Skip to content

apply-engine-direction.md

Latest commit

 

History

History
103 lines (65 loc) · 7.45 KB

apply-engine-direction.md

File metadata and controls

103 lines (65 loc) · 7.45 KB
Raw

Apply-engine direction — diff-shape recipes + agent-in-the-loop substrate

Status: open · substrate growth (Steps 5–7 partial) lives in substrate-extraction.md (tiers 1–6 shipped). This plan owns the apply-engine / diff-shape recipe half of the richer-index synthesis.

Motivator: extend codemap apply from recipe-driven single-file hunks toward agent-in-the-loop row contracts (apply --rows -, diff-input, fixpoint loops) without crossing Moat A (SQL/recipe API) or Moat B (no re-extraction in apply).

Source: consolidated from research/codemap-richer-index-synthesis-2026-05.md § 6–8 (lifted 2026-05). Full triangulation matrices remain in that research note until all steps close.

Shipped (do not re-litigate)

Step Work Canonical home
1 Doc reframe (Floors split; apply vs verdict lints) roadmap.md § Floors, why-codemap.md
5 calls.{line_start, column_start, column_end} + call-shape metadata substrate-extraction.md Tier 1.A, architecture.md § Schema
7 (substrate) exports position columns + re_export_chains substrate-extraction.md Tiers 1.B / 2.2

Open steps (implementation sequence)

Step 2 — Three diff-shape recipes (S × 3)

Ship replace-marker-kind, migrate-import-source, add-jsdoc-deprecated — pure SQL + frontmatter; no engine/schema change.

Open: sequential PRs per tracer-bullets; hold organize-imports / missing-exports for wave 2.

Step 3 — Per-row actions[].command template (S)

Extend recipes-loader.ts / query-recipes.ts so actions[] carries a rendered command template (reuse recipe-params.ts; block-list shape only).

Step 4 — auto_fixable gating (S)

Enforce actions[].auto_fixable: true (or --force) before write. --force bypasses gate only, not phase-1 conflict detection.

Step 6 — App-wide rename recipe (S; depends Step 5 substrate ✓)

Extend rename-preview.sql with call_rows CTE. Bias: extend existing recipe id, not a second rename recipe. Preserve honest v1 gap list (re-exports, JSX, default-import binds).

Step 7 (recipe) — rename-preview re_export_rows CTE (S)

Substrate shipped; recipe extension open. Single-hop alias chains in v1; recursive CTE if gap appears.

Step 8 — apply --rows - + apply_rows MCP/HTTP tool (M)

JSON array of {file_path, line_start, before_pattern, after_pattern} from stdin/args; existing phase-1/2 validation. Separate MCP tool (not polymorphic with recipe id).

Step 9 — apply --diff-input <file> (S)

Unified diff → row contract → same engine. Hand-roll parser; per-file atomicity; cross-file rollback deferred.

Step 10 — apply --commit "<msg>" (S)

git add only files apply touched, then commit. Message templating deferred.

Step 11 — --until-empty + --max-passes N (S)

Apply → reindex → re-run recipe loop. Abort whole loop on phase-1 conflicts. Extend result envelope with passes + terminated_by.

Step 12 — apply.autoApplyRecipes allowlist (S)

Config list of recipe ids allowed with --yes without interactive confirm. --force bypasses allowlist.

Preserved moats (apply path)

  • No severity on apply rows — recipes propose; codemap executes.
  • No suppression-by-default — existing codemap-ignore-* substrate only.
  • No JS execution at apply time — rows are static data; never eval.
  • No codemod-tool ambition — SQL API + substrate executor; AST work stays Path B / external tools.
  • No telemetry upload — reliability metrics stay local/opt-in per roadmap.md § Floors.

The "no fix engine" floor was about product class (no ESLint-style verdict engine), not forbidding a substrate-shaped apply executor.

Rejected items (grep-able; revisit only on trigger)

Item Why rejected Revisit when
Curated CLI write verbs (codemap rename, …) Moat A — premature verb sprawl ≥3 diff-shape recipes + clear agent-host demand beyond actions[].command
Parallel AST apply engine (Path A) Competes with ts-morph/jscodeshift; maintenance burden ≥2 external teams hit substring wall + concrete AST-shape demand
Trust tiers (safe/review/risky) Taxonomy debt; auto_fixable + allowlist suffice Allowlist insufficient + ≥2 consumers ship trust filters in CI
Per-row confidence scores No consensus on computation Recipe needs per-site ranking when before_pattern is ambiguous
Verifier as product surface (typecheck/lint gate) Consumer CI owns orchestration Consumer plan with concrete examples
Reliability loop telemetry No upload floor Self-hosted observability request
--branch / --output-patch flags --commit priority --commit insufficient in practice
Multi-line kind-tagged row contract After single-line path stable Multi-line edits required and workarounds fail
Cross-file moves in one apply Higher risk Alternative two-step ops insufficient
Cross-file atomic apply (50+ files) Per-file atomicity sufficient today Real 50+ file apply + partial failure leak

Full trigger wording: research/codemap-richer-index-synthesis-2026-05.md § 7.

Cross-references