devloom 1.0.0 → 1.2.0

package/skills/build/simplify.md

@@ -0,0 +1,50 @@
+---
+name: simplify
+description: Behavior-preserving simplification — reduce complexity for readability and maintainability without changing what the code does.
+---
+
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+## When to use
+
+After behavior is understood and tests exist. During review when code is harder to read than it needs to be. Never before correctness is proven.
+
+## Non-negotiable invariant
+
+**Behavior does not change.** Same inputs → same outputs, same side effects, same error behavior. If a simplification changes any observable behavior, it is a refactor with risk — flag it, do not apply silently.
+
+## Protocol
+
+1. **UNDERSTAND**: read the code + its tests. If no tests cover the code, write characterization tests FIRST (lock current behavior)
+2. **IDENTIFY** complexity (in priority order):
+   - duplication that can be extracted once
+   - nesting depth > 3 (early returns, guard clauses)
+   - names that hide intent (rename only, no logic change)
+   - dead code (provably unreachable — delete)
+   - over-abstraction (wrappers that add indirection without value — inline)
+   - conditionals expressible as data/lookup tables
+3. **APPLY** one simplification at a time, smallest diff possible
+4. **VERIFY** after EACH change: run the tests that cover the touched code — all must pass before the next simplification
+5. **STOP** when further changes would trade clarity for cleverness
+
+## What NOT to do
+
+- No behavior changes disguised as simplification
+- No new abstractions "for future flexibility" (YAGNI)
+- No reformatting-only diffs mixed with logic simplification (separate commits)
+- No renaming public APIs without flagging the breaking change
+- No simplifying generated code, migrations, or vendored files
+
+## Output
+
+```
+SIMPLIFY_REPORT:
+changes:
+  - <file:line> <what was simplified> | risk: none|low
+  - <file:line> <what> | risk: none|low
+skipped:
+  - <what> | reason: <why not safe to simplify>
+tests: <ran> → <result>
+```
+
+RULES: tests before touching | one change at a time | verify after each | behavior-preserving or flagged | smallest diff.

package/skills/build/vision-analysis.md

@@ -0,0 +1,98 @@
+---
+name: vision-analysis
+description: Analyze images and return structured, detailed descriptions tailored by image type for downstream agent consumption.
+---
+
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+DETECT_TYPE:
+- ui_screenshot: browser/app UI, windows, buttons, forms, navigation
+- wireframe: low-fidelity mockup, boxes, labels, layout sketch
+- diagram: flowchart, ERD, sequence, architecture, network, UML
+- chart: bar/line/pie/scatter data visualization
+- code_screenshot: code editor, terminal output, stack trace, logs
+- design_mockup: high-fidelity design, colors, typography, components
+- photo: real-world photograph, product, person, environment
+- document: text document, form, invoice, specification
+- unknown: none of the above
+
+ANALYZE by type:
+
+ui_screenshot:
+- app/page name; layout structure (nav/sidebar/main/footer)
+- all visible UI elements: labels, inputs, buttons, links, icons, placeholders
+- state: loading/error/empty/success indicators; disabled states; active selections
+- content: text values, data shown, counts, statuses
+- UX issues visible: broken layout, overflow, missing states, a11y problems
+
+wireframe:
+- page/screen intent; sections and their purpose
+- element inventory: inputs, labels, CTAs, content blocks
+- user flow implied; open questions or ambiguities
+
+diagram:
+- diagram type and subject
+- nodes/entities: name, role, attributes if shown
+- relationships: direction, labels, cardinality
+- flow or sequence: start>steps>end in order
+- constraints or annotations noted
+
+chart:
+- chart type; data topic; axes labels + units
+- data range, trends, peaks, anomalies
+- legend items; what each series represents
+- key insight in one sentence
+
+code_screenshot:
+- language/runtime if identifiable
+- file name / path if visible
+- code structure: functions, classes, logic flow
+- errors or warnings: exact message, line, type
+- relevant context: imports, configs visible
+
+design_mockup:
+- component inventory with names and variants
+- color palette: primary/secondary/accent hex if readable
+- typography: font names, sizes, weights if visible
+- spacing and grid patterns
+- interactive states shown
+- brand/style notes
+
+photo:
+- subject and context
+- key objects, people, text visible
+- environment, lighting, composition notes
+- any technical or domain details visible
+
+document:
+- document type and purpose
+- sections and their content
+- key data fields and values
+- requirements, rules, or constraints stated
+
+OUTPUT_FORMAT:
+```
+IMAGE_TYPE: <type>
+SUMMARY: <one sentence — what this image shows and its purpose>
+
+DETAILS:
+<type-specific structured analysis — complete, no omissions>
+
+FOR_AGENTS:
+<actionable insights downstream agents need:
+- developer: what to implement/fix based on this image
+- planner: requirements or constraints visible
+- qa: acceptance criteria derivable from the image
+- security: any visible data exposure, auth flows, sensitive inputs
+- documenter: what to document about this screen/flow>
+
+OPEN_QUESTIONS:
+<anything ambiguous or missing that the requesting agent should clarify>
+```
+
+RULES:
+- describe what IS visible, not what you assume
+- quote exact text seen in the image
+- if image is unclear or low-res, state confidence level
+- no hallucination of unvisible elements
+- output must be self-contained — other agents have no access to the image

package/skills/design/app-design.md

@@ -0,0 +1,31 @@
+---
+name: app-design
+description: Design for dashboards, enterprise software, product UI, forms, settings, CRUD, mobile/desktop-style apps — hierarchy and efficiency over art direction.
+---
+
+LOAD: ~/.config/opencode/devloom-ai/core.dsl|~/.config/opencode/devloom-ai/workflow.dsl
+DOMAIN: PRODUCT_APP (dashboard, enterprise, forms, settings, CRUD, mobile-style, desktop-style)
+
+PRIORITIZE:
+- hierarchy: primary action / primary data always wins the eye; progressive disclosure for secondary detail
+- efficiency: minimal clicks/keystrokes to outcome; keyboard paths for repeated actions
+- scanability: aligned tabular data, consistent label placement, status readable at a glance (never color-only)
+- density: choose an explicit density mode (comfortable/compact) per surface and keep it consistent
+- predictable interaction: same control = same behavior everywhere; conventions over novelty
+- stable visual language: tokens for color/spacing/type; components reused, not reinvented per page
+- accessibility: WCAG-AA, visible focus, form errors associated + recoverable, touch targets >= 44px on mobile-style UI
+
+AVOID:
+- over-art-directing productivity software (no hero typography inside admin tools)
+- turning every piece of content into a card — use lists, tables, dividers, plain rows where they scan better
+- pill soup (every label a badge), decoration that competes with data
+- hidden critical actions, silent failures, spinner-without-context
+- modal stacking; wizards for simple tasks
+
+STATES (every data surface): loading | empty (with next action) | error (with recovery) | success (with confirmation) | disabled (with reason).
+
+RESPONSIVE: tables reflow (priority columns + row detail), nav collapses deliberately, forms single-column on small screens. Viewports: 1440x1000, 1024x900, 390x844 or project targets.
+
+CHECK: primary action obvious | density mode chosen | all five states designed | no card-soup violation | a11y pass
+
+OUT: APP_DESIGN_SPEC (hierarchy | density | interaction rules | state matrix | DON'Ts) — implementation stays with devloom-developer

package/skills/design/design-grounding.md

@@ -0,0 +1,30 @@
+---
+name: design-grounding
+description: Convert vague aesthetic language into explicit executable design constraints; optional OpenDesign grounding, existing design systems first.
+---
+
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+PURPOSE: eliminate vague aesthetic prompting. Every vague term must become constraints a developer can implement and a critic can verify.
+
+TRANSLATE (vague -> explicit):
+- premium -> type scale with clear ratios, restrained palette (<=3 hues + neutrals), generous whitespace rhythm, refined surface treatment, zero default templates
+- modern -> current typography conventions, flat-or-restrained depth, high contrast hierarchy, purposeful micro-interaction; no skeuomorphic leftovers
+- beautiful (unqualified) -> ask nothing: apply hierarchy + composition + spacing discipline; beauty is an outcome, not an input
+- futuristic -> forward type/layout choices with legibility intact; motion explains state; NO gratuitous glow/scanlines/hexagons
+- Awwwards-like -> strong art direction + compositional bravery + fast load; NEVER clone a specific product
+- clean -> fewer competing elements, consistent spacing scale, limited palette, strong alignment grid
+- playful -> rounded shape language, warmer palette, expressive motion budget — still WCAG-AA
+
+GROUNDING ORDER:
+1. existing project design system / design.md (always first)
+2. OpenDesign MCP if available: typography relationships, spacing rhythm, layout density, surface treatment, palette relationships, motion, interaction, explicit DON'T rules
+3. internal reasoning + domain skill (web-design / app-design / game-design)
+
+RULES:
+- OpenDesign is OPTIONAL — its absence never fails or delays normal work; do not install MCPs to satisfy this skill
+- never clone another product's identity; extract relationships (ratios, rhythms, contrast), not surfaces
+- every direction ends with explicit DON'T rules (3-5), not just DO rules
+- constraints must be checkable by visual-critique (e.g. "heading/body ratio >= 2" not "big headings")
+
+OUT: GROUNDED_CONSTRAINTS (type | spacing | palette | density | shape | motion | interaction | DON'Ts)

package/skills/design/design-system.md

@@ -0,0 +1,27 @@
+---
+name: design-system
+description: Define or extend a project design system — tokens, components, states, iconography — as the single visual source of truth.
+---
+
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+WHEN: EXISTING_DESIGN_SYSTEM domain, or repeated components appearing across tickets.
+
+DO:
+- inventory what exists first (tokens, themes, component library, icon set) — extend, never fork
+- tokens (as project conventions allow): color ROLES (primary/secondary/accent/surface/text/border/danger/success/warning), not raw hues; type scale; spacing scale; radii; shadows/elevation; breakpoints; motion durations
+- components: name, purpose, variants, and the full state matrix (default/hover/focus/active/disabled/loading/error/empty where applicable)
+- iconography: one family/style, size grid, stroke weight consistency; no mixed icon sources
+- accessibility baked in: contrast pairs per role, focus ring token, touch target sizes
+- document DO/DON'T per token group so misuse is detectable in visual-critique
+
+RULES:
+- design.md (project design memory) is the narrative; the system is the executable form — keep them consistent
+- no one-off values that bypass tokens ("magic numbers" are defects)
+- a new component needs a reason to exist — reuse before inventing
+- game projects: HUD tokens include attention class (CRITICAL/ACTIVE/CONTEXTUAL/AMBIENT)
+- never redesign the system inside a feature ticket unless the ticket IS the system change
+
+CHECK: roles not raw hues | state matrix complete | focus/contrast tokens present | DO/DON'T recorded | no magic values introduced
+
+OUT: DESIGN_SYSTEM_DELTA (tokens added/changed | components added/changed | inconsistencies found) 

package/skills/design/game-design.md

@@ -0,0 +1,35 @@
+---
+name: game-design
+description: Design for game HUDs, menus, inventories, skill trees, loadouts, maps, settings and overlays — attention budget and state recognition first.
+---
+
+LOAD: ~/.config/opencode/devloom-ai/core.dsl|~/.config/opencode/devloom-ai/workflow.dsl
+DOMAIN: GAME_UI (HUD, menu, inventory, skill tree, loadout, map, settings, overlay)
+
+PRIORITIZE:
+- gameplay context: UI must never obscure the moment-to-moment read of the game
+- attention budget: only one CRITICAL element may demand attention at a time
+- information urgency classification — assign EVERY element:
+  - CRITICAL — survival/progression breaking (health critical, objective timer)
+  - ACTIVE — current engagement state (ammo, cooldowns, active buffs)
+  - CONTEXTUAL — on demand / when relevant (map, quest log, minimap detail)
+  - AMBIENT — flavor/world-building (subtle meters, decorative readouts)
+- readability: legible at gameplay distance/speed; high contrast against varied game backgrounds (backing plates, shadows, or adaptive color)
+- fictional/world identity: the UI belongs to THIS world's materials, typography, and logic
+- input method: controller/keyboard/touch focus order and hover-vs-focus parity
+- state recognition: idle/hover/active/disabled/error states readable in peripheral vision
+
+AVOID (generic game-UI defaults — reject unless the world demands them):
+- cyan sci-fi HUD everywhere, random hexagons, scanlines, glowing boxes on everything
+- dark SaaS dashboard disguised as game UI (generic cards + badges is not a game UI)
+- fake futuristic shapes with no information function
+- animation on every readout; screen shake/flash spam
+- menu hierarchies deeper than the player's patience
+
+RESPONSIVE/SCALABLE: safe-area aware (notches, TV overscan), scale presets for HUD size, 16:9 + ultrawide framing, handheld-class small screens.
+
+A11Y: no color-only status; text scaling where platform allows; reduced-motion respected; photosensitivity rules (no rapid flashing).
+
+CHECK: every element classified CRITICAL/ACTIVE/CONTEXTUAL/AMBIENT | one critical at a time | world identity stated | input method covered | banned defaults absent
+
+OUT: GAME_UI_SPEC (element priority table | layout zones | state matrix | world-identity notes | DON'Ts) — implementation stays with devloom-developer

package/skills/design/motion-design.md

@@ -0,0 +1,34 @@
+---
+name: motion-design
+description: Motion only where it materially contributes — state, hierarchy, causality, continuity, navigation, feedback; reduced-motion respected.
+---
+
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+USE THIS SKILL ONLY WHEN MOTION MATERIALLY CONTRIBUTES. Default answer for most UI work: minimal or none.
+
+MOTION MAY COMMUNICATE (pick the job first, then the animation):
+- state — enter/exit, expand/collapse, enable/disable
+- hierarchy — what appeared above what, where focus moved
+- causality — this happened BECAUSE you did that (confirmation, consequence)
+- continuity — spatial relationship between two views (shared element, push/pop)
+- navigation — where I am going / where I came from
+- feedback — progress, loading, success, failure
+
+RULES:
+- 1 property per animation where possible; short durations (150-300ms UI, <=500ms large transitions)
+- easing: ease-out for entrances, ease-in for exits, no linear except progress
+- never animate layout-critical elements in ways that cause reflow jank; prefer transform/opacity
+- respect prefers-reduced-motion: reduce to opacity fades or instant cuts — content must never depend on motion to be understood
+- no heavy animation dependencies for trivial effects (CSS first; libraries only with existing project precedent)
+- loading: indeterminate motion must still communicate "working", not decorate
+
+AVOID:
+- animating everything on page load (motion fatigue)
+- parallax that fights scrolling, endless ambient loops in productivity UI
+- bounce/spring overuse, shake-on-error without alternative feedback
+- motion that delays the user's next action
+
+CHECK: each animation has a stated job | reduced-motion path exists | no animation blocks interaction | no new dependency added without need
+
+OUT: MOTION_SPEC (trigger | property | duration | easing | reduced-motion fallback) or MOTION_NONE

package/skills/design/visual-direction.md

@@ -0,0 +1,32 @@
+---
+name: visual-direction
+description: Route visual tasks and establish design direction before implementation — domain, hierarchy, three directions, one selection, design memory.
+---
+
+LOAD: ~/.config/opencode/devloom-ai/core.dsl|~/.config/opencode/devloom-ai/workflow.dsl
+
+INPUT: requirements/ticket + optional existing .opencode/devloom/context/design.md + optional screenshots (via devloom-vision)
+
+DO:
+- classify domain: WEB_CREATIVE | PRODUCT_APP | GAME_UI | DATA_VISUALIZATION | GENERIC_UI | EXISTING_DESIGN_SYSTEM
+- EXISTING_DESIGN_SYSTEM wins when a real system exists: extend it, never invent a parallel language
+- extract user goal + audience; set information hierarchy (web/app: primary>secondary>tertiary; game: CRITICAL>ACTIVE>CONTEXTUAL>AMBIENT)
+- inspect existing visual system: tokens, palette, type, spacing, components, icons, motion
+- for substantial NEW graphical work define three internal directions, each with: concept | hierarchy | typography | layout | palette | density | shape | motion | interaction | responsive | explicit anti-pattern
+  - SAFE: coherent evolution of the existing language
+  - BOLD: pushes one axis hard (type / palette / composition), stays legible
+  - UNEXPECTED: re-frames the concept, still fits the domain and goal
+- SELECT EXACTLY ONE direction for implementation — never three complete apps (unless max-quality mode explicitly requested)
+- define responsive strategy (target viewports) + motion budget (or NONE)
+- read existing design.md before subsequent visual work; update ONLY intentionally changed sections; never rewrite from zero without cause
+- emit DESIGN_FILE for the orchestrator to persist (this skill has no write access to app code)
+
+RULES:
+- no implementation code — decisions only
+- vague adjectives are not a direction: translate premium/modern/futuristic/Awwwards-like into executable constraints (see design-grounding)
+- reject default AI aesthetics unless the brief demands them
+- OpenDesign MCP optional: use for grounding when present, never required
+
+CHECK: domain chosen | hierarchy explicit | one direction selected | design.md delta described
+
+OUT: VISUAL_DIRECTION_COMPLETE (DOMAIN | DIRECTION_SELECTED | DESIGN_FILE | CHANGES_SUMMARY)

package/skills/design/web-design.md

@@ -0,0 +1,33 @@
+---
+name: web-design
+description: Art direction for landing pages, portfolios, brand/marketing sites, editorial and campaign pages — composition and typography over component soup.
+---
+
+LOAD: ~/.config/opencode/devloom-ai/core.dsl|~/.config/opencode/devloom-ai/workflow.dsl
+DOMAIN: WEB_CREATIVE (portfolio, landing, brand, marketing, editorial, campaign)
+
+PRIORITIZE:
+- art direction: one governing visual idea per page, expressed consistently
+- typography: deliberate scale ratios, contrast of voice (weight/size/case), typographic rhythm; type IS the layout
+- composition: asymmetry allowed, alignment systems explicit, negative space as a designed element
+- visual pacing: variation of density down the page — breathe, compress, breathe; no uniform stack
+- motion: staged entrances/scroll only when they direct attention; content readable with JS disabled
+- coherent originality: a visitor should recognize the site from a cropped screenshot
+
+AVOID (default AI patterns — reject unless the brief demands them):
+- hero + three cards + CTA as the entire page
+- card soup and pill soup
+- generic purple/blue gradient hero
+- meaningless glassmorphism and arbitrary glow
+- decorative blobs / abstract 3D filler
+- giant text with no compositional purpose
+- carousels hiding content; parallax that breaks scrolling
+- stock-photo dependence instead of designed layout
+
+RESPONSIVE: design mobile-first framing decisions, not desktop shrunk down; reflow preserves hierarchy (never "stack everything"); test 390x844 + 1440x1000.
+
+A11Y: WCAG-AA contrast for all text-over-image; focus visible; landmarks + heading order; reduced-motion honored.
+
+CHECK: governing idea stated | type system deliberate | pacing varies | no banned pattern used | hierarchy survives 390px
+
+OUT: WEB_DESIGN_SPEC (concept | type system | section rhythm | motion budget | DON'Ts) — implementation stays with devloom-developer

package/skills/loop/changelog-drafter.md

@@ -0,0 +1,32 @@
+---
+name: loop-changelog-drafter
+description: Draft release notes from commits since last tag. L1 read-only.
+---
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+DRAFT:
+- find last release tag: git describe --tags --abbrev=0
+- get commits since tag: git log --oneline <tag>..HEAD
+- categorize by conventional commit prefix:
+  - feat: → Features
+  - fix: → Bug Fixes
+  - docs: → Documentation
+  - chore: → Chores/Maintenance
+  - refactor: → Refactoring
+  - test: → Testing
+
+OUTPUT:
+- write draft to .opencode/devloom/reports/changelog-YYYY-MM-DD.md
+- format per template below
+
+```markdown
+# Changelog DRAFT YYYY-MM-DD
+## Features
+- commit messages...
+## Bug Fixes
+- ...
+## Chores
+- ...
+```
+OUT: .opencode/devloom/reports/changelog-YYYY-MM-DD.md
+CHK: commits categorized correctly | tag found | report written

package/skills/loop/ci-sweeper.md

@@ -0,0 +1,24 @@
+---
+name: loop-ci-sweeper
+description: Detect and fix CI failures with worktree-based changes (L2).
+---
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+DETECT:
+- check latest CI run status via gh CLI or commit status API
+- parse failure output — classify as test/lint/build/typecheck
+- create worktree branch for each fix
+
+FIX:
+- lint failures: auto-fix with linter, no manual changes
+- test failures: inspect failing test, fix implementation (not the test)
+- build failures: check for missing imports, syntax errors, config changes
+- typecheck failures: fix type errors, never use `any` as workaround
+
+VERIFY:
+- run full test suite: npm test
+- run typecheck: npx tsc --noEmit
+- log outcome to run-log.json
+
+OUT: fixed CI | worktree merged | run-log updated
+CHK: test suite passes | typecheck clean | no new defects

package/skills/loop/constraints.md

@@ -0,0 +1,21 @@
+---
+name: loop-constraints
+description: Load and validate loop-constraints.md before each tick, build guard string for agent injection.
+---
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+LOAD_CONSTRAINTS:
+- read .opencode/devloom/loop/loop-constraints.md
+- parse numbered rules (lines matching /^\d+\.\s/)
+- build [loop-constraints] guard string
+- inject into agent system prompt before each delegation
+
+DEFAULT_RULES:
+- run tests before any commit
+- never push without human review on denylisted paths
+- pause if daily token budget exceeded
+- L1 level: report-only, no source code edits
+- all loop agents must log outcomes to run-log.json
+
+OUT: constraints guard string injected into agent system prompt
+CHK: rules parsed correctly | guard string non-empty or empty gracefully

package/skills/loop/dependency-sweeper.md

@@ -0,0 +1,24 @@
+---
+name: loop-dependency-sweeper
+description: Scan and patch outdated dependencies (L2 — patch-only auto-updates, minor/major reported).
+---
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+SCAN:
+- run npm outdated to list outdated packages
+- categorize updates: patch (safe), minor (report), major (report)
+- check vulnerabilities: npm audit
+
+UPDATE_RULES:
+- patch updates (1.2.3 → 1.2.4): auto-apply
+- minor updates (1.2.3 → 1.3.0): log to report, no apply
+- major updates (1.2.3 → 2.0.0): log to report, no apply
+- vulnerability fixes: apply regardless of semver if patch available
+
+VERIFY:
+- run full test suite after each patch batch
+- log updated packages to run-log.json
+- report non-patch updates to .opencode/devloom/reports/deps-YYYY-MM-DD.md
+
+OUT: patches applied | deps report written | run-log updated
+CHK: test suite passes | only patch auto-applied | vulnerabilities fixed

package/skills/loop/design-audit.md

@@ -0,0 +1,67 @@
+---
+name: loop-design-audit
+description: "Two-mode full audit: L1 catalogs UI/UX + functional issues (broken/missing actions, dead-ends, API mismatches); L2 fix applies unified design + functional fixes in worktrees until all verified"
+---
+LOAD: ~/.config/opencode/devloom-ai/core.dsl|~/.config/opencode/devloom-ai/verify.dsl
+
+AUDIT:
+- enumerate: read target URL list from args, normalize each to canonical page-id
+- visit each page, wait for full render
+
+UI_CHECK:
+- catalog: buttons, links, inputs, modals, forms, tables, tabs, navigation, cards, badges, icons, avatars, tooltips, progress, notifications, loading/skeleton, error, empty states
+- record per-element: position, CSS (color, spacing, font, border-radius, shadow)
+- note responsiveness: breakpoints, layout shifts, overflow
+
+FUNCTIONAL_CHECK:
+- test every button: click → verify expected action fires (not dead, not error)
+- test every link: navigate → verify target exists (not 404, not dead-end)
+- test every form: submit → verify validation + submission + success/error feedback
+- test every modal: open → verify close mechanism + overlay + content
+- test every table: sort/pagination/row-action → verify functional
+- test every tab: switch → verify content loads + active state
+- test every API call: verify endpoint exists + params correct + error handling present
+- test every async action: verify loading state + success state + error state present
+- catalog MISSING actions: CRUD where needed, validations absent, feedback missing
+- catalog DEAD-END flows: journeys leading nowhere, missing back/return, no success/failure feedback
+- catalog STATE issues: forms not resetting, stale data, race conditions, data not refreshing after mutation
+
+CATALOG INCONSISTENCIES:
+- UI: color mismatches, spacing violations, typography drift, border-radius, shadow, icon style, button variants, form field asymmetry, table styling, modal differences, nav inconsistency, hover/focus/active states
+- Functional: broken actions (click does nothing), broken links (404), form validation missing, API error handling absent, loading/error/empty states missing, dead-end flows, stale state
+
+PRODUCE SPEC:
+- distill canonical UI values per component variant
+- define functional contract: every action has handler, every form validates, every API has error handling, every flow has feedback
+- write unified spec to .opencode/devloom/reports/design-system-spec.md
+- write audit report to .opencode/devloom/reports/design-audit-YYYY-MM-DD.md
+- tag each issue with category (ui|functional|missing-action|broken-action|api|state) + severity (critical|high|medium|low)
+
+FIX:
+- init: read progress from .opencode/devloom/loop/progress.json
+- pick: select next item with status "pending" (ordered by severity: critical first)
+- implement per item:
+  - create worktree branch: devloom/design-audit/{PAGE-ID}
+  - apply design system spec values to UI elements on the page
+  - fix broken actions: wire up handlers, fix API calls, add error handling
+  - add missing actions: implement missing CRUD, validations, feedback states
+  - resolve dead-end flows: add navigation, success/failure feedback
+  - fix state issues: form resets, data refresh, race condition guards
+  - commit changes in worktree
+- verify each item:
+  - re-scan page against spec + functional contract
+  - confirm: UI matches + actions work + forms validate + links resolve + API errors handled + states present
+  - if failures: log, increment retry, go back to implement (max 3)
+  - if pass: mark item "verified", write progress.json
+- loop: if not all verified and budget remains → continue (next tick)
+- stop: all verified OR budget exhausted OR max retries hit OR user pauses
+
+CONSTRAINTS (binding):
+- L1 mode: no source file edits
+- L2 mode: edits only in worktrees (branch: devloom/design-audit/{page-id})
+- max 3 retries per item
+- each item must pass verifier before progressing
+- log all outcomes to run-log.json
+
+OUT: .opencode/devloom/reports/design-audit-YYYY-MM-DD.md | .opencode/devloom/reports/design-system-spec.md
+CHK: audit produced | audit=no-source-edits | fix=all-items-verified | every-action-works | every-form-validates | every-link-resolves | every-api-has-error-handling | progress-written-to-disk

package/skills/loop/issue-triage.md

@@ -0,0 +1,27 @@
+---
+name: loop-issue-triage
+description: Triage new issues with labels and priority. L1 read-only.
+---
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+TRIAGE:
+- list open issues without labels: gh issue list --state open --limit 20
+- for each issue, read title and body
+
+CLASSIFY:
+- Bug: unexpected behavior, crash, error message
+- Feature: new capability request
+- Question: how-to, configuration help
+- Docs: documentation gap or error
+
+ASSIGN_PRIORITY:
+- High: security, data loss, production blocker
+- Medium: feature request, non-critical bug
+- Low: cosmetic, nice-to-have, question
+
+REPORT:
+- write to .opencode/devloom/reports/triage-YYYY-MM-DD.md
+- per issue: #N: [type] priority=priority suggested_labels=[...]
+
+OUT: .opencode/devloom/reports/triage-YYYY-MM-DD.md
+CHK: all unlabeled issues processed | classification correct | report written

package/skills/loop/post-merge-cleanup.md

@@ -0,0 +1,24 @@
+---
+name: loop-post-merge-cleanup
+description: Clean up stale branches and orphaned worktrees. L1 report-only, no destructive actions.
+---
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+FIND:
+- list all worktrees from .devloom-worktrees/.registry.json
+- for each branch, check if merged into main/master
+- check branch age — flag >30 days as stale
+- check for orphaned worktree directories (not in registry)
+
+CLASSIFY:
+- merged worktrees: safe to remove, recommend in report
+- stale unmerged (>30 days): flag for human review
+- orphaned directories: list paths, recommend manual cleanup
+- active worktrees (clean, recently used): no action needed
+
+REPORT:
+- write to .opencode/devloom/reports/cleanup-YYYY-MM-DD.md
+- include per entry: branch name, age, merge status, recommended action
+
+OUT: .opencode/devloom/reports/cleanup-YYYY-MM-DD.md
+CHK: report written | no destructive actions taken | no false positives

package/skills/loop/pr-babysitter.md

@@ -0,0 +1,26 @@
+---
+name: loop-pr-babysitter
+description: Watch open PRs, push stale ones forward, fix blocked CI, merge approved. L2 with verifier review.
+---
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+ANALYZE:
+- list open PRs: gh pr list --state open
+- check each PR's last activity timestamp
+- flag PRs idle >2 hours as candidates for action
+- read PR diff and comments to understand status
+
+ACT:
+- stale PR: rebase or resolve conflicts
+- blocked PR: identify blocker, ping reviewer or fix deps
+- draft PR with failing CI: fix CI in linked worktree
+- approved PR: merge if CI passes
+
+SAFETY:
+- max 3 attempts per PR per tick
+- never force-push
+- never close PRs automatically
+- log all actions to run-log.json
+
+OUT: PRs advanced or merged | run-log updated
+CHK: max 3 attempts enforced | no force-push | no auto-close

package/skills/loop/triage.md

@@ -0,0 +1,25 @@
+---
+name: loop-triage
+description: Daily triage loop — check PRs, CI, branches, worktrees, budget. L1 report-only.
+---
+LOAD: ~/.config/opencode/devloom-ai/core.dsl
+
+TICK:
+- check open PR count and age (gh pr list / git log)
+- check CI status from recent commits (git log --oneline -5)
+- list stale branches older than 7 days (git branch -r)
+- check unmerged worktrees (.devloom-worktrees/.registry.json)
+- check token budget remaining (.opencode/devloom/loop/budget.json)
+
+REPORT:
+- write markdown to .opencode/devloom/reports/daily-YYYY-MM-DD.md
+- include: summary table, recommendations, budget status
+
+CONSTRAINTS (binding):
+- no source file edits
+- no test creation or modification
+- no push or PR creation
+- log outcome to run-log.json
+
+OUT: .opencode/devloom/reports/daily-YYYY-MM-DD.md
+CHK: report written | no source files touched | budget checked