ADR 0065: Autonomous Delivery With Less Ceremony

Status: Accepted for the 0.7 redesign; release acceptance remains separate

Date: 2026-09-19

Context

STAGE's safety obligations remain useful, but the operating instructions grew around individual failures. Seven delivery stages, a second Director state machine, repeated gameplay references, routine worktree and handoff guidance, and wording-sensitive tests increased maintenance without demonstrating better outcomes. Local checkpointing also depended ambiguously on human product acceptance. Release proposals left deterministic clerical work to the human.

The 0.7 audit found concrete defects alongside that process debt: Git inspection could refresh an index or misreport observation failure, revision inventory omitted root or merge semantics, output writers accepted linked paths, and manual publication could lack current successful quality evidence. Simplifying prose alone would leave those defects intact.

Decision

Recommend Understand → Change → Verify → Deliver for a solo director and an autonomous engineering agent. These are working steps, not forms or narration. Preserve all eight Safety Kernel obligations and optional-contract semantics.

Complete authorized engineering without repeated approval. After relevant mechanical checks pass, review the diff and make a task-only local commit while preserving unrelated staged and unstaged state. If safe separation is uncertain, leave the diff and explain the pending checkpoint. Human product acceptance, push, merge, release, and publication are separate decisions; authorization already given carries forward. Stop only an affected branch for material unresolved intent, authority, capability, or unauthorized operations.

Load gameplay, visual, quality, continuity, and Director references only when their triggers apply. Consolidate packaged gameplay guidance. Keep the Director's Flow as an optional attention board and its starter at five fields. Raw ideas remain distinct from approved work. New or materially changed gameplay needs an approved hypothesis and human play before broad expansion; ordinary fixes do not reopen settled approval. Choose isolation and handoffs for actual needs.

Use one packaged read-only Git adapter and shared path validation where needed. Treat observation failure as failure, preserve external outputs when publication fails, and never restore someone else's concurrent changes. Keep successful CLI fields, flags, exact supported schemas, and artifact migration behavior.

Use a shared current-main quality eligibility guard for both documentation publication and draft release proposals, including manual recovery. Check first with read-only credentials and recheck freshness before consequential writes. Keep the scoped public mirror, fast-forward pushes, and draft-only Release Please. The agent completes deterministic release edits and evidence before human review.

Test executable and structural contracts, not incidental prose or copied tool versions. Keep immutable action pins, publisher and permission boundaries, package closure, compatibility, and version consistency. Pin the resolved maintainer Python environment, retaining consumer validator ranges. Report the website's complete dependency graph, distinguishing valid advisories from a failed query; advisory findings remain maintainer input.

Mapping, Unity bootstrap, and research tooling receive correctness, compatibility, and maintenance support. New capability needs a named present consumer. Keep engine-specific mechanics in profiles and dated evidence in release records.

Relationships To Earlier Decisions

• Clarifies 0020, 0023, and 0050: ordinary delivery stays primary; optional infrastructure and guidance add no universal obligations.
• Preserves 0022 and 0051: claims include real consumers and authority is assessed at the actual operation.
• Preserves 0052, 0053, and 0054, with one conditional packaged gameplay reference.
• Supersedes the required-looking state transitions in 0055, routine isolation defaults in 0057, and routine handoff/recovery-test ceremony in 0058.
• Preserves 0056 and 0059: gate strength and explicit companion mutation boundaries remain intact.
• Clarifies 0062 and supersedes the human-clerical-completion requirement in 0064. Release Please remains proposal-only; a finished proposal precedes human acceptance.

Consequences And Evaluation

Release as plugin 0.7.0 / method 0.7, retaining existing artifact compatibility. Instructions remain model-neutral. Paired fresh-context GPT-6 Astra probes compare 0.6.11 and the candidate on preservation, progress, stopping, truthful evidence, and interruptions. Preserve every run and investigate failures before retesting. These support bounded release confidence, not general performance. The 0.7 release record owns dated results and limitations.

For the next 8–12 real game changes, use the existing evaluation context to record brief observations of human attention, fair inspection, useful decisions, rework, escaped defects, and STAGE maintenance. This observation period does not block ordinary delivery or 0.7 release. The owner has not decided 1.0.