If you discover a security vulnerability, please report it responsibly:

1. For sensitive vulnerabilities, email vishnu@hookwise.dev with a description, reproduction steps, and impact assessment. Do not open a public issue for exploitable or unpatched vulnerabilities.
2. For low-severity or already-patched issues, you may use the Security Vulnerability issue template.
3. You will receive an acknowledgment within 48 hours
4. We aim to release a fix within 7 days for confirmed vulnerabilities

hookwise undergoes a structured security review pipeline on every release:

1. Automated analysis -- 4 parallel review agents scan the full codebase (~80 source files) across isolated domains: core engine, CLI layer, feed producers, and Python TUI
2. False-positive filtering -- Every candidate finding is independently validated against a strict exclusion list and concrete exploitability criteria (confidence >= 8/10 required)
3. PR-level review -- Every pull request is diffed against origin/HEAD and reviewed for security regressions before merge

The most recent full-package review (v1.3.0) covered all source files and returned zero confirmed vulnerabilities.

hookwise operates under the following trust boundaries:

The codebase enforces the following security practices by design:

• Parameterized SQL everywhere -- All SQLite queries use ? or @param placeholders. No string concatenation in query construction.
• Safe YAML parsing -- Go uses gopkg.in/yaml.v3 (safe by default, no code execution tags); Python TUI uses yaml.safe_load().
• JSON-only deserialization -- All external data is parsed with Go's encoding/json, which cannot execute code.

• No code execution from untrusted or semi-trusted input -- External API responses and Claude Code hook payloads are parsed as data only (encoding/json) and are never evaluated as code.
• Handler dispatch uses no shell -- The dispatcher runs handler executables via exec.Command with arguments passed as array elements (not a shell string), so a hook payload cannot inject shell commands.
• Custom feeds run a shell by design, scoped to trusted config -- The optional feeds.custom mechanism executes its configured command via sh -c (internal/feeds/custom.go). This is intentional and is scoped to the Trusted hookwise.yaml -- the same trust level as a Makefile target or an npm scripts entry. A custom feed command runs with your privileges, so only configure commands you would run yourself; treat a project's hookwise.yaml the way you would treat its build scripts before enabling custom feeds. hookwise never derives a shell command from any semi-trusted or untrusted input.

• Restrictive permissions -- Database files use 0o600; directories use 0o700 (owner-only).
• Atomic writes -- Config, cache, and state files use temp-file-plus-rename to prevent corruption.
• PID files with O_EXCL -- Prevents TOCTOU race conditions in daemon startup.

• Fail-open design -- Internal errors in hookwise never block Claude Code tool calls. The safeDispatch wrapper ensures graceful degradation.
• Read-only access to Claude Code data -- The insights producer reads ~/.claude/usage-data/ but never writes to it.
• No network listeners -- hookwise does not open any ports or accept inbound connections. The daemon communicates via the filesystem (cache bus).

hookwise maintains a minimal dependency footprint:

• Runtime (Go): modernc.org/sqlite, cobra, yaml.v3 — compiled into a single static binary (CGO-free, ~19 MB stripped release)
• Test-only (Go): testify — not compiled into the shipped binary
• Runtime (Python TUI): textual, pyyaml, rich, anthropic
• No native HTTP server dependencies -- External data is fetched via Go's net/http or Python urllib
• Binary releases -- Go binaries are published via GitHub Releases with Dagger-based CI builds