Skip to content

josembuitron/bridge-agentic-pipeline

Repository files navigation


License: MIT Claude Code Pipeline

BRIDGE Agentic Pipeline

Turn business requirements into delivered technical solutions, automatically.

Most AI development tools help you write code faster. Bridge does something different: it takes a messy meeting transcript, a client email, or a rough product brief and runs it through a structured pipeline that translates requirements, researches technologies, designs architecture, builds the solution, and validates everything before delivery. You get working code, client-ready proposals, and architecture diagrams -- not just autocomplete suggestions. It is the difference between an AI assistant and an AI development team.

BRIDGE Pipeline Architecture

Who This Is For

You are a... Bridge helps you...
Development agency Go from client call to delivered proposal in hours, not weeks
Consultancy / advisory firm Generate technology assessments and solution proposals with real cost models
Fractional CTO / engineering lead Run a full development pipeline solo -- Bridge acts as your analyst, researcher, architect, and QA team
Startup founder Build your MVP with enterprise-grade process without the enterprise-grade team
Freelance developer Deliver professional proposals and architecture docs that justify premium rates
Data engineering team Design ETL pipelines, API integrations, and dashboard architectures with validated tech stacks
System integrator Connect platforms (NetSuite, Salesforce, Dynamics 365) with researched, documented approaches

How It Works

 INPUT                    PIPELINE                                    OUTPUT
 ─────                    ────────                                    ──────

 Meeting transcript   ┌─────────────────────────────────────────┐   Client-ready
 Client email        │  Phase 1: TRANSLATE (BRIDGE B-R-I-D)    │   deliverables
 Product brief  ───> │  Phase 2: RESEARCH technologies          │ ───> Solution proposals
 Chat messages       │  Phase 3: ARCHITECT solution             │   Architecture diagrams
 Requirements doc    │  Phase 4: BUILD (vertical slices + TDD)  │   Working code + tests
                     │  Phase 5: VALIDATE and deliver           │   Deployment guides
                     └─────────────────────────────────────────┘   Quality reports

 Human approval gates at EVERY phase. Stop at any point and get deliverables.

The Five Phases

Phase What Happens What You Get
1 -- Translate Raw input analyzed through the BRIDGE framework. Business challenges, root causes, impact metrics, and data context extracted. Technical Definition + BRIDGE Analysis
2 -- Research APIs, platforms, and technologies investigated using live documentation. Claims from Phase 1 validated against real-world data. Research Report with verified tech stack
3 -- Architect Complete solution designed with architecture diagrams, cloud cost models, specialist team breakdown, and vertical slices. Solution Proposal + Mermaid diagrams
4 -- Build Dynamic specialist agents execute each vertical slice using TDD. Walking skeleton first, then incremental hardening. Working code, tests, build manifest
5 -- Validate Goal-backward verification, 6-pass code review, SAST security scanning, secrets detection, and quality scoring. Validation report + client deliverables

You control the pipeline at every step. Approve, modify, go back, or stop and generate deliverables from whatever is complete. The most common exit point is Phase 3 -- perfect for generating client proposals without writing code.

Agent Interaction Map

BRIDGE Agent Interaction Map

Commands

All commands are invoked through Claude Code's slash command system.

Pipeline Commands

Command What It Does
/bridge Start a new project or continue an existing one. The pipeline detects returning projects and offers to resume.
/bridge help Show the setup and configuration guide: folder structure, brand assets, tools, phases, and all available commands.
/bridge list List all projects across all clients with their current phase and status.

Maintenance Commands

Command What It Does
/bridge health Run 5 diagnostic checks on the most recent active project: state coherence, knowledge freshness, dependency vulnerabilities, test suite health, and documentation drift.
/bridge health {client}/{project} Run health checks on a specific project.
/bridge dream {client} Consolidate the knowledge graph for a client. Merges decisions, resolves contradictions, prunes stale entries, and archives old data. Recommended after 3+ completed projects.
/bridge dream List all clients with knowledge graphs and choose which one to consolidate.
/bridge dream all-tooling Consolidate global tooling patterns across all clients. Only tool success rates cross client boundaries -- never business data.

During Pipeline Execution

At every phase gate, you have these options:

Option When to Use
Approve Phase output meets your standards. Advance to next phase.
Modify Provide specific feedback. The responsible agent re-runs with your corrections.
Stop and deliver Accept what exists so far. Deliverables generated from completed phases.
Reject Start the phase over. Previous attempt feedback is included to prevent repeating mistakes.

Phase 4 Build Options (per slice)

Option Effect
Approve and continue Accept this slice, move to next.
Approve all remaining Skip per-slice review for remaining slices.
Request changes Re-run this slice with your feedback.
Skip remaining slices Accept thin functionality, move to next specialist.
Review code Inspect specific files before deciding.
Pause pipeline Save progress and resume later with /bridge.

Configuration (Phase 0)

During initialization, the pipeline asks for configuration preferences:

Setting Options Default
Mode Interactive (confirm each step) / YOLO (minimal confirmations) Interactive
Model profile Quality (Opus-heavy) / Balanced / Budget (Sonnet-heavy) Balanced
Budget cap Dollar amount or unlimited Unlimited
Critical review Enable Ojo Critico skeptical reviewer at phase gates Enabled
Security gate Blocking (critical findings prevent delivery) / Advisory Blocking
De-Sloppify Run garbage collection after build phase Enabled
Build granularity Fine (1 feature/slice) / Medium (2-3) / Coarse (full specialist) Auto-detect

The BRIDGE Framework

BRIDGE is a structured methodology for translating business problems into technical solutions. It ensures every project starts from the real problem, not from a premature technology choice.

Letter Phase Owner Purpose
B Business Challenge Translator What was said vs. what is actually needed
R Root Causes Translator Causal analysis -- why the problem exists
I Impact & Symptoms Translator KPIs, financial exposure, operational friction
D Data & Context Translator + Researcher Systems, APIs, constraints (preliminary, then validated)
G Generate Use Cases Architect 3-5 technical use cases mapped to root causes
E Evaluate Feasibility Architect Viability, complexity, timeline, risk for each use case

Rather than concentrating all analysis in one agent, BRIDGE phases are distributed where each agent has the right expertise:

B --- R --- I --- D(prelim)          D(validated)          G --- E
|--------------------------|         |--------------|      |--------------|
      Phase 1: Translator             Phase 2: Researcher    Phase 3: Architect
      (business analysis)             (tech validation)      (solution design)

Read the original article: Why Some AI Projects Start with the Wrong Problem


Complete Architecture

Modular Orchestrator Design

The pipeline orchestrator is modular -- files are loaded on-demand as each phase begins, never preloaded.

skills/bridge/
├── SKILL.md                          # Entry point -- triggers on /bridge
├── orchestrator/
│   ├── core.md                       # Pipeline flow, rules, guardrails, gate enforcement
│   ├── phases/
│   │   ├── 00-initialization.md      # Tool discovery, input collection, workspace setup
│   │   ├── 00b-codebase-analysis.md  # Brownfield/existing codebase support (conditional)
│   │   ├── 01-translate.md           # Phase 1: BRIDGE B-R-I-D requirements translation
│   │   ├── 02-research.md            # Phase 2: Technology research with live docs
│   │   ├── 03-architect.md           # Phase 3: Solution design + specialist team spec
│   │   ├── 04-build.md              # Phase 4: Dynamic agent creation + vertical slice execution
│   │   └── 05-validate.md           # Phase 5: Validation, security, code review, delivery
│   └── modules/
│       ├── available-plugins.md      # Full catalog of 35 ToB skills + all plugins/MCPs/CLIs
│       ├── tool-matrix.md            # Agent-to-tool assignment matrix + dependency resolution
│       ├── cross-skill-activation.md # When to activate each skill per phase
│       ├── doc-access-strategy.md    # 6-tier documentation access chain
│       ├── model-routing.md          # Cost-aware model selection (Opus/Sonnet/Haiku)
│       ├── context-budget.md         # Context window management (9 rules + enforcement)
│       ├── cost-tracking.md          # Token/cost estimation with budget caps
│       ├── pipeline-state.md         # State file for cross-session resumability
│       ├── flexible-execution.md     # Out-of-order phases, parallel execution, resume
│       ├── rollback.md               # Git tag-based phase rollback
│       ├── deliverable-generation.md # Internal + client deliverable generation
│       ├── sanitization-checklist.md # Client deliverable sanitization rules
│       ├── milestone-delivery.md     # Incremental milestone delivery
│       ├── client-knowledge-graph.md # Per-client knowledge graph (strict isolation)
│       ├── issue-tracker.md          # External issue tracker integration
│       ├── structural-linter.md      # 5-check post-build architectural compliance
│       ├── garbage-collector.md      # 5-check codebase hygiene (extends De-Sloppify)
│       ├── harness-hooks.md          # Project pre-commit hooks + pipeline protection hooks (5 guards)
│       ├── adversarial-verifier.md   # Independent execution-based verification (Phase 5)
│       ├── dream-consolidation.md    # Client knowledge graph consolidation between projects
│       ├── health-check.md           # On-demand pipeline and project health diagnostics
│       ├── proposal-fast-track.md    # 3-phase collapsed pipeline for proposals/decks (30-45 min)
│       ├── pptx-engine.md            # Coordinated multi-tool PPTX generation (python-pptx + pptxgenjs + PresentationGO)
│       ├── design-enforcement-hook.md # Deterministic hooks enforcing design rules (em dashes, NODE_PATH, local installs)
│       ├── pixel-agent.md            # Agent description naming convention
│       ├── sbom-generator.md         # CycloneDX 1.5 SBOM generation (Phase 0 + Phase 4)
│       ├── supply-chain-gate.md      # Pre-install security scanning for pip/npm packages
│       ├── persistent-security-log.md # Structured JSONL security event logging with hash chain
│       └── self-test.md              # Structural validation dry-run checklist
├── references/
│   ├── ojo-critico.md                # Critical reviewer prompt template
│   ├── tool-risk-matrix.md           # Risk classification + taint tracking protocol
│   ├── bridge-tool-versions.json     # Pinned tool versions for reproducible installs
│   ├── security-guardrails.json      # GPO-like security policies (AI-SAFE2 aligned)
│   ├── security-checklist.md         # Mandatory checklist for Phase 0, Phase 5, and changes
│   ├── config-schema.json            # JSON Schema for pipeline/config.json validation
│   ├── ai-safe2-alignment.md         # AI-SAFE2 v2.1 framework alignment (79% all / 87% technical)
│   ├── SDLC-policy.md               # Formal SDLC governance document
│   └── module-anatomy.md             # Standard structure, naming, and patterns for BRIDGE modules
├── ct/methodologies/
│   └── catalog.json                  # 24 dev frameworks with bridge_compatibility scores
└── memory/
    ├── program.md                    # Karpathy Loop: what to evaluate (editable)
    ├── evaluate.ts                   # Post-project correlation script
    └── insights.json                 # Patterns from past projects (auto-updated)

Core Agents (6 persistent agents)

Agent Role Model Key Tools
requirements-translator Extracts structured requirements from unstructured input Sonnet Context7, sequential-thinking, memory
researcher Fetches live docs, evaluates tools, APIs, MCPs Sonnet Context7, Playwright, memory, crawl4ai CLI
solution-architect Designs architecture, specifies agent team Opus Context7, Playwright, Excalidraw, Serena, Greptile, azure/aws-pricing, uml, memory
validator Goal-backward requirements verification Opus Context7, gitguardian, Serena, Greptile, code-review-graph, memory
code-reviewer Code quality, test coverage, documentation Sonnet memory, eslint CLI
adversarial-verifier Tries to BREAK the implementation by executing code independently Opus Bash (curl, servers), Playwright MCP
security-auditor SAST, secrets, dependencies, OWASP Top 10 Opus gitguardian, memory, semgrep CLI

Dynamic Specialist Agents (created per project)

The Architect specifies which specialists are needed. The orchestrator creates them dynamically in Phase 4. Examples:

  • spec-netsuite-integrator, spec-quickbooks-integrator, spec-salesforce-integrator
  • spec-etl-pipeline-engineer, spec-data-warehouse-engineer, spec-fabric-engineer
  • spec-python-backend, spec-typescript-frontend, spec-fullstack-developer
  • spec-azure-deploy, spec-terraform-engineer, spec-ml-engineer

Each specialist includes: task definition, tools, methodology (TDD, security awareness), documentation access chain, completion signal, and quality checklist. Specialists persist between runs and accumulate knowledge via project memory.


Complete Tool Stack

MCP Servers (Model Context Protocol)

MCP Server Purpose Phases Required?
context7 Code library documentation (React, Node, Python packages) 2, 3, 4 Recommended
playwright Browser automation, E2E testing, interactive doc sites 2, 3, 4 Recommended
excalidraw Mermaid to PNG/SVG architecture diagrams with cloud icons 3 Optional
sequential-thinking Structured step-by-step reasoning for Phase 1 1 Optional
uml Formal C4, BPMN, ERD, sequence diagrams 3 Optional
memory Persistent knowledge graph across sessions and agents All Recommended
azure-pricing Real Azure service pricing for cost models 3 Optional
aws-pricing Real AWS service pricing for cost models 3 Optional
gitguardian Secrets detection and credential scanning 5 Recommended
serena LSP code intelligence: find_symbol, replace_symbol_body, rename_symbol 3, 4, 5 Optional
greptile AI semantic code search (requires API key) 3, 5 Optional
deepwiki AI-generated documentation from GitHub repos 2, 3, 4 Optional
code-review-graph Codebase knowledge graph, blast radius, call graph 4, 5 Optional

CLI Tools

CLI Tool Purpose Phases Install
crawl4ai (crwl) Web scraping to clean markdown -- free, no auth 2, 3, 4 pip install -U crawl4ai && crawl4ai-setup
semgrep SAST static analysis (OWASP Top 10, custom rules) 4, 5 pip install semgrep
vitest Fast JS/TS test runner with coverage 4 npm install -D vitest
eslint JavaScript/TypeScript linting and auto-fix 4, 5 npm install -D eslint
lighthouse Performance, accessibility, SEO, best practices audit 4, 5 npm install -g lighthouse
gh GitHub CLI for repos, PRs, issues, releases All brew install gh / winget install GitHub.cli
stryker Mutation testing -- verifies tests catch real bugs 5 Optional
pixelmatch Visual regression via screenshot comparison 4 Optional
pandoc Markdown to Word/PDF document conversion 5 pip install pandoc
pptxgenjs PowerPoint generation from pipeline data 5 npm install -g pptxgenjs
exceljs Excel generation from pipeline data 5 npm install -g exceljs

Claude Code Plugins

Plugin Purpose Phases
superpowers Methodology gateway: TDD, brainstorming, writing-plans, debugging, code review, verification, branch finishing All
pr-review-toolkit 6-pass deep code review (code, tests, silent failures, types, comments, simplification) 5
code-review Auto-post review findings to GitHub PRs (Haiku scoring, Sonnet filtering, 80+ confidence) 5
code-simplifier Post-build code cleanup and clarity improvements 4
frontend-design Production-grade UI design guidance (not generic AI aesthetics) 4
commit-commands Git workflow automation 4
security-guidance Security warnings on file edits (hook) 4, 5
feature-dev Guided feature development with quality gates 4

Trail of Bits Security Skills (32 of 35 active)

Always Active (8 skills -- every run)

Skill Purpose Phase
static-analysis Deep SAST with CodeQL + Semgrep + SARIF integration 5
supply-chain-risk-auditor Audit deps for CVEs, typosquatting, malicious packages 5
entry-point-analyzer Map attack surface -- all APIs, endpoints, user inputs 3
audit-context-building Ultra-granular code analysis: modules, actors, storage, cross-function flows 3, 5
sharp-edges Dangerous API patterns, risky library usage 4
differential-review Compare final code vs original architecture plan 5
insecure-defaults Flag insecure default configurations 3
fp-check Systematic false positive verification for all SAST findings 5

Triggered by Context (9 skills)

Skill Trigger Phase
property-based-testing Critical business logic 4
testing-handbook-skills Critical business logic (fuzzing, sanitizers) 4
spec-to-code-compliance Brownfield projects or final validation 3, 5
variant-analysis Vulnerability found -- search for same pattern everywhere 5
semgrep-rule-creator Vulnerability found -- create project-specific rule 5
semgrep-rule-variant-creator Multi-language project + custom rule created 5
ask-questions-if-underspecified Ambiguous requirements 1
second-opinion External LLM CLI available (Codex, Gemini) 5
agentic-actions-auditor GitHub Actions CI/CD with AI agent steps 5

Domain-Specific (5 skills)

Skill Trigger Phase
building-secure-contracts Blockchain/Web3 -- 20+ weird token patterns, platform-specific vulns 3, 4, 5
constant-time-analysis Cryptographic operations -- timing side-channels 5
zeroize-audit Secrets/keys in memory -- missing zeroization 5
firebase-apk-scanner Android + Firebase -- security misconfigurations 5
seatbelt-sandboxer macOS/iOS -- minimal Seatbelt sandbox profiles 4

Supply Chain & Artifact Security (3 skills)

Skill Trigger Phase
yara-authoring External scripts/tools/packages installed or artifacts scanned 2, 4, 5
burpsuite-project-parser Pentest engagement results available 5
dwarf-expert Compiled binary verification (C/C++/Rust) 5

Development Tooling (6 skills)

Skill Trigger Phase
modern-python Python project -- enforces uv, ruff, ty, pytest 4
devcontainer-setup Reproducible .devcontainer/ for team onboarding 4, delivery
gh-cli GitHub URL access -- enforces authenticated rate limits All
git-cleanup Post-pipeline branch cleanup Post-5
workflow-skill-design Pipeline self-improvement and quality review Meta
skill-improver Quality refinement of dynamically created specialists 4

Not Used (3 skills -- genuinely out of scope)

Skill Reason
let-fate-decide Entertainment (tarot spreads)
culture-index HR/organizational -- outside pipeline scope
debug-buttercup Trail of Bits internal Kubernetes tool

Complete Process Detail

Phase 0: Initialization

0.0  Tool & Resource Discovery (cross-platform)
     ├── Multi-fallback detection chains per tool (binary → module → import)
     ├── Windows: pip user-site, npm globals, Git Bash PATH handled
     ├── macOS/Linux: venv isolation, Homebrew/apt paths handled
     ├── Smart Plugin Check -- compare installed vs recommended
     ├── Supply Chain Pre-Install Gate -- risk-score packages BEFORE installing
     ├── SBOM Generation -- CycloneDX 1.5 inventory of all installed tools
     ├── Security Guardrails Load -- read GPO-like policies from security-guardrails.json
     ├── Hookify Rules Verification -- check if security hooks are installed (WARN if missing)
     ├── Config Schema Validation -- validate config.json against JSON Schema
     └── Auto-install missing CLIs (platform-aware: choco/brew/apt)

0.0b Smart Plugin Check
     └── Report gaps: "Missing: semgrep (CRITICAL), lighthouse (MEDIUM)"

0.1  Collect Input
     ├── Paste text, provide file path, or describe project
     └── Support: meeting transcripts, emails, chats, specs, URLs

0.2  Validate Understanding (MANDATORY before folder creation)
     └── Confirm: client name, project name, problem interpretation

0.3  Create/Reuse Client/Project Folder
     └── clients/{client-slug}/{project-slug}/

0.3b Load Client Knowledge Graph (if returning client)
     └── Technology decisions, constraints, anti-patterns from prior projects

0.4  Initialize Configuration
     ├── Interactive or YOLO mode
     ├── Model profile (quality/balanced/budget)
     ├── Budget cap (optional)
     └── Feature flags for every workflow step

0.5  Discuss Phase (optional -- resolves ambiguities before pipeline starts)
0.6  Initialize Todo List

Phase 0b: Codebase Analysis (conditional -- brownfield projects)

Only if user references existing codebase:
├── Scan project structure and conventions
├── Identify technology stack and patterns
├── Map existing architecture
└── Feed findings into Phase 1 as constraints

Phase 1: Translate

1.1  Spawn Requirements Translator
     ├── BRIDGE B-R-I-D analysis with Fishbone/Ishikawa root cause categorization
     │   (People / Process / Technology / Data / Environment / Measurement)
     └── Produces: 01-technical-definition.md + 01a-bridge-analysis.md

1.2  Ojo Critico Review (if config.critical_review=true)
     └── Skeptical reviewer challenges output → 01c-critical-review.md

1.3  Human Approval Gate
     └── Options: Approve / Modify / Stop and deliver / Reject

Phase 2: Research

2.1  Spawn Technology Researcher
     ├── 6-tier doc access: llms.txt → Context7 → DeepWiki → crawl4ai → Playwright → Context Hub → WebSearch
     ├── Validates D-preliminary from Phase 1 (marks [CONFIRMED], [CORRECTED], [NOT AVAILABLE])
     ├── Force-Field analysis per technology: driving forces vs restraining forces (scored 1-5)
     ├── Security & Taint Assessment:
     │   ├── Classify taint sources by trust level (TRUSTED / SEMI-TRUSTED / UNTRUSTED)
     │   ├── Map critical sinks (SQL, file writes, command exec, HTML render)
     │   └── Tool risk assessment per references/tool-risk-matrix.md
     └── Produces: 02-research-report.md (includes Security & Taint Assessment section)

2.2  Ojo Critico Review
2.3  Human Approval Gate

Phase 3: Architect

3.1  Spawn Solution Architect
     ├── Architecture diagrams (Mermaid + optional Excalidraw PNG/SVG)
     ├── Cloud cost models (azure-pricing, aws-pricing MCPs)
     ├── File manifest for every file to create
     ├── Specialist team specification with dependencies
     ├── Vertical slice decomposition (walking skeleton methodology)
     ├── SCAMPER analysis: Substitute/Combine/Eliminate to prevent over-architecture
     ├── Security Guardrails (Section H): guardrails for HIGH-risk integrations from Phase 2 taint assessment
     ├── Project Quality Hooks (Section I): pre-commit hooks for the project's tech stack
     └── Produces: 03-solution-proposal.md (Sections A-I)

3.2  Ojo Critico Review
3.3  Plan Checker (7 dimensions: req coverage, deps, integration, scope, tests, gaps, BRIDGE)

3.6  Methodology Selection (CT-driven)
     ├── Reads catalog of 24 development frameworks (11 traditional + 13 AI-powered)
     ├── Filters by bridge_compatibility > 0.6
     ├── Six Thinking Hats analysis on top 5 candidates:
     │   White (data) / Red (intuition) / Black (risks) / Yellow (benefits) / Green (creative) / Blue (process)
     ├── Force-Field analysis on top 2: driving forces vs restraining forces (scored 1-5)
     ├── Adjusts Phase 4 config (gate frequency, parallelization, testing rigor)
     └── Produces: 03c-methodology-selection.md

3.7  Human Approval Gate (includes methodology justification)

Phase 4: Build

PRE-PHASE: Skill Invocations (cached, reused across specialists)
├── superpowers:test-driven-development → embed TDD in all prompts
├── sharp-edges (ToB) → dangerous API patterns warning
├── property-based-testing (ToB) → if critical business logic
├── testing-handbook-skills (ToB) → if critical business logic
├── frontend-design → if frontend work
└── building-secure-contracts (ToB) → if blockchain/smart contracts

4.1  Create/Update Specialist Agents
     ├── Read Architect's specialist specifications
     ├── Resolve dependencies:
     │   ├── CLI tools → auto-install via setup script
     │   ├── npm/pip packages → install before spawning
     │   ├── MCP servers → add to agent tools or degrade gracefully
     │   ├── Trail of Bits skills → invoke and embed, or embed from reference docs
     │   └── Custom scripts → create mock servers, data generators, etc.
     ├── Compose agent with workflow pattern (sequential/safety-gate/task-driven/routing)
     ├── Quality check: clear task, all tools, doc access chain, <750 words prompt
     └── Write to .claude/agents/spec-{role}.md

4.2  Human Approval Gate (Team Review)
     └── Review specialist team before building starts

4.3  Execute Build Groups (Vertical Slice Execution)
     For each specialist, slice by slice:

     ┌─────────────────────────────────────────────────┐
     │  BUILD (TDD: red → green → refactor)            │
     │    ↓                                            │
     │  TEST (vitest run / pytest / framework-specific) │
     │    ↓                                            │
     │  HARDEN (2-4 additional tests:                  │
     │    error paths, boundaries, concurrency,        │
     │    invalid input)                               │
     │    ↓                                            │
     │  E2E (if frontend: Playwright smoke test)       │
     │    ↓                                            │
     │  POST-SLICE SECURITY SCAN (MANDATORY):          │
     │    semgrep scan --config auto src/              │
     │    ├── CRITICAL → BLOCK next specialist         │
     │    └── WARNING → log, present at gate           │
     │    ↓                                            │
     │  STRUCTURAL LINTER (orchestrator, no LLM):     │
     │    ├── File manifest compliance                 │
     │    ├── Import direction enforcement             │
     │    ├── Naming convention compliance             │
     │    ├── File size guard (>300 WARN, >500 ERROR)  │
     │    └── Test file presence check                 │
     │    ↓                                            │
     │  VERIFY: all tests pass + acceptance criteria?  │
     │    ├── YES → BRIDGE_SLICE_COMPLETE → next slice │
     │    └── NO → RETRY (max 3, then escalate)       │
     └─────────────────────────────────────────────────┘

     Abductive reasoning: when data is incomplete, list 2-3 hypotheses
     ordered by plausibility, implement most plausible with verification step.

4.4  Human Approval Gate (Per Slice or Per Specialist)
     └── Milestone Delivery: optionally generate client deliverable per execution group

4.5  De-Sloppify Pass + Garbage Collector
     ├── De-Sloppify: dead code, naming, YAGNI, debug statements, unused imports (Haiku)
     └── Garbage Collector (5 checks):
         GC-1: Dead code / orphaned files | GC-2: Pattern consistency
         GC-3: Architecture drift | GC-4: Documentation freshness | GC-5: Duplicate code

4.7  Specialist Archival
     └── Successful specialists archived to .claude/agents/library/ for future reuse
4.6  Update Build Manifest
4.7  Archive Successful Specialists (for future reuse)

Phase 5: Validate and Deliver

PRE-PHASE: Security Skill Invocations
├── superpowers:verification-before-completion
├── static-analysis (ToB) → deep SAST (CodeQL + Semgrep + SARIF)
├── supply-chain-risk-auditor (ToB) → dependency audit
├── differential-review (ToB) → code drift vs architecture plan
├── spec-to-code-compliance (ToB) → evidence-based spec alignment
├── audit-context-building (ToB) → ultra-granular code analysis
├── fp-check (ToB) → false positive verification gate
├── variant-analysis (ToB) → if vulnerability found
├── semgrep-rule-creator (ToB) → if vulnerability found
├── agentic-actions-auditor (ToB) → if GitHub Actions CI/CD
├── zeroize-audit (ToB) → if code handles crypto/secrets
├── constant-time-analysis (ToB) → if timing-sensitive crypto
├── building-secure-contracts (ToB) → if blockchain
├── firebase-apk-scanner (ToB) → if Android+Firebase
└── second-opinion (ToB) → if external LLM CLI available

5.1a Validator Agent (Requirements & Architecture)
     ├── Goal-backward verification from business goal
     ├── Stub detection (empty bodies, TODO, orphaned components)
     ├── Requirements traceability matrix (REQ-XXX → file:line)
     ├── BRIDGE alignment check (R, I, D-validated, G+E)
     ├── Locked constraints verification
     ├── Doc-Architecture sync checks (DOC_DRIFT: doc says X, code does Y)
     └── Produces: 05-validation-report.md (APPROVE/REJECT)

5.1b Code Reviewer Agent
     ├── Clean code: naming, SRP, error handling
     ├── Test quality: meaningful tests, edge cases, no assert(true)
     ├── YAGNI violations
     ├── ESLint scan
     └── Produces: 05a-code-review.md (PASS/FAIL)

5.1c Security Auditor Agent (BLOCKING)
     ├── SAST: semgrep scan --config auto --json
     ├── Secrets: gitguardian MCP + pattern grep (AKIA, sk-, passwords)
     ├── Dependency audit: npm audit / pip-audit
     ├── OWASP Top 10 review on all endpoints
     ├── Insecure defaults check
     └── Produces: 05c-security-audit.md (SECURE/BLOCKED)

5.1d Multi-Pass Code Review (pr-review-toolkit)
     ├── Pass 1: Code reviewer -- guidelines, bugs
     ├── Pass 2: Test analyzer -- coverage quality
     ├── Pass 3: Silent failure hunter -- empty catches, missing logs
     ├── Pass 4: Type design analyzer -- encapsulation
     ├── Pass 5: Comment analyzer -- accuracy
     ├── Pass 6: Code simplifier -- simplification opportunities
     └── Produces: 05b-pr-review.md

5.1e Optional: Mutation Testing (stryker)
     └── Score >80% strong, 60-80% warning, <60% critical

5.2  Validator Consensus Protocol
     ├── Cross-check verdicts from all validation agents
     ├── Security gate overrides all other verdicts (BLOCKED = BLOCKED)
     ├── Adversarial FAIL vs Validator APPROVE → conflict escalated to human
     └── Cross-LLM contradictions require explicit human resolution

5.2b Quality Score Calculation
     ├── requirements_coverage × 0.35
     ├── test_pass_rate × 0.25
     ├── security_score × 0.20
     ├── code_quality × 0.10
     └── documentation × 0.10
     Threshold: >=0.80 APPROVE, 0.60-0.79 CONDITIONAL, <0.60 REJECT

5.3  Rejection Loop (max 2 cycles)
     ├── Route issues to responsible agent with targeted feedback
     ├── Auto-fix or manual instructions
     └── Same issue 2+ times → escalate immediately

5.4  Human Approval Gate (Final)

5.5  Generate Deliverables
     ├── Internal: pipeline/ (full details)
     └── Client: deliverables/ (sanitized -- no AI/agent references)

5.5b Decision Logging & Self-Improvement Evaluation
     ├── Log all key decisions to pipeline/ct-decisions.json:
     │   phase, agent, CT framework used, confidence, human override status
     └── Karpathy Loop: evaluate.ts correlates decisions with quality outcomes
         └── Patterns from 3+ projects → insights.json (feeds future methodology selection)

5.6  Cross-Run Lesson Capture (failures requiring 2+ attempts → lessons)
5.7  Update Client Knowledge Graph
5.7b Final Integration Checklist (superpowers:finishing-a-development-branch)
5.8  Final Summary + Cost Report

Security Gate (BLOCKING by default)

config.security_gate = "blocking" (default):
  ANY CRITICAL finding → BLOCKED (stronger than REJECT)
  Options:
    a) Auto-fix -- re-spawn responsible specialist with security feedback
    b) Manual fix -- user provides guidance
    c) Accept risk -- user must type "I accept the risk for: {finding}" (per finding)
    d) Abort delivery

config.security_gate = "advisory":
  Findings logged but do not block. User was warned.

Agent-to-Tool Matrix

Agent Base Tools MCP Tools CLI Tools Model
Requirements Translator Read, Write, Glob, Grep, Bash, WebSearch, WebFetch Context7, sequential-thinking, memory -- Sonnet
Technology Researcher Read, Write, Glob, Grep, Bash, WebSearch, WebFetch Context7, Playwright (5 tools), memory crawl4ai Sonnet
Solution Architect Read, Write, Glob, Grep, Bash, WebSearch, WebFetch Context7, Playwright (2), Excalidraw (4), Serena, Greptile, azure-pricing, aws-pricing, uml, memory crawl4ai Opus
Code Specialists Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch Context7, Serena, code-review-graph, memory vitest, eslint Sonnet
Python Specialists Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch Context7, Serena, code-review-graph, memory uv, ruff, ty, pytest Sonnet
Frontend Specialists Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch Playwright (5), Serena, code-review-graph, memory vitest, eslint, lighthouse Sonnet
Blockchain Specialists Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch Context7, Serena, memory hardhat/foundry/anchor Sonnet
Infrastructure Specialists Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch Context7, memory terraform/kubectl/docker/az/aws Sonnet
Validator Read, Write, Glob, Grep, Bash, WebSearch, WebFetch Context7, gitguardian, Serena, Greptile, code-review-graph, memory semgrep, lighthouse Opus
Code Reviewer Read, Write, Glob, Grep, Bash, WebSearch, WebFetch memory eslint Sonnet
Security Auditor Read, Write, Glob, Grep, Bash, WebSearch, WebFetch gitguardian, memory semgrep Opus
De-Sloppify Read, Write, Edit, Glob, Grep, Bash -- eslint Haiku

Dynamic Dependency Resolution

When a specialist needs tools not in the base matrix:

Dependency Type Resolution Blocking?
CLI tools Auto-install via setup script (scripts/setup-{role}.sh) Yes
npm packages npm install {package} (project-local) Yes
pip packages pip install {package} (or uv pip install) Yes
MCP servers Add to agent tools if available; degrade gracefully if not No
Trail of Bits skills Invoke if installed; embed from reference docs if not No
Helper scripts Orchestrator creates before spawn; agent can create more at runtime Yes

Documentation Access Chain (6 tiers)

llms.txt quick check (try FIRST)
  ↓ not found
Tier 1: Context7 MCP          → Code libraries (React, Node, Python packages)
  ↓ not a code library
Tier 2: DeepWiki MCP           → GitHub repo documentation (optional plugin)
  ↓ not a GitHub repo or not installed
Tier 3: crawl4ai CLI           → ANY online docs (NetSuite, Azure, Salesforce, SAP) -- free
  ↓ can't render page
Tier 4: Playwright MCP         → Interactive/JS-heavy/auth-gated sites
  ↓ no browser needed
Tier 5: Context Hub CLI        → Curated API docs (Stripe, Twilio, AWS, 68+ APIs)
  ↓ all else fails
Tier 6: WebSearch + WebFetch   → Fallback

What Gets Delivered

clients/{client}/{project}/
├── README.md                          # Project metadata and progress
├── input/
│   └── original-input.md             # Original requirements (preserved)
├── pipeline/                          # Internal artifacts (your team only)
│   ├── config.json                    # Pipeline configuration
│   ├── state.json                     # Pipeline state for resumability
│   ├── cost-log.json                  # Token/cost estimation per agent
│   ├── 00-constraints.md             # Locked decisions from discuss phase
│   ├── 01-technical-definition.md    # Requirements translation
│   ├── 01a-bridge-analysis.md        # Full BRIDGE B-R-I-D-G-E analysis
│   ├── 01c-critical-review.md        # Ojo Critico review (Phase 1)
│   ├── 02-research-report.md         # Technology research
│   ├── 02c-critical-review.md        # Ojo Critico review (Phase 2)
│   ├── 03-solution-proposal.md       # Architecture design
│   ├── 03b-plan-check.md            # Pre-build plan validation
│   ├── 03c-critical-review.md        # Ojo Critico review (Phase 3)
│   ├── 04-build-manifest.md          # Build status per specialist/slice
│   ├── 05-validation-report.md       # Validator assessment
│   ├── 05a-code-review.md           # Code review results
│   ├── 05b-pr-review.md             # 6-pass PR review results
│   ├── 05c-security-audit.md        # Security audit (BLOCKING)
│   ├── quality-score.json            # Composite quality score
│   ├── 03c-methodology-selection.md  # CT-selected development methodology
│   ├── ct-decisions.json             # Decision audit trail for self-improvement
│   ├── feedback-routing.json         # Issue routing for fix cycles
│   ├── improvements.tsv              # Fix attempt tracking
│   ├── security-events.json          # JSONL security event log (hash chain integrity)
│   ├── sbom.json                     # CycloneDX 1.5 Software Bill of Materials
│   ├── risk-acceptances.json         # Explicitly accepted security risks
│   ├── error-log.md                  # Pipeline error history
│   ├── internal-summary.md           # Final summary with cost report
│   └── lessons/                       # Cross-run learnings
├── src/                               # Built solution code
├── tests/                             # Test suites
├── scripts/                           # Auto-generated setup/mock scripts
├── deliverables/                      # Client-facing documents (sanitized)
│   ├── README.md                     # Table of contents
│   ├── solution-proposal.md          # Architecture + recommendations
│   ├── client-report.md              # Full technical report
│   ├── architecture-diagrams.md      # Mermaid diagrams + descriptions
│   ├── deployment-guide.md           # Step-by-step deployment
│   ├── api-reference.md              # API documentation (if applicable)
│   ├── {project}-report.html         # Interactive HTML report
│   └── images/                        # Diagram exports (PNG/SVG)
└── brand-assets/                      # Your brand guidelines
    ├── brand-config.json             # Colors, fonts, logo
    └── templates/                     # PPTX, DOCX, CSS templates

Client deliverables are fully sanitized -- no agent, pipeline, or AI system references. The deliverables/ folder is independently shareable.


Critical Thinking Integration

CT frameworks are applied as reference knowledge at key decision points -- not as separate scripts or modules, but embedded in agent prompts where they add the most value.

Phase CT Framework What It Does
1 -- Translate Fishbone/Ishikawa Categorizes root causes into 6 dimensions: People, Process, Technology, Data, Environment, Measurement
2 -- Research Force-Field Analysis Scores driving forces vs restraining forces (1-5) for each technology recommendation
3 -- Architect SCAMPER Substitute, Combine, Eliminate checks prevent over-architecture before finalizing design
3c -- Methodology Six Thinking Hats 6 perspectives (data, intuition, risks, benefits, creative, process) evaluate 24 development methodologies
3c -- Methodology Force-Field Final scoring of top 2 methodology candidates with driving vs restraining forces
4 -- Build Abductive Reasoning When data is incomplete, formulate 2-3 hypotheses ordered by plausibility, implement most plausible with verification
1-3 Gates Ojo Critico Skeptical review combining Paul-Elder intellectual standards with Watson-Glaser evaluation
5 -- Validate Goal-Backward Dialectical: what conditions must be TRUE vs what IS true in the code

Self-Improvement (Karpathy Loop)

After each completed project, the pipeline evaluates its own decisions:

  • Logs key decisions to pipeline/ct-decisions.json with phase, CT framework, confidence, and human override status
  • evaluate.ts correlates CT decisions with quality outcomes
  • Patterns confirmed across 3+ projects become insights in memory/insights.json
  • Future projects use insights to select better methodologies and calibrate confidence
  • program.md defines what to evaluate -- editable, so you control the direction of improvement

24 Development Methodology Catalog

The methodology selector chooses from 24 frameworks based on project characteristics:

Traditional (11): Agile, Waterfall, Scrum, Kanban, DevOps, Lean, XP, RAD, FDD, Spiral, Hybrid

AI-Powered (13): Agentic AI-Driven, Hybrid VSM, Context-First, Platform Engineering, AgentSecOps, CTEM, Data-Centric AI, Human-in-the-Loop, Predictive Sprint Planning, Shift Intelligence Left, Automated QC, AI Governance at IDE, Upskilling

Each framework has a bridge_compatibility score (0.0-1.0), best_for project types, and config_adjustments that change how Phase 4 operates (gate frequency, parallelization, testing rigor).


Harness Engineering

Two independent hook systems protect code quality and pipeline integrity.

Project Pre-Commit Hooks

The Architect generates hooks appropriate for the project's tech stack (Section I of the Solution Proposal). Walking Skeleton installs them in Slice 1.

Hook Node.js/TypeScript Python
Lint staged eslint ruff
Type check tsc --noEmit ty / pyright
Test affected vitest related pytest -x
Secret scan grep AKIA, sk-, password= grep AKIA, sk-, password=
File size >500 lines = WARN >500 lines = WARN

Three modes: Off (default) | Warn (detect, never block) | Enforce (block on violations)

Hookify Enforcement (always-on, mechanical)

9 hookify files provide regex-based enforcement that runs automatically on every tool call. Two scopes:

Global hooks (~/.claude/) - active in ALL Claude Code sessions:

Hook Event Detects
bridge-destructive-commands bash rm -rf, force push, DROP TABLE, kubectl delete + composite command chains
bridge-scope-escape file write Writes to .claude/settings, .git/objects, node_modules
bridge-secrets-detection file write AWS keys (AKIA), API tokens (sk-), private keys, hardcoded passwords
bridge-em-dash-titlecase file write Em dashes and 3+ consecutive Title Case words in headings

Project hooks (.claude/ in workspace) - active only during BRIDGE:

Hook Event Detects
bridge-zero-assumptions file write Hedging language: "probably", "I assume", "likely", "I think this is"
bridge-completion-check stop Verification checklist before agent stops
bridge-client-install-guard bash npm/pip/yarn install targeting clients/ folders
bridge-node-path file write .js files using global npm packages without NODE_PATH preamble
bridge-taint-cleanup file write [EXTERNAL-UNVERIFIED] tags in deliverables/

Layered enforcement model:

  • Hookify = always-on baseline (warn mode, never blocks)
  • Harness hooks OFF = only hookify coverage
  • Harness hooks WARN = hookify + settings.json warnings
  • Harness hooks ENFORCE = hookify + settings.json BLOCKING

Pipeline Protection Hooks (Claude Code settings.json)

Optional additional hooks with configurable enforcement (off/warn/enforce):

Hook Detects
Destructive Command Guard rm -rf, git push --force, DROP TABLE, kubectl delete
Secrets in Output Guard AWS keys, API tokens, passwords in Write/Edit
Scope Escape Guard File writes outside project path
Composite Action Guard Chained commands (&&, ;) where any part is destructive
Written-File-Execution Guard Files written then executed - content checked for destructive patterns

Taint Tracking & Tool Risk Matrix

External data sources classified by trust level:

  • TRUSTED: Internal, controlled sources
  • SEMI-TRUSTED: Partner APIs with SLA
  • UNTRUSTED: User input, public web, file uploads

Critical sinks mapped (SQL, file writes, command exec). HIGH-risk integrations get architectural guardrails in Phase 3.


Resilience Features

Feature How It Works
Pipeline State File state.json tracks completed phases, specialist status, last checkpoint. Resume picks up exactly where you left off.
Git Rollback Git tags after each phase approval. "Go back to Phase 2" restores pipeline state via git checkout.
Stall Detection Missing BRIDGE_SLICE_COMPLETE signal, error keywords, or timeout → auto-escalate. Walking skeleton failures escalate immediately.
Rejection Loops Max 3 retries per slice, max 2 validation cycles. Same issue appearing 2+ times escalates immediately.
Context Budget 10 rules: file-bridge (never accumulate), phase refresh (re-read core.md at key points), prompt size guard (<750 words), emergency recovery, rejection loop memory.
Cost Tracking Per-agent token/cost estimation (chars/4). Optional budget cap with 80% warning and 100% pause.
Flexible Execution Run phases in any order, skip ahead, run in parallel with reconciliation.
Cross-Run Lessons Failures requiring 2+ attempts generate lessons loaded automatically in future runs.
Client Knowledge Graph Per-client isolation. Technology decisions, constraints, anti-patterns persist across projects.
Ojo Critico Skeptical reviewer after Phases 1-3 catches issues before expensive build work. Default: REJECT.
Adversarial Verifier Independent agent that EXECUTES code and tries to break it -- boundary values, idempotency, type confusion. Anti-rationalization guards prevent "the code looks correct" shortcuts.
Anti-Rationalization Tables Each phase includes 8-10 rationalizations agents use to skip steps, paired with factual rebuttals. Defensive prompting based on observed failure modes. 45 entries across 5 phases prevent shortcuts before execution starts.
Red Flag Indicators 52 observable deviations the orchestrator checks before approval gates. Catches problems during execution, not after delivery. Complements Anti-Praise Guard (Phase 5), Ojo Critico (Phases 1-3), and hookify (mechanical). Four layers: prevention, detection, critical review, mechanical enforcement.
Module Anatomy Reference references/module-anatomy.md documents standard module structure, naming conventions, size guidelines, loading patterns, and anti-patterns. Facilitates consistent module authoring.
Dream Consolidation /bridge dream {client} -- consolidates, reconciles, and prunes a client's knowledge graph across projects. Detects contradictions, archives stale decisions.
Proposal Fast Track Deliverable-only projects (proposals, decks, assessments) get a collapsed 3-phase pipeline: Understand > Generate Assets > Assemble. 4-5 agents instead of 12+, 30-45 minutes instead of 2-3 hours. Design Director agent with professional visual standards.
Visual-First Presentations Enforced deck design rules: 7 slides max, visual-first (every slide leads with imagery), stat cards over bullet lists, cascading timelines, PresentationGO searched by exact diagram type, editable architecture shapes in appendix, no em dashes, sentence case.
Image Selection Protocol Cover images generated through competitive comparison: one Remotion candidate vs. up to 5 stock photos, scored on industry relevance/quality/brand fit, best wins. Under 5 minutes.
Smart Deliverable Folders Typed subfolder structure under deliverables/ (proposals, reports, code, data, images, scripts) instead of flat file dumps.
Coordinated PPTX Engine python-pptx as master builder, pptxgenjs for editable shapes, PresentationGO for layout reference, Remotion for visuals. Brand templates and PresentationGO slides combined with brand colors. 6-level fallback chain.
Design Enforcement Hooks Deterministic shell hooks enforce design rules: no em dashes, no local installs in clients/, NODE_PATH preamble required, PresentationGO generic search blocked. Mechanical rules get hooks, subjective rules get agent prompts + human gates.
Cross-Platform Tool Detection Fallback chain per tool: CLI binary > package manager > Python import > npm global. Caches NPM_GLOBAL_PATH for Windows compatibility. Never re-installs tools already present.
No Local Installations Enforced guardrail: never install packages inside client folders. All tools are global. Temp projects go in system temp directory.
Rejection Loop Memory Re-run agents receive explicit feedback on WHY the previous attempt was rejected, preventing repeated mistakes.
Standardized Return Contract All validation agents report in a 5-field format (Scope, Findings, Fixes, Validated vs. Unverified, Verdict) for consistent orchestrator parsing.
Analysis Paralysis Guard 5+ consecutive reads without writing → must explain or report BLOCKED.
Deviation Rules Auto-fix bugs/safety; escalate architecture changes; skip scope creep.
Self-Test bridge self-test validates all referenced files, templates, agents, and docs exist.

Security Governance

BRIDGE implements defense-in-depth security aligned with the AI-SAFE2 v2.1 framework -- an open-source governance standard for autonomous AI systems. Coverage: 79% of applicable controls (87% of technical controls, excluding organizational training programs).

Supply Chain Security

Control What It Does
Pre-install scanning Every pip/npm package is risk-scored BEFORE installation. Score >= 10 = BLOCKED.
Version pinning bridge-tool-versions.json pins exact versions for all 20+ tools. No floating ranges.
SBOM generation CycloneDX 1.5 Software Bill of Materials at Phase 0 and Phase 4. Every dependency tracked.
Supply chain gate Checks PyPI/npm metadata for typosquatting, low adoption, empty descriptions, known CVEs.

Runtime Protection

Control What It Does
Taint tracking All external content marked [EXTERNAL-UNVERIFIED]. Never written to config files or executed.
Config protection Hookify rule BLOCKS agent writes to .claude/settings and hooks (prevents AI poisoning of core.md).
Scope escape guard Detects writes outside project directory.
Secrets detection Regex patterns for AWS keys, API tokens, private keys, passwords. Blocks before commit.
Client data isolation Per-client directories with cross-contamination warnings. Knowledge graphs strictly isolated.

Audit Trail

Artifact Format Purpose
security-events.json JSONL (append-only, hash chain) Every security event with SHA-256 integrity chain
approval-log.json JSON Every human decision at every gate
sbom.json CycloneDX 1.5 Complete dependency inventory
risk-acceptances.json JSON Explicitly accepted risks with user statements

Security Gate

The security gate defaults to BLOCKING -- any critical finding prevents delivery. The human must explicitly accept each risk with a documented reason. Risk acceptance is intentionally friction-heavy.

Compliance Alignment

Framework Coverage
AI-SAFE2 v2.1 79% applicable / 87% technical controls
OWASP Top 10 (2021) Audited, all findings addressed
OWASP ASVS v4.0 Audited, findings addressed
OWASP WSTG v4.2 Audited, findings addressed

See references/ai-safe2-alignment.md for the full 72-control tactic-by-tactic assessment and references/SDLC-policy.md for the formal SDLC governance document.


Quick Start

Prerequisites

Installation

git clone https://github.com/josembuitron/bridge-agentic-pipeline.git
cd bridge-agentic-pipeline && bash install.sh

Restart Claude Code, then type: /bridge

To update:

cd bridge-agentic-pipeline && git pull && bash install.sh

To uninstall:

rm -rf ~/.claude/skills/bridge

Alternative: Plugin Install (for teams)

If your team needs managed distribution with namespaced commands:

# Inside Claude Code:
/plugin marketplace add josembuitron/bridge-agentic-pipeline
/plugin install bridge-dev-pipeline@josembuitron-bridge-agentic-pipeline
/reload-plugins

Plugin invocation: /bridge-dev-pipeline:bridge (namespaced per Claude Code convention).

Alternative: Test Without Installing

claude --plugin-dir ./bridge-agentic-pipeline

First Run

When you invoke /bridge, the pipeline will:

  1. Discover tools -- cross-platform detection with fallback chains (binary → module → import), auto-install via platform package manager (never blocks on optional tools)
  2. Collect input -- paste text, provide file paths, or describe your project
  3. Confirm understanding -- validate client name, project name, and problem interpretation before creating folders
  4. Run phases -- each phase produces artifacts and waits for your approval

BRIDGE auto-installs missing CLI tools (crawl4ai, semgrep) on first run. See SETUP.md for optional MCP servers and plugins that enhance the pipeline.

Other Commands

Command What It Does
/bridge Start a new project or continue an existing one
/bridge help Show setup and configuration guide
/bridge list List all projects with status
/bridge dream {client} Consolidate a client's knowledge graph across projects
/bridge dream all-tooling Consolidate global tooling patterns (non-client data)

Configuration

The pipeline creates pipeline/config.json in each project folder:

{
  "mode": "interactive",
  "granularity": "standard",
  "model_profile": "balanced",
  "workflow": {
    "discuss_phase": false,
    "critical_review": true,
    "plan_checker": true,
    "de_sloppify": true,
    "mutation_testing": false,
    "visual_regression": false,
    "auto_advance": false
  },
  "parallelization": { "enabled": true, "max_concurrent_specialists": 3 },
  "security_gate": "blocking",
  "budget_cap_usd": null,
  "gates": {
    "after_translate": true,
    "after_research": true,
    "after_architecture": true,
    "per_slice": true,
    "after_validation": true
  }
}
Setting Options Effect
mode interactive / yolo Approval gates at every phase vs. auto-advance
granularity coarse / standard / fine Fewer slices (faster) vs. more slices (thorough)
model_profile quality / balanced / budget Opus everywhere vs. mixed routing vs. Sonnet everywhere
security_gate blocking / advisory Critical findings block delivery vs. log only
budget_cap_usd number / null Cost cap with 80% warning and 100% pause

Project Type Presets

Preset Granularity Key Flags
api-integration standard plan-checker ON, de-sloppify ON
data-pipeline standard plan-checker ON, security-gate blocking
dashboard coarse plan-checker OFF (simpler scope)
enterprise-feature fine discuss phase ON, all gates ON
mvp-rapid coarse plan-checker OFF, de-sloppify OFF, per-slice gates OFF

Brand Assets

Place your brand guidelines in brand-assets/ to customize all deliverable output:

{
  "company": "Your Company",
  "colors": { "primary": "#003366", "secondary": "#0066CC", "accent": "#FF6600" },
  "fonts": { "heading": "Georgia, serif", "body": "Calibri, sans-serif" },
  "logo_path": "logo.png"
}

Quality Scoring

Phase 5 computes a composite quality score for every project:

quality_score = (requirements_coverage * 0.35)
              + (test_pass_rate         * 0.25)
              + (security_score         * 0.20)
              + (code_quality           * 0.10)
              + (documentation          * 0.10)
Component Weight Measurement
Requirements Coverage 35% REQs addressed / total REQs in Technical Definition
Test Pass Rate 25% Tests passing / total tests (vitest run --reporter=json)
Security Score 20% Semgrep SAST findings + GitGuardian secrets scan
Code Quality 10% ESLint errors, complexity, duplication checks
Documentation 10% Documented APIs / total APIs + README completeness

Decision thresholds:

  • >= 0.80 -- APPROVE
  • 0.60 - 0.79 -- CONDITIONAL APPROVE (specific improvements listed)
  • < 0.60 -- REJECT (routed back to responsible agent with targeted feedback)

Repository Layout

bridge-agentic-pipeline/
├── .claude/agents/      # Core agent definitions (6 agents)
├── .claude/commands/    # Slash command entry point (/bridge)
├── .claude-plugin/      # Plugin metadata (for marketplace distribution)
├── agents/              # Agent definitions (plugin distribution copy)
├── skills/bridge/       # Pipeline orchestrator (SKILL.md + modular phases/modules)
├── templates/           # Output format templates (5 templates)
├── docs/                # Domain knowledge (3 docs) + CLI reference (6 docs)
├── CLAUDE.md            # Project-level Claude Code instructions
├── SETUP.md             # Detailed setup, installation, and usage tips
├── DISCLAIMER.md        # Legal disclaimer for pipeline-generated outputs
├── LICENSE              # MIT License
└── README.md            # This file

License

MIT License.


Support

If BRIDGE saves you time or helps you deliver better work, consider supporting the project. Your contribution keeps this pipeline open-source and actively maintained.


Repository: github.com/josembuitron/bridge-agentic-pipeline

Questions or feedback? Open an issue on GitHub.

About

BRIDGE Development Pipeline. From business requirements to delivered solutions using AI agents. Built for development agencies, consultancies, and engineering teams.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors