otto-impact-assessment

Shared impact assessment — every resolution path

Before claiming success (Resolve: fixed) or approving a PR, check what the full final diff could break beyond the requested change. This applies to every mode (Resolve: author, existing-PR review, ticket-only, and review-remediation), including skip_push and workflow-owned publication. (Resolve: ticket-only and review-remediation still do not require an inherited repro; their publication/branch permissions are unchanged.)

1. Inspect the candidate and its consumers. Start the assessment while investigating, then refresh it after the last edit. Record the target branch's full tip SHA as base_sha and the candidate's full HEAD SHA as head_sha. Inspect git diff <base_sha>...<head_sha> plus staged, unstaged, and untracked source/test changes. Include configuration, dependencies, and workflows; runtime prompts count as behavior changes even when stored in Markdown. In review modes, inspect the whole PR, not only edits made in this run. Follow shared helpers to their callers and existing tests. For each meaningful risk, identify the changed files, affected behavior, consumers, invariant to preserve, and focused check. A passing original repro is not coverage for every consumer.

2. Exercise the relevant boundaries. Select checks from the actual change:

   Changed behavior	Required focused coverage when applicable
   Shared startup, configuration, auth, or serialization	Existing/default configurations and other affected consumers; exercise configuration creation, transport/decoding, and real process startup together. Correctly constructed objects supplied directly to a unit test do not cover that path.
   Sandboxing	The applicable real OS sandbox with its intended gates enabled. A disabled gate or an unsupported/skipped sandbox test leaves that boundary uncovered.
   UI state or interaction	The reported journey and affected adjacent journeys on the candidate build, including relevant selection, persistence, or reconnect behavior.
   Environment-derived defaults	Ordinary configuration and relevant ambient variables populated; apply the agent's hermetic check (Resolve: 2B.5).
   Performance or timing	Measure the claimed quantity under the relevant workload. Distinguish simulated-clock tests from elapsed-time measurements; functional success alone does not establish latency or throughput.

   Use existing tests where they cover the invariant; add a behavioral regression test only for a meaningful gap. Tests of the requested change follow their mode's baseline requirements (Resolve: review-remediation still skips fail-before proof). Checks protecting previously correct behavior may pass on both base and candidate; do not manufacture a failing baseline for them. Run the focused checks for each affected consumer/boundary, even when they live in a different module. Do not run the full repository suite locally. Do not weaken isolation or change global rollout settings to obtain a passing result.

3. Record observed results in impact_assessment. Each risk names its check, result (passed, failed, blocked, or not_run), and retained evidence. Reference actual command/output or CI run artifacts; include the tested revision/build, dependency pins, and relevant environment/feature gates. Confirm the test imports and exercises this candidate, not another checkout, installed copy, stale SPA, or older CI runtime. Record worktree_state as clean only when the tested source/tests are clean; otherwise name the uncommitted source/test/support files and their content hashes. Artifact-only dirt can be noted separately. A HEAD SHA alone does not identify dirty code.

   Distinguish component tests, real process/sandbox checks, and live bot checks. An unrun, skipped, xfailed, or setup-failing required check is not a pass: name the missing boundary and reason in uncovered_boundaries. A large green test count or an unrelated green CI job does not fill that gap. Evidence references describe what was retained; never invent an execution record.

4. Refresh before delivery. Re-read the final diff, Git status, and current base/head before every final handoff or review verdict. For an existing PR, compare with the live GitHub base/head SHAs, not just local refs. If the author moved the head, follow the mode's update/retry procedure without approving from old results. After a retry, new PR head, merge/rebase, CI/Polly fix, or changed code, assertions, dependencies, or relevant environment, rebuild the assessment and rerun affected checks. Reuse a result only when its tested contents and context still match; retain its original identity rather than relabeling old evidence as a new execution. Commit hooks can change tested files too. Keep earlier evidence and the saved checkpoint; never reset work to make the identities match.

5. Let gaps affect the verdict. Success (Resolve: fixed) requires the original behavior to be proven as its mode requires, all identified regression risks checked, and no uncovered required boundary. A discovered regression must be repaired and checked before success or approval. If a change is only partly verified, preserve the work, use the partial outcome (Resolve: partially_fixed), and name the failing/unrun checks in remaining_work; use needs_more_info when verification cannot establish a resolution. Setup failures are blockers, not proof of a product regression. A missing recording alone still follows the agent's recording rules.