Home How It Works Skills Skills Catalog Capabilities Roadmap Changelog FAQ Pricing
The Protocol

How Scaffold OS
builds software

Not a chatbot. Not a copilot. A coordination engine that defines exactly what every AI agent must do, check, and verify before each session even begins.

v5.5-v5.6 Operating Layer

Preflight now covers context, delivery intent, and release readiness.

Scaffold now operates with six flows, not five. What changed across v5.5 and v5.6 is the layer before and around those flows: the engine can begin from trusted context, coordinate deployment handoff, enter recommendation mode for the next version, surface richer product signals, and verify that the environment is truly ready before serious project work starts.

Context Intake
Uploads, URLs, repo snapshots, and prior bundles become durable inputs
Parser choices become easier to track and trust
Later sessions begin from a compact ready-context summary
Delivery Intelligence
Deploy handoff becomes a clearer engine-to-surface contract
Rollback paths are planned more explicitly before release follow-through
Recommendation mode can route a live product into next-version planning
Release Readiness
Secret checks become explicit workflow state
Environment doctor checks reduce setup drift before real work
Unsafe release attempts can pause instead of slipping through silently
Evolution of the Protocol

How the protocol evolved.
From build coordinator to protocol platform.

Version 2.2 was already beyond simple prompting, but the protocol still had visible ceilings. v3 solved multi-system coordination and first-class build targets. The later v3 releases added ML workflow coverage, better project structure, tiered audits, UI-readable state, and safer upgrades. v4 then made planning sharper, specialization real, and product-surface compatibility much cleaner. v4.8 through v5.0 turned the engine into a reusable protocol platform, and v5.1 through v5.6 made it more queryable, delivery-aware, recommendation-capable, and operationally ready for real project work.

v2.2 — What it couldn't do
User chose the wrong file to start

Multiple entry point files caused friction. Users didn't know which to use. The wrong choice wasted the session.

External platforms were APIs only

Platforms like n8n and WordPress were integration targets. Agents called them, but never built on them as first-class surfaces.

Manual contract changes without protocol

API contract changes required manual user coordination. No automatic dependency scanner. No enforcement window.

Single backend assumed

The system assumed one backend. Multi-system architectures (microservices, external platforms, ML pipelines) had to be hacked in.

No ML/AI pipeline support

ML projects were treated as regular code. No training pipeline types, no evaluation gates, no model drift tracking.

v3 — What we fixed
Single entry point with intelligent routing

One file. The system reads your project state and auto-routes to the correct flow. Zero user decision required. Zero wasted sessions.

External platforms become BUILD TARGETS

Salesforce, Snowflake, n8n, WordPress — the agent builds artifacts directly on these platforms and drift-tracks them identically to code.

8-step Contract Change Protocol

Any contract change triggers auto-dependency scan, alert propagation, 24-hour ACK window, and build block until all dependents are updated.

True multi-system orchestration

Coordinate backend, frontend, cloud infrastructure, external SaaS platforms, ML pipelines — all under one protocol, one health score, one drift system.

First-class ML/AI pipeline support

Auto-detected from architecture. Training + evaluation as first-class step types. Model drift tracked as a structural contract alongside code drift.

6 Auto-Detected Operating Flows
Flow 1
Fresh Start

No project detected → full brainstorm + architecture + planning + audit + build protocol activated.

Flow 2
Continue Build

Architecture + plan exist → resume briefing, drift check, next build steps executed.

Flow 3
Architecture Only

Architecture exists, no plan → validate architecture, generate phased build plan, enter audit.

Flow 4
Archaeology Mode

Code exists, no plan → read codebase, reconstruct full planning document set by inference.

Flow 5
Continuous Mode

Live product detected → start from drift, health, backlog, and current operating condition before new work begins.

Flow 6
Recommendation Mode

Live product asks what comes next → study shipped work, backlog, health, and priorities to recommend the next version.

v3.2 → v4.0 — The Maturity Bridge
v3.2 — ML / AI Workflows
Training and evaluation stop being treated like normal code steps
Model operations become part of the protocol surface
AI-heavy products get first-class execution paths instead of hacks
RAG, model drift, and pipeline thinking start becoming native concepts
v3.3 → v3.4 — Structure and Audit Discipline
Project organization becomes easier to navigate over long sessions
Review stops being one-size-fits-all and becomes tiered
Light, medium, and deep audit intensity become deliberate choices
The engine gets more maintainable for both humans and wrappers
v3.5 → v3.6 — State and Upgrade Safety
Structured state becomes easier for products and dashboards to read
Clarification and architecture handling get more UI-friendly
Versioning and migration infrastructure make upgrades safer
Long-running projects stop paying such a heavy cost for engine evolution
v4.0 — Product-Surface Readiness
The engine starts behaving better as a platform boundary, not just a local protocol
Compatibility surfaces can read progress and state more safely
Standalone usage stays intact while product integration gets cleaner
This becomes the bridge into the later platform and wrapper era
v4.1 → v4.6 — What Changed in the Skills Era
v4.1
Fixed list of specialist roles — always the same set
Advisory-only quality reporting
No project type awareness
All roles load for every project
v4.6
36 curated skills activate only when the project actually needs them
1,300+ extended skills become available as an on-demand catalog instead of dead weight
11 project profiles make specialization start at brainstorm, not halfway through the build
PASS / ADVISORY / FAIL enforcement turns quality from advice into an actual gate
Lazy invocation keeps builds lean while still unlocking much deeper expertise when needed
v4.8 → v5.0 — The Platformization Shift
v4.8 — Update Cycle
Post-build work gets a named, structured cycle
Each round: request → plan → build → audit → result
Active workspace + formal archive per cycle
No more rolling plan file as the only record
v4.9 — Canonical Workflow State
Engine owns workflow state — canonical and machine-readable
Session handoff state becomes consistent at every boundary
Decision tracking, review readiness, deploy readiness
Generic wrapper execution contract for any surface
v5.0 — Protocol Platform Foundation
Target registry + integration registry
Multi-track work queue management
Lifecycle event history — wrapper-safe, append-only
Any product surface can sit on top of the same engine protocol
v5.1 → v5.6 — The Operational Intelligence Era
v5.1 — Machine-Readable Completeness
Drift, audits, setup state, backlog, and update condition become queryable
Wrappers stop depending on markdown parsing for critical signals
Session handoff and freshness become easier to read directly
Dashboards and automation surfaces get a cleaner state contract
The engine becomes much easier to interrogate, not just execute
v5.2 — Protocol Completeness
Brainstorm state becomes resumable and wrapper-visible
Pause reasons and resume semantics become engine-native
Phase skips and lifecycle transitions stop being informal behavior
Deploy handoff becomes more structured per target and surface
Wrapper capability contracts become clearer and less ad hoc
v5.3 — Git Strategy & Intelligence
Git strategy makes branching model drive delivery behavior
Project health score becomes a trusted engine-computed signal
Complexity signal estimates project weight before heavy execution starts
Fast session warmup reduces restart waste between sessions
Existing repos get stronger archaeology and continuity handling
Wrappers get more deterministic follow-up and delivery signals
v5.4 — Context Brain & Release Safety
Trusted context summaries let the next session start warm
Spreadsheet intelligence makes workbook structure shape planning
Repo intelligence exposes hotspots, ownership clues, and dependency risk
Archaeology decisions become easier for products to render honestly
Release safety adds explicit pre-push and pre-publish protection
Context handling becomes more engine-owned and less wrapper-invented
v5.5 — Deploy, Recommend, Modularize
Deploy handoff becomes a clearer protocol-owned contract
Rollback handling becomes sharper and more deliberate
A sixth recommendation flow can propose the next version
Feature, coverage, and follow-up signals become easier for surfaces to expose
The generation engine becomes more phase-focused and context-efficient
v5.6 — Readiness Hardening
Environment doctor checks catch setup drift before serious work begins
Release validation becomes stricter about drift and coverage
Golden-path fixtures make the core starts easier to trust
The shortest successful install path gets cleaner for fresh teams
Real-project readiness becomes a clearer go or no-go state
The Full Build Protocol

Every project. Every time.

Each phase is a defined protocol — not a suggestion. An agent executing Scaffold OS cannot skip a phase, cannot reorder steps, and cannot proceed without confirmed state.

0
Intelligent Entry — State Detection
Automatic

Every session starts here. The agent reads the project folder, detects the current state, and routes to the correct flow — automatically. No user decision required. No ambiguity about where to start.

If: No project exists
→ Route to brainstorm intelligence protocol
If: Architecture exists, no build plan
→ Validate architecture, generate phased build plan
If: Build interrupted mid-session
Read the recovery and context summary first, then resume from the next incomplete step with current readiness signals.
If: Code exists, no plan (existing repo)
→ Standalone archaeology session — full analysis before any building begins
If: Live product asks what's next
→ Route to recommendation mode and rank the next version options
1
Structured Brainstorm Intelligence
Foundation Phase

The system runs a structured cognitive protocol to extract architecture. It infers from everything in the folder first — existing docs, vocabulary, tech signals — before asking a single question. It detects your technical level and adapts the entire interview on the fly.

Before touching architecture, the system validates real demand: it asks who actually needs this product, what they do today without it, and what the minimum useful version looks like. Then an adversarial review argues against the plan before it's locked — surfacing the strongest objections while you can still change course, not after weeks of building.

Output: a complete production architecture with confirmed tech stack, full system topology, every external service with access method and agent role, security model, error and failure path map, observability plan, environment variables, and deployment target. Zero TBDs. Zero vague notes. Any gap found = critical finding.

Planning Mode: You choose how the engine approaches your plan — expand and surface what's missing, hold scope and make it bulletproof exactly as specified, or cut ruthlessly to the narrowest useful version. Same project, different modes for different moments.

2
Multi-Dimensional Pre-Build Audit
Quality Gate

Before a single line of code is written, the system runs two rounds of consistency checks across 13 audit dimensions. These range from schema-interface alignment and endpoint-feature mapping to deployment pipeline consistency and error code registry coherence.

Critical findings are fixed and re-audited. The build only begins when audit verdict is CONSISTENT. Issues found here take minutes to fix. The same issues found in production take days.

Round 1 — 7 Dimensions
1.Internal consistency — no contradictions across architecture files
2.Completeness — zero TBD, UNKNOWN, or placeholder values
3.Verification command quality — no "echo done"-style commands
4.Security model review — auth requirements on every endpoint
5.Build sequence validation — dependency ordering correctness
6.Setup executability — every setup step is achievable by a human or agent
7.Feature spec quality — acceptance criteria are specific and testable
Round 2 — 6 Focus Areas
A.End-to-end data flow trace — data travels correctly from ingestion to output
B.Round 1 fix verification — all flagged issues confirmed resolved
C.Spec/dependencies cross-validation — features reference what they actually use
D.Agent prompt dry-run — agent could execute each step without ambiguity
E.Security end-to-end review — auth flows coherent across all layers
F.Machine column audit — each automated setup step is actually automatable
3
Contracted Multi-System Build
Build Phase

The build executes against a phased plan where every step is a clearly defined unit of work with explicit entry conditions, success criteria, and completion signals. Agents don't interpret the plan — they execute it.

Shared interfaces between systems are governed by contracts. If a contract changes — an API endpoint is renamed, a database schema shifts — an 8-step protocol fires: alert published, dependency scan across every system, 24-hour acknowledgment window. Unacknowledged = build blocked.

Build Step Types
code • schema • infrastructure • automation • training • evaluation
Contract Alert Lifecycle
CREATED → ACTIVE → ACKNOWLEDGED → CLOSED
4
Continuous Drift Detection
Health Monitoring

Every session begins with a mandatory drift check across every feature and every external system. This cannot be skipped. The system compares what was specified against what was built — and produces a per-feature drift state.

IN_SYNC

Code matches spec exactly. Proceed normally.

DRIFTING

Minor deviation. Score -5. Investigate before new work.

DIVERGED

Significant gap. Score -20. Must resolve first.

Project Health Score
Health = 100 - (DIVERGED x 20) - (DRIFTING x 5) - (open_debt x 2) - min(10, stale_days)
Post-MVP Work Routing — Four Tracks

After drift is assessed, every post-launch work request routes to one of four structured tracks — each with its own execution protocol:

Track A — Bug Fixes

Timed hypothesis protocol with a hard 3-strike limit. After 3 failed attempts, the agent stops and presents a structured escalation block — hypotheses tested, outcomes, best current guess — rather than continuing to guess.

Track B — New Features

Full mini-spec cycle before any code is written. Each new feature gets a spec, acceptance criteria, and drift tracking from day one — same standard as the original build.

Track C — Tech Debt

Debt items are weighed against active feature drift states and sprint capacity before scheduling. Debt that creates new risk gets elevated priority automatically.

Track D — Feature Evolution

Any change to existing feature behavior requires a brief evolution gate before execution — purpose review, side-effect analysis, spec update. Prevents silent drift accumulating through "quick changes."

5
Session Recovery Protocol
Resilience

If a session crashes, context is lost, or a new agent enters mid-build — the system recovers without human intervention. A structured recovery summary and progress markers become the source of truth for the restart.

Step 1

Read the latest recovery summary and progress markers - the engine's explicit record of what finished cleanly and what comes next.

Step 2

Read current project state, active queue context, and pending decisions to confirm the exact status of the build before doing more work.

Step 3

Resume from the last completed step, then validate any in-flight work before continuing. Completed work is not blindly rerun.

Step 4

Present the exact next action and continue. No re-explaining. No lost decisions. No starting over from phase zero.

Works across agent crashes, context window resets, tool restarts, and new agents joining an active build mid-session.

Intelligent Routing

One entry point. Six flows.

Every project is different. Scaffold OS detects its current state and routes automatically. You never have to know which file to run or which mode to enable.

New Project

No project detected. Routes to full brainstorm protocol — adaptive interview, challenge phase, complete architecture output.

Architecture Ready

Architecture file found, no build plan. Validates architecture for completeness, then generates phase-by-phase build plan.

Resume Build

Active build found. Runs a structured resume protocol: drift check, state reconstruction, and pickup at the exact interruption point.

Archaeology Mode

Code exists without plan. System reverse-engineers the architecture from code, then manages it under full Scaffold OS protocol going forward.

Continuous Mode — Build Already Live

Project is in production. Session begins with full drift audit, health score calculation, and backlog prioritization. Then every work request routes to the right track: Track A for bugs (3-strike escalation), Track B for new features (mini-spec required), Track C for tech debt (capacity-weighted), Track D for changes to existing behavior (evolution gate). Drift resolved before new work begins.

Recommendation Mode — Plan the Next Version

Project is live and the question is no longer "what are we building now?" but "what should we build next?" The engine studies shipped work, backlog, current health, and open decisions, then proposes ranked next-version options with one clear recommendation.

Skills Active Per Project

Domain Skills Active.
One coordinated protocol.

Scaffold OS activates the right skills for your project type — not a fixed list of roles that always run. 36 curated skills across 4 categories. 11 project profiles. Skills that match what you're building. Explore the skills system →

Solutions Architect
Phase 1 — Brainstorm

When the session begins on a new project, Scaffold OS acts first as a Solutions Architect. It reads everything already in the folder — existing documents, tech stack signals, vocabulary — and infers as much as possible before asking a single question. The structured brainstorm protocol extracts tech stack, system topology, every external service with its access method and agent role, security model, data flows, environment variables, deployment target, and scalability approach. The output is never vague. Every field is declared. Zero TBDs. Any gap found during architecture creation is a critical finding.

Tech stack selection System topology design Service dependency mapping Deployment target decision Challenge phase (8 structured challenges)
Pre-Build Auditor / QA Lead
Phase 2 — Audit

Before a single line of code is written, the system converts to a rigorous QA auditor. Two full rounds of consistency checks run across 13 compliance dimensions — schema-interface alignment, endpoint-feature coverage mapping, pipeline script validation, error code registry coherence, deployment artifact consistency, and more. Each issue is ranked by severity: CRITICAL (build blocked) or WARNING (flagged but continuable). Any CRITICAL finding must be fixed and re-audited before the build begins. Issues found here take minutes. The same issues found in production take days to weeks.

13 audit dimensions Two-round audit process CRITICAL / WARNING severity ranking Schema-interface alignment checks Automatic re-audit after fixes
Backend Engineer
Phase 3 — Build

Builds the complete server-side system — API architecture and routing, database schema and migrations, authentication flows (JWT, OAuth, session handling), role-based access control, middleware chains, background job systems, caching layers, and rate limiting. Every endpoint is declared against the system's architecture specification before it's built. The agent knows exactly which endpoints belong to which feature phase and cannot implement ahead of plan. Completion signals are written to disk. The session doesn't move on until they are.

REST / GraphQL API design DB schema + migrations Auth flows (JWT, OAuth) RBAC middleware Background jobs + caching
Frontend Engineer
Phase 3 — Build

Builds the complete client-side application — component architecture, routing, state management, API integration layer, form handling, authentication flows, responsive layouts, and performance optimization. The frontend build step knows which backend endpoints are live and which are being built simultaneously — and never calls endpoints outside the declared contract. UI components are built against feature specifications, not developer interpretation.

Component architecture State management design API integration layer Auth flow implementation Contract-aware API calls
Integration Specialist
Phase 3 — Build

Manages all external system connections as build artifacts — not as afterthought integrations. Every external service (API providers, SaaS platforms, webhook targets, third-party data sources) is declared in the architecture with its access method, authentication mechanism, rate limit constraints, and error handling strategy. The integration specialist role provisions access, builds the connection layer, and tests each endpoint against declared behavior — all as tracked build steps with completion signals.

External API contract definition Auth mechanism setup per service Webhook handler scaffolding Rate limit + retry logic Integration drift tracking
Data Engineer
Phase 3 — Build

Handles the full data layer — relational schema design with normalized relationships and proper indexing strategies, migration sequencing, data transformation pipelines (ETL/ELT), analytics view construction, data contract definition across services, and seeding strategy. When the project includes data warehouse components (Snowflake, Databricks, dbt), this role provisions those schemas and transformation models directly inside those platforms through their APIs — with the same drift tracking applied to code.

Schema design + migrations Indexing + query optimization ETL/ELT pipeline construction Data contracts across services dbt / Snowflake model generation
Security Engineer
Cross-Phase

Security isn't a single phase — it's enforced throughout. The security model is declared in the architecture (auth approach, permission model, data isolation rules, encryption strategy, secret management). The pre-build audit includes security-specific dimensions: checking that every API endpoint has an explicit auth requirement, that IAM roles follow least-privilege principles, that secrets are never hardcoded. For enterprise projects with AWS or similar infrastructure, IAM policies are generated and tracked as build artifacts — not copy-pasted from docs and forgotten.

Security model declaration Endpoint auth audit IAM role + policy generation Secret management protocol Least-privilege enforcement
ML / AI Engineer
Phase 3 — ML Build (v3.2+)

When a project includes machine learning components — detected automatically from architecture signals like PyTorch, TensorFlow, XGBoost, LangChain, or vector database declarations — the system activates ML-specific build protocols. A separate specification layer defines model architecture, training data sources, performance thresholds, and evaluation criteria. Training pipelines are built as first-class build step types. Evaluation gates enforce pass/fail against declared metrics before promotion. Model drift is tracked separately from code drift: BASELINE → DEGRADED → RETRAIN_REQUIRED.

Auto-detection from architecture Training pipeline construction Evaluation gate enforcement Model drift tracking (3-state) Vector DB / LLM infrastructure
DevOps / Infrastructure Engineer
Phase 3 — Build

Provisions cloud infrastructure as structured, versioned artifacts — not as one-off CLI commands that can't be reconstructed. AWS IAM roles, S3 bucket policies, Lambda functions, API Gateway routes, Cloudflare WAF rules, Kubernetes manifests, Docker configurations — all declared against the architecture spec and built as tracked steps. Infrastructure changes trigger the same contract protocol as code changes: scan for dependencies, alert affected systems, acknowledgment window before proceeding.

AWS / Cloudflare / K8s provisioning Infrastructure as build artifacts CI/CD pipeline scaffolding Environment config management Infrastructure contract enforcement
Automation Engineer
Phase 3 — Build

Creates automation workflows on external platforms as version-controlled build artifacts via their APIs — not manual GUI point-and-click configurations. n8n workflows are generated from structured JSON specs. Zapier Zaps, Make scenarios, and similar automation logic are built programmatically, tracked in the project's drift system, and compared against declared trigger/action logic on every session. A workflow that diverges from its spec is flagged as drift — same as any other feature.

n8n / Make / Zapier via API Workflow-as-code approach Trigger/action drift detection Scheduled job orchestration Event-driven architecture implementation
Technical Project Manager
Every Session

Every session, every time — a structured state check runs before any build activity. This is the Technical PM function: what is the current health of the project? What features are IN_SYNC, DRIFTING, or DIVERGED? What's the exact phase of the build plan? What remains? The health score is calculated every session. Backlog is reprioritized based on current debt. No build session starts without knowing exactly where things stand. This is what prevents "I thought we built that" moments months into a project.

Per-session health score Build plan phase tracking Backlog prioritization Drift-before-build enforcement Session handover documentation
CMS Architect
Phase 3 — Build

For projects with content management requirements, the CMS Architect role structures the entire CMS programmatically via CLI and REST API — not a human clicking through admin panels. WordPress custom post types, taxonomies, ACF field groups, metabox configurations. Strapi content types and permission policies. Ghost membership and tag structures. All scaffolded from code, tracked as build artifacts, and compared against declared content schema on each subsequent session. CMS schema drift is as detectable as API drift.

Custom post type scaffolding Field group generation (ACF) Taxonomy + relation structure CMS schema-as-code CMS drift detection
Cross-System Contract Enforcer
Always Active

The most unique role — and the one no other system has. Every shared interface between systems (an API endpoint consumed by the frontend, a message queue schema used by two backend services, a database table read by the data pipeline) is governed by an explicit, versioned contract file. If any contract changes — an API endpoint renamed, a field added to a shared schema — an 8-step protocol fires automatically: alert published, dependency scan across all consuming systems, 24-hour acknowledgment window opened. Any unacknowledged system blocks the entire build. No silent breaking changes. Ever.

Cross-system interface contracts 8-step change alert protocol Dependency scanner 24-hour acknowledgment window Build block on unacknowledged change
v3.1
Data & Systems Engineer — Infrastructure Layer
Auto-Detected (v3.1)

For projects that fit the standard stack profile, the system auto-detects eligibility and generates a fully configured local infrastructure setup — database, authentication provider, blob storage, and API gateway — all pre-wired and MCP-connected. This role is unique in what it eliminates: the "infrastructure tax" that typically consumes 30–50% of an AI build session before business logic is ever written. Non-technical users skip Docker entirely. The infrastructure materializes, the agents connect to it, and every session starts with a fully operational environment.

Why this matters: A typical raw agent spends hours debugging why a local database won't accept connections. That time is budget consumed building zero features. The infrastructure layer redirects 100% of the build budget toward functionality.

Stack profile auto-detection Docker Compose generation Auth + DB + Storage wiring MCP connection params output Human/Machine column setup file
Unique to Scaffold OS
Code Archaeologist — Reverse Engineering Mode
Flow 4 — No Other Tool Has This

You have an existing codebase. No planning files. No architecture document. Maybe it was built before Scaffold OS existed. Maybe it was handed to you from a previous team. Maybe you built it yourself months ago and the context is gone. The Code Archaeologist reverses the problem: instead of planning then building, it reads what was built, infers what the planning must have been, and reconstructs a complete set of planning files from the codebase.

Every inferred section is marked with a verification flag. Depth scales with codebase size: full read for smaller codebases, intelligent sampling for larger ones. Output: complete architecture specification, schema reconstruction from migrations/models, feature folders for detected feature areas, and API contract reconstruction from routes.

What archaeology produces
→ Complete architecture specification (from code)
→ Schema reconstruction (from migrations/models)
→ Feature folders per detected area
→ API contract file (from existing routes)
→ All inferred sections flagged for human verification
What a raw agent does instead
→ Asks you to explain everything from memory
→ Makes assumptions that go undocumented
→ Begins building without context
→ Cannot restart if the session dies
→ The next session starts from zero again
Full or sampled codebase read Architecture inference Schema reconstruction [VERIFY] flags on inferences Complete planning file set output
Recovery Protocol

When a session dies mid-build.
Scaffold OS recovers without you.

No other AI build system has a formal recovery protocol. When a session crashes mid-build — context limit hit, timeout, agent error — raw systems require you to explain everything again from scratch. Scaffold OS has a defined recovery sequence.

Recovery Steps (Automatic)
1

Read the latest recovery markers and locate the last safe progress point

2

Read the current recovery summary ? reconstruct current build position, open blockers, and active alerts

3

Run drift check — verify what was built since last session matches what was planned, surface any divergence

4

Present a concise resume summary ? what completed, what was in progress, and what comes next

5

Continue. No re-explanation. No guessing. The next agent picks up exactly where the last one stopped.

Without a recovery protocol

"We were 60% through building the order management feature when the session crashed. The next session didn't know what had been done. I had to manually read through 800 lines of code and re-explain what was built. Even then, the agent made conflicting assumptions about what the next feature should touch."

What actually happens without recovery
Agent re-invents incorrect state about what was built
Duplicate code written for half-completed features
Contract decisions made in previous session forgotten
Session budget consumed on archaeology instead of building

Every build step leaves behind structured recovery markers so the system can recover even if an agent crashes mid-feature, not just between phases. Future sessions inherit a clean paper trail instead of reconstructing state from scratch.

Intelligence Engine

Not just structured prompts.
Specialized intelligence, deployed precisely.

Scaffold OS is built on top of advanced AI capabilities — but uses them with surgical precision. The right cognitive function is invoked at exactly the right moment in the protocol. Not general-purpose chat. Purpose-built intelligence firing at the moments it actually matters.

Adaptive Intelligence at Brainstorm

The brainstorm protocol reads your vocabulary, response patterns, and the confidence with which you describe technical concepts — and calibrates the entire interview on the fly. Detected as NON_TECHNICAL: the agent takes on full architectural decision-making — it doesn't ask you to choose between Postgres and MySQL. Detected as TECHNICAL: peer-level debate and structured disagreement, right up to the challenge phase.

Three technical levels. One automatic detection. No configuration.

Structured Challenge Protocol

Before any architecture is finalized, the system enters a formal challenge session. It raises up to 8 structured challenges across blind spots, scaling risks, security concerns, and architectural contradictions. The agent is explicitly designed to disagree here — not to validate your choices. Every challenge that survives makes the architecture stronger. Every challenge that fails gives you confidence your original decision was sound.

Disagreement as a feature. Not a bug.

Precision Skill Invocation

Certain protocol moments require specialized cognitive depth: architectural validation, deep security review, ML model specification, infrastructure topology reasoning. At these exact moments, Scaffold OS invokes purpose-built intelligence functions — not the same general model that writes code. The brainstorm intelligence that extracts architecture is not the same function that audits it or builds it. The right cognitive tool at the right moment.

Specialized modules. Precise invocation timing.

Context Distillation Between Phases

AI context windows are finite. Scaffold OS solves this not by ignoring it, but by compressing aggressively at phase boundaries. Between brainstorm and build, between sessions, between major milestones — the system distills the full session history down to exactly the information subsequent phases need. Each new session starts with precise, dense, relevant state — not raw conversation history from three sessions ago. This is why sessions don't degrade as a project grows.

Phase-boundary compression → consistent quality across all sessions.

Parallel Agent Architecture COMING SOON

The contract enforcement layer was built specifically to enable this. Once active, multiple specialized agents run simultaneously — a backend agent and frontend agent building in parallel, each with their own context slice, governed by the same contract system. When the backend changes a shared endpoint, the contract protocol fires — the frontend agent is notified, awaits acknowledgment, and the build continues. Parallel builds. Zero silent breaking changes. Build time drops 60-70% on complex projects.

The contract system was designed from day one for this moment.

See what it can build.

From SaaS products to ML platforms. Every complex software system is supported.

Explore Capabilities → See Roadmap