Roadmap

The orchestrator that landed on chad-dev was, by design, a minimum viable contract — enough surface area to prove the sandbox + sub-agent + autonomy story end-to-end, no more. Several improvements were deliberately held for follow-up; some have shipped since the initial cut. Status is current as of the Changelog head.

Shipped

Item	Where	Notes
Nested-spawn isolation	gha substrate	Sub-agents can spawn onto a per-job GitHub Actions runner instead of sharing Chad's container. Phases A–C all shipped 2026-05-06. See Substrates.
Async sub-agent spawn	chad-spawn --async --substrate gha	Returns task id immediately. chad-spawn-poll cron (every 5 min) reconciles result.json from the runner branch. Branch retention via chad-spawn-gc weekly cron.
Hermes-style memory curator	chad-memory-curator weekly	Inactivity-gated, snapshot-first, draft-only consolidation pass over the LTM. See Memory.
Self-modify safe-list (cron tunings)	chad-proposal-apply daily	Reads structured proposals from chad-budget-audit, applies narrow safe-list (timeoutSeconds, maxOutputTokens within ±2× current). Gated by chad-action-gate chad_self_modify_cron.
NVIDIA fallback for OpenAI-compat kinds	codex and opencode kinds	Falls through to integrate.api.nvidia.com when OPENAI_API_KEY is absent.
Premium escalation	chad-premium + AuthContext	Anthropic Claude routing for tasks where Nemotron's reasoning depth feels shallow. See Premium escalation.
Open WebUI front-end	chad-shim.py + docker-compose	Chad as an OpenAI-compat model. See Front-ends.

Phase 2 — held for follow-up

These are tracked in chad-readme.md §10. Each is a deliberate defer, not an oversight.

k3s-pod substrate

A per-spawn pod in a local k3s cluster would give the isolation of the GHA runner and the L7 policy enforcement of the local substrate. Today you pick one or the other. The blocker is operating a k3s control plane on the host machine — possible, just real overhead for a single-operator setup.

When it'll matter: the moment a sub-agent kind genuinely needs both isolation and L7 enforcement. The fitness kind doesn't (read-only). The codex/opencode kinds don't (in-flight, sandboxed by GHA). A coder kind that writes to multiple repos in one spawn would.

Cron DSL

chad-intake --from cron takes a task file today. A YAML DSL would let Chad register new crons at runtime — "every Tuesday 9 a.m., spawn a reviewer against my open PRs."

Why deferred: the autonomy surface for a self-registering cron is non-trivial. The current set is operator-curated; allowing Chad to register his own crons means another chad-action-gate action type and a careful safe-list.

openclaw-sandbox.yaml split

The single 346-line policy file is at its complexity ceiling. Splitting per-domain (inference / github / messaging / browsing) would make policy review tractable and let presets compose more cleanly.

Why deferred: the merge tool that compiles per-domain files back into a single OPA policy doesn't exist yet. The cost is higher than the benefit until the file gets larger.

Multi-Chad scheduling

One k3s cluster, multiple Chads, a shared scheduler that routes by kind + load. Needed the moment a second user shows up.

Why deferred: there is no second user yet. When there is, the Open WebUI session split (one Chad per identity) is the first thing that has to land.

MCP hub

Expose the orchestrator helpers (chad-spawn, chad-budget, chad-collect, …) as MCP tools so a non-Chad agent (Claude Code, a local editor) can drive the same sub-agent contract without Chad in the middle.

Why deferred: the orchestrator helpers' contract is binary-stable today (stdout sentinels, JSON in result.json). Wrapping them in MCP is straightforward; the open question is whether non-Chad callers should be allowed to bypass the action gate, and the answer is "no" but the implementation needs care.

Diff-checked, compressed backups

chad-backup-to-github does per-file sha-diff today. A Phase-2 consolidation would do a single tree-sha diff per backup run — roughly one API call instead of one per file. Saves ~400 calls per day on the current backup set.

Why deferred: the per-file path is correct, just inefficient. The cron schedule (every 6 h) absorbs the inefficiency for now.

Webhook-based completion for async spawns

Today chad-spawn-poll polls every 5 minutes. A webhook receiver in Chad's gateway would replace the poll with push semantics — zero reconciliation lag, no work when nothing's in flight.

Why deferred: the gateway doesn't speak inbound webhooks yet. Adding that surface is a non-trivial security review (incoming HTTP from GitHub is its own threat model).

Out of scope, on purpose

A few things that won't be roadmapped because the design says no:

• Chad-as-a-product. This is a single-operator agent built against a private state repo. "Reproducible by anyone" is in Reproducing, but a hosted multi-tenant version is not the design.
• github_pr_merge: auto. The autonomy roadmap says block is likely permanent. A merge is irreversible and synthesizes CI signal that humans should still gate on.
• chad_self_modify_identity: auto. Identity files (SOUL.md, IDENTITY.md, USER.md) are operator-owned. Chad can propose; only the operator applies.

How to read this page

The shipped list is what chad-dev does today. The Phase 2 list is what might land next, in some order, when there's reason. The "out of scope" list is what won't, and the reasoning is the interesting part — every "no" is a place where the design intentionally chose limits over capability.