← Back to Payloads
AI News2026-09-18

OpenAI Codex v0.155.0 Adds Touch ID Verification for MCP Requests, Real Bedrock Credential Rotation, and an Experimental /voice Mode You Should Ignore Until the Flag Is Off

OpenAI Codex v0.155.0 (2026-09-17 23:16 UTC) is the new stable CLI replacing the 0.155.0-alpha churn. Six capability additions (experimental /voice with live WebRTC transcripts, TUI reasoning summaries, agents-overview task lifecycle with worktree ownership, Touch ID verification for MCP requests on supported Macs with Secure Enclave signing, configurable app-server daemon update schedules, Amazon Bedrock credential rotation with caching and expiration refresh) plus six fix clusters (turn-acceptance, tmux/viewport restoration, MCP OAuth expired-credential reporting, automatic approval review fidelity, account-switch state hygiene, WSL Windows-process escape + brokered shell credential exposure). Documentation comparison; no firsthand test.
Quick Access
Install command
$ mrt install codex
Browse related skills

OpenAI Codex v0.155.0 Adds Touch ID Verification for MCP Requests, Real Bedrock Credential Rotation, and an Experimental /voice Mode You Should Ignore Until the Flag Is Off

Hey guys, Mr. Technology here.

rust-v0.155.0 shipped 2026-09-17 23:16 UTC as the new stable Codex CLI, replacing the 0.155.0-alpha.1..alpha.9.x pre-release churn we've been tracking for the last week. Six new capabilities, six bug-fix clusters of consequence, one chore that quietly tightens the publishing contract between the Python SDK and the CLI runtime. The headline capability is Touch ID verification for MCP requests in local TUI sessions on supported Macs — every MCP request can now require a biometric prompt at the Secure Enclave level before the tool runs. The headline security fix is blocked Windows-process escapes from restricted WSL sandboxes and hardened brokered shell snapshots against credential exposure. The headline operational improvement is Amazon Bedrock now obtains AWS credentials from configured commands, with caching, expiration-based refresh, and authentication recovery — i.e., real Bedrock credential rotation on the CLI, not the placeholder rotation the SDK was doing.

This is a documentation-surfacing report. Every claim below is verifiable verbatim in the v0.155.0 release notes on GitHub. I have not run v0.155.0 against a production harness; treat every operational claim as documentation comparison, not firsthand test. (Release 0.155.0, Compare v0.154.0...v0.155.0)

What Happened

OpenAI shipped rust-v0.155.0 on 2026-09-17 23:16 UTC. Six new capabilities (experimental /voice, TUI reasoning summaries, agents-overview task lifecycle, Touch ID verification, configurable app-server daemon updates, Amazon Bedrock credential rotation) plus six documented bug-fix clusters (turn-acceptance reliability, tmux/viewport restoration, MCP OAuth reporting, automatic approval review fidelity, account-switch state hygiene, WSL sandbox escapes). Plus a chore that aligns Python SDK and runtime publishing to the stable CLI version. The 0.155.0-alpha series has been visible on the atom feed since the prior desk sweep, but this is the first stable tag in the series and the first one suitable for production pin.

What Actually Changed

Verbatim from the rust-v0.155.0 release page (fetched 2026-09-18 14:09 UTC), grouped by what each item actually does. Six capability additions, six fix clusters, one chore.

Capability additions

  • Experimental /voice conversations with live transcripts and microphone controls on supported builds, enabled through /experimental. This is the headline capability and the one to be careful about. Per the release notes, /voice is a WebRTC-based live-voice conversation mode that surfaces the live transcript and microphone controls in the TUI composer strip, with split-flap-tile animation, mute shortcut, and configurable mute key. It is experimental and lives behind the /experimental flag — i.e., not on by default. For a CLI whose core proposition has always been text-only terminal-first interaction, this is a meaningful expansion of surface area. The release page also adds an explicit warning to the user if their connected Codex service is older than the CLI — a small but useful guard for mixed-version environments.
  • The TUI now shows live reasoning summaries in the status row and completion timestamps after successful turns. Operationally useful for anyone running Codex unattended: a failed turn is now visible by its timestamp in the status row, and the live reasoning summary surfaces what the model is currently reasoning about. (#43558, #43921)
  • Task hiding, archiving, and deletion in the agents overview, plus worktree ownership details and confirmed deletion of clean managed worktrees. Substantive agents-overview upgrade — three missing affordances are now shipped. Hidden tasks are filtered out of the default view but recoverable; archived tasks are a separate bucket; deletion of a clean managed worktree (i.e., a worktree with no uncommitted changes) now confirms and deletes, removing the manual git worktree remove step. Worktree ownership is now visible on each task. (#43942, #44424, #44433)
  • Touch ID verification for MCP requests in local TUI sessions on supported Macs. This is the substantive security addition. On Macs with a Secure-Enclave-capable fingerprint reader, the bundled TUI can now gate MCP requests on a biometric prompt — the macOS user-verification provider signs the request with a Secure Enclave-backed credential, and the agent verifies the signature before the MCP call runs. The release adds three PRs (#43624, #43712, #43715) covering macOS user verification with Secure Enclave signing, the TUI user verification prompt component, and enabling user verification for the bundled TUI on supported devices. For any developer running Codex on a Mac with sensitive MCP servers (filesystem, email, calendar, shell, code-exec), this is the gating primitive that was missing. Important caveat: this is a local-TUI feature on supported Macs. It does not gate CLI, app-server, cloud, or non-Mac workflows.
  • Configurable daemon update schedules and codex app-server daemon update; saved threads and active goals can recover after daemon restarts. Two related items: app-server daemon automatic updates are now configurable (so an org that wants to pin the daemon to a specific version for stability can do so), and there is now an explicit app-server daemon update command. Saved threads and active goals can recover across daemon restarts. (#43542, #43562, #44314) Also: the standalone updater now runs on managed daemon starts (so daemon-update cadence is honored for managed installs), and the standalone release pins are preserved during daemon updates.
  • Amazon Bedrock can now obtain AWS credentials from configured commands, with caching, expiration-based refresh, and authentication recovery. The substantive Bedrock-credential-rotation story. Pre-v0.155.0, Codex against Bedrock relied on whatever the ambient AWS credential chain resolved to — a static ~/.aws/credentials profile, an environment variable, an instance role. Post-v0.155.0, Codex can be configured to run a credential-refresh command (e.g., a vault query, a federation exchange, an STS AssumeRole) and cache the result with expiration. If the cached credential expires mid-session, Codex refreshes and continues; if the refresh fails, it surfaces a recovery path instead of failing the turn. (#44028) For any team running Codex against Bedrock in a regulated environment (SOC2 / HIPAA / FedRAMP) where ambient credentials are forbidden, this closes the "you can't actually run Codex on Bedrock in production" gap.

Bug-fix clusters

  • Accepted prompts are now saved even when compaction fails before a turn starts. Reliability fix — a session that crashes between prompt acceptance and turn-start now preserves the prompt for --resume, rather than orphaning it.
  • Fixed missed tmux resizes, transcript viewport restoration, and stale history appearing after switching threads. Three TUI fixes in one bullet — the tmux-resize notification was being missed on some sessions, the transcript viewport was not restoring correctly after a thread switch, and stale history from a previously-selected thread could appear after a thread switch.
  • MCP servers now report expired OAuth credentials accurately and provide reconnect guidance when token refresh fails. Two related fixes — pre-v0.155.0, an MCP server with an expired OAuth token would surface a generic connection error without indicating that the token was the cause, sending the developer on a debug wild goose chase. Post-v0.155.0, the error explicitly names "OAuth credential expired" and points to the reconnect flow.
  • Automatic approval reviews now preserve complete actions and authorization evidence more reliably, retry transient failures, and distinguish review failures from unsafe-action findings. Three fixes in one — the automatic approval review (a Codex feature that asks the Guardian-class safety system to evaluate tool calls before execution) was silently dropping authorization evidence on transient failures, which would cause the review to fail with an opaque error. Post-v0.155.0, the review retries transient failures, preserves evidence across retries, and explicitly distinguishes "the review process itself failed" from "the review correctly identified an unsafe action."
  • Switching accounts now invalidates remote-control sessions, cached WebSocket state, and model catalogs belonging to the previous identity. Account-switch state hygiene — pre-v0.155.0, switching the Codex account on a machine did not invalidate remote-control session tokens, cached WebSocket connections, or cached model catalogs from the previous account, meaning a new account on the same machine could inherit those surfaces. Post-v0.155.0, account switch invalidates all three.
  • Blocked Windows-process escapes from restricted WSL sandboxes and hardened brokered shell snapshots against credential exposure. Two security fixes in one. The Windows-process escape was a path where a process inside a restricted WSL sandbox could escape to the host Windows side via a sandbox-misconfiguration; closed. The brokered shell snapshot hardening prevents credential exposure in shell snapshots (snapshots of brokered shell sessions that Codex records for debugging); closed.

Chore

  • Aligned Python SDK and runtime publishing with stable CLI releases, using matching versions and verifying runtime assets before SDK publication. A publishing-contract improvement: the Python SDK and the CLI runtime are now published with matching version numbers, and runtime assets are verified before SDK publication. This is plumbing, but it closes a class of "Python SDK is one version ahead of CLI runtime" mismatches that have been a recurring pain point in 2026.

Why Developers and Founders Should Care

Four buckets of impact, ranked by who is most affected.

Anyone running Codex on a Mac with sensitive MCP servers (most affected). The Touch ID verification for MCP requests is a substantive security addition. If you are running Codex on macOS with filesystem, email, calendar, shell, or code-exec MCP servers, this is the gating primitive that lets you keep the agent running with sensitive MCP access while requiring a biometric prompt for every MCP call. The pre-v0.155.0 posture was either "trust the agent" or "manually approve every MCP call from the UI"; the post-v0.155.0 posture is "the operating system prompts you for a fingerprint every time the agent wants to use an MCP tool." This is a step-change for any solo developer running Codex on a personal Mac, and a major step for any founder running Codex on a fleet of Macs with sensitive MCP integrations. Caveat: local-TUI only on supported Macs; does not gate CLI, app-server, cloud, or non-Mac workflows.

Anyone running Codex on Amazon Bedrock in a regulated environment (next most affected). The Bedrock credential rotation with caching + expiration + refresh + recovery is the substantive capability addition. Pre-v0.155.0, Codex against Bedrock relied on ambient AWS credentials, which is a non-starter for SOC2 / HIPAA / FedRAMP environments where ambient credentials are forbidden. Post-v0.155.0, Codex can run a configured credential-refresh command and cache the result with expiration — closing the "you can't actually run Codex on Bedrock in production" gap for regulated workloads. (#44028)

Anyone managing the agents overview at scale (next most affected). Task hiding, archiving, deletion, worktree ownership, and confirmed deletion of clean managed worktrees are the substantive agents-overview upgrade. The agents overview is now actually usable as a long-running task dashboard rather than a firehose of one-off sessions. (#43942, #44424, #44433)

Anyone running account-switch workflows (next most affected). The account-switch invalidation of remote-control sessions, WebSocket state, and model catalogs is a state-hygiene fix. Pre-v0.155.0, switching accounts on a Codex install could leak state from the previous account; post-v0.155.0, the new account starts clean.

Skip /voice unless you are testing it. It is experimental, gated behind /experimental, and the release page explicitly warns about version-mismatch issues. Useful for design exploration, not for production pin.

Evidence and Verification

All capability and security claims above are taken verbatim from the rust-v0.155.0 release page on GitHub, fetched 2026-09-18 14:09 UTC. The release tag is rust-v0.155.0, atom feed timestamp 2026-09-17T23:16:01Z, comparison range rust-v0.154.0...rust-v0.155.0. The six capability additions, the six fix-cluster bullets, and the one chore are taken verbatim from the release page.

This is a documentation comparison, not a firsthand test. I have not installed v0.155.0 against a production harness. I have not exercised the Touch ID verification on a real Mac with a Secure-Enclave-capable fingerprint reader. I have not run a Bedrock credential-rotation cycle end-to-end. I have not validated the /voice experimental flag. Treat every operational claim as documentation comparison. Cross-check against your own staging harness before adopting any new flag or relying on any fixed behavior.

Cost, Risk, and Limitations

Bedrock credential rotation requires a configured refresh command. The release does not include a built-in credential provider — you have to supply the command. For a team running Codex against Bedrock with no existing refresh mechanism, this is a precondition, not a turnkey solution. Plan the integration.

Touch ID verification is local-TUI only. It does not gate CLI, app-server, cloud, or non-Mac workflows. For teams running Codex in CI (where there is no fingerprint reader), this fix is irrelevant; the underlying MCP-request-verification machinery is platform-bound to macOS.

/voice is experimental. Behind /experimental. Version-mismatch warnings between CLI and Codex service will appear if the connected service is older than the CLI. Treat as design exploration, not production pin.

Account-switch invalidation may break legitimate cross-account workflows. If you currently rely on Codex inheriting state from a previous account (e.g., a long-running remote-control session that survives an account refresh), that workflow now requires explicit re-establishment. Test the account-switch path against your staging harness before adopting v0.155.0.

The Bedrock credential rotation has not been benchmarked in the release notes. The release does not publish a credential-rotation-cycle latency number. If your Codex-on-Bedrock workload is latency-sensitive (and they all are), benchmark the refresh path against the pre-v0.155.0 ambient-credential baseline before adopting.

Mr. Technology Verdict

v0.155.0 is a substantive stable release for the production Codex user. The Touch ID verification is the most consequential capability addition for the developer-and-founder audience — it is a step-change in the practical security of running Codex on a personal Mac with sensitive MCP servers. The Bedrock credential rotation is the most consequential capability addition for the enterprise audience — it is the gating primitive for any regulated workload that needs Codex against Bedrock in production. The agents-overview task lifecycle is the most consequential workflow improvement — it makes the agents overview usable as a real task dashboard.

The /voice experimental feature is interesting but should be ignored until the flag is off the experimental track. The Windows-WSL sandbox escape fix and the brokered shell snapshot hardening are quietly important for any team running Codex on Windows. The Python SDK + CLI runtime publishing alignment is plumbing but it closes a real class of mismatches.

The release is a documentation comparison; I have not run it against a production harness. Treat every operational claim as documentation comparison, cross-check against your own staging harness before adopting any new flag or relying on any fixed behavior.

Recommended Action

Today: pin v0.155.0 on your staging harness. Run the Touch ID verification path on a supported Mac with at least one sensitive MCP server configured. Run a Bedrock credential-rotation cycle end-to-end against a sandbox AWS account. Validate the agents-overview task lifecycle against your long-running workflows. Test the account-switch path against your staging harness.

This week: if the staging validation passes, promote v0.155.0 to production. Pin the version explicitly in your install scripts. Re-test any MCP-driven workflow that relied on the pre-v0.155.0 ambient behavior.

Skip: the /voice experimental feature until it leaves the experimental track. Useful for design exploration, not for production pin.

Mr. Technology

Sources


Originally published: 2026-09-18 14:08 UTC Last verified: 2026-09-18 14:09 UTC (sources fetched 2026-09-18 14:09 UTC; rust-v0.155.0 release timestamp verified 2026-09-17T23:16:01Z via the GitHub atom feed) No corrections at this time.

Related Dispatches