Impact-aware execution

Impact analysis gives coding agents a smaller validation loop without weakening ayni check as the repository completion gate.

Commands

Explain the checks affected by the current checkout relative to an explicit Git base:

ayni impact show --base <revision>

Run that plan:

ayni impact run --base <revision>

impact run executes in the current workspace and execution environment. impact show plans without running quality tools. Both commands require an explicit local Git base and resolve it directly to a commit. The candidate is the final local working-tree state relative to that base: commits through HEAD, tracked index/worktree changes, and non-ignored untracked files. Index and worktree deltas are not reported as separate hosted review concepts. The deterministic fingerprint includes tracked binary/mode changes and framed untracked path, type, executable-bit, link-target, and content evidence. There is no hidden baseline, implicit merge-base, remote fetch, pull-request lookup, or GitHub/GitLab/Bitbucket integration.

Conservative planning

Ayni maps changed paths through configured language roots and adapter-owned package topology. Rust Cargo workspaces and npm Node workspaces include the full reverse-dependency closure, so changing a library also selects checks for workspace packages that depend on it.

Each selected check records machine-readable inclusion reasons. When ownership, topology, or tool scoping is uncertain, Ayni broadens from file or package scope to the configured root and records the uncertainty. Contract changes broaden all configured targets. Environment-lock changes conservatively broaden every configured target because the lock is one repository environment contract. Missing impact capability never silently drops required work.

Signal strategies remain conservative:

• tests include directly affected and reverse-dependent packages when supported;
• coverage and mutation broaden whenever narrower evidence would be unsafe;
• size and complexity use changed files when the adapter can measure them exactly;
• dependency checks follow manifest and internal-topology changes;
• deleted, renamed, copied, ambiguous, or configuration-sensitive inputs broaden scope.

Local Git requirements

Impact uses the local git executable and ordinary worktree data only. The configured repository may be a subdirectory of a larger worktree; paths and the candidate fingerprint are scoped to that configured repository. Ignored untracked files are excluded, while tracked files remain visible regardless of ignore rules. Unresolved conflicts, unsafe or non-UTF-8 paths, and unsupported filesystem entries fail closed. Ayni never contacts or interprets a hosting provider.

Results are not completion evidence

impact run writes its separate result below:

.ayni/impact/last/impact.json

The impact envelope contains the resolved base and candidate, changed inputs, selected checks and reasons, uncertainties, typed signal rows, and execution issues. It explicitly states that repository completion is still required. Impact evidence never overwrites .ayni/last/signals.json or focused verification evidence. Ayni refuses impact planning unless this generated slot is ignored by Git. Managed runs resolve and fingerprint the complete local-Git candidate on the host, then create a bounded source snapshot and hand the container an immutable, read-only serialized plan; the copied managed workspace therefore does not need Git metadata. Before checks, the inner run proves the snapshot fingerprint matches the host handoff. It fingerprints the same manifest again after checks, while the host recomputes the Git plan after the managed launch. The inner artifact and output remain provisional until those checks and full artifact/exit validation pass; copy mutation, process interruption, or host drift therefore cannot publish evidence for a different candidate. impact run also removes prior impact evidence when a new invocation starts, preventing an earlier result from remaining current after setup or execution failure.

A successful impact run means only that its selected plan passed. Before calling work complete, always run:

ayni check

Output

Human output emphasizes selected scopes, reasons, broadening, and the final-gate reminder. --output json emits one deterministic document on stdout; progress and diagnostics use stderr. Markdown output is suitable for summaries while retaining the same prominent non-completion marker.