fix(typos-format): remove opt-in config-gate, run typos unconditionally

[kyle-sexton]

Summary

typos-format's hook gated its entire spell-fix behavior behind a pre-existing consumer typos config (typos.toml/_typos.toml/.typos.toml/Cargo.toml/pyproject.toml), mirroring the ruff-format/biome-format opt-in pattern. That pattern is wrong for typos specifically: it ships a built-in spelling dictionary and runs with zero configuration, so the gate made the hook a silent no-op on any repo without a hand-authored typos config — i.e. most repos on day one, defeating the plugin's zero-config auto-fix purpose.

Fix

Removed the opt-in config-gate entirely. typos --write-changes now runs unconditionally on every edited file (no extension filter, same as before), matching markdown-format's existing unconditional pattern instead of ruff-format's config-gated one. This is a removal of an activation gate only — typos's own file-anchored config discovery still applies automatically when a config is present, to widen its allowlist (extend-words/extend-identifiers/extend-ignore-re) and excludes ([files] extend-exclude); the hook no longer implements its own ancestor config walk-up at all, since it was only ever used to decide whether to invoke typos, never to pass the config to it.

Updated:

• plugins/typos-format/hooks/typos-format.sh — removed the CONFIG_FOUND walk-up loop and its gate check; reworded the missing-binary skip notice (no longer references "a typos config governs this repo"); rewrote the header comment block to describe unconditional behavior.
• plugins/typos-format/hooks/typos-format.test.sh — replaced "opt-in gate OFF -> file left untouched" with a test proving the hook fixes a typo with no config present anywhere; reframed the config-precedence test as proving typos' own native discovery (not the hook's); flipped the "gate OFF -> telemetry status skipped" case to "no-config -> status ok"; updated the missing-binary test's message assertion.
• plugins/typos-format/README.md, plugins/typos-format/skills/setup/SKILL.md — updated to describe unconditional behavior; config presence in setup check is now informational only (INFO), never a FAIL/PASS gate.
• plugins/typos-format/.claude-plugin/plugin.json — bumped 0.1.0 → 0.2.0 (minor, not patch): this changes runtime behavior for existing installs — a repo with no typos config goes from inert to actively rewriting files on edit, a surprise-on-upgrade consumers should see flagged in the version.
• plugins/typos-format/CHANGELOG.md — new [0.2.0] entry.

Verification

All 43 hook contract tests pass (bash plugins/typos-format/hooks/typos-format.test.sh) against a real typos-cli 1.44.0 binary.

Empirical before/after against a zero-config repo (no _typos.toml anywhere), same input this has teh typo:

• Before (old gated hook, checked out at the parent commit): file left untouched — this has teh typo.
• After (this PR's hook): file fixed in place — this has the typo.

Closes #884

Related

• Supersedes the typos-format half of typos-format/ruff-format: opt-in config grep misses TOML inline-table form for [tool.X] #876 (a grep-pattern issue against the now-removed gate check) — the gate no longer exists, so that concern is moot.
• Origin: feat(typos-format): add per-file typos autofix hook plugin #872 shipped the gate this PR removes.

[claude]

Claude finished @kyle-sexton's task in 2m 33s —— View job

---

PR Review

• Gather context
• Read changed files
• Analyze hook script logic
• Analyze test coverage
• Analyze docs/README changes
• Post review findings

---

Summary: The change is correct and well-scoped. Removing the 32-line CONFIG_FOUND walk-up loop and its single gate check is the right fix — the walk-up was pure activation logic that the hook implemented itself; typos already handles config discovery natively, and the tool's zero-config built-in dictionary makes the gate actively harmful on unconfigured repos. The diff is clean, tests are comprehensive, documentation is accurate, and the minor version bump is justified.

---

Hook script (typos-format.sh)

No correctness issues. The removal is surgical: CONFIG_FOUND was only ever consumed by the gate, never passed to typos itself, so deleting the walk-up loop + [[ -n "$CONFIG_FOUND" ]] || emit_skipped leaves all downstream logic intact. The emit_skipped function now only fires on the missing-binary path, which is correct.

The skip notice wording change is accurate:

• Before: "a typos config governs this repo but no 'typos' binary was found on PATH"
• After: "no 'typos' binary was found on PATH"

The "governs this repo" qualifier no longer applies since activation is unconditional.

--force-exclude with no config is a no-op, not a hazard — typos simply has no excludes to honor, so it runs against its built-in dictionary. No regression there.

---

Tests (typos-format.test.sh)

Tests are thorough and well-structured. A few observations:

Missing-binary test still uses a config-bearing repo — REPO_NT gets a _typos.toml written to it (line 333), even though the PR text says the notice fires "regardless of whether a typos config exists." The logic is provably config-independent (binary check precedes any config logic, and CONFIG_FOUND is gone), so it's low risk, but a no-config variant of this case would close the gap completely if thoroughness is desired.

Case 1d (config precedence) tests typos itself, not the hook — the case is correctly labeled as "typos' NATIVE config discovery (Config::from_dir), not anything the hook implements," so this is a deliberate regression guard. The assertion relies on fooone being absent from typos' built-in dictionary (so it's left alone when _typos.toml is suppressed by typos.toml). fooone is a sufficiently synthetic word that this should be stable, but it could theoretically collide with a future dictionary entry. No action required; just worth flagging.

Case 1c (pyproject.toml with [tool.typos]) — confirms that a [tool.typos] section doesn't accidentally block the unconditional run. Good coverage.

Telemetry / no-config test at line 303–315 is the correct flip: previously the gate caused emit_skipped (status "skipped"); now the hook runs and exits cleanly (status "ok"). The assertion jq -r '.status' == "ok" confirms the new path end-to-end.

---

SKILL.md (setup check probe 4)

Minor doc inconsistency. Probe 4 includes the parenthetical:

(read the hook for the exact names and the section test — its test is the authority)

The config file names are still present in the hook's header comment (lines 16–21 of typos-format.sh), so reading the hook still surfaces them. However, the "section test" — the grep -qE '^[[:space:]]*\[(workspace|package)\.metadata\.typos regex — no longer exists in the hook (it was part of the removed walk-up loop). A model following this instruction would read the hook comment and find the names but not a "section test."

Since probe 4 is now informational-only (not a FAIL gate), this is low priority, but the parenthetical could be updated to reflect that the hook's header comment is the reference, not a grep-pattern test that no longer exists.

Fix this →

---

Version bump

0.1.0 → 0.2.0 (minor) is correct. Per semver and the repo's own convention, a behavior change that actively rewrites files for existing installs that previously got a silent no-op is a consumer-visible surprise. The CHANGELOG entry is clear about the blast radius.

---

Verdict

Approve. The core logic change is correct, the test coverage is solid, and the documentation accurately reflects the new behavior. The SKILL.md parenthetical is a polish nit, not a blocker.

[claude]

Claude finished @kyle-sexton's task in 2m 57s —— View job

---

Security Review — PR #901

• Gather PR diff and changed files
• Audit typos-format.sh hook for injection/unsafe shell patterns
• Audit typos-format.test.sh for unsafe test patterns
• Audit plugin.json, README.md, SKILL.md, CHANGELOG.md
• Post findings

---

Findings

No CRITICAL or IMPORTANT security issues found.

Reviewed surfaces: command injection, path traversal, unsafe variable handling, JSON injection, supply-chain risk, secrets/credential exposure, and dangerous sink execution patterns. Details below.

---

Hook invocation — typos-format.sh line 169

OUTPUT=$(cd "$RUN_DIR" && "$TYPOS_BIN" --write-changes --force-exclude --format json "$TYPOS_ARG" 2>&1)

$TYPOS_BIN comes from command -v typos (PATH resolution only — no download). $TYPOS_ARG is always double-quoted, preventing word-splitting. It resolves to either the absolute $FILE (always starts with / on POSIX, drive letter on Windows — never --) or a relative path produced by stripping the common repo-root prefix from an absolute path (also can never start with --). No -- separator is needed here in practice. $RUN_DIR is canonicalized through cd … && pwd before use. No command injection surface.

File path input — hook::read_file_path

The path comes from Claude Code's JSON envelope, parsed via jq -r '(.tool_input.file_path // empty)' (proper jq flag injection prevention), then validated with [[ -f "$file" ]] (must exist) and optionally bounded to CLAUDE_PROJECT_DIR via a symlink-resolved prefix check. No untrusted path ever reaches a shell eval or unquoted expansion. No path traversal.

Output from typos fed back into context

while IFS= read -r line; do
  [[ -n "$line" ]] || continue
  hook::ctx_append "  $line"
done <<<"$OUTPUT"

$OUTPUT (typos stdout+stderr) is appended to _HOOK_CTX_BUFFER, which is then flushed via hook::emit_additional_context using jq -n --arg ctx "$ctx". The --arg flag means jq handles all JSON escaping — a typos output line containing "},"systemMessage":"injected" is inert. No JSON injection.

Telemetry sink execution — hook-utils.sh line 455

printf '%s\n' "$envelope" | ("$sink" >/dev/null 2>&1) &

$HOOK_TELEMETRY_SINK is an executable path configured by the consumer. Execution is intentional by design; stdout/stderr are both redirected to /dev/null so the sink cannot write to the hook's additionalContext channel (fd1). The sink path is resolved relative to $CLAUDE_PROJECT_DIR when not absolute (not user-controlled per-edit input). By design; no new risk introduced by this PR.

Supply-chain

$TYPOS_BIN is resolved from PATH only and never downloaded at hook time. No new remote sources added. plugin.json introduces no new MCP servers or remote dependencies. No supply-chain risk.

Behavioral security note (informational)

The primary change — removing the config gate — means typos --write-changes now runs unconditionally on every edited file in every repository where the plugin is installed. --force-exclude is passed, so typos' own [files] extend-exclude config is honored. The consequence: repos with intentional-misspelling fixtures (e.g. test vectors for spell-checker unit tests) that lack a _typos.toml extend-words entry will have those fixtures silently rewritten. This is a behavioral surprise, not a security vulnerability. It is adequately addressed by the version bump to 0.2.0 (consumers see the version change on upgrade) and the CHANGELOG entry.

---

Summary: The hook code is well-structured — all subprocess calls use properly-quoted variables, JSON construction uses jq --arg throughout, file path input is validated before use, and the binary is PATH-resolved only. The gate removal does not introduce any new injection, traversal, or credential-exposure surfaces.

[kyle-sexton]

Closing as a duplicate of #900 — both PRs independently implemented the same fix for #884 (a genuine race: two workers ended up on the same task). #900 landed first with a fully clean review pass (2 minor non-blocking nits, no correctness/security issues) and zero pending checks. Not merging this one.