dev-rust

Rust Implementation

Overview

Implement Rust changes against the repository's real toolchain, project class, crate boundaries, error model, and validation contract. Select validation by risk rather than applying native/FFI-level gates to every routine change.

Consume urn:skills:rust-change-request:v1; the portable output is urn:skills:source-change-result:v1. When supplied, consume typed Task/requirement refs, authority refs, a compatible Asset Graph snapshot/query result, and an exact input PackageManifest; Rust source, manifests, and native interface/schema contracts remain authoritative.

Workflow

1. Read effective repository guidance, including AGENTS.md, CLAUDE.md, and host-provided instructions when present, then run git status --short before edits.
2. Identify the Rust project class: library workspace, application workspace, HTTP service, CLI, Tauri/native backend, or compact single package.
3. Read the approved requirement/specification when one exists. Confirm requested behavior, acceptance criteria, non-goals, affected crates/modules/files, compatibility, and validation seams; for complex work without a usable specification, use the host's built-in planning and effective repository instructions before editing.
4. Inspect the relevant Cargo.toml, lockfile, toolchain, formatter, lint, command source, modules, tests, architecture docs, and API/interface docs.
5. Consume a compatible graph asset/consumer/impact query or perform the same targeted live search across route registration, handlers, services, repositories, traits/impls, types/DTOs, errors, migrations, callers, tests, and analogous features. Reject a stale/mismatched snapshot, and never treat a query miss or derived Markdown render as proof of absence.
6. When the requested Rust change crosses reachable runtime/configuration, packaging, API, persistence, compatibility, security, deployment, or cross-repository boundaries, load references/project-grounding.md and build the smallest task-scoped record before edits. A Rust file, Cargo manifest, or framework match alone does not activate grounding; leave unrelated risk classes Not applicable and report unavailable runtime proof as Not verified.
7. Start with the Baseline validation contract, then select every applicable risk overlay. Overlays are composable, not severity levels:
   • Protocol automation: an existing OpenAPI/generated-client pipeline or an explicitly requested contract migration. Ordinary HTTP/API changes keep the repository-native route/DTO/client/test boundary under Baseline.
   • Agent Runtime: a stateful local-agent workflow, typed agent protocol, durable operation history/recovery, approval/policy/sandbox enforcement, or Tauri/local app-server IPC. A Tauri app, SQLite crate, or async function alone does not activate this profile; load references/agent-runtime-profile.md only when the runtime lifecycle is reachable.
   • Concurrency/runtime: Tokio tasks, channels, locks, cancellation, blocking work, overload, or shutdown.
   • Persistence/SQLite: migrations, transactions, schema/query changes, durable compatibility, backup, or recovery.
   • Unsafe/FFI: unsafe, ABI/layout, raw pointers, callbacks, allocators, native handles, or cross-language resource ownership.
   • Porting/parity: language port or large rewrite that must preserve observable behavior and release semantics.
   • Target/platform: cfg, target-specific APIs, packaging, native linkage, or supported-platform behavior. A routine change uses Baseline with no overlays. An Agent Runtime change composes this profile with Concurrency/runtime, Persistence/SQLite, Protocol automation, and Target/platform only when each boundary is reachable. A mixed FFI plus SQLite change selects both overlays; a target-only change selects Target/platform without inheriting unrelated heavy tools.
8. Decide in order: directly reuse, extend an existing contract, adapt the nearest reference, or create new. Record why existing interfaces are insufficient before adding an endpoint, trait, type family, or module.
9. Trace ownership, dependency direction, and the complete interface chain before adding or moving code. If Protocol automation applies, identify one code-first or contract-first authority; otherwise preserve the repository-native API authority without introducing OpenAPI.
10. When behavior is stable and a durable public seam exists, confirm that seam, then work one external behavior at a time: run one red-capable check, make the minimum green change, and continue as a vertical tracer bullet. Load references/behavior-first.md; do not force it onto exploratory work, generated code, or behavior without an honest seam.
11. Implement the smallest idiomatic change that follows local ownership, borrowing, module, error, async, persistence, FFI, configuration, logging, documentation, and test patterns. When Agent Runtime applies, load references/agent-runtime-profile.md and keep the smallest explicit Thread/Turn/Operation, typed protocol, lifecycle, authorization, persistence, and IPC contract that the requested behavior needs. When the task materially involves duplication, dead/unused code, abstraction, coupling, or maintainability, load references/code-quality.md with implementation semantics and remove only declarations made obsolete by the authorized change after resolving Rust reachability.
12. Update manifests, module exports, tests, commands, docs, CI/deploy paths, migrations, generated files, and indexes when the structural or public boundary changes.
13. Run focused checks after each slice, then the repository's baseline gates and every selected overlay. Use Miri, sanitizers, fuzzing, stress, or repeated-operation tools only when both supported by the target repository/environment and relevant to the changed invariant.
14. When Forgeway delivery integration is active, require an immutable Run with input refs, exact scope, and input PackageManifest before mutation. Let the package producer fingerprint each Attempt result and bind commands, tests, runtime checks, and artifacts as typed Observations to that exact result package. Retries do not overwrite earlier Attempts; this implementation owner emits no Receipt.

Modes

• Targeted implementation: add or fix Rust behavior without broad architecture or toolchain changes.
• Structure alignment: align modules, crates, manifests, commands, and docs to an explicit repository standard.
• Contract migration: change an API, DTO, feature, persistence, or consumer boundary with compatibility and rollout evidence.
• Native/porting implementation: preserve observable behavior, ABI/resource semantics, supported targets, and release behavior before idiomatic cleanup.
• Implementation self-check: verify the edited Rust surface for ownership, errors, async behavior, safety, tests, dependencies, and structural drift before repo-review assesses commit readiness.

Do Not Use For

• First-pass repository discovery; use repo-map.
• Planning-only requests without authorized Rust source changes; use the host's built-in planning.
• Shared cross-functional business language/rule or lifecycle conflicts; use domain-modeling. Route feature-local behavior, states, and acceptance to product-spec.
• Diagnosis-only requests without authorized Rust source changes; use the host's built-in diagnosis under effective instructions.
• Dirty-tree ownership, staging plans, or commit grouping; use repo-review. Use repo-delivery for actual staging or commits after review.
• Systematic Rust architecture, performance, memory, concurrency, SQLite, unsafe, or FFI audit without requested edits; use audit-rust.
• Review of a fixed Rust change basis, including authorization or token risks; use repo-review.
• Frontend or webview UI changes; use dev-frontend.

Hard Rules

• Follow repository-pinned toolchain, layout, dependency, lint, and command contracts. Reuse or extend the established interface chain before adding a contract, and do not introduce OpenAPI or copy another project class without an explicit selected overlay.
• Preserve local ownership, dependency, error, and persistence boundaries. Apply best practices through the selected overlays rather than treating generic Rust doctrine as evidence for this change.
• Load and apply only references for the selected Agent Runtime, protocol, concurrency, persistence, unsafe/FFI, porting, target/platform, behavior-first, conditional code-quality, or codebase-design overlays. Do not inherit heavy gates from an unselected overlay.
• Treat new warnings in the touched surface as defects. Do not broaden scope to clean unrelated legacy warnings; report them separately. Prefer a local justified #[expect(...)] over weakening workspace lints.
• Move behavior to a shared crate only after real reuse, stable API, named ownership, and consumer validation. Keep every affected manifest, export, command, test, doc, CI/deploy path, generated output, and index synchronized with structural changes.
• Preserve unrelated local changes and generated files not owned by the task.
• When implementing destructive cleanup or deletion behavior, discover candidates first, prove ownership and bounded scope, preserve a recovery path when the product contract allows it, require the applicable user confirmation or policy authorization, perform only the accepted action, and rescan to reconcile the actual result. This governs the Rust implementation; it does not authorize the agent to delete user data during development.
• Do not add speculative abstractions. Delete apparently unused items only after resolving public API, feature/target configuration, generated reachability, FFI exports, and downstream use.

Validation Model

• Baseline: repository-defined format/check plus focused behavior tests; Clippy only when it is part of the repository baseline.
• Selected overlays: add only the Agent Runtime, contract, concurrency/runtime, persistence/SQLite, unsafe/FFI, porting/parity, and target/platform evidence required by the changed surface. Combine overlays when risks interact; do not let one overlay erase another.
• Agent Runtime: when selected, validate the legal state machine and uncertain-operation recovery, the single typed protocol/schema authority, bounded task/channel and cancellation behavior, distinct approval/policy/sandbox decisions, and durable log/projection or IPC gates that are actually in scope. Static types, generated schemas, local compilation, or host declarations do not prove target runtime, sandbox, client, or recovery behavior.
• Protocol-automation overlay: only when selected, validate/rebuild the OpenAPI artifact, check clean idempotence and compatibility, regenerate owned clients, and run applicable backend conformance. Otherwise use repository-native API tests.
• Optional heavy tools: Miri, sanitizers, fuzzing, stress, leak, and repeated-operation gates are never inherited merely from a target-specific change. Run them only when supported and relevant, or record why they were excluded.
• Full gate: reserve workspace-wide builds/tests and release gates for merge, release, deployment, final fixed-basis acceptance, explicit user requests, or when no credible focused check exists for the actual risk.

Do not claim Baseline or an overlay passed when a required tool, target, runtime, dataset, or external dependency was unavailable; mark the exact gap Not verified.

Output Contract

Report capability rust.source.implement, Run/input/result PackageManifest refs when integration is active, typed source-change result/Observation refs, scope; detected project, crate/module, toolchain, and ownership boundaries; applicable authorities and existing owners; selected Rust risk overlays; reuse/extension/reference decision; changed files and contract chain; validation mapped to Baseline and each selected overlay; Worktree drift; excluded work and optional checks; and Not found or Not verified gaps. Add new-interface justification, manifest/docs lifecycle, failures, and target/runtime evidence only when applicable. If the user explicitly requests independent external review/research, hand one fixed basis/question to ask-ai; never send implicitly.

References

• See references/usage.md for trigger guidance and examples.
• See references/checklist.md for implementation and review checks.
• See references/protocol-contracts.md only when Protocol automation is selected.
• See references/behavior-first.md when a stable public seam supports vertical red-green slices.
• See references/codebase-design.md only when the change materially affects a public module/interface, seam, cross-caller abstraction, or testability.
• See references/best-practices.md for idiomatic Rust API, ownership, error, test, docs, performance, dispatch, and concurrency rules.
• See references/code-quality.md when the requested change materially involves duplication, dead/unused code, abstraction quality, hidden coupling, or maintainability.
• See references/project-grounding.md when reachable runtime/configuration, packaging, API, persistence, compatibility, security, deployment, or cross-repository boundaries affect the Rust change.
• See references/agent-runtime-profile.md only when the change reaches a stateful local-agent workflow, typed agent protocol, durable operation history/recovery, approval/policy/sandbox enforcement, or Tauri/local app-server IPC.
• See references/bun-production-patterns.md only when Porting/parity or Unsafe/FFI needs source-derived prompts for cross-language semantics, resource lifetime, or local invariant enforcement.
• See references/eval-cases.md for trigger and quality evals.