04.product
private betaIgnix67
The platform that operationalises AI-first context-driven development. Ignix67 is currently in private beta with a small set of partner teams; the five capabilities below are the production scope for v1.
01.pr-review
PR Review
AI-assisted code review on every pull request, across Azure DevOps, Gitea, and GitHub from one set of rules. Reads the diff, applies your per-repo review instructions, posts threaded comments and a reviewer vote — running through a predictable, audit-friendly pipeline rather than an open-ended agent loop.
- Works with Azure DevOps, Gitea, and GitHub — same review rules across all three
- Per-repo review rules from .ignix/review-instructions.md, loaded into every review
- Severity-gated voting — HIGH blocks (-10), MEDIUM advises (+5), clean approves (+10)
- Acceptance-criteria check against linked work items or issues
- Diff-only by design — no test execution, no AST, no SAST. Pairs with your CI, doesn't replace it.
02.work-items
Work-items intelligence
Natural-language access to your Azure DevOps work items. Query by ID or in plain English — the chatbot translates to WIQL, surfaces missing acceptance criteria, and shows what's blocking what.
- Today: Azure DevOps (work items, queries, linked PRs)
- AC compliance check during PR review against linked work items
- Roadmap: Jira and Linear connectors
03.service-catalog
Service catalog
A living catalog of who owns what — services, contacts, tech stack — maintained by an agent that reads your code, not a wiki page someone forgot to update.
- Read the catalog in plain English ("who owns the payment service?")
- Feeds the planner so agents land work in the right repo with the right contacts
- Stays in sync with the code; drift detection on the roadmap
04.semantic-layer
Engineering semantic layer
Your engineering knowledge — wikis, code, work items, service catalog — woven into a single layer the assistant and other agents query. Updates flow event-driven: a wiki edit, a new PR, a doc change — all surface without a manual refresh.
- Today: Confluence and Azure DevOps Wiki, plus per-service docs the platform writes itself
- Tiered: definitions, facts, runbooks — surfaced when the question warrants
- Knows your team's terminology, including aliases and short forms
- Roadmap: Notion, GitLab Wiki, GitHub markdown
05.platform-assistant
Platform assistant
A chatbot that knows your platform — work items, code, wikis, service catalog, and your Ignix session history. Each turn is grounded in your engineering semantic layer before it reaches the model, so answers stay rooted in your team's actual context, not the model's training data.
- Asks the right tool, escalates to a reasoning model only when the question deserves it
- Can spawn a coding task from a description — same orchestrator that runs scheduled work
- Web UI today; in-IDE and in-Slack on the roadmap
// how it runs
How it actually runs.
A few things that change what AI-assisted delivery feels like in production rather than a demo.
One platform, three VCS hosts.
Azure DevOps, Gitea, GitHub — same review rules, same chatbot, same orchestrator. Move a repo between hosts without re-onboarding the assistant.
A pipeline you can read, not a black box.
PR Review runs as a fixed sequence of stages with LLM calls bounded to specific ones. Predictable cost; auditable trail.
Best model per job.
Reasoning models for review and analysis; fast models for tool calls and code edits. Provider-agnostic — switched per agent without vendor lock-in.
Documentation that maintains itself.
Per-service API and database docs are generated from the code by a built-in agent, then feed back into the assistant. Drift is harder to hide.
// trust model
Trust model.
Everything is a pull request, with logs, diffs, and artefacts attached. Runner pods are ephemeral and network-isolated — egress to Azure DevOps only. Scoped PATs, no broad credentials. Bounded autonomy: at most two correction cycles before a human is paged, and at most three clarification rounds before the planner forces a decision. The decider routes the happy path with deterministic code; LLMs only enter for genuinely ambiguous calls.
This is how Context-Driven Development looks operationalised. See the methodology →
// what it isn't
What it isn't.
- Not an IDE copilot.
- Not an auto-merge bot — humans merge.
- Not a replacement for your test pyramid or your SAST.
- Not a black box — every step lands as a PR with logs.
// access
Ignix67 is invite-only during private beta. If your team is delivering with AI assistance and you'd like access, send us a brief — we'll respond within a business day.
Request beta access →