The methodology layer for AI-driven development
Coding agents made code cheap.Skix AIDD makes code trustworthy.
By making specs and verification the rails agents run on. A spec channel's life — discussion, intent, generation, verification, signed approval, shipping — is never stored. It is projected from an append-only log of Ed25519-signed events. The events are the truth; the screen is one way of reading them.
- spec.definedIntent fixed2 acceptance criteria3f9c…a21
- spec.compiledAC → test skeleton2 compiled 1:1 (red)8b12…07d
- agent.activitykaji · EDITinternal/notify/quiet.goc4e0…9f5
- gate.resultThree gatesSpec ✓ Debt ✓ Provenance ✓1a77…b30
- digest.postedVerdict digestthe only thing a human reads02de…4c1
- approval.signed@owner signedover sha256(digest)e59b…88a
The problem
Whatever generation got cheaper, verification billed back.
The bottleneck moved from writing code to reviewing it. That is not a feeling; it has been measured.
+154%
Pull-request size. Review time is up 91%.
Faros AI, “The AI Productivity Paradox” (Jul 2025, telemetry from 10,000+ developers)
90% / 30%
About 90% of technology professionals use AI at work — while about 30% of developers trust AI-generated code little or not at all.
DORA 2025 (n≈5,000)
7.5–14.8%
Share of AI-generated commits that introduced at least one statically detectable issue, per AI coding assistant (five studied). 89.3% of the issues found are code smells — maintainability debt, not bugs.
arXiv 2603.28592 v3 (Oct 2026, 302.6k commits)
For the record: that study is a preprint with no human-authored baseline. You cannot conclude from it that AI writes buggier code than people. What you can conclude is that generated code nobody verifies accumulates maintainability debt.
DORA's amplifier hypothesis
AI raises throughput and instability at the same time. Which way it tips is decided not by which agent you picked but by the system around it — specs, verification, measurement, governance. Skix AIDD ships that system as a product.
How it works
The intent–verification loop, closed so it cannot be skipped.
Now that generating code is cheap, the expensive and essential parts are the two ends: deciding precisely what to build (intent → spec), and confirming that what was built is that thing (verification → trust). A channel moves through seven stages.
Discussion
messageboard.updateHumans and agents talk as peers. An AI facilitator sorts the thread into open questions, decisions and unknowns; only approved to-dos enter the loop.
Intent
spec.definedspec.compiledAcceptance criteria are fixed. Non-engineers can file a spec — the AI drafts criteria from a title and a background. Criteria compile 1:1 into test skeletons that fail until the feature exists.
Generation
member.assignedagent.activitySpec-first is enforced: running an agent against an uncompiled spec is refused. On assignment the agent is hydrated with approved skills and verified knowledge — nothing unverified reaches it.
Verification
gate.resultdigest.postedMachines first, humans last. Three gates — spec conformance, debt, provenance — run before a verdict digest is posted. A failure is turned into repair instructions and handed back to the agent.
Signed approval
approval.signedA human reads the verdict digest, not a 2,000-line diff, and signs the sha256 of that digest with their own key. What they read and what they approved is cryptographically bound. Self-approval is refused.
Shipping
merge.completeddeploy.signedMerge into the mainline. Deployment is designed to wait on a signature over the artifact digest, and every stage is itself a signed event.
Archive
channel.archivedThe channel is sealed. Spec through deployment survives as one signature chain, so a change-management trail falls out of ordinary work instead of being assembled for an auditor.
Channels are not stored. They are rebuilt every time as a projection of the signed event log — the events are the truth, the channel is a view.
Capabilities
What actually runs today.
Everything below is implemented and under test. What is still a design lives in “What's built”, further down.
Spec-first, enforced
Acceptance criteria compile 1:1 into test skeletons. A skeleton fails until the feature is implemented (red before green), which is what makes “tests pass” mean “the spec is met”. The CLI refuses to run an agent against a spec that was never compiled.
spec compile · 9 languages / 14 frameworks
Three gates and a verdict digest
Spec conformance (AC-derived tests, uncovered criteria, PR size ceiling), debt (static analysis, smells, duplication, coverage delta) and provenance (attestation chain, secret scanning, provenance trailer). Nothing that fails reaches a human review queue.
gate run → digest.posted
A signed event log
Append-only, prev_hash-chained, Ed25519-signed. Approval, acceptance and deployment are all “the principal's key over the hash of the thing”. Chain verification and tamper-evidence anchoring are built in.
log verify · sig check
Self-repair loop
When a gate fails, the reason is injected into the agent's context and the same agent runs again. Past the retry limit it does not quietly give up — it escalates to a human. Never go dark.
execute --repair N → repair.escalated
Permission-gated equipment
A skill contributes only its name and description to context; the body opens only when the agent asks for it, and the opening is recorded. Unapproved skills are invisible to agents, and only knowledge carrying a verification signature is ever hydrated.
skill.opened · knowledge.verified
Attestation revocation, enforced
An agent's authority is inherited from its human owner. Revoke the owner and every agent under them is revoked instantly; from then on the provenance gate and signature verification reject their signatures. No offboarding checklist to run.
member revoke → gate FAIL
Bring your own agent
Point Claude Code, Codex, Cursor or Zed at the MCP server and every file operation and shell command the agent issues is executed and audited by the harness. The agent asks; the harness writes.
skix-aidd mcp (skix-dev-mcp)
Two interfaces, one ledger
Engineers never leave the terminal: 38 CLI commands. Product, leadership, QA and audit get a 23-page console for governing development without reading code. Both are looking at the same events.
CLI + console (single binary)
Get started
One trip around the smallest loop, in ten minutes.
Write a spec, compile it into test skeletons, drive an agent, run the gates, sign the digest.
# create the .aidd/ starter kit and the event log$ skix-aidd init --code myproj --name "My Project" # state the intent — the criteria are the contract$ skix-aidd spec new "Notification quiet hours" \ --ac "AC-1: no notifications between 22:00 and 07:00" \ --ac "AC-2: the user's timezone setting decides" # criteria → test skeletons (red until implemented)$ skix-aidd spec compile SPEC-100 --into ./pkg # spec-first + hydration + the self-repair loop$ skix-aidd execute SPEC-100 --agent kaji --repair 3 --cli "claude -p" # three gates → verdict digest$ skix-aidd gate run SPEC-100 # sign the digest's sha256 with a real Ed25519 key$ skix-aidd approve SPEC-100 --as @owner # verify the signed event chain$ skix-aidd log verify && skix-aidd sig check
Today, creating channels, filing specs and running agents are CLI-only. The console is the reading, approving, document-review and audit surface.
Read the CLI guideWhat's built
What does not exist yet, stated first.
A product that sells audit trails cannot oversell itself on its own landing page. This is where things stood on 2026-09-18.
Working today
- 38 CLI commands and a 23-page Next.js console (embeddable in a single binary)
- Ed25519 signed event log, prev_hash chaining, chain verification, tamper-evidence anchors
- AC → test-skeleton compilation across 9 languages / 14 frameworks, with spec-first enforced
- Three gates, verdict digest, signed approval (self-approval refused)
- Self-repair loop (inject gate feedback → re-run → escalate)
- Permission-gated skills and knowledge; enforced attestation revocation
- Codebase analysis inside a Firecracker MicroVM
- Evidence-package export, cross-cutting search, per-event signatures, a client portal for customer acceptance
Not there yet
- Real CI and real deployment — the pipeline advances as events, but nothing runs and nothing deploys
- Starting a loop from the web — creating channels, filing specs and running agents are CLI-only
- MicroVM isolation for agent execution — real on the analysis and repair paths; direct dispatch still runs on the host
- An open-source release, a LICENSE, and brew distribution
- Self-serve signup and published pricing
Drawing this line is the point of the HONESTY-1 phase: governance that is displayed but not enforced counts as a breach of trust, not a missing feature.
FAQ
Frequently asked questions
What people ask before adopting it, answered only with what runs today.
- What does Skix AIDD do?
- It is an AI-driven development platform that puts a coding agent's output through spec (acceptance criteria) → test skeletons → three gates (spec conformance, debt, provenance) → verdict digest → human signed approval. Each step is recorded in an event log signed with Ed25519 and chained by prev_hash.
- Can I use it today?
- It is in development (Developer Preview). There is no self-serve sign-up and no public pricing yet. To try it, contact [email protected].
- Which coding agents does it work with?
- skix-aidd mcp runs as an MCP server that you connect to MCP clients such as Claude Code, Codex, Cursor and Zed. The agent's file operations and shell commands are executed and audited by the harness. It can also drive a print-mode CLI (e.g. claude -p).
- Which languages and test frameworks are supported?
- Test skeletons compiled from acceptance criteria can be generated for Go, TypeScript, JavaScript, Python, Ruby, Java, Rust, Elixir, PHP and C#, across 14 test frameworks.
- What can I do in the web console?
- Read, approve, review deliverables and inspect evidence. Creating channels, writing specs and running agents is currently CLI-only.
- Does it connect to CI or production deployment?
- Not yet. The pipeline advances as events, but it runs no real CI and deploys nothing.
- Is agent execution isolated?
- The codebase-analysis and repair paths run in a MicroVM (Firecracker). Agent runs via direct dispatch still execute on the host.
- Is it open source?
- Not yet (no LICENSE and no brew distribution).
Watch one trip around the loop.
The console is a reading surface assembled from the signed event log. Channel timelines, gate verdicts, approval signatures and chain verification are all visible exactly as recorded.