Code Factory

6 mandatory audit lanes. 143 coded rejection conditions. One human-owned release decision.

Your code passed. But did it pass the right tests?

An AI agent—or a human developer—can write the code, write the tests, and set the rules used to approve its own work. Everything can turn green while the result still misses the request, leaks another tenant's data, breaks during a retry, damages a migration, or slows down in production.

Code Factory is an engineering audit and orchestration system: the independent review layer between “the build passed” and “a person should approve this.”

It seals the original intent before work begins, observes what changed, challenges the implementation and its tests, and connects every decision to inspectable evidence. If evidence is missing, contradictory, stale, or based on a weakened rule, the result does not quietly become a pass.

0.46.8 release preview

This update connects the original request to typed agent handoffs, verification routes, candidate changes, fresh receipts, and the human review decision. It also reduces repeated context and repair work with searchable audit rules, exact-hash evidence reuse, bounded receipt indexing, and lazy CLI loading. Optional loopback-only stateless MCP HTTP joins the existing local stdio surface. The six audit lanes remain intact; agents still cannot approve, merge, publish, or deploy. See the 0.46.8 preview and channel notes.

One review path, six questions

Those lanes contain 81 lane-specific and 62 cross-cutting coded rejection conditions. The source-bound inventory recomputes the total from the implementation; it is not a claim that every project executes 143 tests.

What the Code Factory upgrade improves

Code Factory is the core audit and orchestration system. FactoryLine is one CLI/IDE/MCP surface for using it; the performance gains below come from the CF workflow itself, not from a JetBrains-specific adapter:

• Less repeated work: the dependency-aware proof DAG routes each change to RUN, REUSE, SKIP, or BLOCK, then compares the incremental route with a full baseline when that comparison is required.
• Fewer wasted agent retries: Proof-Delta stops a loop when the candidate did not change or fresh evidence did not improve the result, instead of spending another repair cycle on the same idea.
• Faster review orientation: hash-bound receipts, Graph Ops, and the failure briefing keep the exact diff, proof gap, consequence, evidence, and next action together instead of making a reviewer reconstruct them from agent chat and scattered logs.
• Bounded resource cost: evidence reads, workspace measurements, and process-bound replays have explicit limits and fresh-workspace rules, so the audit does not silently scan or mutate an entire project.
• Fast, repeatable agent context: factory efficiency pack emits a priority-ordered, hash-verified packet with bounded excerpts and exact cache reuse, so unchanged context need not be reassembled for each hand-off. Token counts are explicitly estimated—not provider-usage or savings claims— and drift or secret-shaped material fails closed.
• Clear agent accountability: the proof-coupled Junie handoff records the known tool, changed-file rationale, hashes, and explicit unknowns against a taxonomy digest, so an agent receives credit only for evidence a reviewer can inspect.
• Release evidence that travels with the artifact: the optional supply-chain gate binds source and lockfiles, SBOM/VEX and licence policy, identical rebuilds, and package secret/path scans to one candidate digest. It blocks drift without pretending that a local receipt is a marketplace or security approval.

These are workflow-efficiency improvements, not a promise that application code or tests execute faster. Actual time or cost savings remain unknown until your project supplies a baseline and an evidence window.

Agents can start with factory first-lap status or the read-only factory.first_lap_status tool over local MCP; browser-capable clients see the same bounded status through WebMCP. All three surfaces report the same integrity state and next action without executing a journey or granting release authority.

Start with the smallest useful path

Six audit lanes turn a plausible change into a reviewable engineering decision before the first proof command is run.

1. Run factory first-proof --root . to see a deliberately hollow check caught by the local auditor.
2. Run factory first-lap status --root . to inspect the human-observed, evidence-bound activation path.
3. Run factory architecture health --root . --json to check repository complexity budgets before adding another command, module, or document.

Only then move to the six-lane audit, Graph Ops, or specialist packs. Each step keeps the same evidence and authority boundaries; advanced controls add coverage, not permission to approve or release.

Mission Control and IDE clients can also request factory.agui_review_events or run factory agui review-events --json to render the same state as a small, hash-bound controlled event stream (RUN_STARTED → STATE_SNAPSHOT → REVIEW_CARD → RUN_FINISHED). This is a presentation bridge for review cards and human interrupts—not open-ended generated UI, provider transport, or agent authority. See AGUI and MCP2 review surface.

New: First Lap activation control

The new factory first-lap path makes the safest part of Code Factory the easiest to start. One command creates three plain-language files: MISSION.md for intended and forbidden outcomes, END-TO-END.md for observable journeys, and verifier-only .factory/holdouts/HOLDOUT.md for independent failure cases. Before supervised autonomy is considered, every critical verifier must pass a three-point calibration (approved candidate passes, a deliberate defect fails, and the wrong candidate is inconclusive or blocked), one human must observe the complete intake-to-handoff lap, and any builder access to a holdout blocks the run. A new incident can be promoted through incident → failure code → invariant → reproducer → mutation → permanent gate → owner, turning past failures into executable regression controls. Retry policy is typed and fail-closed: only a narrowly classified transient provider failure is retryable; stale evidence, product failures, identity mismatches, and setup failures remain review events. See First Lap.

Built for the way you work

• Solo developers and vibe coders: get a second opinion before a plausible AI-generated change becomes your problem in production.
• Engineering teams: compare the request, code delta, independent checks, repair evidence, and remaining unknowns in one reviewable handoff.
• Enterprise teams: keep policy provenance, exceptions, approvals, tenant boundaries, and evidence traceable without giving the coding agent authority to redefine success.
• Connected agents: use the CLI, FactoryLine IDE surfaces, MCP, WebMCP, or A2A handoffs to follow the same bounded process and report the next action to a human.

Specialist modules join the same review when needed: Deep Defect Mesh consumes analyzer evidence, SaaS checks examine customer journeys, and AppForge organizes mobile build, design, privacy, signing, and store-submission evidence. AppForge does not replace the other audit lanes or promise store approval.

Full-Stack UX Harness contract

For mobile and full-stack releases, the Full-Stack UX Harness turns the release brief into a strict, local contract that AppForge can consume. It normalizes six evidence categories—visual media, privacy-to-listing, release-chain, design-system, production-signal, and Android parity—and keeps the source YAML, normalized projection, and receipt hash-bound. Validate the repository contract before running an audit, then replay the receipt in a fresh workspace:

python -m factoryline.cli quality-harness spec-validate full-stack-ux-harness-v1.ssat.yaml --root .
python -m factoryline.cli quality-harness spec-verify .factory/receipts/full-stack-ux-harness-v1.json --root .

The CI workflow at .github/workflows/validate-ux-harness.yml runs the same fail-closed checks and focused tests on pull requests. It validates intent and evidence structure only; it does not access credentials, run provider uploads, or mark an App Store/Play submission approved. See docs/FULL_STACK_UX_HARNESS.md for the direct and legacy SSAT forms and the exact evidence boundary.

Choose your review path

pip install factoryline-code-factory
factory guide

factory guide is a read-only orientation: it shows the smallest suitable path without running tests or agents. When you are ready to produce evidence, follow the engineering review workflow. Its result shows what was checked, what failed, why it matters, what evidence supports the finding, and what must happen next. Code Factory does not promise perfect software. It makes the basis for approval visible and keeps the final decision human-owned.

Install the CLI, then follow the engineering review workflow for a full review:

pip install factoryline-code-factory

Want a small demonstration first? factory first-proof --root . challenges whether a test can actually reject its declared failure in a disposable sandbox. It writes a receipt and a privacy-safe Proof Card; it does not assess or change your project. First Proof is one entry point—not the scope of the full audit.

LLM-friendly product context: Canonical Product Card states the supported capabilities, boundaries, audiences, integration terms, and exact vocabulary that documentation assistants and agent tools may safely reuse. It is intentionally claim-bounded: local evidence is not a release, provider approval, remote execution, or a guarantee of correctness.

Pick the path that matches your work

Beyond hollow tests: Deep Defect Mesh evaluates signed analyzer-report contracts, flags missing negative controls, traces findings into Graph Ops, and compares repair outcomes. It coordinates supplied evidence; it does not execute the analyzers or guarantee that a clean report means safe code.

Choose your review path below. The current factory guide command still lists the existing onboarding paths, including the solo First Proof demo; it is not the full engineering-review workflow linked above.

• Individual developer or vibe coder: What did my change miss? Use the engineering review workflow to inspect intent, behavior, and evidence. Try factory first-proof --root . for a contained demonstration first.
• Engineering team: Did the agent build what we asked for? Start with factory oracle status --root .. Bind the original intent, observe the exact file delta, challenge the tests independently, and keep approval human-owned.
• Enterprise evaluator: Can we govern agent work without trusting the agent's story? Start with factory graph ops --root . --json. Inspect the source-to-decision chain, blockers, unknowns, and authority boundaries in a controlled pilot.

Choose the relevant audit scope. Specialized SaaS, mobile, multi-repository, and enterprise controls remain available in the advanced workflow guide only when their trigger applies. AppForge is an optional mobile-delivery capability; it surfaces avoidable gaps but cannot guarantee store approval.

For a release-bound service change, use the six-lane runtime assurance workflow. It coordinates stateful invariants, tenant isolation, fault recovery, consumer compatibility, database migration, and equivalent-load performance/resource checks under one signed plan. Every failed lane returns the consequence, evidence digest, exact replay, and next repair. A six-lane pass means ready for human review—not approved for release.

Senior engineering controls

When a team needs stronger evidence than a self-reported green build, use the senior-engineering integration:

• verify an independently collected, candidate-bound execution attestation;
• measure a reviewed buggy/fixed defect corpus with case-level recall and precision; and
• route a dependency DAG through RUN, REUSE, SKIP, or BLOCK, then compare the incremental plan with a full baseline;
• reproduce a failure in a fresh, contract-bound process workspace, compare the original failure with a repair and negative controls, explain every reuse decision across policy/dependency/toolchain/environment fingerprints, and hand a reviewer a runnable, evidence-linked failure briefing.

These adapters are provider-neutral and review-only. Replay and repair execute only when the operator supplies --execute; they use a fresh temporary process workspace with a secret-free environment, not a kernel/container sandbox. They never publish a package or approve a release. Missing independent evidence remains an evidence gap.

Do not take the word verified on faith. The Capability Evidence Map separates the locally verified core, controlled-pilot workflows, reference enterprise controls, and candidate-bound mobile preflight—and links each claim to the tests you can inspect and run yourself. For one agent-neutral human review path, use factory proof-review to seal intent, audit trajectory evidence, route the change, learn confirmed regressions, inspect the team inbox, and export a tamper-evident Proof Card. To create a contained starter later, run factory mvp "Build an approval tracker" --root .. That starter is never called production-ready until the relevant proof exists.

AppForge: make the review queue the last surprise, not the first

You have a build, a launch date, and a review queue that can take days to come back around. The costly failure is often not a crash—it is discovering after submission that the selected build has no exercised restore path, an iPad dead end, unrepresentative screenshots, an accessibility gap, or reviewer access that does not match the app.

AppForge turns that uncertainty into a local, candidate-bound worklist. Start with the user’s design direction, keep a story-led storyboard, then bind real iPhone/iPad media, App Review checks, quality evidence, SaaS journey evidence, and a final Markdown/PDF dossier to the same build. In JetBrains, AppForge Mission Control makes the missing lane visible before a human decides to enter the queue.

For iOS and Android delivery, factory revenue appforge-mobile-evidence normalizes supplied, hash-bound Xcode/XCTest, Android Gradle/ADB, Fastlane, device-cloud, and production-monitoring exports into one local receipt. It requires visual, privacy-to-listing, release-chain, design-system, production-signal, and Android-parity evidence to be explicit; it does not run those tools, access a store, submit a build, or guarantee approval.

That can prevent avoidable rework and repeat review cycles; it does not submit to Apple, access credentials, replace TestFlight, provide legal advice, or guarantee approval. Start with the AppForge Evidence Kit, then use the credential-free EAS handoff preflight before a separately authorized build or submission.

Support posture

The open-source core is free. Community support runs through public GitHub issues and has no guaranteed response time. For teams that need an operating commitment, the repository includes a transparent proposed Support SLA with P1/P2/P3 response targets, availability, recovery, exclusions, and an activation checklist.

Those targets are planning terms, not a current contract: the SLA remains proposed and non-effective until Code Factory has a named support owner, monitored escalation, production telemetry, dependency and recovery evidence, legal/security terms, and a signed enterprise order. This keeps the GitHub page useful to evaluators without implying that a free repository already carries a managed-service guarantee.

See actual Factory Studio and the current FactoryLine visual set or open the live Hugging Face Space.

Advanced workflows

• Start a real project. Turn one outcome into a contained web, mobile, API, CLI, worker, MCP, or agent-UI starting state.
• Review what AI produced. Turn requirements, diffs, proofs, and blockers into receipts, Graph Ops, and one fact-derived next action.
• Refresh the evidence that matters. Watch a local Assembly while it runs, refresh the next-proof brief, and see privacy-bounded observed Git contribution context without pretending it is a verified directory or billing roster.
• Keep "done" honest. Challenge declared validators for hollow tests; a green-looking scaffold is never called production-ready by default.
• Stop vague work before it starts. Use Intake Grill to record framework, exact intent, observable acceptance evidence, and the external-effects boundary before a Product Mission begins.
• Make a behavior survive its declared failure cases. Use the supervised Gauntlet to turn human-written E2E sabotage cases into an offline-verifiable Survival Card. A hollow negative check stays visible; a card is never a production-readiness certificate.
• Let autonomy be earned, not assumed. Use factory license to derive an expiry-bound local tier from independently verified governed runs. A severe hollow-test, hollow-validator, or scope-escape result demotes the declared agent automatically; it never authenticates identity or grants execution.
• Compare agents with receipts, not vibes. Use factory combine to rank completed, sealed, independently verified runs on the same task. It never starts an agent or turns a local scoreboard into a vendor-quality claim.
• Make every agent session feed the evidence loop. Use factory wrap around Claude Code, Codex, DeepSeek, or another local CLI. It verifies a sealed admission before launch, observes the exact file delta, runs declared independent validators, and records the result for Agent License and Combine. It stores hashes and bounded facts—not prompts or raw output—and observes the process without pretending to sandbox it.
• Verify an explicit agent workflow without adopting its runtime. Use factory atomic import to bind a secret-free exported stage DAG, scoped handoffs, checkpoint facts, and immutable source preconditions to a current Oracle Contract. Graph Ops, MCP, and WebMCP show the receipt read-only; Code Factory never starts Atomic, resumes a checkpoint, or treats a declared workflow as release authority.
• Keep decisions from becoming tribal knowledge. Use factory judgment to track a human-promoted design decision with exact paths, a named owner, review date, and hash-bound proof obligations. A Change Safety Case routes one explicit diff to its owner; it never infers intent, promotes a decision, runs a test, or approves a release.
• Prove the whole journey—and audit the repairing agent. Use factory journey to compare declared and observed product journeys, preserve a hash-bound failure capsule, prove state flow and cleanup, and challenge a proposed healing with a negative mutation. Choose human-controlled or bounded supervised-auto mode; every autonomous attempt receives a separate FactoryLine audit and neither mode can self-approve.

What the latest upgrade resolves

These are local evidence and supervision tools, not guarantees of performance, security, productivity, production readiness, or an automatic repair service.

FactoryLine by role

Expected operational outcomes

FactoryLine is designed to reduce avoidable review loops, evidence hunting, and the frustration of discovering late that an AI-generated test never exercised the claimed behavior. The expected outcome is a shorter path to a defensible human decision—not a promise that every project becomes faster or cheaper.

FactoryLine never converts an unmeasured estimate into a savings claim. Use the Savings Tracker to report observed time or cost only when the repository has a valid baseline and evidence window.

For planning, a transparent starting band is 25–120 minutes of avoidable review and rework per AI-assisted PR: roughly 10–30 minutes of reviewer context reconstruction plus 15–90 minutes of one preventable clarification, validator, or scope-rework loop. At 50 AI-assisted PRs per month, that models about 21–100 hours; at a team-supplied loaded cost of $75–$150/hour, the illustrative range is about $1,560–$15,000 per month. These are adjustable scenario inputs, not observed FactoryLine results. See the outcomes and savings model before using the range.

Works with your existing AI development stack

FactoryLine is the proof and control layer around generation, orchestration, and review tools. The status column distinguishes implemented adapters from clean workflow fits; it does not imply a vendor partnership.

See the complete compatibility and handoff guide for the recommended flow and the exact authority boundary for each stack.

For teams: use the Teams and Enterprise Operations Manual to run the same proof-first loop with named reviewers, approved AI-change scope, and explicit Proof Debt—without giving Code Factory merge, release, or provider authority. The commercial packaging guide keeps the free core separate from proposed Team and Enterprise services that are not purchasable yet. For a human-selected, customer-managed reference pilot, the local Team Pilot readiness gate hash-binds the operating evidence for owner review; it does not accept a customer or activate a paid service.

Design is part of the review. For UI-scoped work, add the optional Prestige Design Review: a purpose-led design brief plus review artifacts for hierarchy, responsive behavior, affordances, consistency, and declared design tokens. It makes design quality visible; it does not claim a conversion result, WCAG certification, or production readiness.

Reuse a proven decision without reusing stale context. Factory Continuity keeps a local, purpose-bound record of the evidence behind prior work. Graph Ops can replay only redacted, current, independently promoted metadata; it does not store private source, prompts, embeddings, or transcripts, and it cannot execute a repair.

Revenue-ready iOS scaffolds

factory revenue build turns one reviewed products.yaml into StoreKit 2, a transparent SwiftUI paywall, an Apple-JWS-verifying entitlement-server scaffold, privacy/review artifacts, and a local evidence page. Then replay, testflight-sync, failure-matrix, and policy-watch expose where observed purchase behavior, beta feedback, negative paths, or policy sources diverge—without turning unknowns green. Evidence Memory carries human-approved, receipt-backed lessons into the next exact-app journey while expiring stale knowledge and quarantining contradictions. IDEs and coding agents can inspect the same local facts through MCP; compatible browser agents can inspect bounded Graph Ops status through progressive WebMCP. App Store writes, pricing, experiments, offers, publication, and credentials stay human-controlled. See RevenueForge and the MCP/WebMCP handoff.

Install

# No account, model key, or cloud connection is required for this local run.
pip install factoryline-code-factory
factory first-proof --root .
factory mvp "Build an approval tracker" --root .
factory studio --root .\my-mvp

If Code Factory helps you find a proof gap or makes an AI-assisted change easier to review, star Code Factory so other developers can find it. This optional link only opens the repository.

Choose the job in front of you

For the short product map, read the overview. For a two-minute first run, follow Start Here. For full command and contract reference, browse the documentation directory.

Use Code Factory with CodeRabbit or another AI reviewer

They solve different parts of the review problem. CodeRabbit can supply AI findings and suggestions; Code Factory makes declared local proof gaps, coverage, and the next review action explicit. Enable the opt-in GitHub Proof Review workflow to put one neutral, commit-bound FactoryLine Check and walkthrough beside existing CodeRabbit comments. It uses no CodeRabbit account, API, credential, or output as proof.

For agent-created pull requests, add a small human-approved factory.agent_plan.v1 envelope. Plan-to-Proof compares that plan with the exact diff and exposes Proof Debt: unresolved scope, declared-test, human-routing, and existing proof obligations that must be settled before a team makes its own merge decision.

That makes Code Factory a standalone proof layer for teams that do not use an AI reviewer, and a complementary evidence layer for teams that do. For agent-created pull requests, Code Factory does not replace human review, auto-approve a pull request, or merge code.

Use Code Factory with LangGraph

LangGraph teams can keep their own graph and checkpoint runtime, then use the LangGraph Assurance Bridge to compare recorded reference and resumed transitions. The free local adapter produces hash-only parity evidence and a reviewable incident capsule when a run diverges; it does not import LangGraph, invoke a graph, or claim production resilience. The optional GitHub Action puts the same Proof Card in a pull-request job summary.

For Codex, Claude Code, and Deep Agents, install the Code Factory LangGraph plugin to add the proof workflow and read-only local MCP facts to the coding-agent surface.

Use Code Factory with DeepSeek Harness

The optional DeepSeek Harness adapter starts the same local read-only MCP proof surface through Harness's official generic MCP client. It lets a Harness agent inspect Graph Ops, current proof gaps, and Earned Autonomy / Combine facts without sending source to a hosted endpoint or gaining permission to execute, repair, approve, release, deploy, sign, or use credentials.

Use it where you work

Code Factory keeps the same local, receipt-bound workflow across the command line, VS Code, and the JetBrains plugin. It also provides local stdio Cursor or OpenCode MCP without handing an AI client permission to publish, deploy, approve, sign, or access credentials.

The same local proof surface is discoverable in the Official MCP Registry as io.github.zrk222/code-factory; registry setup starts the public PyPI package over local stdio and never creates a hosted service or write authority.

FactoryLine's core local proof workflow remains free. The planned JetBrains Freemium transition starts January 1, 2027, subject to Marketplace and activation gates. Early adopters may retain Founding Proof Pro at $5.95 USD/month or $60 USD/year; standard Proof Pro is planned at $9/month or $90/year, and AppForge Builder at $24/month or $240/year. Team Assurance is a separate $20-per-active-contributor annual-commit hypothesis. None is active today; no checkout, entitlement, or license enforcement exists. See the Marketplace control-room guide for the exact feature boundary and approval gates.

The matching GitHub Assurance Seat is also planned for January 1, 2027 at the same future price. It is for maintained, customer-managed proof operations (commit-bound review, Proof Debt, policy drift, governed exceptions, and evidence packets)—not source access or opaque AI-token resale. The source license and free core are unchanged. See the GitHub per-seat plan.

For Open VSX, the extension and local proof core remain free. All capabilities shipped before the transition are free through December 14, 2026. From December 15, 2026, Founding Proof Pro may remain $5.95/month or $60/year for early adopters; standard Proof Pro is planned at $9/month or $90/year, AppForge Builder at $24/month or $240/year, and Team Assurance at $20 per active contributor/month annually or $24 monthly, subject to explicit activation gates. See the Open VSX service plan.

Release owners can use the read-only release candidate preflight to bind source, commit, platform artifacts, and active metadata before any upload. The Codex metadata integrity audit keeps archival claims out of active proof, and Windows process binding closes the bind-before-run process-tree race.

The proof boundary

Code Factory creates and inspects local artifacts. It does not silently call a model, discover credentials, publish, deploy, sign, approve, message, or grant a connector. The Gauntlet executes only caller-declared E2E pairs after a named, expiry-bound, one-run admission; all other Gauntlet paths are read-only. Its deterministic proof receipts bind supplied byte bindings, declared identities, and evidence; an external runner must separately prove runtime isolation and network policy. Token, cost, and productivity claims remain unknown until a bound measurement exists.

flowchart LR
  intent["Plain-language outcome"] --> mvp["Local MVP"]
  mvp --> evidence["Receipts and declared checks"]
  evidence --> review["Graph Ops / review packet"]
  review --> decision["One evidence-backed next action"]

Use Intake Grill to pin down intent before a mission, PRD Grill before code exists, the deterministic contradiction gate when requirements collide, Proof Review when a diff arrives, and the Verifier Plane when a worker claims it is finished. Use the supervised Gauntlet when a specific behavior needs to survive explicitly reviewed failure cases. The local MCP contract and generated Mermaid output map make the same proof context reusable by a client you choose.

Go deeper when you need it

• Read why I built Code Factory for the founder story behind catching passing tests that fail in real use.
• Start with PRD Grill, Proof Review, or Verifier Plane when the job calls for it.
• Browse Graph Ops, Gauntlet, Proof-Delta Loop, Graph Portfolio and Run Admission, Evidence Frontier, Factory Reality Check, proof reuse, and LangGraph Assurance, and savings boundaries for advanced evidence workflows.
• Use IDE Health and Index Continuity when a JetBrains IDE slows down and you need locally observed signals before deciding what to review.
• For UI work, read Prestige Design Review for the optional design-quality lane and its explicit review boundaries.
• Read The approval signal decays when AI-written code becomes routine for the design and limits of the habituation gate.
• See the 0.46.8 release preview, the grilling ladder, CHANGELOG.md, release channels, and publication guide for versioned release detail.

License

MIT OR Apache-2.0.