- Extract buildApp() from server.ts into app.ts for reuse
- Create api/index.ts as Vercel serverless function entry point
- Add vercel.json with rewrites (static assets + catch-all to API)
- Build script now copies static assets to public/ and templates to dist/
- Include src/templates/** in serverless function bundle
- Local dev and tests continue to work unchanged (60/60 passing)
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Fastify v5 + TypeScript + Nunjucks + Tailwind CSS v4 application that
guides Brazilian lawyers through a step-by-step flow to collect case
details and assemble optimized prompts for Jus IA.
Includes:
- Flow engine with registry pattern for area/subtipo flows
- Complete trabalhista/horas-extras flow (2 steps, legal references)
- MPA state management via hidden form fields
- Chip-selector UI with 44px touch targets
- Optional LLM integration for contextual refinement questions
- Prompt builder with template interpolation and URL delivery
- Progressive enhancement JS (copy-to-clipboard)
- Deep link support (/:area/:subtipo)
- Brand design tokens from Jusbrasil guidelines
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Replace assumed blue/gold palette with real Jus IA colors extracted via
Firecrawl: primary green #007A5F, text #5C6F8A, border #B3C0D0.
Keep gold #D4A843 for legal badges only. Add brand-guidelines.md reference.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Incorporates 5 key decisions from multi-agent discussion:
- Invisible AI transition: AI questions indistinguishable from structured ones
- Quantified speed: max 4-5 screens grouped by mental moment
- Preview as legal trust contract: show CLT articles/TST precedents
- Recovery moment as Critical Success Moment: back button + state preservation
- Effortless interactions: one tap per decision, contextual placeholders
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
12 NFRs across 4 categories: performance (loading/LLM latency),
security (no data persistence, no logging), integration (Jus IA URL
format, overflow validation), compatibility (evergreen browsers,
mobile-first, WhatsApp OG tags). Lean approach - no SLAs.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
MPA mobile-first, evergreen browsers, no SEO priority,
request-response with loading state for LLM calls,
no accessibility priority for MVP. Stateless architecture.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
No breakthrough innovation detected. Product is smart execution of
existing concepts (dynamic forms + LLM + parameterized URL + verified
legal base). Innovation is in the approach, not the technology.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Legaltech high-complexity concerns (OAB ethics, data retention,
attorney-client privilege, court integration) don't apply to Start Kit:
no data persistence, no login, no court integration, redirect-only.
Only residual constraint: domain knowledge for prompt template creation.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Web App | Legaltech | Medium complexity (high domain knowledge) | Greenfield.
Party Mode revised complexity from High to Medium based on stateless/anonymous
architecture with no data retention or court integration.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
- Updated all JusPrompt references to Jus IA Start Kit in product brief
- Updated Party Mode insights to reflect decisions made (name, AI in MVP)
- Created PRD document from template with Step 1 initialization complete
- Discovered and loaded 7 input documents (1 brief, 1 research, 1 brainstorm, 4 project docs)
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Completed Step 4 with North Star metric (1000 redirects/month),
user success metrics, business objectives, and funnel KPIs.
Product is one-shot by design - no retention loop.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Completed Step 3 (Target Users Discovery) with three personas:
- Dra. Carla (never used AI, pragmatic resistor)
- Dr. Rafael (uses AI poorly, digital overload)
- Dr. Marcos (office owner, Shadow AI concerns)
Includes detailed user journey for labor law petition flow.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Completed Step 2 (Vision Discovery) of the BMAD product brief workflow
for JusPrompt - a conversational assistant that guides lawyers through
building optimized requests for Jus IA without prompt engineering knowledge.
Includes Party Mode insights on naming, URL limits, and MVP strategy.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Set up the product brief document with frontmatter tracking workflow state
and input documents discovered during Phase 1 Analysis (brainstorm, research,
and 4 detailed idea explorations).
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Research covers:
- Prompt Library: 8+ BR competitors mapped, white space in verified-data-connected prompts
- Calculator: 10+ competitors including Jusfy (R$50M+ revenue, 118% YoY), gap in AI end-to-end
- Dashboard: No BR competitor exists but better as feature (80% don't measure AI ROI)
- Community: 4-5 paid communities exist (personality-driven), better absorbed as Library feature
Final prioritization: (1) Prompt Library, (2) Calculator, (3/4) Dashboard+Community as features
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Incorporate proprietary survey data from Jusbrasil (Aug 2025, n=2861; Mar 2026, n=977)
covering adoption rates, tool usage, trust levels, non-user barriers, and conversion
factors. Add section 3 with data-driven implications for each of the 4 selected ideas.
Key findings: Prompt Library has highest relevance (ease of use +4.2pp as conversion
factor), Dashboard is more retention than acquisition, Community is better as a feature
than standalone product.
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Add structured documentation for the Jus IA external tools brainstorming project:
- Research context on AI legal tools in Brazil and lawyer pain points
- Full brainstorm with 23 ideas across 5 categories
- 4 selected ideas with detailed descriptions: Prompt Library, Legal Calculator,
Productivity Dashboard, and Prompts Community
- README with project overview, principles, and next steps
https://claude.ai/code/session_01CvrcMDqfCKWV2hC3xpRbx3
Align documentation with the skills-based architecture migration by
replacing references to "commands", "slash commands", and legacy command
names with the new "skills" terminology and skill names.
* refactor(sprint-status): convert workflow.yaml + instructions.md to single workflow.md
Merge workflow config and instruction content into a unified workflow.md
with YAML frontmatter, following the established convention for converted
workflows. Update module-help.csv reference accordingly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(sprint-status): restore no-time-estimates rule dropped during conversion
The <critical> preamble removal incorrectly classified this behavioral
rule as boilerplate. It is an actual output constraint.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* test: add comprehensive workflow conversion test results
- 14 test fixtures covering data and validate modes
- Tested across Opus, Sonnet, and Haiku models
- OLD format (yaml+md) vs NEW format (workflow.md)
- Confirms zero regressions in conversion
- Includes reproduction instructions for future sessions
* fix(sprint-status): consolidate no-time-estimates into role line
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(qa-generate-e2e-tests): convert workflow.yaml + instructions.md to single workflow.md
Task 2 of yaml-to-md conversion plan. Merges config variables into
INITIALIZATION section, inlines instructions into EXECUTION section.
Drops non-consumed yaml keys (required_tools, tags, execution_hints).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(retrospective): convert workflow.yaml + instructions.md to single workflow.md
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: update workflow.yaml references to workflow.md for converted workflows
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: remove test results file from version control
SPRINT_STATUS_CONVERSION_TEST_RESULTS.md contains hardcoded local
filesystem paths and is a session-specific test artifact. Added to
.bare/info/exclude to keep it ignored across all worktrees.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(skills): migrate Roo Code installer to native skills format
Move Roo Code from legacy `.roo/commands/` flat files to native
`.roo/skills/{skill-name}/SKILL.md` directory output. Verified
skill discovery in Roo Code v3.51 with 43 skills installed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* test(skills): add native skills tests for Claude Code, Codex, and Cursor
Add dedicated test suites covering config validation, fresh install,
legacy cleanup, and ancestor conflict detection for Claude Code, Codex
CLI, and Cursor. Updates migration checklist to reflect verified status.
84 assertions now pass (up from 50).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* test(skills): add Roo Code reinstall/upgrade test
Verify that running Roo setup over existing skills output succeeds
and preserves SKILL.md output. Checks off the last Roo checklist item.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(skills): migrate GitHub Copilot to config-driven native skills
Replace 699-line custom installer with config-driven skill_format.
Output moves from .github/agents/ + .github/prompts/ to
.github/skills/{skill-name}/SKILL.md. Legacy cleanup strips BMAD
markers from copilot-instructions.md and removes old directories.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* docs: update migration checklist with Copilot and Roo verified results
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(skills): migrate Cline to config-driven native skills
Move Cline installer from .clinerules/workflows to .cline/skills with
SKILL.md directory output. Add legacy cleanup and 9 test assertions.
* feat(skills): migrate CodeBuddy to config-driven native skills
Move CodeBuddy installer from .codebuddy/commands to .codebuddy/skills
with SKILL.md directory output. Add legacy cleanup and 9 test assertions.
* feat(skills): migrate Crush to config-driven native skills
Move Crush installer from .crush/commands to .crush/skills with
SKILL.md directory output. Add legacy cleanup and 9 test assertions.
* feat(skills): migrate Trae to config-driven native skills
Move Trae installer from .trae/rules to .trae/skills with SKILL.md
directory output. Add legacy cleanup and 9 test assertions.
* feat(skills): migrate KiloCoder to config-driven native skills
Replace 269-line custom kilo.js installer with config-driven entry in
platform-codes.yaml targeting .kilocode/skills/ with skill_format: true.
- Add installer config: target_dir, skill_format, template_type, legacy_targets
- Add cleanupKiloModes() to strip BMAD modes from .kilocodemodes on cleanup
- Remove kilo.js from manager.js customFiles and Kilo-specific result handling
- Delete tools/cli/installers/lib/ide/kilo.js
- Add test Suite 22: 11 assertions (config, install, legacy cleanup, modes, reinstall)
- Update migration checklist with verified results
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(skills): migrate Gemini CLI to config-driven native skills
Replace TOML-based .gemini/commands output with native SKILL.md output
in .gemini/skills/. Gemini CLI confirms native skills support per
geminicli.com/docs/cli/skills/.
- Update platform-codes.yaml: target_dir, skill_format, legacy_targets
- Add test Suite 23: 9 assertions (config, install, legacy, reinstall)
- Add Gemini CLI section to migration checklist
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(skills): migrate iFlow, QwenCoder, and Rovo Dev to native skills
Complete the native skills migration for all remaining platforms:
- iFlow: .iflow/commands → .iflow/skills (config change)
- QwenCoder: .qwen/commands → .qwen/skills (config change)
- Rovo Dev: replace 257-line custom rovodev.js with config-driven
.rovodev/skills, add cleanupRovoDevPrompts() for prompts.yml cleanup
All platforms now use config-driven native skills. No custom installer
files remain. Manager.js customFiles array is now empty.
- Add test suites 24-26: 20 new assertions (173 total)
- Update migration checklist: all summary gates passed
- Delete tools/cli/installers/lib/ide/rovodev.js
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(installer): preserve bmad-os-* skills during cleanup
The cleanupTarget method removed all entries starting with "bmad" from
IDE skills directories, which would also wipe version-controlled
bmad-os-* skills from the BMAD-METHOD repo. Add exclusion for the
bmad-os- prefix so those skills survive reinstalls.
* docs: flag all unverified platforms for manual IDE testing
Add NEEDS MANUAL IDE VERIFICATION to KiloCoder, Gemini CLI, iFlow,
QwenCoder, and Rovo Dev checklists. CodeBuddy, Crush, and Trae already
had the flag.
* fix(installer): suspend Kilo Code and add verified Gemini/Crush results
Kilo Code does not support the Agent Skills standard — the migration
from modes+workflows to skills was based on a false fork assumption.
- Add suspended field to platform-codes.yaml, hiding Kilo from the IDE
picker and blocking setup with a clear message
- Fail the installer early (before writing _bmad/) if all selected IDEs
are suspended, protecting existing installations from being corrupted
- Still clean up legacy Kilo artifacts (.kilocodemodes, .kilocode/workflows)
when users switch to a different IDE
- Mark Crush and Gemini CLI as manually verified (both work end-to-end)
- Replace Suite 22 install tests with suspended-behavior tests (7 assertions)
* docs: update KiloCoder checklist to reflect suspended status
* fix(skills): add canonicalIds for BMM research and PRD workflows
Drop the bmm module prefix from 6 workflow skill names so they
install as bmad-create-prd, bmad-domain-research, etc. instead of
bmad-bmm-create-prd, bmad-bmm-domain-research, etc.
* fix(installer): address PR review findings from automated reviewers
Triage of 18 findings from Augment and CodeRabbit reviews on PR #1841:
Source code fixes:
- Exclude bmad-os-* from findAncestorConflict to match cleanupTarget
- Wrap cleanupCopilotInstructions in try/catch (best-effort, not fatal)
- Wrap suspended-platform cleanup in try/catch (failure boundary)
- Clean up temp backup dirs in catch block when install aborts
- Normalize IDE keys to lowercase before suspended lookup
- Delete dead loadCustomInstallerFiles method and stale references
- Rename "Roo Cline" to "Roo Code" in both platform-codes.yaml files
- Fix Gemini CLI package name (@google/gemini-cli, not @anthropic-ai)
Test improvements:
- Add name/frontmatter invariant check to 6 missing platform suites
- Assert stale bmad-architect skill is removed after cleanup
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
These manifests were missed during the all-is-skills migration (#1834),
leaving 6 workflows undiscoverable by the native skills installer.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(skills): add canonical bmad- naming via skill manifests
Add bmad-skill-manifest.yaml sidecars to all 38 capabilities (tasks,
agents, workflows) declaring canonicalId as the single source of truth
for skill names. Update Claude Code and Codex installers to prefer
canonicalId over path-derived names, with graceful fallback.
- 24 manifest files covering 38 capabilities
- New shared skill-manifest.js utility for manifest loading
- resolveSkillName() in path-utils.js bridges manifest → installer
- All command generators propagate canonicalId through CSV manifests
- Drops bmm module prefix from all user-facing skill names
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat(skills): claude-code installer outputs .claude/skills/<name>/SKILL.md
Refactor the config-driven installer to emit Agent Skills Open Standard
format for Claude Code: directory-per-skill with SKILL.md entrypoint,
unquoted YAML frontmatter, and full canonical names.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* refactor(installer): migrate codex to config-driven pipeline
Delete the custom codex.js installer (441 lines) and route Codex
through the config-driven pipeline via platform-codes.yaml. This
fixes 7 task/tool descriptions that were generic due to bypassing
manifests, and eliminates duplicate transformToSkillFormat code.
Key changes:
- Add codex entry to platform-codes.yaml with skill_format + legacy_targets
- Remove codex from custom installer list in manager.js
- Add installCustomAgentLauncher() to config-driven for custom agent support
- Add detect() override for skill_format platforms (bmad-prefix check)
- Set configDir from target_dir for base-class detect() compatibility
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(installer): guard codex skill installs in nested directories
* fix(installer): warn on stale global legacy skill dirs
* feat(installer): migrate cursor to native skills
* Migrate Windsurf installer to native skills
* Clarify Windsurf skill invocation in checklist
* feat(installer): migrate kiro to native skills
* docs: record kiro skill visibility verification
* Migrate Antigravity installer to native skills
* Document Antigravity ancestor skill verification
* Synchronize native skills migration checklist
* Migrate Auggie installer to native skills
* Migrate OpenCode installer to native skills
* Document live skill verification for Auggie and OpenCode
* fix(test): replace _bmad filesystem dependency with self-contained fixture
The installation component tests walked up the filesystem looking for a
pre-installed _bmad directory, which exists locally but not in CI. Replace
findInstalledBmadDir() with createTestBmadFixture() that creates a minimal
temp directory with fake compiled agents, making tests fully self-contained.
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Brian <bmadcode@gmail.com>
The staleness warning in sprint-status always fired after 7 days because
it checked the 'generated' timestamp, which is only set once during
sprint-planning. Other workflows (dev-story, create-story, code-review,
retrospective) update statuses but never touched 'generated'.
This adds a 'last_updated' field that is:
- Set initially alongside 'generated' during sprint-planning
- Bumped to current date by every workflow that modifies sprint-status.yaml
- Used by the staleness check (with fallback to 'generated' for backward
compatibility)
Fixesbmad-code-org/BMAD-METHOD#1820
Co-authored-by: Oz <oz-agent@warp.dev>
Prevent the agent from treating detailed, plan-like input as a validated
plan and short-circuiting the workflow. The new rule ensures the full
workflow is followed regardless of input specificity.
* feat(i18n): add zh-cn locale support with Starlight routing
Reorganize Chinese translations from docs_cn/ into docs/zh-cn/
following Starlight's i18n content structure. Configure Starlight
with root locale (en, unprefixed) and zh-cn (/zh-cn/ prefix).
- Move 28 files from docs_cn/*_cn.md to docs/zh-cn/*.md
- Fix 21 internal links to remove _cn suffixes
- Add defaultLocale + locales config to astro.config.mjs
- Add .gitignore exception for docs/zh-cn/ (overrides z*/ rule)
- Language switcher activates automatically via existing Header
- hreflang tags auto-generated by Starlight on all pages
* feat(i18n): add zh-CN UI translations and sidebar labels
- Add website/src/content/i18n/zh-CN.json with Starlight UI strings
- Add sidebar group label translations (欢迎, 路线图, 教程, etc.)
- Register i18n collection in content config to fix deprecation warning
* fix(i18n): exclude 404 pages from sitemap
Custom 404 pages (root and zh-cn) were indexed as regular content in
the sitemap. Add a filter to the @astrojs/sitemap integration that
matches the /404 path segment precisely via regex on the URL pathname.
* feat(skills): add edge case hunter as parallel review layer in PR review
Wire review-edge-case-hunter.xml into bmad-os-review-pr as a second
review layer running in parallel with the adversarial review. Both
subagents receive the same PR diff concurrently. Findings are merged,
deduplicated, and tagged by source before tone transformation.
* fix(core): resolve contradictions in edge case hunter task spec
- Show array wrapper [{}] in output-format example to match JSON array
contract, and document empty array [] as valid output
- Consolidate empty-content handling: step 1 now defers to halt-conditions
instead of defining separate "ask and abort" behavior
- Zero-findings halt no longer contradicts JSON contract: re-analyze once,
then return [] instead of ambiguous "HALT or re-analyze"
- Soften "Execute ALL steps" to acknowledge halt-conditions can interrupt
* fix(review): address PR #1791 review feedback
- Remove "Task tool" reference per maintainer; use generic "subagents"
- Fix nested triple-backtick fencing with four-tick outer fence
- Widen location format to support multi-line ranges and hunk refs
- Add JSON-safety constraint to guard_snippet field
- Tighten input loading to "strictly from provided input"
- Replace vague "unreadable" with "cannot be decoded as text"
- Replace vague "increased scrutiny" with concrete re-analysis checklist
- Resolve HALT-immediately vs re-analysis conflict in LLM instructions
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
These config keys were added in efc69ffb but never consumed by any
template, conditional logic, or runtime code. They represent platform
capabilities that the agent can determine at runtime — persisting a
stale user guess adds installation friction with zero execution value.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Team-based skill that orchestrates human-in-the-loop triage of review
findings using parallel Opus agents. One agent per finding researches
autonomously, proposes a plan, then holds for human conversation before
a decision is recorded. Team lead maintains scorecard and lifecycle.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(quick-flow): add quick-dev2 unified workflow
Add the quick-dev2 workflow that unifies clarify, plan, implement,
review, and present into a single flow. Register it in the agent
menu, module-help catalog, and test fixtures.
* fix(quick-flow): rename QD2 trigger to QQ for schema compliance
COMPOUND_TRIGGER_PATTERN only allows uppercase letters in shortcuts.
Rename to QQ so quick-dev2 passes agent schema validation.
* fix(quick-flow): address PR review findings for quick-dev2
- step-04-review: fix copy-paste fallback text to say "perform all
three reviews inline sequentially" instead of "implement directly"
- workflow.md: add missing planning_artifacts to initialization list,
matching quick-spec and quick-dev siblings
- quick-flow-solo-dev.agent.yaml: change QD and QQ menu entries from
workflow: to exec: for .md files, matching the exec-for-md convention
* fix(quick-flow): use human-in-the-loop fallback for review without subagents
Sequential inline reviews in the same context suffer from anchoring
bias and context blowout. Instead, generate separate review prompt
files and ask the human to run each in a separate session.
* refactor(quick-flow): rename quick-dev2 to quick-dev-new-preview
Rename directory, update all references in agent menu, module-help,
and workflow internals.
Claude
Alex Verkhovsky
cccczl
Chandan Veerabhadrappa
Brian Madison