Autonomous-runner repo charter: timing and scope boundary

[kyle-sexton]

Greenfield now vs after the 2→3 trust loop is earned (Boris's trap: no agent-count scaling before the loop earns trust). Scope boundary vs plugins (interactive-upstream / autonomous-downstream split). Evidence: RESEARCH-headless-agents.md 8 seams + human-merge-gate finding; RESEARCH-peer-frameworks.md composition recommendation; work-items lease layer reuse; interview-checklist deferred-branch entry.

---

Reference: Boris Cherny, "Steps of AI Adoption" (Google Doc)

[kyle-sexton]

🔒 claim: wayfind-session-20260717-cont-8acf2066

[kyle-sexton]

Resolved: charter now, build on trigger — charter locked (scope + constraints), no build before the 2→3 trust loop earns it.

Q1 timing: charter now; build trigger = C2 promotion fires (per #243) + clean autonomous drain on existing executors behind the #240 contract, OR existing executors hit an isolation/concurrency/platform wall.

Q2 scope boundary: queue-contract split (interactive-upstream / autonomous-downstream). Plugins own all interactive stages + trigger adapters (#240: adapters live where signals land). Runner = autonomous drain only: lease-claim → execute in isolation → verification gates → merge policy per class → escalation to humans. Runner never decides what to build; it is one executor behind the queue contract.

Q3 substrate: self-run primary, hosted via adapter. Self-operated CLI/SDK surface class is the only place merge policy is ownable (every hosted issue-to-PR agent keeps a deliberate human-merge-gate — RESEARCH-headless-agents), so C2 auto-merge promotion is only reachable self-run. Hosted surfaces reachable via the invocation-adapter seam but capped at human-gated merge; hosted vendor-managed isolation remains a legitimate L3 instance (e.g. C5). Contract names surface classes, never vendors.

Inherited constraints (unchanged): ≥L2 fail-closed substrate (#245); per-class gates/merge/cost/escalation (#243); queue+lease reuse from work-items — no second claim mechanism; no queue bypass, audit trail is the trust loop (#240). Abstraction bar: the 8 seams (invocation adapter, output envelope, queue+lease, isolation policy, verification gate, merge-policy toggle, observability+cost, session/resume+caps). Anti-goals: set-and-forget, ungated autonomy, pull_request_target-class footguns.

Deferred to design/architect: adopt-vs-reimplement composition spine, multi-backend isolation at launch, lifecycle depth, escalation UX, repo/packaging topology.

Sharpens #243: escalation column mechanism = the runner's escalation seam (severity/class-routed); UX stays fog → design stage.

Boris alignment verified against session captures: build-on-trigger = step-3 trap honored (no agent-count scaling before trust); queue-drain = "let Claude kick off Claude" with audit trail; self-run = only path to the step-4 closed loop, chartered at 2→3, no step-skipping.

Full charter: .work/ai-adoption-ladder/design-threads.md T4.

[kyle-sexton]

Source anchors (added for durability — the design threads re-anchor to this artifact):

• Boris Cherny, "Step & your role" (Jul 16, 2026) — Google Doc (reliable copy): https://docs.google.com/document/d/1R91ayvj7uvlxgNi--__2-Bf3w8x5r1nF-xIBN7ds8Ns/edit
• Original claude.ai artifact (intermittently erroring as of 2026-07-17): https://claude.ai/code/artifact/bfdfaef9-bc62-4dfe-ba9e-c58a26c9accf