v0.3.1

askcodex 0.3.1 is a dependency-only security patch. It has no CLI, output or behavior changes compared with 0.3.0. Upgrade if you use the v0.3.0 binaries, which link rustls 0.23.43.

Security fixes (bundled TLS stack)

Crate	0.3.0	0.3.1	Advisories fixed
rustls	0.23.43	0.23.46	GHSA-2mjx-qc3c-rqvc (medium): TLS 1.3 handshake messages accepted across encryption-level boundaries · GHSA-c8m5-c952-6cm4 (medium): session Debug formatting exposes TLS secrets · GHSA-w44r-93j7-qv28 (low): an unauthenticated ServerHello can downgrade later key-exchange shares to classical X25519 · GHSA-h2v8-394m-p7vv (low): Brotli certificate decompression memory (the Brotli feature is not enabled in askcodex)
rustls-webpki	0.103.13	0.103.16	RUSTSEC-2026-0382: certificate paths through distrusted emSign intermediates
webpki-roots	1.0.9	1.0.10	RUSTSEC-2026-0381: bundled roots allowed the distrusted emSign intermediates

h2 also moved 0.4.15 → 0.4.16 (RUSTSEC-2026-0258). It is a test-only dependency (httpmock) and is not in the binary.

Also since 0.3.0

• The skill's text-prompting reference defines its extraction example's fields and null policy (#17). It is installed only with ./install.sh --skills ..., and the binary does not change.

What's Changed

• fix(deps): patch rustls and TLS root advisories in Cargo.lock (#19)
• docs(skill): the extraction example defines its fields and null policy (#17)
• chore(release): askcodex 0.3.1 (#20)

Binaries are built with --locked from the tagged commit. Verify the archive against SHA256SUMS before you install it.

Full Changelog: v0.3.0...v0.3.1

v0.3.0

What's Changed

• fix(install): keep skill backups outside every skills directory by @JairoTorregrosa in #15
• feat: agent-first output and request controls (0.3.0) by @JairoTorregrosa in #14
• docs(skill): teach the askcodex 0.3.0 agent controls by @JairoTorregrosa in #16

Full Changelog: v0.2.0...v0.3.0

v0.2.0

What's Changed

• feat!: GPT-6 catalog, gpt-6-sol default, per-model prompting guides by @JairoTorregrosa in #3
• docs: qualify prompt-driven transparency; Free/Go default guidance by @JairoTorregrosa in #4
• skill: Luna example declares its null policy and validates the schema by @JairoTorregrosa in #5
• skill: withhold a conclusion that rests on missing evidence by @JairoTorregrosa in #6
• skill: report only the provenance a command actually has by @JairoTorregrosa in #7
• skill: report blocked requests instead of rewording around them by @JairoTorregrosa in #8
• skill: redact sensitive data before it leaves the machine by @JairoTorregrosa in #9
• skill: Sol's proceed line withholds fixes that depend on missing code by @JairoTorregrosa in #10
• skill: preflight the askcodex version the skill describes by @JairoTorregrosa in #11
• skill: label the Luna check as structural by @JairoTorregrosa in #12
• feat: Codex 0.160.0 catalog, gpt-6.1-sol default and guide by @JairoTorregrosa in #13

Full Changelog: v0.1.0...v0.2.0

v0.1.0

askcodex 0.1.0 refactors command execution, credential handling and machine output while preserving the existing command paths and raw backend escape hatch.

Migration

• Semantic --json responses now use {"schema_version":1,"command":"...","result":{...}}. Change selectors such as .text to .result.text. Usage, models and transcription retain original responses in .backend.
• --events emits NDJSON. Ask streams text deltas and emits exactly one final result only after successful completion. Operational errors are structured on stderr.
• ./install.sh installs only the binary by default. Use --skills agents|claude|all to synchronize skills and --check-auth for an explicit no-refresh auth check.
• Internal library modules are now private. Public parsing, execution, errors, data models and configuration remain available.
• Stdin is limited to 16 MiB; WAV and combined PNG references to 25 MiB. Media must be regular files.

Reliability and security

Local inputs are prepared before authentication. Credential sessions retain their store; runtime credential documents have no generic serialization. Persistent kernel locks release automatically on process death, and OAuth redirects are disabled per request. Stream write failures stop response consumption immediately. Image publication retains atomic replacement behavior.

The lock coordinates participating new askcodex processes only, not Codex or older unlink-lock versions. Restoring old files cannot reverse a server-side token rotation. Keyring support and a native Codex credential-owner adapter remain unimplemented; the latter is plausible but unverified as a drop-in replacement.

Verification

366 offline Rust tests, 84 governance tests and 7 installer tests pass, along with format, Clippy, private rustdoc, release build, generated-reference comparison, and explicit archived command/help parity. Fresh synthetic text and WAV transcription calls passed with --no-refresh and unchanged credential bytes. Optional full live/image tests were not rerun.

See PR #2, command reference, and output contract.

Release archives cover Linux and macOS on x86_64 and ARM64. Verify downloads against the included SHA256SUMS.

askcodex 0.0.1

First public cut of askcodex: ask OpenAI's models from your terminal.

• Nine commands: ask, image create/edit, models, usage, whoami, raw, auth status/refresh.
• Auth comes from the credentials codex login already stores. Nothing to configure.
• Agent-native --json output (model, effort, text, token usage in one document).
• Binaries for Linux and macOS on x86_64 and aarch64; verify with SHA256SUMS.