ADR-0008: Configurable Review and Merge Policy per Ticket

Status:

Accepted

Date:

2026-10-01

Context

Different work warrants different oversight: a security-sensitive change needs a human review; a documentation typo does not. Ballet must support strict projects and highly autonomous ones, and the decision to reduce oversight must stay with humans.

Decision

Each ticket has an execution policy with:

  • review mode: agent (the pipeline’s Review stage only) or agent+human (additionally a human approval on the pull request);

  • merge mode: auto (merge when all gates pass) or manual.

The default for new projects is agent + auto: Ballet’s goal is unattended operation, with humans involved through questions (Questions and Human Involvement). The agent Review stage always runs (ADR-0014: Separate Agent Session per Pipeline Stage).

Reviews and merges take place on the git platform; Ballet applies the policy through forge adapters (ADR-0007: Code Review on Git Platforms via Forge Adapters).

Defaults come from the project’s process profile. Only a human with the engineer role (or higher) in scope may change a ticket’s policy. The planner may suggest a policy; the suggestion is applied only when a human approves it.

Alternatives Considered

Always human review

Advantages:

  • Maximum safety.

Disadvantages:

  • Humans become the bottleneck; work stops whenever nobody is awake.

Project-level policy only

Advantages:

  • Simpler.

Disadvantages:

  • Cannot distinguish risky from trivial tickets.

Decision Criteria

Safety, throughput, human control.

Rationale

Autonomy by default serves the core goal of unattended operation; quality is protected by independent pipeline stages rather than by human availability. Per-ticket policy still lets humans demand oversight where risk warrants it.

Consequences

Positive

  • Work continues without humans; strict oversight remains available per ticket.

Negative

  • More states and combinations to test.

Risks

  • Auto-merge with agent review can merge flawed changes; mitigated by gates (tests, write-back) and audit.

Follow-up

  • Define the reviewer-agent skill.

References