Wiki · Backlog
Extract Aider — git-first coding, repository context, and validation
Shelf
Backlog
Source
docs/backlog/archive/WI-022-extract-aider.md
Status
done

Proposal (rationale)

SWE-agent supplies a minimal coding loop and OpenHands supplies a sandboxed, multi-run product, but neither was selected to answer the local git lifecycle as its primary question. rungs writes into a developer’s existing repository, so the boundary between model-selected context, an applied edit, validation, commit, diff, undo, and human control is unusually relevant to its own product.

Aider is selected as a git-first local counterpoint. Its public project documents repository mapping, automatic lint/test integration, and git commits and undo. The extraction must establish how those features actually compose and what repository state they assume; the project description alone is not implementation evidence.

Decision

accepted then planned — 2026-08-15. The user directed the follow-on research to continue after WI-019 and WI-020/WI-021 established the product and evaluation extraction workflow. This item is the git-native product child of WI-018; its written requirements and acceptance criteria are now the execution contract.

Plan

Requirements

  • Pin SHA, licence, date, measured scale, default configuration, and the coder/git/test packages read.
  • Trace one change request through repository-map or context selection, model request, edit format, patch/application, lint or test feedback, and commit or user-visible diff.
  • Establish the dirty-tree, branch, ignored-file, generated-file, and undo boundaries, including what is automatic versus opt-in.
  • Trace malformed edits and failing validation back into the loop, including termination budgets.
  • Identify where the user approves, steers, rejects, or recovers a change and what state survives restart.
  • Map findings against narrow-anchor, structural-gate, land/candidate, agent-facing-interface, and prompt-writes-artifact patterns without editing the catalogue.

Impacts

  • One durable/local-product extraction and index row.
  • Direct evidence about repository ingress/egress and validation that can be compared with WI-016 in WI-028, without changing either source extraction.

Approach

Use one ordinary edit path and the default git integration. Treat repository mapping as a context selection mechanism, not a claim about semantic understanding. Read architect/editor or alternate edit formats only where they materially change ownership or validation boundaries.

Acceptance criteria / tests

  1. One request is traced from context selection through repository mutation and its git result.
  2. Dirty-tree, commit, undo, ignore, and validation-failure semantics cite named implementation and tests.
  3. The extraction states which safeguards are defaults, optional settings, or user workflow conventions.
  4. Repository-map claims distinguish measured structure from model-facing summaries.
  5. Opinion is labelled; catalogue work remains WI-028; rungs check and site links pass.

Out of scope

  • Model leaderboards, benchmark scores, or choosing an edit format by performance.
  • IDE integrations beyond the boundary needed to trace the core local path.
  • Comparing Aider as a product against OpenHands or SWE-agent inside this extraction.
  • Catalogue, module, or CLI changes.

Execution

Completed 2026-08-15 on feature/WI-022-extract-aider.

  • Pinned Aider-AI/aider at 5dc9490bb35f9729ef2c95d00a19ccd30c26339c, recorded Apache-2.0, and measured 691 tracked files, 147 Python files, 88 test-named/test-directory files, 160 text/docs files, and 38 coder implementations.
  • Added aider.md and the follow-on index row. The extraction traces repo detection/map/context, model request, exact edit application, dirty and ignore gates, lint/test feedback, automatic/user commit, diff, undo, and restart boundaries.
  • Validation: node src/cli.ts check — 20 pass, 0 fail; site npm run build — 82 pages; site npm run check — 0 Astro diagnostics and 1,048 internal links, 0 broken. No real model request or repository mutation was run; cited Aider tests provide executable evidence at the pinned SHA.

Review

Self-review completed 2026-08-15 against every acceptance criterion:

  1. Pass. The extraction traces one ordinary edit from repo-map context through edit parsing, mutation, validation, and Git result with pinned implementation paths.
  2. Pass. Dirty-tree pre-commit, Git/Aider ignore handling, auto/user commit, guarded undo, malformed edits, lint/test failure, and reflection limits cite source and named tests.
  3. Pass. Defaults (auto_commits, dirty_commits, auto_lint), opt-ins (auto_test, shell confirmation), and workflow conventions are distinguished.
  4. Pass. Repo-map syntax/tag measurements are explicitly separated from semantic understanding.
  5. Pass. Opinions are labelled, catalogue work is deferred to WI-028, and gates/site links pass.