#GMMWorks / How GMMCode works
How GMMCode works
GMMCode is the delivery product inside #GMMWorks. Requests arrive through ITIL-style service tiers, the engagement's shape decides which specialists join the pod and which gates apply, a governance mesh checks every artefact, and everything lands in an evidence bundle you can export.
The LLM is swappable.The governance is not optional.
A front door, an AI pod, a governance mesh, an evidence bundle.
- Engagement shapes for build, review, audit, documentation, migration, forensics, continuous compliance, RFPs, hotfix, maintenance, problem management and release trains. Each shape declares its roles, gates and deliverables.
- Specialist roles including business analyst, architect, developer, SDET, reviewer, security, DBA, UI developer, release manager and SRE. No role approves its own work.
- Works with the LLM of your choice. The LLM is swappable without touching governance, standards or the audit trail. Today it runs on AWS.
The build gates, in the order a story meets them.
A failed gate stops the story and routes it back with the reason on record. Work continues only when the gate passes or a person decides.
Design
Architecture, data and acceptance criteria agreed before any code.
Implementation
Built test-first. A red integrated build blocks this gate.
QA, review, security
Three independent checks of the same change, each with its own verdict.
Scope adherence
Built what the story asked for, and only that.
Release readiness
Deployable, observable and reversible.
Change authorisation
A person approves, with the evidence bundle in front of them.
Every role reads what the customer actually asked.
An AI role that sees only a one-line ticket guesses. In GMMCode every role decides from the full context of the work.
The customer's own words
The original request, not a paraphrase of it.
The spec and the scope
Acceptance criteria, design notes and what is out of scope.
The ticket history
Earlier verdicts, rejections, follow-ups and approvals.
The conversation
What the pod and the people have already said and decided.
Checks that cite the rule, and judge only what applies.
Every artefact goes through a mesh of specialist checkers: tools first (scanners, SBOM, licence policy), then small governance language models that must cite the standard they apply. One registry decides which checks may block a delivery.
- 35 specialist checks run on every code change. Nine objective standards checks may block; 26 judgement checks advise.
- In scope, or a follow-up. A finding outside the story's scope does not silently reject the work. It is recorded as a follow-up, filed in Jira, and listed to you when the work closes.
- Provenance on every verdict: model, prompt hash, registry version and tool versions.
- Supply chain in the bundle: OSV scanning with KEV and EPSS, a CycloneDX SBOM, licence policy.
Built in parallel. Merged safely. Fixed properly.
The defects that hurt are the ones between features that each passed their own tests. GMMCode treats integration testing as part of delivery.
It fixes the app, not the test
When a test fails, the developer agent diagnoses the failure from the evidence (test output, logs, the diff) and fixes the application, rather than weakening the test until it passes.
Every story ships an integration test
GMMPlan writes the test plan; the QA gate refuses a story without its planned test.
A merge train
Parallel work is integrated in dependency order, and a red integrated build blocks the implementation gate.
Results on the board and in Jira
Integration results ride on the board's cards and are written back to Jira, with a project CI gate and a coverage census that can only go up.
The approval is the product.
AI does the work; people make the calls. GMMCode opens on what is waiting for you, explains it in plain words, and records your decision in a way an auditor can check.
- Policy, not habit. Approval policies per decision type, segregation of duties enforced at decision time, quorum groups.
- Nothing silently expires. Due times from the tenant SLA, reminders, an escalation chain, delegation and out-of-office.
- Signed record. Every decision is pinned to the artefact version it approved and lands in the evidence bundle.
Dev and test environments can auto-decide approvals; every such decision is labelled "bypassed". Production never bypasses.
It approves, and it rejects.
Unedited captures from a live run: roles signing their own gates, and an architect and a code reviewer sending a change back because the design missed an edge case.




#GMMWorks builds itself by its own rules.
The platform's own backlog runs on the same board. Status as of 3 Oct 2026: 509 of 624 stories and bugs done (82%).
Headline counts are live from the GMMCOD, GMMUI and GMMDSK projects on 3 Oct 2026; the per-area bars are the 1 Oct 2026 snapshot. Ops, defects and self-hosting work (105 of 134 done) is left out of the chart. Integration testing is the newest work and the furthest from done.
See GMMCode on your own backlog.
A pilot starts from your Jira backlog. GMMPlan turns one requirement into stories, GMMCode builds them on your board, and you approve each gate.