Agentic engineering: Tier 2 / Tier 3 backlog (lints, test runner, supply-chain, errors) #7

Context

Tier 1 of the agentic-engineering hardening has landed (lint policy in the manifest, clippy.toml,
rustfmt.toml, deny.toml + cargo-deny CI, #![forbid(unsafe_code)] on deckard-core, closed CI
gaps, CLAUDE.md/AGENTS.md). The full rationale and the deliberately-rejected rules are in
docs/AGENTIC-ENGINEERING.md.

This issue tracks the Tier 2 / Tier 3 backlog. Each item was researched against paradigm Rust
repos (reth, alloy, Zed) and adversarially challenged by a cross-model (codex GPT-5.5) review —
that verdict is recorded per item below.

⭐ Promoted to "do now" (highest leverage)

Both models rank these above the rest because they harden the wallet engine and make the Tier-1
unused_must_use = "deny" actually bite.

1. Restriction lints in deckard-core

Clippy rules that make the security engine's production code unable to .unwrap() / panic! on
bad input (tests still unwrap freely, via the allow-*-in-tests keys already in clippy.toml).

Two paths — pick one:

• Safe (zero code change, lands green):

  # crates/deckard-core — via lib.rs inner attrs or a non-workspace [lints] block
  unwrap_used = "deny"   # clean: all real unwraps are in #[cfg(test)]
  get_unwrap  = "deny"
  mem_forget  = "deny"
  expect_used = "allow"  # 2 non-test expects in eth.rs
  panic       = "allow"
  indexing_slicing = "allow"

• Hardened (codex-preferred, ~3 small edits): also deny expect_used / panic /
  indexing_slicing, and fix the real sites rather than blanket-allow — those sites (the keystore
  byte-parser + RPC decode) are exactly where future agent-written indexing should be reviewed:
  • eth.rs:48 / eth.rs:97 — startup expect()s → documented fatal boundary or fallible spawn.
  • keystore.rs Reader (~492–505) — raw slice indexing → bounds-checked .get() + explicit error.
  • balances.rs:80 — results[0] → .first().ok_or(...)?.

codex verdict: MODIFY → PROMOTE. "Promote the corrected core lint set now, with expect_used,
panic, and indexing_slicing denied and the current indexing/expect sites fixed rather than exempted."

2. #[must_use] audit on secret types

Tag the secret-bearing types/methods so the compiler errors if a caller creates one and ignores it:
Vault, UnlockedVault, Vault::unlock, account_signer, account_address, primary_address,
reveal_phrase.

codex verdict: PROMOTE. Low churn, directly useful; unused_must_use = deny only fully bites
once these are annotated.

Tier 2 — high value, some effort

3. cargo-nextest

A faster, nicer test runner than cargo test (parallel, clear per-test output, per-test timeouts).
Add .config/nextest.toml, install via a pinned taiki-e/install-action, run
cargo nextest run --locked --workspace; keep a separate cargo test --locked --doc (nextest skips
doctests).

codex verdict: MODIFY — use retries = 0, NOT 2. Retries normalize flaky security tests,
which is dangerous for a wallet. fail-fast = false. Pin the install action version.

4. profile.dev tuning (faster debug crypto)

Make debug builds optimize the slow crypto so wallet create/unlock and tests aren't painful at
opt-level = 0.

codex verdict: MODIFY — do NOT optimize all deps (opt-level=2 for * slows the first
gpui/alloy compile and the agent feedback loop). Target just the crypto crates:

[profile.dev.package.argon2]
opt-level = 3
[profile.dev.package.blake2]
opt-level = 3

Drop [profile.dev.build-override] unless measured.

5. Dependency + spelling hygiene

• cargo-machete — find unused dependencies. Start manual / non-blocking: it false-positives the
  cfg-gated objc2* + gpui-component-assets, but it correctly flags that the app's root
  Cargo.toml lists alloy-signer-local / alloy-primitives as direct deps it may not use directly
  (they're re-exported from deckard-core) — worth investigating.
• typos (crate-ci) in CI with a crypto-jargon allow-list + extend-exclude = ["Cargo.lock"].
• lefthook pre-commit hooks (fmt/clippy/typos) — local convenience only, NOT a CI gate (agents
  bypass with --no-verify; CI is the real gate).

codex verdict: MODIFY/SPLIT (as above).

6. Doctests + rustdoc warnings

RUSTDOCFLAGS="-D warnings" cargo doc --locked --no-deps -p deckard-core
cargo test --locked --doc -p deckard-core

Catches broken intra-doc links and ensures any future doc examples compile.

codex verdict: KEEP — sensible, not urgent (no doc examples exist yet).

Tier 3 — revisit later

• Typed errors (thiserror) for the keystore. Structured errors instead of anyhow.

  codex verdict: MODIFY, do LAST. ⚠️ Keep one opaque UnlockFailed — distinguishing
  "wrong passphrase" vs "tampered" is an attacker oracle (surfaced via short_err in
  shell.rs). Only the keystore; keep network/provider errors as anyhow until there's real
  programmatic handling.

• proptest round-trips for the keystore binary format (to_bytes/from_bytes, wrong-passphrase
  rejection). Reuse the existing KdfParams::FAST_TEST profile.
• MSRV / toolchain-drift CI job — cargo +1.95.0 build --locked + rust-version = "1.95.0" in
  both manifests, to catch a bump-gpui that silently needs a newer compiler.
• Automated monthly bump-gpui PR bot + Dependabot for the registry deps (serde, alloy,
  argon2, tray-icon); Dependabot/Renovate handle git deps poorly.
• Scheduled cargo-audit cron (decouples the advisory time-signal from PR CI).
• merge_group + a required aggregated check for branch protection (one-time repo-admin action).
• cargo-vet / cargo-semver-checks / coverage gates — deferred (heavy / premature for a pre-1.0 app
  with no external consumer).

Deliberately NOT recommended

See the table in docs/AGENTIC-ENGINEERING.md:
forbid(unsafe_code) at the workspace root, clippy::pedantic/nursery/full restriction,
indexing_slicing = deny workspace-wide, edition 2024, nightly rustfmt/build flags, panic = abort,
multiple-versions = deny, mold/lld linker config (1.95 already defaults to the fast linkers).