Roadmap

Roadmap

ai-core-kit builds itself in eight phases, driven by /ack-build over the data-driven config bootstrap/ack.bootstrap.yaml. This page is the honest status of each phase, taken from the done and gate flags in that config. It tells you exactly what works today versus what is a contract or roadmap item.

Accuracy note. A frozen contract is not the same as a working runtime. P3 froze the interface that P4–P6 depend on, and those contracts are stable and documented — but the engines that consume them are not all complete. Where this page says “contract / partial / roadmap,” treat the feature as not fully working today.


Status at a glance

PhaseTitleGateStatus
P1Ground-truth + repo tree + license ledgeryesShipped
P2.claude/ core + team agents + settings.json + frontmatter linternoShipped (skills/agents/commands + linter)
P3Frozen contract — manifest schema + questions.yaml + render-engine contract + /ack-inityesShipped (the contract is frozen)
P6 (telemetry engine)Offline aggregator + pricing.jsonnoShipped (engine), child wiring partial
(bootstrap-config)ack.bootstrap.yaml + schema + /ack-build orchestratorShipped
P4Archetype template sets (deep) + completeness rubric + license NOTICEsyesPartial / roadmap (render engine is a frozen contract; not all template sets complete)
P5Contract template + 3-mode manifest-driven gate hookyesPartial (hook template exists, proof-of-concept only)
P6 (child wiring)Promote orchestrator cost stub to real numbers; wire telemetry into childrennoPartial
P7Discovery — sources.yaml + proposals + scheduled-PR Action + /discovernoRoadmap
P8E2E + META-hygiene + idempotency + per-feature cost (final gate)yesRoadmap

Shipped

These are complete and verifiable in the repo today.

P1 — Ground-truth and license ledger

Reference repos cloned and verified against their actual LICENSE / LICENSE.txt files; the license ledger (docs/REFERENCES.md + THIRD_PARTY_NOTICES.md) is authoritative. Gate phase, marked done.

P2 — .claude/ primitives and the linter

The kit’s own agents, commands, and skills, plus scripts/lint-frontmatter.py (the frontmatter linter). The full set of CHILD-payload skills, language packs, RPI agents, and the RPI/product commands is authored and present under templates/. See the Skills Catalog and Commands Reference.

P3 — The frozen contract

The stable interface every downstream phase depends on. Four frozen artifacts:

  1. Manifest schemaproject.manifest.yaml shape (schema_version: 2), with JSON-Schema validator. See Manifest & Interview.
  2. Question banktemplates/interview/questions.yaml, 39 deterministic questions with ask_if/skip_if gating. See Manifest & Interview.
  3. Render-engine contract${VAR} substitution, _when.* conditional inclusion, structural idempotency. See Render Engine Contract.
  4. /ack-init — the producer command. See Commands Reference.

Gate phase, marked done. The contract is frozen; the engines that consume it are not all complete — see Partial / Roadmap below.

Offline cost telemetry (engine)

telemetry/aggregate.py (stdlib-only) and telemetry/pricing.json compute cost offline from transcript assistant.message.usage lines times a versioned pricing map. Cost is always offline — there is no live token/cost API in hooks (issue #11008; PostToolUse fires on tool turns only, missing roughly 61% of spend). See Offline Cost Telemetry.

Bootstrap config + /ack-build

bootstrap/ack.bootstrap.yaml (the machine-readable build plan) with its JSON-Schema validator, and the /ack-build orchestrator that reads it, validates it, and drives the phases. See META Build Configuration and /ack-build: META Build Orchestrator.


Partial

Authored as a contract or proof-of-concept, but not a complete runtime. Document and use these with that caveat.

P4 — Render engine and archetype template sets

The render-engine contract is frozen and documented (Render Engine Contract). What remains: rendering backend-api and fullstack deep against a per-archetype completeness rubric, wiring branch-aware interview logic end to end, and carrying a LICENSE.txt + THIRD_PARTY_NOTICES entry per vendored Apache-2.0 skill. The other three archetypes (monorepo, library-sdk, infra-iac) stay schema-known minimal-core in v1. Gate phase, not yet done.

P5 — Contract gate runtime

The 3-mode (block / warn / off) manifest-driven contract-gate hook exists as a template / proof-of-concept, not a finished runtime. The load-bearing footgun is baked into the contract: block mode must use exit code 2 plus hookSpecificOutput.permissionDecision: "deny" — the wrong form silently no-ops the guard. The gate is fail-open at runtime (corrupt/missing manifest -> mode off + stderr) and fail-closed at author-time (/ack-init validates before render). See Design-Contract Gate (P5). Gate phase, not yet done.

P6 — Telemetry child wiring

The offline aggregator engine is shipped (above), but promoting the orchestrator’s cost stub to real numbers and wiring telemetry into forked children (features.cost_telemetry: true) is partial. See Offline Cost Telemetry.

MCP and agent-teams wiring

Optional per-project MCP server wiring and the experimental agent-teams flag are authored as interview toggles, but their end-to-end provisioning is partial / TBD. See MCP Wiring; the experimental agent-teams flag (features.agent_teams: true) sets CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and has no dedicated page yet.


Roadmap

Not yet implemented. Presented as a contract / design only.

P7 — Discovery engine

The propose-not-adopt discovery loop: discovery/sources.yaml (seed sources), a hardened scheduled GitHub Action that opens proposal PRs, a proposal schema (format TBD), and /discover modes. Discovery proposes and never adopts — humans manually move proposals from proposals/ to adopted/. discovery.enabled defaults off in children, and the META engine is never copied into a child (forkability, invariant I7). Permission/security hardening (isolation, no self-merge, kill switch) and the kill/expiry/reject/dedupe mechanisms are deferred. See Discovery Engine (P7). Not yet done.

P8 — End-to-end verification

The final gate: fork -> /ack-init per archetype -> gate blocks -> /discover -> per-feature telemetry cost. Adds the META-hygiene invariant (root and .claude/ contain only META artifacts; no concrete child manifest or contract at the META root), two-run idempotency, and golden-fixture branch tests. The cost acceptance test runs against a captured transcript, not live orchestration, per the decoupled-cost rule. Depends on P4–P7. Not yet done.


Deferred / not ported

Tracked so nothing disappears silently. The full deferred / not-ported list is below.

  • Language/framework/DB packs planned for v2: fastapi-patterns, gin-patterns, axum-patterns, sveltekit-patterns, nuxt-patterns, mysql-patterns, sqlite-patterns, mongodb-patterns, sqlalchemy-patterns, drizzle-patterns, gorm-patterns.
  • docker-patterns wiring is blocked by a missing interview question (recommend always-render for backend-api/fullstack until the schema gap is closed).
  • Anthropic example skills available but not ported (adopt behind feature flags if needed): algorithmic-art, brand-guidelines, canvas-design, frontend-design, slack-gif-creator, theme-factory, web-artifacts-builder, webapp-testing, doc-coauthoring.
  • Proprietary doc skills permanently excluded: docx, pdf, pptx, xlsx — source-available, never read/copied/derived from. See Licensing.

Recommendation: propose new sources via the discovery engine rather than pre-emptively porting them.


See also