skillpacks 0.1.0

package/global/claude/animation-design-planner/SKILL.md

@@ -0,0 +1,115 @@
+---
+name: animation-design-planner
+description: Plan interactive UI animations before implementation with visible motion contracts, lifecycle ownership, accessibility/performance guardrails, and proof gates
+type: planning
+version: v0.0
+invocation: orchestrator
+argument-hint: "[component, interaction, animation bug, or motion brief]"
+visual_tier: prototype
+---
+
+# Animation Design Planner
+
+Use this skill when a user asks to create, change, debug, or polish an interactive UI animation, transition, or multi-component motion sequence. This is a planning and proof skill, not a direct implementation skill: it prevents local timing tweaks from starting before the visible sequence, lifecycle owner, state phases, and proof method are named.
+
+Use it for drawers, modals, popovers, cards, route transitions, list reordering, shared-element illusions, mount/unmount choreography, staggered content, and visual regressions such as flashing, popping, jumping, duplicated layers, or broken close/open sequencing.
+
+For generated video, Remotion timelines, or non-interactive animation, produce only the transferable storyboard/timing and proof expectations, then route execution to the relevant video or media skill when available. Do not expand this skill into a full video-production workflow yet.
+
+## Workflow
+
+1. **Identify the motion surface**
+   - Name the component, route, screen, or interaction being planned.
+   - Identify the trigger: click, hover, focus, drag, route change, async state, mount, unmount, reorder, or error/recovery state.
+   - Read the relevant local component, style, animation, and test files when the user is asking about an existing codebase.
+   - Treat ambiguous requests such as "smooth this out" or "fix the animation" as requests to define the visible sequence before implementation.
+
+2. **Write the visible motion contract**
+   - Define the start state and end state in visible user terms.
+   - State what the user should see at each transition boundary.
+   - Name what must never happen: flash, pop, jump, duplicate, clipped content, stale layer, focus loss, scroll jump, layout shift, or incorrect z-order.
+   - Define accepted interruption behavior: repeated clicks, quick close after open, route away, async data arrival, resize, reduced-motion preference, or component unmount.
+   - Define the focal point and continuity target: what should keep the user's eye anchored.
+
+3. **Storyboard the timeline**
+   - Break the motion into phases or key frames with approximate timing, easing, and visible state per phase.
+   - Mark focal element, supporting elements, continuity/shared elements, and any staggered content.
+   - Include anticipation, follow-through, or settling only when it improves comprehension rather than decoration.
+   - For shared-element or layout illusions, state which element is visually continuous and which elements are entering, exiting, or temporarily hidden.
+   - Keep the storyboard implementation-neutral until the lifecycle ownership map is clear.
+
+4. **Map lifecycle ownership**
+   - Assign one explicit owner for any multi-component sequence. This can be a page-level controller, state machine, reducer, parent component, or animation coordinator.
+   - Name every state-machine phase. Use phase names that describe visible lifecycle, such as `closed`, `opening`, `drawer-open`, `closing-collapse`, `sheet-exiting`, `layout-morph-out`, and `sealed`.
+   - Map which component owns mount/unmount, visibility, identity, focus, layout measurement, and callbacks for phase completion.
+   - Define stable keys and identity rules. State which key must persist through exit and which key may change only after the exit or morph is complete.
+   - Define callback handoff points: animation complete, exit complete, layout measurement ready, focus restored, overlay dismissed, or state cleared.
+   - If ownership is split across siblings without a named coordinator, stop and redesign the ownership map before proposing timing changes.
+
+5. **Choose implementation guardrails**
+
+   **Framework detection and routing:** detect which animation framework the codebase uses, then route to the appropriate framework subskill for framework-specific guardrails. Detection order:
+   - Grep for `framer-motion` or `motion` in package.json / imports → route to `/animation-design-planner frameworks/motion-framer`
+   - Grep for `gsap` in package.json / imports → route to `frameworks/gsap` subskill
+   - Grep for `@angular/animations` → route to the appropriate Angular animation subskill when available
+   - Grep for `three` or `@react-three/fiber` → route to `frameworks/threejs` subskill
+   - Grep for `element.animate(` or Web Animations API usage → route to `frameworks/web-animations-api` subskill
+   - Grep for CSS `@keyframes`, `transition:`, `animation:` in stylesheets → route to `frameworks/css-transitions` subskill
+   - If multiple frameworks detected or ambiguous, ask the user which framework to target
+   - If no subskill exists for the detected framework, use the baseline guardrails below
+
+   **Baseline guardrails (all frameworks inherit these):**
+   - Prefer transform and opacity animation. Avoid animating layout, paint, filter, shadow, or expensive size properties unless the plan names the cost and proof.
+   - Define reduced-motion behavior for interaction-triggered motion. Nonessential motion must be suppressed or converted to instant/fade behavior when the user prefers reduced motion.
+   - Specify focus management, pointer-event behavior, scroll locking, z-index/layering, and content clipping rules when they affect the motion.
+
+6. **Define the proof gate**
+   - Write acceptance checks tied directly to the visible motion contract.
+   - Include slow-motion or debug-duration review steps for sequencing.
+   - Include Playwright screenshot or video capture when feasible, especially for close/open sequences, shared-element motion, drawer/modal transitions, and regression-prone visual states.
+   - Include reduced-motion verification.
+   - Include manual review steps for what automated screenshots cannot prove, such as perceived continuity, focal attention, and interruption feel.
+   - State the exact proof artifacts expected before the implementation is considered done.
+
+7. **Hand off implementation only after the plan**
+   - If the user asked only for planning, stop with the plan and proof gate.
+   - If the user asked for implementation too, make the implementation step conditional on the plan and proof gate. Then route to `/exec` or the appropriate implementation skill with the plan as the contract.
+   - For bug-fix requests, route to `/investigate --ui` only after this skill has named the visible motion contract and likely lifecycle owner, unless the bug is a crash or data issue.
+
+## Output
+
+Produce a structured animation plan with these sections:
+
+- **Visible Motion Contract:** trigger, start state, end state, expected user-visible sequence, focal point, continuity target, forbidden artifacts, and accepted interruption behavior.
+- **Storyboard / Timeline:** phases or key frames, timing, easing, focal element, supporting elements, shared elements, stagger, anticipation, follow-through, and settling.
+- **Lifecycle Ownership Map:** single sequence owner, state phases, component mount/unmount responsibilities, stable key rules, identity persistence, callback handoffs, focus/scroll responsibilities, and interruption handling.
+- **Implementation Guardrails:** framework-specific guardrails (from the routed subskill), reduced-motion behavior, transform/opacity-first performance constraints, layout/paint risks, layering, and pointer/focus rules.
+- **Proof Gate:** slow-motion/manual review steps, Playwright screenshot or video capture plan where feasible, reduced-motion verification, regression checks, and acceptance checklist tied to the visible motion contract.
+- **Implementation Handoff:** whether implementation is ready, blocked, or should route to `/exec`, `/investigate --ui`, or another skill.
+
+End with:
+
+```md
+**Next work:** <specific implementation, investigation, or review step>
+**Recommended next command:** <one command or "none">
+```
+
+## Constraints
+
+- Do not start by tweaking durations, easing constants, CSS classes, or Motion props before the visible motion contract, lifecycle owner, state phases, and proof method are named.
+- Do not propose a multi-component animation without one lifecycle owner or explicit state-machine coordinator.
+- Do not clear identity, active item state, or stable keys before exit, shared layout, or close-morph phases have completed.
+- Do not rely on "looks better" as proof. Tie proof to visible contract acceptance checks and concrete artifacts.
+- Do not ignore reduced-motion behavior for interaction-triggered motion.
+- Do not use layout/paint-heavy animation properties as the default path. Prefer transform and opacity unless the plan justifies and proves otherwise.
+- Do not create or modify GitHub Actions workflows.
+
+## Alignment Page
+
+When this skill produces durable deliverables (research, specs, plans, reports, prototypes, or any document output), build a full-depth HTML alignment page following `ALIGNMENT-PAGE.md` in this skill's directory. Output: `alignment/animation-design-planner-{topic}.html`.
+
+## Default Shipping Contract
+
+- This skill is planning-first. Normally it should not mutate implementation files directly.
+- If the user explicitly asks this skill to create or modify tracked files, finish by committing and pushing all intended changes to the repository primary branch (`main` when present, otherwise `master`) before stopping.
+- This contract does not override stricter safety rules about secrets, destructive history changes, release publication/tag confirmation, production deploy confirmation, paid actions, or public visibility changes.

package/global/claude/animation-design-planner/frameworks/css-transitions/SKILL.md

@@ -0,0 +1,48 @@
+---
+name: css-transitions
+description: CSS transitions and keyframe animation guardrails — transition/animation properties, will-change, class-based sequencing
+type: planning
+version: v0.0
+invocation: sub-skill
+parent: animation-design-planner
+---
+
+# CSS Transitions Implementation Guardrails
+
+This subskill provides CSS transition and keyframe animation guardrails for step 5 of `/animation-design-planner`. It is invoked automatically when the parent detects CSS-based animation patterns in the project. All baseline guardrails from the parent (transform/opacity preference, reduced-motion, focus/pointer/scroll/z-index rules) still apply.
+
+## Guardrails
+
+### Sequencing and orchestration
+
+- Use `transition` for two-state toggles; use `@keyframes animation` for multi-waypoint or looping motion.
+- Chain sequences via `transitionend`/`animationend` — always filter by `propertyName` or `animationName` because shorthand declarations fire one event per sub-property.
+- Stagger children via `animation-delay` combined with a CSS custom property per child (e.g. `--i`), not JS `setTimeout`.
+- Class toggle order: apply in the same frame. If phase B depends on phase A completing, chain via the end event — do not assume matching durations.
+
+### `will-change` lifecycle
+
+- Apply `will-change` just before the animation starts and remove it after the animation ends.
+- Permanent `will-change` forces a compositor layer and creates a new stacking context — document explicitly if intentionally kept.
+- Never use `will-change: contents` or `will-change: scroll-position` speculatively.
+
+### Forced reflow and paint triggers
+
+- The reflow trap: reading a layout property then writing a style in the same synchronous block forces synchronous layout. Batch reads before writes.
+- Void-to-value trick: force a reflow between DOM insertion and class addition with `el.offsetHeight` to ensure the browser registers the initial state before animating.
+- Double-rAF pattern (`requestAnimationFrame(() => requestAnimationFrame(() => { ... }))`) for maximum reliability on enter transitions when the single-frame trick is insufficient.
+
+### Exit animation patterns
+
+- CSS has no animate-then-remove primitive: add an exit class, listen for the end event, then remove the element from the DOM.
+- Race guard — the end event never fires if the property isn't actually transitioning. Always set a `setTimeout` fallback matching the expected duration.
+- Race guard — shorthand fires per sub-property. Filter by `propertyName` or designate a single canonical property for the event listener.
+- Use `animation-fill-mode: forwards` to hold the final state while DOM removal runs. Remove the element promptly — do not rely on `forwards` indefinitely as it holds the compositor layer.
+
+### Proof expectations
+
+- `transitionend` fires exactly once for the intended property on each transition.
+- Rapid toggle (open/close, show/hide) produces no stuck mid-animation states.
+- `will-change` is removed after completion (verify in DevTools Layers panel).
+- `prefers-reduced-motion: reduce` suppresses or collapses all motion.
+- Slow-motion review (set durations to 2–3s) confirms sequencing matches the animation contract.

package/global/claude/animation-design-planner/frameworks/gsap/SKILL.md

@@ -0,0 +1,51 @@
+---
+name: gsap
+description: GSAP-specific animation guardrails — timeline composition, overwrite modes, ScrollTrigger pinning/scrub, cleanup on unmount
+type: planning
+version: v0.0
+invocation: sub-skill
+parent: animation-design-planner
+---
+
+# GSAP Implementation Guardrails
+
+This subskill provides GSAP-specific implementation guardrails for step 5 of `/animation-design-planner`. It is invoked automatically when the parent detects GSAP in the project. All baseline guardrails from the parent (transform/opacity preference, reduced-motion, focus/pointer/scroll/z-index rules) still apply.
+
+## Guardrails
+
+### Timeline composition
+
+- Use standalone tweens (`gsap.to`/`gsap.from`) for single-element fire-and-forget. Switch to `gsap.timeline()` as soon as two or more tweens need to coordinate.
+- Position parameter: prefer labels or relative offsets (`"+=0.2"`, `"<0.1"`) over bare absolute seconds — they survive duration changes.
+- Nest timelines only for reusable sub-sequences. Flat timeline with labels is easier to debug and inspect.
+- Never create timelines inside rapid-fire event handlers. Create the timeline once, then control it with `.play()` / `.reverse()` / `.progress()`.
+
+### Overwrite modes
+
+- GSAP 3 defaults to `overwrite: false` — competing tweens on the same target stack and cause jitter.
+- `"auto"`: kills only the conflicting properties on other tweens. Use as the standard default for shared targets.
+- `true`: kills all active tweens on the target. Use for full takeover scenarios (e.g. reset animation on unmount).
+- `false`: only for intentional additive behavior that has been visually verified.
+
+### ScrollTrigger patterns
+
+- `scrub: true` for 1:1 scroll-linked progress; use a number (e.g. `scrub: 0.5`) for smoothed scrub. State the chosen value and the reason.
+- `pin`: breaks inside flex/grid layouts. Pin the outermost section container, not flex children. Use `pinSpacing: false` when the generated spacer div is unacceptable.
+- `anticipatePin: 1` when there is a visible snap or jump at the pin start boundary.
+- Wrap breakpoint-specific ScrollTriggers in `ScrollTrigger.matchMedia()` so they destroy and recreate cleanly on resize.
+- Enable `markers: true` during development. Remove before shipping.
+
+### Cleanup on unmount
+
+- React: wrap all tweens in `gsap.context(() => { ... }, containerRef)` inside `useLayoutEffect` and return `ctx.revert()` from the cleanup function.
+- Non-React: call `.kill()` on tweens and `ScrollTrigger.kill()` on each instance in the teardown hook.
+- Never rely on timeline completion as implicit cleanup — the user may navigate away mid-animation.
+- `matchMedia` instances still need explicit kill on unmount.
+
+### Proof expectations
+
+- Timeline completes and `onComplete` callback fires.
+- `markers: true` confirms ScrollTrigger start/end alignment matches the intended scroll positions.
+- After unmount: `gsap.globalTimeline.getChildren()` contains no orphaned references from the destroyed component.
+- Resize across breakpoints — ScrollTriggers destroy and recreate without duplicates or stale pins.
+- Slow-motion review via `timeline.timeScale(0.2)` to confirm sequencing matches the animation contract.

package/global/claude/animation-design-planner/frameworks/motion-framer/SKILL.md

@@ -0,0 +1,47 @@
+---
+name: motion-framer
+description: Motion/Framer-specific implementation guardrails — AnimatePresence placement, sequencing modes, LayoutGroup, layout animation boundaries
+type: planning
+version: v0.0
+invocation: sub-skill
+parent: animation-design-planner
+---
+
+# Motion/Framer Implementation Guardrails
+
+This subskill provides Motion/Framer-specific implementation guardrails for step 5 of `/animation-design-planner`. It is invoked automatically when the parent detects `framer-motion` or `motion` in the project. All baseline guardrails from the parent (transform/opacity preference, reduced-motion, focus/pointer/scroll/z-index rules) still apply.
+
+## Guardrails
+
+### Sequencing mode selection
+
+Choose the `AnimatePresence` `mode` intentionally:
+
+- **`sync`** (default) — entering and exiting children animate simultaneously. Use when the exit and enter are independent or when you need the entering child to begin immediately.
+- **`wait`** — the exiting child completes its exit before the entering child mounts. Use for sequential transitions where the user should see one thing leave before the next arrives (route transitions, modal swap, step wizards).
+- **`popLayout`** — the exiting child is removed from layout flow immediately while still animating out. Use when you need the entering child to take its layout position without waiting for exit, but the exiting child should still visually animate away (overlapping fade, slide-out while new content slides in).
+
+State the chosen mode and the reason in the animation plan.
+
+### AnimatePresence placement
+
+Place `AnimatePresence` at the conditional render boundary that owns the exit lifecycle. The component that decides whether the child is mounted must be the one wrapping it in `AnimatePresence`. Do not bury `AnimatePresence` inside the child or in a wrapper where the exiting child is already unmounted by the time the exit animation would run.
+
+### LayoutGroup and layout animation
+
+- Use `LayoutGroup` only when multiple sibling components share layout identity and the measurement boundary is clear.
+- Set `layout` or `layoutId` only on elements whose layout position genuinely changes. Avoid blanket `layout` on containers that don't move.
+- When using `layoutId` for shared-element transitions, ensure the identity string is stable across the mount/unmount boundary and that only one element with that `layoutId` is mounted at a time (or both are within the same `LayoutGroup`).
+- Define the measurement boundary explicitly. If a scrollable container or transform parent exists between the layout-animated element and the viewport, the measurement may be wrong — test and document.
+
+### Interruption patterns
+
+- **Exit cancellation:** if the user triggers a re-entry while an exit animation is in progress, define whether the exit should reverse, snap to start, or complete before re-entering. Motion's default is to cancel exit and reverse — state whether that default is acceptable.
+- **Re-entry during exit:** when using `mode="wait"`, a re-entry trigger during exit queues the enter. Document whether this queue behavior is acceptable or whether the plan needs `mode="sync"` or `mode="popLayout"` to avoid perceived lag.
+- **Rapid toggle:** for open/close patterns (drawers, accordions, dropdowns), verify that rapid toggling does not produce ghost layers, stuck exit states, or layout jumps.
+
+### Proof expectations
+
+- Verify `onExitComplete` fires and any cleanup (focus restore, scroll unlock, state clear) runs after exit, not during.
+- Slow-motion review (set `transition.duration` to 2-3s) to confirm sequencing mode matches the visible contract.
+- Verify `layoutId` transitions measure correctly across the actual DOM hierarchy (not just in isolation).

package/global/claude/animation-design-planner/frameworks/threejs/SKILL.md

@@ -0,0 +1,54 @@
+---
+name: threejs
+description: Three.js/R3F animation guardrails — useFrame loop, tween libraries, shader-based animation, render-loop timing
+type: planning
+version: v0.0
+invocation: sub-skill
+parent: animation-design-planner
+---
+
+# Three.js / R3F Implementation Guardrails
+
+This subskill provides Three.js and React Three Fiber animation guardrails for step 5 of `/animation-design-planner`. It is invoked automatically when the parent detects `three`, `@react-three/fiber`, or R3F in the project. All baseline guardrails from the parent (transform/opacity preference, reduced-motion, focus/pointer/scroll/z-index rules) still apply.
+
+## Guardrails
+
+### Render loop and useFrame
+
+- All animation must run inside the render loop. Never use `setInterval`, `setTimeout`, or manual `requestAnimationFrame` with R3F.
+- `useFrame((state, delta) => ...)` — multiply all increments by `delta` for frame-rate independence.
+- Use `state.clock.elapsedTime` for oscillations (sin/cos), `delta` for velocity-based movement.
+- The `priority` parameter orders multiple `useFrame` hooks — document any non-zero values and the reason.
+- Plain Three.js: use `renderer.setAnimationLoop(callback)` with `Clock.getDelta()`.
+- Never mutate React state inside `useFrame` — mutate refs directly to avoid re-renders every frame.
+
+### Tween library integration
+
+- Simple A→B interpolation: `maath/easing` (`damp`, `damp3`, `dampE`) inside `useFrame` — always pass `delta`.
+- Complex multi-step sequences: GSAP `.to()` targeting ref properties is acceptable for visual-only props, not for state-dependent logic.
+- Avoid tween.js with R3F — it requires a manual `TWEEN.update()` call inside the render loop.
+- Physics-driven animation: use `@react-three/rapier` or `cannon-es` instead of easing curves for physically plausible motion.
+
+### Disposal and cleanup
+
+- Geometries, materials, and textures created outside the scene graph need manual `.dispose()` in `useEffect` cleanup.
+- Particle systems with dynamic geometry: track instances in a ref array and dispose on despawn or unmount.
+- Swapping materials: dispose the old material before replacing — property assignment does not auto-dispose.
+- `useLoader` / `useGLTF` cache by URL (safe for reuse), but `.clone()`'d materials are your responsibility to dispose.
+- InstancedMesh: geometry and material are shared — dispose once on full unmount, not per instance.
+
+### Frame budget
+
+- 16.6ms at 60fps for all JS + animation + GPU work. Every millisecond counts.
+- Profile with R3F `<Perf>` or Stats.js. Plan fallback strategies: reduce particle count, lower LOD, disable shadows.
+- Use `state.performance.current` (from `@react-three/drei`'s `AdaptiveDpr`/`AdaptiveEvents`) for adaptive quality.
+- Frame-skip pattern for periodic updates (UI labels, health bars): run expensive logic every N frames using a counter, not every frame.
+- Never allocate objects inside `useFrame` — pre-allocate `Vector3`, `Quaternion`, and `Matrix4` in refs outside the loop.
+
+### Proof expectations
+
+- Sustained target framerate verified via `<Perf>` overlay or Stats.js.
+- Delta-time independence: animation speed is consistent at 30fps (throttled DevTools) and 60/144fps.
+- After unmount: `renderer.info.memory.geometries` and `.textures` return to pre-mount counts.
+- Disposed objects produce no visible artifact (no white/pink squares, no console warnings).
+- Interrupted tweens on unmount do not throw or attempt to update disposed objects.

package/global/claude/animation-design-planner/frameworks/web-animations-api/SKILL.md

@@ -0,0 +1,53 @@
+---
+name: web-animations-api
+description: Web Animations API guardrails — Element.animate(), Animation object lifecycle, commitStyles, polyfill considerations
+type: planning
+version: v0.0
+invocation: sub-skill
+parent: animation-design-planner
+---
+
+# Web Animations API Implementation Guardrails
+
+This subskill provides Web Animations API guardrails for step 5 of `/animation-design-planner`. It is invoked automatically when the parent detects `Element.animate()` or Web Animations API usage in the project. All baseline guardrails from the parent (transform/opacity preference, reduced-motion, focus/pointer/scroll/z-index rules) still apply.
+
+## Guardrails
+
+### Animation lifecycle
+
+- `Element.animate()` returns an `Animation` object — store the reference if you need play/pause/cancel/reverse control.
+- `.cancel()` reverts styles to pre-animation values. `.finish()` jumps to the end state and resolves the `finished` promise.
+- The `finished` promise rejects on cancel — always handle rejection if chaining with `await` or `.then()`.
+- For interruption: decide between cancel+restart, reverse (`playbackRate = -1`), or finish+new animation. Document the chosen strategy.
+
+### Fill mode vs commitStyles
+
+- `fill: "forwards"` holds the compositor layer and blocks garbage collection of the `Animation` object — avoid for one-shot transitions.
+- Prefer `commitStyles()` followed by `.cancel()` to bake the final state into inline `style` and release the animation resources.
+- Gotcha: `commitStyles()` writes inline style, which has the highest specificity — it will override stylesheet rules until explicitly cleared.
+- Use `fill: "forwards"` without commit only for reversible or seekable animations. State when and how the animation will be cancelled.
+- Never use `fill: "both"` as a default — it combines the pitfalls of both fill directions.
+
+### Composite modes
+
+- `composite: "replace"` (default): one animation owns the property value.
+- `composite: "add"`: additive — the effect is added to the base value (e.g. hover scale on top of an entrance animation).
+- `composite: "accumulate"`: combines values before transform application — prefer for stacking translates or rotates.
+- `iterationComposite: "accumulate"` only when each iteration should build on the previous one's end value.
+- Additive composition requires the element to have a defined base value for the property.
+
+### Sequencing and grouping
+
+- No built-in timeline — chain animations with `await animation.finished`.
+- Parallel: fire multiple `.animate()` calls, then `Promise.all(animations.map(a => a.finished))`.
+- Synchronized start: set `animation.startTime = document.timeline.currentTime` on all animations for frame-exact alignment.
+- GroupEffect and SequenceEffect proposals are not shipped in any browser — do not plan around them.
+- Stagger: loop with incremental `delay` values. Store all `Animation` references for collective pause/cancel/finish control.
+
+### Proof expectations
+
+- `finished` promise resolves for every animation in a sequence.
+- After `commitStyles()`: inline `style` contains the expected values with no leaked extra properties.
+- `element.getAnimations()` returns an empty array after completion (unless the animation is intentionally kept alive).
+- `.cancel()` reverts styles cleanly — no visual flash or stuck state.
+- Interruption mid-animation produces no visual artifacts or unhandled promise rejections.

package/global/claude/autoresearch/CHANGELOG.md

@@ -0,0 +1,8 @@
+# Changelog
+
+## v0.0
+
+- Initial skill contract: autonomous experiment loop with ratchet mechanism.
+- program.md format for metric, sandbox, directions, and budget.
+- Git branch isolation per experiment, ff-merge for winners.
+- TSV result log and summary report.

package/global/claude/autoresearch/SKILL.md

@@ -0,0 +1,185 @@
+---
+name: autoresearch
+description: Autonomous experiment loop — iteratively mutate code, measure a metric, keep only improvements (hill-climbing ratchet)
+type: orchestrator
+version: v0.0
+invocation: orchestrator
+argument-hint: "[path/to/program.md]"
+---
+
+# Autoresearch
+
+Autonomous experiment loop inspired by Karpathy's autoresearch pattern. Reads a `program.md` research program, proposes one hypothesis per iteration, implements it within a declared sandbox, measures a single metric, and keeps only improvements. Git branches isolate experiments; a TSV log tracks all results.
+
+## When to Use
+
+When the user wants to autonomously optimize a measurable codebase property: benchmark throughput, bundle size, test coverage, load time, memory usage, build speed, or any metric expressible as a shell command that prints a number.
+
+## Preconditions
+
+- A `program.md` file exists (default: `./program.md`, or path from `$ARGUMENTS`). If missing, recommend `/autoresearch-prep` to scaffold one.
+- Git working tree is clean (no uncommitted changes)
+- The metric command runs successfully and prints a parseable number
+
+## program.md Format
+
+The user creates this file. Required sections:
+
+```markdown
+# Research Program
+
+## Metric
+command: <shell command that prints a single number to stdout>
+direction: higher-is-better | lower-is-better
+
+## Budget
+max_iterations: <number>
+
+## Sandbox
+files:
+  - <glob patterns of files the agent may modify>
+exclude:
+  - <glob patterns to never touch>
+
+## Research Directions
+1. <direction one>
+2. <direction two>
+```
+
+Optional sections:
+
+```markdown
+## Test Command
+command: <shell command that must exit 0 before measuring>
+required: true
+
+## Budget
+metric_timeout_seconds: <default 300>
+
+## Context
+<free-form notes about the codebase, prior work, constraints>
+```
+
+## Process
+
+### 0. Validate Preconditions
+
+1. Resolve `program.md` path from `$ARGUMENTS` (default: `./program.md`). If the file does not exist, stop and recommend `/autoresearch-prep`.
+2. Parse and validate required fields: `Metric.command`, `Metric.direction`, `Budget.max_iterations`, `Sandbox.files`.
+3. Verify git working tree is clean (`git status --porcelain` is empty). Stop if dirty.
+4. Create `.autoresearch/` directory if it does not exist.
+5. Record the current branch as the **base branch**.
+6. Add `.autoresearch/` to `.gitignore` if not already present.
+
+### 1. Establish Baseline
+
+1. Run the metric command from the project root.
+2. Parse the last number from stdout (strip non-numeric lines).
+3. Record as iteration 0 in `.autoresearch/results.tsv`:
+   ```
+   iteration	timestamp	hypothesis	metric_value	delta	delta_pct	status	branch	commit	notes
+   0	<ISO-8601>	baseline	<value>	0	0.00%	baseline	<base-branch>	<HEAD-sha>	initial measurement
+   ```
+4. Set `best_value = baseline_value`.
+
+### 2. Check Stop Conditions
+
+Before each iteration, check in order:
+
+1. **Stop file**: if `.autoresearch/stop` exists, log "stop file detected" and go to step 9.
+2. **Iteration limit**: if current iteration > `max_iterations`, go to step 9.
+3. **Re-read program.md**: pick up any user edits to directions, budget, or context mid-run.
+
+### 3. Propose Hypothesis
+
+1. Read: research directions, past results from `.autoresearch/results.tsv`, current sandbox file contents, and context section.
+2. Propose exactly **one** specific, testable hypothesis. Prefer untried directions. Avoid repeating failed approaches.
+3. The hypothesis must name: what to change, which file(s), and the expected effect on the metric.
+4. Log the hypothesis before implementing.
+
+### 4. Create Experiment Branch
+
+1. Ensure base branch is checked out and clean.
+2. Create and checkout: `autoresearch/iter-{N}-{slug}` where `{slug}` is a 2-4 word kebab-case summary of the hypothesis.
+
+### 5. Implement the Change
+
+1. Modify **only** files matching the sandbox glob patterns.
+2. Never modify files matching exclude patterns.
+3. **Never modify**: the metric command/script, benchmark suite, test harness, evaluation code, or anything outside the sandbox.
+4. Keep changes minimal and focused on the single hypothesis.
+5. Commit on the experiment branch with message: `autoresearch: iter {N} — {hypothesis summary}`.
+
+### 6. Validate Build/Tests
+
+1. If a test command is configured and `required: true`:
+   - Run the test command.
+   - If it fails: log `status: test-failed`, checkout base branch, delete experiment branch, record in results.tsv, continue to next iteration.
+2. If no test command is configured, skip this step.
+
+### 7. Measure
+
+1. Run the metric command with the configured timeout (default 300s).
+2. Parse the last number from stdout.
+3. If the command fails or times out: log `status: measure-failed`, checkout base branch, delete experiment branch, record in results.tsv, continue to next iteration.
+
+### 8. Evaluate (Ratchet)
+
+Compare measured value to `best_value` using the configured direction:
+
+**If improved:**
+1. Checkout base branch.
+2. Fast-forward merge the experiment branch: `git merge --ff-only autoresearch/iter-{N}-{slug}`.
+3. Update `best_value`.
+4. Record `status: kept` with the delta and delta percentage in results.tsv.
+5. Delete the experiment branch (it's merged).
+
+**If not improved (or equal):**
+1. Checkout base branch.
+2. Record `status: reverted` with the delta in results.tsv.
+3. Delete the experiment branch.
+
+Always ensure the working tree is clean before proceeding to the next iteration.
+
+### 9. Report Final Status
+
+When the loop ends (stop file, iteration limit, or all directions exhausted):
+
+1. Write `.autoresearch/summary.md`:
+   - Baseline value → final best value and total improvement (absolute + percentage).
+   - Per-iteration table (from results.tsv).
+   - Directions attempted vs. untried.
+   - Top 3 most impactful kept changes.
+2. Print a summary table to the terminal.
+3. Do **not** push to remote — all work stays local for user review.
+
+## Output
+
+- `.autoresearch/results.tsv` — append-only experiment log
+- `.autoresearch/summary.md` — written at loop end
+
+## Constraints
+
+- **Sandbox is absolute**: never modify files outside declared sandbox globs.
+- **Never touch eval**: never modify the metric command, benchmark suite, or test harness.
+- **One hypothesis per iteration**: no bundling multiple changes.
+- **Clean state between iterations**: never leave dirty working tree.
+- **No remote push**: all work stays local.
+- **No GitHub Actions**: do not create CI workflows.
+- **Autonomous**: no user approval per iteration. User steers by editing `program.md` or creating `.autoresearch/stop`.
+
+## Failure Recovery
+
+If the agent finds itself on an experiment branch with uncommitted changes between iterations:
+1. `git checkout -- .` to discard changes.
+2. `git checkout <base-branch>`.
+3. Delete the orphaned experiment branch.
+4. Record `status: crashed-recovered` in results.tsv.
+5. Continue from the next iteration.
+
+## Shipping
+
+This skill does **not** follow the shared shipping contract. All work stays local and unpushed. The user reviews results and decides what to keep and push.
+
+**Next work:** none — loop is self-contained
+**Recommended next command:** `git log --oneline` to review kept changes, then push when satisfied

package/global/claude/autoresearch-prep/CHANGELOG.md

@@ -0,0 +1,5 @@
+# autoresearch-prep Changelog
+
+## v0.0
+
+- Initial skill: auto-detect codebase signals, light interview, scaffold `program.md` for `/autoresearch`