flutter_pruner 1.4.0

Safety-first semantic auditor for Dart and Flutter. Finds unreachable declarations, unused assets, and duplicate files with reversible cleanup.

Flutter Pruner

[Flutter Pruner — semantic cleanup for Flutter and Dart projects]

Safety-first semantic cleanup for Dart and Flutter.
Find unreachable Dart declarations, unused Flutter assets and byte-identical duplicates, then apply only what the configured safety policy allows.

Quick start · What it finds · Safety model · Reports · Documentation

Flutter Pruner connects Dart declarations, libraries and assets in one semantic graph. Every finding receives an evidence-backed confidence tier, while the apply workflow adds planning, byte snapshots, quarantine and verified rollback.

Safety first: start from a clean Git worktree. scan and apply --dry-run do not change project sources or assets; scan does write a report under .flutter_pruner/. Only run apply after reviewing its plan and the documented boundaries.

Quick start

Flutter Pruner requires Dart SDK 3.9 or newer.

Install the CLI

dart pub global activate flutter_pruner

If flutter_pruner is not found, add Pub's executable directory to your PATH. In environments where a global executable is unavailable, such as CI, use:

dart pub global run flutter_pruner:flutter_pruner <command>

Run the safe workflow

From the root of the Flutter or Dart project you want to analyze:

# 1. Detect the project and create .flutter_pruner/config.yaml
flutter_pruner init

# 2. Analyze without changing project sources or assets
flutter_pruner scan

# 3. Preview the eligible plan without applying it
flutter_pruner apply --dry-run

# 4. Apply only the findings authorized by the configured mode
flutter_pruner apply

Step	Purpose	Project sources/assets
init	Declare project type, targets, entrypoints and verification	Unchanged
scan	Build the graph, classify findings and save a report	Unchanged
apply --dry-run	Validate and preview the eligible plan	Unchanged
apply	Quarantine originals, mutate, rescan and verify	May change

The init wizard never marks target coverage complete by inference; completeness is an explicit project-owner assertion. Reusable packages default to a non-actionable open-world mode, so their dry-run and apply commands stay blocked until an appropriate local boundary is explicitly configured.

To work from another directory, add --project path/to/project to any command.

Upgrade or uninstall the global executable with:

dart pub global activate flutter_pruner
dart pub global deactivate flutter_pruner

What it finds

Area	Detection	Apply policy
Dart	Unreachable top-level declarations and empty libraries	Confidence and mode controlled
Assets	Exact references and assets unreachable from live Dart code	Confidence and mode controlled
Routes	go_router declarations, redirects, nested routes, direct path/name navigation, and path navigation through resolved local wrappers	Review only
Dependency injection	Direct base-scope GetIt registrations and exact resolved lookups	Review only
Localization	ARB keys and current real-source gen-l10n accessors reachable from configured app targets	Review only
Duplicates	Byte-identical files grouped with SHA-256	Review only

Why Flutter Pruner

• Semantic, not grep-based. Dart references are resolved through the analyzer's element model instead of source-text matching.
• Cross-domain reachability. Dead code does not keep an otherwise unused asset alive merely because its source contains the asset path.
• Fail-closed confidence. Dynamic references, incomplete targets, dangling graph facts and unsupported edits downgrade findings instead of being ignored.
• Declared-scope analysis. Every configured entrypoint is analyzed; conditional branches the graph cannot model keep coverage incomplete.
• Verified whole-run recovery. Every touched path is quarantined before mutation. If an apply run cannot finish, its mutations are restored to the run baseline and accepted only after byte and verification checks pass.
• Audit-ready output. Completed scans and handled apply outcomes save a structured report; HTML reports are self-contained and work offline.

Safety model

Choose the right analysis mode

The mode defines the analysis boundary and apply policy. It is not a shortcut for increasing confidence.

Mode	Use when	Result and actionability
application	The project owns the complete application boundary	Complete hard gates can produce SAFE; apply selects only SAFE
package	Unknown applications or packages may consume the public API	All non-protected findings are capped at REVIEW; apply and dry-run are disabled
package-internal	You intentionally want to clean only the local package boundary	Private local findings can be SAFE; external-consumer candidates may be HIGH only when that is their sole risk

Reusable and hybrid projects are conservatively detected as package. Choose package-internal explicitly when external consumers are intentionally outside the scan:

flutter_pruner init --type package-internal

The interactive wizard always lists all three modes. Pressing Enter keeps the detected conservative default; package-internal must be selected explicitly. On supported terminals, semantic color and text weight distinguish sections, defaults, confirmations, and risk warnings; icons and labels keep the same meaning when ANSI styling is unavailable.

Both package modes require at least one public_entrypoints entry. Public exports and their dependency closure remain reachable, so an exported API is not reported unused merely because the package does not call it internally.

A generated package configuration contains the explicit boundary:

analysis:
  mode: package-internal
  public_entrypoints: [lib/my_package.dart]

init writes the full target-matrix section and verification policy around it.

Every package-internal run warns that external consumers were not scanned. A mutation containing eligible HIGH findings asks for [y/N]; CI can accept that one specific risk with flutter_pruner apply --yes. The flag does not bypass coverage, planning, verification or quarantine.

Understand the confidence tiers

Tier	Meaning
SAFE	All hard gates passed and no manual risk remains
HIGH	All hard gates passed and exactly one known manual risk remains
REVIEW	Evidence is incomplete, ambiguous, unsupported or has multiple risks
PROTECTED	Framework naming rules or dependency ownership forbid removal

PROTECTED always wins. Conditional imports/exports, unresolved or dynamic analysis, generated-code uncertainty, duplicate groups and unsupported editor actions stay non-actionable rather than being silently promoted.

See the confidence model for the complete policy.

Preview, apply and rollback

Apply takes a project-local operation lock shared with rollback and quarantine maintenance. Before mutation it rejects incomplete or non-terminal historical quarantines, captures the verifier baseline, and builds a dependency-closed plan. The plan may use internal atomic units, but the command is all-or-nothing: if any unit, rescan, convergence check, canonical report, or recovery check fails, it restores every mutation from this run to the original baseline and verifies that recovery.

Use repeatable --finding-id values when the reviewed scope is smaller than the full eligible set:

flutter_pruner apply --dry-run \
  --finding-id 'dart:my_app/lib/example.dart#_unusedHelper'
flutter_pruner apply \
  --finding-id 'dart:my_app/lib/example.dart#_unusedHelper'

IDs are exact and case-sensitive. The complete requested batch must exist, be apply-eligible, and already be dependency-closed; the planner never adds an unrequested logical finding. An unknown, non-actionable, or blocked requested ID stops the whole batch before verification or quarantine creation (exit 2). Empty or duplicate CLI values are usage errors (exit 64). Every rescan and the final convergence check remain bound to the same allowlist, while unselected findings stay visible in the report.

recoveryRequired means that restoration could not be proven (for example, a timed-out mutation process could not be confirmed stopped). Stop using the project with this tool, inspect the quarantine and report, and recover the project before another mutating command. A successfully recovered failed run does not retain earlier transactions.

To restore an applied run:

flutter_pruner quarantine list
flutter_pruner rollback <run-id>

Quarantine lives under .flutter_pruner/quarantine/<run-id>. Its manifest is a revisioned, checksummed journal with staged recovery for interrupted replacement; ambiguous or corrupt journal state blocks a new apply. Add --clean to the rollback command only when the restored quarantine should also be removed.

For a regular-file edit, the original source path is renamed into quarantine before a separately staged candidate is published with no-replace semantics. Restore uses the same preserve-then-no-replace pattern: a concurrent recreation of the project path is retained and the run becomes recoveryRequired instead of overwriting it. This protects against observed path races; it is not a claim of crash-durable filesystem transactions (see the boundaries below).

The rollback contract for a regular file is its captured bytes and, where the platform exposes them, POSIX permission bits. It does not preserve extended attributes, ACLs, ownership (uid/gid), or hard-link topology. Use version control or a filesystem backup when those metadata matter.

Reports

Every completed scan and handled apply outcome receives a unique report below .flutter_pruner/reports/:

scan-<run-id>.html
apply-<run-id>.html

No format or output option is required for the default searchable offline HTML report. The terminal always shows the human summary and saved path. Export JSON for CI or other machine consumers when needed:

flutter_pruner scan --format json
flutter_pruner apply --dry-run --report-format json

Use --output for scan or --report-output for apply only when CI needs a stable artifact path. Mutating apply always keeps canonical JSON alongside the quarantine, even when HTML is selected.

JSON schema v3 records coverage, blockers, findings, manual risks, verification and transaction outcomes. See structured run reports for the schema, HTML features and CI selectors.

Important boundaries

• scan never changes project sources or assets, but it does write its report under tool-owned .flutter_pruner/ state.
• target_matrix.complete: true means the owner has declared every supported platform, flavor, entrypoint and Dart-define combination.
• package-internal proves only the local boundary; external apps may still use public symbols, deep imports or package assets.
• Built-in callback-handle boundaries cover PluginUtilities.getCallbackHandle, Isolate.spawn, and Workmanager.initialize; custom runtime registries must be modeled by a future adapter or protected by project policy.
• Asset paths assembled only by custom runtime code remain a trust boundary. Known dynamic or unresolved asset usage blocks affected findings, but the tool cannot prove that an unmodeled runtime API has no consumers.
• GetIt scopes, runtime or dynamic lookup state, and generated Injectable wiring are uncertainty boundaries. ARB/gen-l10n analysis targets current real-source output. These cases add scoped blockers rather than authorizing removal.
• Route and localization consumers are narrowed to the configured application entrypoint closure only when analyzer resolution proves that closure complete. Missing entrypoints, unresolved local directives, and conditional imports or exports keep the broader conservative scan and emit blockers.
• Configured verification and import cleanup run as argv-only subprocesses with deadlines and bounded captured output. An unconfirmed detached process tree produces recoveryRequired; Windows termination behavior is covered by the hosted release CI job.
• A subprocess that deliberately detaches before it can be observed remains a trust boundary; do not configure mutation-capable cleanup through such a tool.
• The local filesystem contract does not currently include directory fsync, renameat2 durability semantics, or a guarantee against a microscopic open-file-descriptor append concurrent with path-level displacement. Treat a power loss or an adversarial writer during that window as requiring manual recovery, not as a proven atomic rollback.
• Rollback is not a blanket metadata restore: for regular files it restores captured bytes and POSIX permission bits where available, not xattrs, ACLs, uid/gid, or hard-link topology.
• Source-byte measurements are not release binary-size claims.

BuildTarget and TargetMatrix are public v1 Dart APIs. Their defensive immutable snapshot behavior is intentional; source-breaking changes require a future major release.

The public reporting and adapter-presentation DTOs likewise defensively snapshot collection arguments, so collection-bearing constructors are intentionally non-const. VerificationPolicy is exported for API callers. Treat AdapterRegistry.builtIn as read-only at runtime; a custom adapter set is passed explicitly to AdapterRegistry.resolve(adapters: [...]), while changes to the built-ins belong in the registry initializer. These public API contracts follow Semantic Versioning; source-breaking changes require a future major release.

When upgrading older configs, remove analysis.root_coverage. The only accepted mode values are application, package and package-internal; the old workspace, --safe and --high interfaces are intentionally rejected.

Documentation and support

Document	Start here for
Architecture	Scan flow, layers and apply transactions
Graph model	Nodes, roots, reachability and blockers
Confidence model	Safety gates and confidence tiers
Project configuration	Config schema, coverage and verification policy
Run reports	JSON v3, HTML and CI selectors
Report schema migrations	Compatibility and deprecation policy
V2 adapters	As-built route, GetIt and localization behavior and boundaries
Release readiness	Replay, real-project and performance gates
Performance profiling	Synthetic fixtures, benchmarks and privacy rules
V2 natural accuracy	Real-project oracle method, corpus pins and retained results
Verified Flutter facts	Framework assumptions and primary sources
Contributor guide	Development setup and pull requests
Adapter guide	Adding a new analyzer

See ROADMAP.md for planned work. Adapters are the main extension point; the adapter guide includes a copyable template.

Found a bug or have a focused feature request? Open an issue. To contribute code, start with CONTRIBUTING.md and include regression evidence for changes that affect confidence, mutation or recovery behavior.

License

MIT — see LICENSE.