Understood. Here’s what the Claude Code CLI / local agent is responsible for—only the client-side mechanics, not what the LLM reasons about.
-
Workspace discovery → Detect monorepos/modules, language/toolchain boundaries.
Input: root dir • Output: workspace graph (projects, package roots, build files). -
Language/toolchain fingerprinting → Infer runtimes and build tools (e.g., Node/PNPM, Python/Poetry, Java/Maven/Gradle, Go, Rust/Cargo).
Output: toolchain spec used to run builds/tests and parse metadata. -
Dependency graphing → Parse manifests/lockfiles to build internal/external dep DAG with versions/scopes.
Output: dep graph for impact, upgrade feasibility, and sandboxing. -
Symbol/xref indexing (ctags/LSP/AST-assisted) → Cross-reference of definitions/usages across files.
Output: queryable index:symbol → {defs, refs, file, line}. -
Call/dataflow summaries (lightweight static analysis) → Function-level call graph; import graph; sometimes basic alias/taint edges.
Output: graphs for impact calculation and change propagation. -
Execution surface mapping → Enumerate CLIs, entrypoints, HTTP routes, handlers, cron/jobs based on framework conventions and annotations.
Output: entrypoint/index maps. -
Schema/config extraction → Detect ORM models, DB schemas/migrations, config/feature flags, env var surfaces.
Output: schema & config maps usable for compatibility checks.
-
Relevant-region retrieval → Hybrid search over the local indices (BM25/embeddings if available, plus xrefs) to shortlist files and symbols likely to change.
Output: ranked candidate files/symbols/lines. -
Change boundary delimitation → Compute a minimal edit set and seam locations (APIs, adapters, public vs internal).
Output: edit scope with blast-radius estimate. -
Impact analysis (forward/backward) → Traverse call/import graphs to find all affected callers, tests, schemas, routes, and docs.
Output: propagation set for synchronized edits. -
Ambiguity scoring (non-interactive) → When multiple target regions match equally well, score/rank them using repo signals (usage density, ownership patterns, test linkage) without asking the user.
Output: best candidate set + tie-break metadata.
-
Patch application → Apply model-proposed hunks/AST transforms to the working tree, handling file creation/deletion/rename.
Output: consistent on-disk changes; rejected hunks reported. -
Cross-file propagation mechanics → Perform deterministic updates derived from indices (e.g., auto-update imports, rename files, fix package paths) when safe to do so.
Output: synchronized changes beyond the raw patches. -
Merge/conflict mediation (semantic where possible) → Attempt tree-aware merges against current HEAD; surface unresolved conflicts with precise spans.
Output: resolved files or conflict report.
-
Format/lint/type/build gates → Run formatters, linters, type checkers, and compilers with repo-native configs; capture structured output.
Output: gate results mapped to files/lines. -
Selective test execution → Compute impacted tests (from xrefs/import graph) and run only those first; then widen if necessary.
Output: fast failure signal + full-suite fallback. -
Coverage delta → Overlay coverage on the diff to ensure changed lines are exercised.
Output: before/after coverage, uncovered regions. -
Performance harness orchestration (if benches exist) → Run micro/bench targets and compare to baseline; capture timing/allocs where supported.
Output: perf deltas with thresholds. -
Security/compliance scans (optional, when configured) → Secret detectors, SAST rule packs, license audits over deps.
Output: findings tied to files and severity.
-
Toolchain normalization → Activate venvs, asdf/nvm/pyenv, or containerized runners; pin versions from lockfiles.
Output: reproducible run context. -
Hermetic sandboxes → Spin up ephemeral envs (venv/containers/docker-compose) to execute tests/servers without polluting host.
Output: isolated execution logs/artifacts. -
Targeted run orchestration → Shard tests, parallelize jobs, retry flaky tests, and capture artifacts (logs, snapshots) for model consumption.
Output: structured evidence bundle.
-
Branching/checkpointing → Create/checkout branches; stage coherent commits aligned to plan steps.
Output: commit series with machine-readable metadata. -
Diffing & semantic summaries → Generate textual and (when supported) AST diffs; tag changes by subsystem/surface.
Output: diff + change map. -
PR construction → Open PRs with templated body (summary, risks, test evidence, coverage/perf deltas), attach artifacts.
Output: ready-to-review PR.
-
Working-set curation → Maintain a rolling “active context” of snippets/files most relevant to the current task; evict stale regions.
Output: compact bundle for round-trips. -
Local caches/index persistence → Persist symbol/embedding indices per repo to accelerate subsequent sessions and improve retrieval quality.
Output: on-disk indices keyed by commit hash.
-
Invariant checks → Enforce repo-specific rules (e.g., public API stability, migration policies, codeowners) before accepting edits.
Output: hard stops or warnings with diffs. -
Rollback scaffolding → Create restore points and scripted reverts (stash/branch/commit hooks) tied to each iteration.
Output: deterministic rollback path.
- Does not “reason” about intent or invent edits itself; it executes, scopes, measures, and validates.
- Does not rely on full-repo uploads to the model; it streams curated slices derived from the local indices and graphs.
- Does not replace your build/test stack; it orchestrates your existing tools and surfaces structured feedback.
- Index repo → build xrefs/graphs.
- Retrieve candidate regions for the task → delimit change boundaries.
- Apply incoming patch hunks → propagate safe deterministic updates (imports/renames).
- Run format/lint/type/build → run impacted tests first.
- Collect gates/coverage/perf deltas → prepare diffs/artifacts.
- Create commits/PR with evidence → provide rollback checkpoints.
That is the entire local surface area of Claude Code’s CLI: indexing, scoping, applying, validating, and packaging—everything deterministic and file-system/CI/VCS-facing—while omitting the LLM’s semantic work as requested.