 design intake that captures problem statement, audience, constraints, success metrics, and scope into .design/BRIEF.md, and bootstraps .design/STATE.md if missing. Use when starting a new design cycle and before /gdd:explore. Activates for requests involving capturing a problem statement, defining audience and constraints, or starting a new design brief."
-argument-hint: "[--re-brief to redo intake on existing project]"
-tools: Read, Write, AskUserQuestion, mcp__gdd_state__frontmatter_update, mcp__gdd_state__set_status, mcp__gdd_state__update_progress, mcp__gdd_state__get
----
-
-# Get Design Done - Brief
-
-**Role:** You are the Brief stage. Stage 1 of 5 in the get-design-done pipeline.
-
-**Purpose:** Capture the design problem before any scanning or exploration. Produces `.design/BRIEF.md`.
-
----
-
-## Step 1 - Check for existing BRIEF.md
-
-1. Read `.design/BRIEF.md` if it exists.
-2. Parse it into sections: Problem, Audience, Constraints, Success Metrics, Scope.
-3. Note which sections are already answered (non-empty).
-4. If `--re-brief` flag is passed, ignore existing answers and ask all five questions.
-5. Otherwise, only ask questions for unanswered sections.
-
-## Step 2 - Interview
-
-Ask the following one at a time using `AskUserQuestion`, only for unanswered sections:
-
-1. **Problem** - "What design problem are we solving? (user-facing outcome)"
-2. **Audience** - "Who is the primary audience? (role, device, context)"
-3. **Constraints** - "What constraints apply? (tech stack, brand, time, a11y requirements)"
-4. **Success Metrics** - "How will we measure success? (specific metrics or outcomes)"
-5. **Scope** - "What is in/out of scope for this cycle?"
-
-Do not proceed to the next question until the current one is answered.
-
-## Step 3 - Write .design/BRIEF.md
-
-Write the brief with these sections, preserving any pre-existing answers:
-
-```markdown
-# Design Brief — <project name>
-
-## Problem
-<answer>
-
-## Audience
-<answer>
-
-## Constraints
-<answer>
-
-## Success Metrics
-<answer>
-
-## Scope
-<answer>
-
-<prior-research>
-<!-- Phase 38: populated by agents/user-research-synthesizer.md from UserTesting/Maze/Hotjar
-     (pseudonymized) — ranked findings {finding · frequency · severity}. Empty on a fresh brief;
-     the verify stage cross-checks each finding (addressed or explicitly deferred). -->
-</prior-research>
-```
-
-Leave the `<prior-research>` block empty on a greenfield brief - it is filled by `/gdd:research-sync` (the `user-research-synthesizer`) when a research source is connected, and re-checked at verify. See `reference/design-variants.md` for the outcome loop.
-
-## Step 4 - Bootstrap STATE.md (if missing)
-
-<!-- BOOTSTRAP EXCEPTION: STATE.md does not exist yet — MCP tools require it to exist. Direct Write is intentional. All subsequent mutations use MCP. -->
-
-If `.design/STATE.md` does not exist, copy the template block from `reference/STATE-TEMPLATE.md` (between `==== BEGIN TEMPLATE ====` and `==== END TEMPLATE ====`) to `.design/STATE.md` via `Write`. Leave the `<ISO 8601 timestamp>` placeholders in-place - Step 5 stamps them via MCP. If STATE.md already exists, skip to Step 5.
-
-## Step 5 - Commit STATE.md initialization
-
-With `.design/STATE.md` seeded from the template:
-
-1. Stamp timestamps + cycle id: call `mcp__gdd_state__frontmatter_update` with `patch: { started_at: <ISO>, last_checkpoint: <ISO>, cycle: <cycle-id> }`.
-2. Mark brief progress: call `mcp__gdd_state__update_progress` with `task_progress: "5/5"`, `status: "brief_complete"`.
-3. Set handoff status: call `mcp__gdd_state__set_status` with `status: "brief_complete"`.
-
-Do NOT call `mcp__gdd_state__transition_stage` from brief - explore calls it on entry, keeping the transition atomic with the stage that owns the new state.
-
-## Step 6 - Inline glossary (CONTEXT.md) + ADR pointer
-
-When a fuzzy phrase is resolved or a new domain concept is named during the briefing
-interview: write to `./CONTEXT.md` IMMEDIATELY per `./../../reference/context-md-format.md`
-(H2 heading + body; lazy-create on first term; no batching). Glossary entries compound
-across cycles - token savings + naming consistency.
-
-Project-shaping decisions surfaced in briefing can be promoted to an ADR - see
-`./../../reference/adr-format.md` for the 3-criteria gate (hard-to-reverse AND
-surprising-without-context AND real-tradeoff). Routine choices stay in STATE.md.
-
-## After Writing
-
-```
-━━━ Brief complete ━━━
-Saved: .design/BRIEF.md
-Next: @get-design-done explore
-━━━━━━━━━━━━━━━━━━━━━━━
-```
-
-## Spec self-review (before transition)
-
-Run this final spec-quality pass over `.design/BRIEF.md` before the brief→explore transition:
-- Placeholder scan: no TBD / TODO / `<placeholder>` / lorem left in the artifact.
-- Internal consistency: sections don't contradict each other.
-- Scope check: nothing in the artifact exceeds (or silently drops) the agreed scope.
-- Ambiguity check: every requirement/decision is specific enough to act on without a follow-up question.
-
-## Optional brief audit (non-blocking)
-
-Before the gate, you MAY spawn `agents/brief-auditor.md` via `Task` to grade the brief against the five
-brief anti-patterns (vague verbs, missing audience, immeasurable success criteria, scope creep, missing
-anti-goals). The auditor reads `.design/BRIEF.md` plus `reference/brief-quality-rubric.md` and writes
-advisory findings to `.design/BRIEF-AUDIT.md`. This step is advisory and MUST NOT block the brief to
-explore transition.
-
-If the auditor reports one or more fired anti-patterns, surface a single-line pointer to the user:
-
-```
-Brief audit flagged N issue(s) - run /gdd:discuss brief to refine, or proceed to explore.
-```
-
-The user decides. Proceeding to explore with a flagged brief is allowed; the pointer is a nudge, not a gate.
-If the auditor reports no fired anti-patterns, or you skip the audit, continue to the gate unchanged.
-
-<HARD-GATE>
-Do NOT transition to explore (or invoke `/gdd:explore`) until the brief artifact (default `.design/BRIEF.md`) is committed AND the user has approved it. If this project uses a custom `.design` location, read the artifact path from `.design/STATE.md` rather than assuming the default.
-</HARD-GATE>
-
-## Rationalizations - Thought to Reality
-
-The excuses an agent invents to skip or shortcut the brief, and what each one actually costs the cycle:
-
-| Thought | Reality |
-|---------|---------|
-| "This brief is too simple to need a problem statement." | Skip the brief = guess at requirements, then redesign mid-design when the real problem surfaces. |
-| "The user told me what to build, I can skip the interview." | Unasked constraints (a11y, brand, stack) become rework - the five questions exist because each one has blown a past cycle. |
-| "I'll capture success metrics later in verify." | Verify has nothing to check against; an un-metricked brief produces an un-verifiable cycle. |
-| "Scope is obvious, I don't need an in/out line." | Undeclared scope is scope creep waiting to happen - the explore scan widens to fill the vacuum. |
-| "I can answer all five questions for the user from context." | AskUserQuestion one-at-a-time exists because batched/assumed answers smuggle in wrong premises that compound downstream. |
-| "STATE.md bootstrap can wait." | Every later MCP mutation requires STATE.md to exist; skipping the bootstrap hard-blocks explore on entry. |
-
-## BRIEF COMPLETE

package/dist/claude-code/.claude/skills/budget/SKILL.md

@@ -1,45 +0,0 @@
----
-name: gdd-budget
-description: "Forecasts GDD design-cycle spend before the bill arrives. Reads .design/telemetry/costs.jsonl (cost per cycle) + .design/budget.json (the project_cap), runs the pure cost-forecast model via agents/cost-forecaster.md, and projects the next N cycles - surfacing 'at the current rate you'll hit your $X project cap in Y cycles.' Supports --scenario best|typical|worst and --cycles N. Read-only - it forecasts and warns; it never spends, edits budget.json, or halts (the budget-enforcer hook halts). Use to sanity-check spend trajectory before a long run."
-argument-hint: "[--cycles N] [--scenario best|typical|worst]"
-user-invocable: true
-tools: Read, Bash, Grep, Glob, ToolSearch, Task
----
-
-# /gdd:budget
-
-Closes the long-horizon cost gap: Phase 10.1 per-task caps + Phase 26 per-runtime telemetry track
-*cost*, but nothing **forecasts** it. This skill projects the next N cycles of spend and tells you how
-many cycles you have before you hit your `project_cap`. **Read-only** - it forecasts and warns; it
-never spends, never edits `budget.json`, and never halts (the Phase 25 budget-enforcer hook is the
-only thing that blocks a spawn). Contract: `../../reference/cost-governance.md`.
-
-## Invocation
-
-| Command | Behavior |
-|---|---|
-| `/gdd:budget` | Typical-scenario forecast over the next 5 cycles + cycles-to-cap. |
-| `/gdd:budget --cycles N` | Forecast over the next N cycles. |
-| `/gdd:budget --scenario best\|typical\|worst` | Pick the projection rate (best / steady / worst). |
-
-## Steps
-
-1. **Check telemetry exists.** No `.design/telemetry/costs.jsonl` (or zero rows) → print
-   `budget: no cost telemetry yet — run a cycle first.` and exit.
-2. **Delegate to `cost-forecaster`** (via `Task`): it groups `est_cost_usd` by `cycle`, runs the pure
-   `scripts/lib/budget/cost-forecast.cjs` model for the requested `--scenario`/`--cycles`, reads
-   `project_cap_usd` from `.design/budget.json`, and computes cycles-to-cap.
-3. **Render.** Show: the scenario + its per-cycle rate, the best↔worst band, the projected total over
-   N cycles, and - when `project_cap_usd > 0` - **"at the `<scenario>` rate (~$X/cycle) you'll reach
-   your $`<cap>` project cap in `<Y>` cycles"** (or "not at this rate" when the trend is flat/down).
-   When no cap is set, show the trajectory and note that `project_cap_usd` is unset (so the hook won't
-   halt).
-4. **Do not act.** Never raise/lower the cap, never spend - GDD forecasts; the human sets the budget.
-
-## Output
-
-End with:
-
-```
-## BUDGET COMPLETE
-```

package/dist/claude-code/.claude/skills/cache-manager/SKILL.md

@@ -1,66 +0,0 @@
----
-name: gdd-cache-manager
-description: "Maintains .design/cache-manifest.json for Layer B explicit cache per D-08. Computes deterministic SHA-256 input-hash from (agent-path + sorted-input-file-paths + input-content-hashes). On spawn: lookup key → return cached blob if within TTL, else miss. On completion: write result + TTL. Consulted by hooks/budget-enforcer.js before every Agent spawn."
-user-invocable: false
-tools: Read, Bash, Write
-disable-model-invocation: true
----
-
-# gdd-cache-manager
-
-## Role
-
-You are the deterministic cache-key computer and cache-manifest writer for the optimization layer. You do not spawn agents, and you do not make model calls. You read agent paths, input file paths, and input file contents; you compute a stable SHA-256 key; you look that key up in `.design/cache-manifest.json`; you return a hit (cached result + `cache_hit: true`) or a miss. On spawn completion, the orchestrator calls you again to persist the result. You are Layer B of the two-layer cache (D-08). Layer A - Anthropic's 5-min prompt cache - is not owned by you; it's owned by the shared-preamble ordering convention in `agents/README.md`.
-
-## Invocation Contract
-
-### Phase 1: compute-key
-
-- **Input**: `{agent_path: string, input_file_paths: string[]}` - `agent_path` is the absolute or repo-relative path to the `agents/<name>.md` file; `input_file_paths` is the sorted-unique list of files the agent will read.
-- **Output**: `{input_hash: string}` - 64-character lowercase SHA-256 hex.
-- **Algorithm**: canonicalize inputs, concat with newline separators, SHA-256 (see Deterministic Input-Hash Algorithm below).
-
-### Phase 2: lookup
-
-- **Input**: `{input_hash: string}`
-- **Output**: `{hit: true, result: string, expires_at: string} | {hit: false}`
-- Reads `.design/cache-manifest.json`. If key absent → miss. If present and `Date.now() >= entry.expires_at` → miss (lazy cleanup: caller may evict but is not required to). Else → hit with `entry.result`.
-
-### Phase 3: short-circuit-or-miss
-
-- On hit: orchestrator short-circuits the spawn. Budget-enforcer hook (Plan 10.1-01) emits `SkippedCached` telemetry with `tokens_in: 0, tokens_out: 0, cache_hit: true, est_cost_usd: 0` (writer lands in Plan 10.1-05). Returns cached `result` as the tool output.
-- On miss: orchestrator proceeds with real spawn. Cache-manager is not involved until Phase 4.
-
-### Phase 4: write-result-on-completion
-
-- **Input**: `{input_hash: string, agent: string, result: string, ttl_seconds: number}` - `ttl_seconds` defaults to `.design/budget.json.cache_ttl_seconds` (3600 if budget.json absent).
-- **Behavior**: open `.design/cache-manifest.json` (create if missing), set key `input_hash` → `{agent, result, written_at, ttl_seconds, expires_at}`, write file. `written_at` is `new Date().toISOString()`. `expires_at` is `written_at + ttl_seconds` as ISO.
-
-## Deterministic Input-Hash Algorithm
-
-The canonical reference implementation (single source of truth; `hooks/budget-enforcer.js` imports the same primitive via a shared helper) lives in `./cache-policy.md#deterministic-input-hash-algorithm-layer-b` - it documents the JS implementation, the maintainer notes (sorted-unique paths, MISSING-file sentinel, agent-path bust behavior), the manifest shape, and TTL semantics in one place. Conform to the algorithm exactly so the hook and any orchestrator agree byte-for-byte.
-
-## Integration Points
-
-- **`hooks/budget-enforcer.js`** (Plan 10.1-01) reads the manifest on every Agent spawn. The hook already calls `cacheLookup(agent, inputHash)` against `.design/cache-manifest.json`. This skill is the authority on how `inputHash` is computed so the hook and any orchestrator agree byte-for-byte.
-- **Orchestrators** (e.g., `skills/map/`, `skills/discover/`, `skills/plan/`) invoke Phase 1 (compute-key) + Phase 4 (write-result-on-completion) around each Agent spawn they launch. Phase 2 + Phase 3 are executed by the hook.
-- **Warm-cache command** (`skills/warm-cache/SKILL.md`, Task 02) does not touch Layer B - it only primes Anthropic's 5-min prompt cache (Layer A). Do not confuse the two.
-
-## Failure Modes
-
-- `.design/cache-manifest.json` missing or malformed → treat every lookup as miss; orchestrator proceeds with full spawn. Do not throw.
-- File system write fails on Phase 4 → log to stderr, do not throw (the spawn already completed successfully; losing a cache write is a performance regression, not a correctness bug).
-- `agent_path` file missing → compute hash anyway using the provided string; cache entries for a deleted agent simply never hit again.
-
-## Non-Goals
-
-Per D-09:
-
-- No semantic / graph-based lookup. Manifest is a dumb KV store keyed by content hash.
-- No cross-project cache sharing. Manifest lives at `.design/cache-manifest.json`, scoped per repo.
-- No eviction beyond lazy expiry on read. A `/gdd:cache-prune` command is a Phase 11 reflector candidate, not 10.1 scope.
-- No hash-algorithm tuning. SHA-256 is fixed; if a future phase wants BLAKE3, bump the manifest schema version (not relevant in v1).
-
-## TTL Semantics
-
-Default `ttl_seconds` = `.design/budget.json.cache_ttl_seconds` = 3600s (1 hour) per D-10. `expires_at` is computed at write time and stored; readers do not recompute. Stale entries are lazily cleaned on read (no eager reaper in v1). Full TTL discussion: `./cache-policy.md#ttl-semantics-layer-b`.

package/dist/claude-code/.claude/skills/cache-manager/cache-policy.md

@@ -1,126 +0,0 @@
----
-name: cache-policy
-type: heuristic
-version: 1.0.0
-phase: 28.5
-tags: [cache, layer-a, layer-b, sha-256, ttl, warm-cache, cache-manager, anthropic-prompt-cache]
-last_updated: 2026-05-18
----
-
-# Cache Policy (Layer A + Layer B)
-
-Extracted from `skills/cache-manager/SKILL.md` and `skills/warm-cache/SKILL.md` per Phase 28.5
-D-10 (extract-then-link, never delete content). The two skills keep their orchestration
-contracts and step-by-step flows; the deeper algorithmic and operational detail moves here
-so the SKILLs stay under the 100-line cap.
-
-The two layers (D-08):
-
-- **Layer A** - Anthropic's 5-min prompt cache (owned by `warm-cache`). Keyed on shared-preamble-first prompt prefix. No project-local state.
-- **Layer B** - explicit `.design/cache-manifest.json` (owned by `gdd-cache-manager`). Keyed on deterministic SHA-256 of `(agent-path, sorted-input-file-paths, input-content-hashes)`. Per-repo state.
-
-## Deterministic Input-Hash Algorithm (Layer B)
-
-The canonical reference implementation (the single source of truth; `hooks/budget-enforcer.js` imports the same primitive via a shared helper):
-
-```js
-// Deterministic cache-key primitive (reference implementation)
-// hash = SHA-256(
-//   agent_path + "\n" +
-//   sorted(input_file_paths).join("\n") + "\n" +
-//   sorted(input_file_paths)
-//     .map(p => sha256(readFileSync(p, "utf8")))
-//     .join("\n")
-// )
-const crypto = require('crypto');
-const fs = require('fs');
-
-function sha256Hex(s) {
-  return crypto.createHash('sha256').update(s, 'utf8').digest('hex');
-}
-
-function computeInputHash(agentPath, inputFilePaths) {
-  const sortedPaths = [...inputFilePaths].sort();
-  const contentHashes = sortedPaths.map(p => {
-    try { return sha256Hex(fs.readFileSync(p, 'utf8')); }
-    catch { return 'MISSING'; }
-  });
-  const canonical = [
-    agentPath,
-    sortedPaths.join('\n'),
-    contentHashes.join('\n')
-  ].join('\n');
-  return sha256Hex(canonical);
-}
-```
-
-Notes for maintainers:
-
-- **Sorted-unique paths** - ordering must be stable; caller is expected to de-duplicate. If the same path appears twice the hash still matches as long as caller pre-dedupes before invoking.
-- **Missing file** - the string `MISSING` is used in place of the content hash so a missing dependency doesn't silently collide with an empty file (empty file's SHA-256 is `e3b0c44...`). Missing-file hashes naturally miss on the next read because the real file has a different content hash.
-- **Agent-path** - agents changing their own body (role, tools, output contract) invalidate all their cache entries automatically because the agent file's content is not hashed; but the `agent_path` string is concatenated. Upgrading agents between versions naturally busts the cache only when the path changes. Plan 10.1-04 (shared preamble extraction) is expected to slightly adjust agent bodies - consumers should treat the first post-10.1 run as a full cache miss, which is the intended behavior.
-
-## Manifest Shape (Layer B)
-
-See `./config-schema.md` §.design/cache-manifest.json Schema (Phase 10.1) for the authoritative schema. Keyed object, flat SHA-256 hex keys. Example:
-
-```json
-{
-  "a3f1e...": {
-    "agent": "design-verifier",
-    "result": "<base64-or-path>",
-    "written_at": "2026-04-18T12:00:00Z",
-    "ttl_seconds": 3600,
-    "expires_at": "2026-04-18T13:00:00Z"
-  }
-}
-```
-
-## TTL Semantics (Layer B)
-
-- Default `ttl_seconds` = `.design/budget.json.cache_ttl_seconds` = 3600s (1 hour) per D-10.
-- `expires_at` is computed at write time and stored; readers do not recompute.
-- Lazy cleanup: stale entries are not actively deleted on read (overhead for no benefit in normal operation). A separate reaper is optional and out of v1 scope.
-
-## Concrete Warm-Cache Command Examples (Layer A)
-
-Full invocation:
-
-```
-$ /gdd:warm-cache
-
-Warming Anthropic prompt cache for 14 agents (5 min TTL)...
-[1/14] design-verifier ... ok (0.3s)
-[2/14] design-planner ... ok (0.3s)
-[3/14] design-integration-checker ... ok (0.3s)
-...
-[14/14] design-reflector ... ok (0.3s)
-
-## Warm-cache complete
-- Agents warmed: 14
-- Skipped (no shared preamble import): 3  (agents/README.md not an agent; 2 agents not yet migrated to shared preamble)
-- Duration: 4.2s
-- Next 5 min: repeated spawns of these agents pay cached_input_per_1m rate
-```
-
-Filtered invocation:
-
-```
-$ /gdd:warm-cache --agents design-verifier,design-planner
-
-Warming Anthropic prompt cache for 2 agents (filtered from 14)...
-[1/2] design-verifier ... ok (0.3s)
-[2/2] design-planner ... ok (0.3s)
-
-## Warm-cache complete
-- Agents warmed: 2
-- Filtered out by --agents: 12
-- Duration: 0.7s
-```
-
-## Cost Model (Layer A)
-
-- Each no-op Haiku ping: ~50 input tokens (shared preamble + "No-op warm: acknowledge..." system+user) + ~5 output tokens ("ok").
-- At Haiku rates (`./model-prices.md`): `(50 / 1e6) * 1.00 + (5 / 1e6) * 5.00 = $0.00005 + $0.000025 = $0.000075` per agent.
-- 14 agents × $0.000075 = **$0.00105** total for a full warm-cache invocation.
-- Payback: a single subsequent Opus spawn with 40k cached input tokens saves `(40000/1e6) * (15.00 - 1.50) = $0.54`. Warm-cache pays for itself ~500× over on the first repeated planner spawn.