Skills inventoried
55 executable; 37 knowledge-only or awaiting reviewed activation contracts. Discovery and RAG indexing never grant execution.
Evolution & roadmap
From one agent
to a platform.
Forge went from a single hardened production dashboard to a cloud-era, multi-vertical autonomous platform in about a month. Drag the slider — or use arrow keys — to walk the timeline: each version, what actually shipped, what is being qualified next, and why it matters.
◷
Interactive timeline
Every version, and what it unlocked
Shipped
Current release
Repository complete / qualification open
Planned
Drag the knob, click a dot, press ◀ ▶, or ▶ Play.
Status rule: green means the milestone's release gates are complete, not merely that its code exists. v3.2 is now a completed local baseline; Cloud Alpha is the current internal release. v3.3–v3.6 remain amber because their qualification, distribution and production-evidence gates are open. Engineering and v3.7 remain active work.
◎
Current truth
Repository foundations vs release gates
The current release candidate converges the architecture around one six-plane contract spine: Intelligence → Context & Knowledge → Capability → Execution → Governance → Assurance. It does not turn external evidence into a repository checkbox.
Shared contracts
Thirty registered contracts with golden/rejection discipline and three compatibility deprecations.
State spine implemented
Context budgeting/diff, World State v2 and compatibility, evidence-derived confidence, normalized Agent/Capability projections, sealed checkpoints and the execution-event read model are repository foundations.
Operational gates now
All seven RAG workspaces completed a verified post-recommission refresh; ongoing canonical edits correctly return the derived indexes to fail-closed canonical_revision_mismatch until the next source freeze and refresh. Protected Google Sheets and one Atlassian Cloud source still need owner access or an explicit retirement decision. The specialist-preview host producer now registers exact source authority; future producers must use the same boundary. Unity completion still needs the independently operated signed feed and pinned public trust. The 2D Iconic Item family is repository-complete, but its exact Unity apply capability remains candidate-only until the approved disposable-project qualification closes.
External qualification remains amber
Environment fingerprints, real disposable and accepted-ticket runs, approved writable scopes, owner activation, independent review, signed/notarized installed builds, certification, publication and commissioning must come from their owning systems and people.
Sequence state before speed.
World State and replanning stay above cache work. Provider/runtime KV or exact prefix caches are performance-only. A semantic answer cache is deliberately out of scope because generated answers cannot safely cross freshness, provenance, workspace, authority, policy and validation changes.
→
Upcoming
The next governed feature sequence
The next releases turn implemented contracts into independently evidenced operations. They do not bypass qualification by adding a green badge.
Trust & source closure
Authorize protected Sheets, decide the remaining Atlassian source, freeze canonical inputs and refresh RAG. Onboard the separately operated Unity Ed25519 attestor with a read-only feed and public trust only.
Qualification Operations
Add one operator view for feed freshness, exact capability/candidate/environment bindings, accepted-case counts, independent review and next governed action. It observes evidence; it never signs, certifies or promotes.
Real Photoshop & Unity campaigns
Qualify one exact capability family at a time on reviewed disposable targets—starting with the bounded 2D Iconic Item Unity apply candidate—then retain 20 distinct accepted cases for the exact corpus and an independent review bound to its digest.
Hybrid runtime proof
PostgreSQL as sole hosted authority, two authenticated workers on distinct hosts, signed failover/dead-letter/backpressure/isolation evidence, then signed and notarized distribution.
Measured model intelligence
Replace unknown model-card fields only with retained cost, latency, cache, privacy, structured-output and tool-call observations bound to exact model/provider/runtime/environment revisions.
Vertical commissioning
Commission the first Production, QA, Product and Art agents only after exact skill, capability, model, permission, qualification, publication and human-owner bindings pass.
Later, without weakening the spine.
Extend Code Intelligence into review and impact planning; expand the Production Knowledge Graph beyond characters; add tenant-qualified team inference and cross-team knowledge federation. Semantic answer caching remains out of scope.
↗
Where it's going
The through-line
One engine, compounding.
Every stop on the timeline is the same Fabric engine gaining a capability — never a new tool. Knowledge and confidence learned in Tech Art carried into Design, into the governed Faction Offers and 2D Iconic Item families, and into the in-development bug-fix engine and incident response. The cloud era didn't replace the local platform; it added a budgeted cloud lane beside it. That is the design bet: new work becomes a new agent, and new intelligence becomes a new capability behind the same contract — so the platform gets more capable without the surface above it moving.
Dive deeper: the architecture and the Fabric engine for how it's built, or the Architecture Master Plan for the whole platform on one page. Decisions are recorded as ADRs — the promoted, default-on governed specialist layer is ADR-0006 · Fabric Core Intelligence, with a persistent Settings kill switch and emergency process override.