The Scanner Writes the Fix, Codex Takes Directions Mid-Answer, and Claude Logs Who Said Yes

At a glance

• GitHub's AI scanner now reviews code CodeQL skips, and Copilot can fix alerts in bulk.
• Codex lets you redirect an answer while it is still being written.
• Claude Code telemetry now records whether a rule or the classifier approved a tool call.
• Copilot is retiring three models next month, so check which ones your team pins.

Thursday was a security and control day, so today's thread is who decides what an agent may do and how you check it later. GitHub now lets an AI find and fix vulnerabilities in code CodeQL never covered, Codex lets you correct an agent mid-turn, and Claude Code records what approved each action. Treat today as the day to turn on AI Scan for one repository and assign one alert to Copilot, update Codex and try steering a long task, and add the new Claude Code permission fields to whatever dashboard watches auto mode.

Top Stories

1. GitHub's scanner now covers more code and writes its own fixes

GitHub's Changelog put two security features into general availability on Thursday. AI Scan for pull requests, posted at 11:04 AM ET, looks for vulnerabilities in pull request changes written in languages CodeQL does not cover, including PHP, Shell and Bash, Terraform HCL, Dockerfiles, JSP for Java and Blazor for C#. Findings appear next to CodeQL results with an "AI" marker and are advisory, so they never fail the code scanning check or block a merge through rulesets. It skips fork and Dependabot pull requests, and private repositories need Code Security or Advanced Security, though no Copilot license.

Agentic autofix, posted at 9:49 AM ET, goes further. When you assign a code scanning alert to Copilot, it explores the relevant files, proposes a fix, reruns CodeQL to confirm the alert closes, and opens a draft pull request explaining the fix and its checks. You can select up to 25 alerts from a repository or a security campaign for one pull request, or assign one through the REST API to copilot-swe-agent[bot]. It handles only default branch alerts, is for Code Security and Advanced Security customers on GitHub Enterprise Cloud, and needs an enterprise admin to enable the Agentic autofix policy plus a Copilot billing method. Both features draw Code Security AI Credits, and autofix also uses Actions minutes.

Practical dev impact: AI Scan is a second reviewer for the infrastructure and shell code that often slips past static analysis, but because it is advisory, someone still has to read the findings before merging. Agentic autofix turns an alert backlog into reviewable pull requests, so in our reading the real work moves to setting a credit budget and deciding who reviews Copilot's fixes.

2. Codex v0.163.0 lets you steer an answer while it is still running

OpenAI tagged Codex v0.163.0 on GitHub at 5:30 PM ET Thursday, and npm moved latest to it at 5:37 PM ET, up from v0.162.1. A follow-up sent while Codex is responding now steers the turn right away, and interrupting a turn also stops active Code Mode cells and their nested tool calls. The /config screen now searches preferences, applies supported changes immediately and flags those that need a restart.

Resuming a saved thread now restores its permissions and network approvals while still honoring explicit overrides and current managed policy. Codex also stops bundling a patched zsh.

Practical dev impact: Steering saves killing a long turn and starting over when the agent heads the wrong way, and the resume and proxy fixes help anyone who fences unattended runs with approvals and network policy. Because of the zsh change, our advice is to check custom shell setups after updating.

3. Claude Code v2.1.301 records who said yes to each tool call

Anthropic published Claude Code v2.1.301 to npm at 10:34 AM ET Thursday, and it holds latest and next, while stable still read v2.1.292 at our 5:54 AM ET check Friday. The OpenTelemetry tool_decision event gains classifier_verdict and permission_mode, so you can tell an auto mode classifier decision apart from a permission rule, and with OTEL_LOG_TOOL_DETAILS=1 it also carries classifier_category, naming the built in rule a classifier block matched. So far these fields are documented only in the changelog; the monitoring docs do not list them yet.

Several fixes stop the tool from acting, or reporting success, behind your back. Quitting no longer removes a worktree made by a WorktreeCreate hook without asking when it may hold unpushed commits, and switching to a model with a smaller context window no longer spends usage credits on a summary by the previous model without asking. Workflow agents no longer skip their step while the run reports success, and Grep and Glob now report a failed search instead of no matches.

Practical dev impact: If your team runs auto mode, the new fields let you count how often the classifier rather than your own rules approved or blocked actions, which in our view is the evidence to gather before widening it.

Practical Impact Analysis

All three stories hand agents more decisions while giving people better ways to check them, so in our view you should switch on the evidence before widening the autonomy. On GitHub, enable AI Scan on one active repository, watch a week of findings and credit use, and pilot agentic autofix on a handful of alerts before selecting 25 at once. Codex v0.163.0 is a reasonable update for individuals, while teams with custom shells should test it first. For Claude Code, individuals can take v2.1.301 from latest, teams that prefer slower change can stay on stable at v2.1.292, and whoever owns telemetry can add the new fields once they confirm them in their own data.

One smaller GitHub item needs a calendar entry. Copilot will deprecate Claude Opus 4.8, Claude Opus 5 and GPT-5.6 Terra on 16 November, with Claude Opus 5.5 suggested for the two Claude models and GPT-6.1 Sol or GPT-6 Luna for Terra, though Opus 4.8 stays for individual subscribers on annual plans. At our 5:54 AM ET check Friday, Copilot CLI latest held at v1.0.98, and prerelease had moved to v1.0.99-2 at 8:50 PM ET Thursday with no changelog entry yet. Gemini CLI latest held at v0.64.0 with preview at v0.65.0-preview.0, and the Cursor changelog had no new entry at 5:58 AM ET.

Tutorial

We have not run this script end to end on every platform, so read it before you run it. It prints channel pointers and local versions, and every line that installs, enables telemetry or assigns an alert is commented out and labeled.

#!/usr/bin/env bash
# We have not run this script end to end on every platform. Nothing below installs or changes anything unless you uncomment it.
# 1. Channel pointers. At 5:54 AM ET Fri 16 Oct npm showed Codex latest v0.163.0,
#    and Claude Code stable v2.1.292, latest and next v2.1.301.
npm view @openai/codex dist-tags --json
npm view @anthropic-ai/claude-code dist-tags --json

# 2. What this machine runs
codex --version
claude --version

# 3. Optional, Codex latest channel, npm installs only, npm latest at 5:54 AM ET Fri 16 Oct:
# npm install -g @openai/codex@0.163.0

# 4. Optional, Claude Code latest and next channel, npm installs only, at 5:54 AM ET Fri 16 Oct:
# npm install -g @anthropic-ai/claude-code@2.1.301

# 5. Optional, documented in Claude Code monitoring docs: print telemetry events to this shell
# export CLAUDE_CODE_ENABLE_TELEMETRY=1
# export OTEL_LOGS_EXPORTER=console

# 6. Optional, needs the Agentic autofix policy enabled. Replace OWNER, REPO and 42.
# gh api -X PATCH repos/OWNER/REPO/code-scanning/alerts/42 -f 'assignees[]=copilot-swe-agent[bot]'

With step 5 enabled, run a short session that triggers a permission check and look at the tool_decision event. Whether classifier_verdict and permission_mode show up there in a short session, especially outside auto mode, is our expectation, since only the changelog describes them so far. Step 6 spends Code Security AI Credits and Actions minutes, so try it on a low risk alert. Do not mix npm and native installs of the same tool on one machine.

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever AI you use.