Agent Patch Weekly — Issue #003

Patch notes for the agent meta · Week of Oct 28, 2026

GG and welcome back to patch day. This one covers Oct 21 to 28: OpenAI shipped a half-price coding model, Copilot now lets a model approve the safe commands, Claude Code fixed a bug that could overwrite a big CLAUDE.md, and Epic made AI optional at its AI game jam. Plus a rules file check we actually ran. Every item links to its source.

🟢 BUFFS

GPT-6.1 Terra: half the API price, now in Codex and Copilot (Oct 26) OpenAI's new fast model is in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.1-terra at $1 per million input tokens ($0.05 cached) and $5 per million output. OpenAI says that's 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. On Terminal-Bench 4.0 at max effort it solves 50.3% of tasks, up from GPT-5.6 Terra's 21.5%. OpenAI's own caveat: "For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice." GitHub made it generally available in Copilot the same day for Pro, Pro+, Max, Business and Enterprise, with a gradual rollout. GPT-5.6 Terra leaves Copilot on Nov 16 (see Issue #002), so this is the obvious swap to test. → OpenAI announcement · GitHub changelog

Copilot assisted permissions are generally available (Oct 26) In Copilot CLI, the Copilot app and VS Code, a model now reviews each request to run a command, access a file system path or open a URL. Clearly safe requests go through; everything else still asks you. In the CLI, run /permissions and pick assisted. Admins can block Allow All while keeping assisted mode, turn assisted off, or set the default mode. GitHub's docs are blunt about the limits: "The model can make mistakes and approve dangerous or unintended actions. This feature is not a substitute for sandboxing or access restrictions." → GitHub changelog · docs

Claude Code 2.1.311 fixes long-session pain (Oct 26) Standard-context sessions no longer stop with "Usage credits required for 1M context" when a request runs past 200K tokens; Claude Code now compacts and keeps going. Also fixed: long freezes at startup and after file edits when a CLAUDE.md holds a very long list item, stale MEMORY.md and CLAUDE.md content right after automatic compaction, and SubagentStop hooks firing about twice a minute instead of once. → Claude Code changelog

Codex CLI 0.165.0 (Oct 27) Adds a built-in Google Cloud Vertex AI provider (model_provider = "gcp") and GPT-6.1 Terra in the model catalogs. You can now change reasoning effort mid-conversation by default on supported models, fresh profiles get memory v2 (which can import memories from other agents), and the session picker has a repository filter. It also closes a marketplace URL validation gap that could let Git connect to a different host than the one checked. → Release notes

🔴 NERFS (and heads-ups)

Claude Code could overwrite a big CLAUDE.md, and 2.1.312 changes a few defaults (Oct 27) The fix line says it all: "Fixed Claude overwriting a very large CLAUDE.md, rules file or @-included file in a subdirectory after seeing only its start." Write now asks Claude to read the whole file or edit in place. If you ran long sessions recently, check git diff on your rules files. The same release stops sending the claude_code.safety_event OpenTelemetry event, which only arrived in 2.1.307 on Oct 21, so a dashboard built on it will go quiet. Prompt and agent hooks now run at most 24 at a time per session, sub-agents included. → Claude Code changelog

Cline changed its default models again (Oct 26) Cline v4.1.26: "Default models change for most other providers, mostly to Kimi K3.1 Pro, MiniMax M3.1 Flash or DeepSeek V4.1 Flash 1017. If you use a provider without pinning a model, expect a different default." With no model chosen, the Cline provider goes back to Claude Sonnet 5.5, and model requests now retry at most 3 times instead of 5. Pin your model if you care which one writes your code. → Cline v4.1.26 release

Atlassian: faster shipping, same friction (Oct 20) Atlassian's 2026 State of Developer Experience report: 87% of developers say AI helps them ship faster, but 72% have the same or more friction than before AI. In 2025, 68% said they saved 10+ hours a week; in 2026 it's 25%. 65% regularly spend time correcting AI output. Only 17% of developers and 23% of leaders say their organization has a clearly communicated and implemented AI strategy, while 60% of developers say AI usage is treated as a productivity signal and 41% sometimes use it more than necessary because they feel expected to. → Atlassian blog · report

Attackers are running agent swarms too (Oct 21) Microsoft says Storm-0940, a China-state aligned actor, used ARTEX, an open-source multi-agent LLM penetration testing framework, to automate attacks on hundreds of organizations. In multiple cases it went from first targeting to privileged cloud access in under 3 hours, and after the ConnectWise Cloud Backup incident it tried to sign in to more than 2,250 tenants. Microsoft says the activity didn't come from a compromise of or vulnerability in its own systems. For affected customers, its guidance starts with keeping the affected ConnectWise app connections disabled and reviewing sign-in and audit records. The takeaway for agent builders: the other side has the same tools, so review what your agents' tokens can reach. → Microsoft Security blog

🎁 NEW LOOT (models and tools)

🛠️ WORKFLOW OF THE WEEK: The Rules File Health Check

The problem: your rules file (AGENTS.md, CLAUDE.md) grows and rots, and nothing tells you. Codex stops adding project rules files once their combined size hits project_doc_max_bytes, 32 KiB by default (OpenAI docs). Claude Code just fixed Claude overwriting a very large CLAUDE.md after reading only its start, and a freeze caused by very long list items. And stale rules get followed: see the study in our AGENTS.md guide. The fix: one script that turns your checkpoint red when a rules file is over budget or got clobbered, and warns about dead paths and giant lines.

  1. Install the free Checkpoint Loop, which gives you scripts/agent-check.sh. Save the script below as scripts/rules-check.sh and run chmod +x scripts/rules-check.sh.
  2. Hook it into the loop. In scripts/agent-check.sh, add this line just above the secret guardrail:
    if [[ -x scripts/rules-check.sh ]]; then run "scripts/rules-check.sh"; fi
  3. Add this to your AGENTS.md:
    ## Rules file hygiene
- Never rewrite AGENTS.md or CLAUDE.md from scratch. Read the whole file, then edit in place.
- If ./scripts/rules-check.sh fails, stop and show me the output. Don't fix it by deleting rules.
- When you rename or delete a path, update any rule that mentions it.
  4. Run ./scripts/rules-check.sh once by hand. In CI, use RULES_STRICT=1 ./scripts/rules-check.sh so warnings fail too.

What it checks, in any git repo:

Here's scripts/rules-check.sh (114 lines), or download it:

#!/usr/bin/env bash
# rules-check.sh: Rules File Health Check (Agent Patch Weekly, Issue #003)
# Checks the agent rules files in a git repo (AGENTS.md, AGENTS.override.md,
# CLAUDE.md, GEMINI.md) before you checkpoint:
#   FAIL  a directory's Codex chain (root -> dir) is over the project-doc budget
#   FAIL  a tracked rules file shrank by more than half, or was deleted, since HEAD
#   WARN  a line is very long
#   WARN  a backticked path in a rules file doesn't exist
# Exit 0 = green, 1 = red. RULES_STRICT=1 turns warnings into failures.
# Settings (env): RULES_MAX_BYTES (default 32768, Codex's default
# project_doc_max_bytes), RULES_MAX_LINE (default 1000 chars),
# RULES_SHRINK_PCT (default 50), RULES_IGNORE (regex of paths to skip).
set -uo pipefail

MAX_BYTES="${RULES_MAX_BYTES:-32768}"
MAX_LINE="${RULES_MAX_LINE:-1000}"
SHRINK_PCT="${RULES_SHRINK_PCT:-50}"
STRICT="${RULES_STRICT:-0}"
IGNORE="${RULES_IGNORE:-}"

root="$(git rev-parse --show-toplevel 2>/dev/null)" || { echo "rules-check: not inside a git repo"; exit 1; }
cd "$root" || exit 1

fails=0; warns=0
fail() { echo "✖ FAIL  $*"; fails=$((fails+1)); }
warn() { echo "⚠ WARN  $*"; warns=$((warns+1)); }
bytes() { wc -c < "$1" | tr -d ' '; }

is_rules() { case "$(basename "$1")" in AGENTS.md|AGENTS.override.md|CLAUDE.md|GEMINI.md) return 0;; esac; return 1; }

# Rules files in the working tree (tracked + untracked, respecting .gitignore)
files=()
while IFS= read -r f; do
  [ -f "$f" ] && is_rules "$f" && files+=("$f")
done < <(git ls-files -co --exclude-standard | sort -u)

echo "rules-check: ${#files[@]} rules file(s) in $(basename "$root")"

# 1) Codex chain budget. Codex reads at most one file per directory from the
#    project root down to the working directory (AGENTS.override.md wins over
#    AGENTS.md, empty files skipped) and stops adding files once the combined
#    size reaches project_doc_max_bytes.
codex_file() { # $1 = dir ("." for root); prints the file Codex would pick, if any
  local d="$1" c
  for c in AGENTS.override.md AGENTS.md; do
    if [ -s "$d/$c" ]; then echo "$d/$c"; return; fi
  done
}
dirs=()
for f in "${files[@]}"; do
  case "$(basename "$f")" in AGENTS.md|AGENTS.override.md) dirs+=("$(dirname "$f")");; esac
done
if [ ${#dirs[@]} -gt 0 ]; then
  while IFS= read -r d; do
    total=0; chain=""
    path="."
    pick="$(codex_file .)"; [ -n "$pick" ] && { total=$((total+$(bytes "$pick"))); chain="$pick"; }
    if [ "$d" != "." ]; then
      IFS='/' read -r -a parts <<< "$d"
      for p in "${parts[@]}"; do
        path="$path/$p"
        pick="$(codex_file "$path")"
        [ -n "$pick" ] && { total=$((total+$(bytes "$pick"))); chain="$chain + ${pick#./}"; }
      done
    fi
    if [ "$total" -gt "$MAX_BYTES" ]; then
      fail "Codex chain for ${d#./}/ is $total bytes (budget $MAX_BYTES): ${chain#./}. Files past the budget won't all be loaded. Trim, or raise project_doc_max_bytes on purpose."
    fi
  done < <(printf '%s\n' "${dirs[@]}" | sort -u)
fi

# 2) Shrink / delete guard against HEAD
if git rev-parse --verify -q HEAD >/dev/null; then
  for f in "${files[@]}"; do
    if git cat-file -e "HEAD:$f" 2>/dev/null; then
      old=$(git cat-file -s "HEAD:$f"); new=$(bytes "$f")
      if [ "$old" -gt 0 ] && [ $((new*100)) -lt $((old*SHRINK_PCT)) ]; then
        fail "$f shrank from $old to $new bytes since the last commit. Possible accidental overwrite; check 'git diff -- $f' (restore: git checkout -- $f)."
      fi
    fi
  done
  while IFS= read -r f; do
    [ -n "$f" ] && is_rules "$f" && fail "$f is tracked but deleted in the working tree (restore: git checkout -- $f)."
  done < <(git diff --name-only --diff-filter=D HEAD 2>/dev/null)
fi

# 3) Very long lines and 4) backticked paths that don't exist
exts='md|json|jsonc|ts|tsx|js|jsx|mjs|cjs|py|sh|bash|toml|yaml|yml|txt|go|rs|rb|java|kt|swift|cs|cpp|c|h|hpp|html|css|scss|sql|lock|cfg|ini|gd|tscn|uasset|umap|uproject|lua'
for f in "${files[@]}"; do
  awk -v max="$MAX_LINE" -v f="$f" 'length($0) > max { printf "%s:%d:%d\n", f, NR, length($0) }' "$f" |
  while IFS=: read -r ff ln len; do
    echo "⚠ WARN  $ff line $ln is $len chars (limit $MAX_LINE). Very long lines are hard for agents and people to scan; split it."
  done
  n=$(awk -v max="$MAX_LINE" 'length($0) > max' "$f" | wc -l | tr -d ' '); warns=$((warns+n))

  dir="$(dirname "$f")"
  while IFS= read -r p; do
    [ -z "$p" ] && continue
    [ -n "$IGNORE" ] && [[ "$p" =~ $IGNORE ]] && continue
    case "$p" in http*|/*|~*|*'*'*|*'$'*|*'<'*|*'{'*|*'..'*|-*) continue;; esac
    if [[ "$p" == */* ]] || [[ "$p" =~ \.($exts)$ ]]; then
      q="${p#./}"
      [ -e "$q" ] || [ -e "$dir/$q" ] || warn "$f mentions \`$p\`, which doesn't exist. Stale rule? (see agent-patch-weekly.vercel.app/guides/agents-md)"
    fi
  done < <(grep -o '`[^` ]*`' "$f" | tr -d '`' | sort -u)
done

if [ "$STRICT" = "1" ] && [ "$warns" -gt 0 ]; then fails=$((fails+warns)); fi
if [ "$fails" -gt 0 ]; then
  echo "✖ RED: $fails problem(s), $warns warning(s). Don't checkpoint yet."
  exit 1
fi
echo "✔ GREEN: rules files look healthy ($warns warning(s))."
exit 0

What we actually tested: 12 cases on throwaway git repos (Linux, git 2.47, Node 20, Python 3.13). A clean repo came back GREEN. A dead path in AGENTS.md warned, and failed with RULES_STRICT=1. A 30 KB root AGENTS.md plus a 4 KB services/api/AGENTS.md failed at 34,941 bytes; a small AGENTS.override.md in that folder made it pass, and so did raising the budget. A 2,989-character line warned. A committed 13.9 KB CLAUDE.md overwritten with 3 lines failed, git checkout brought it back to GREEN, and deleting it failed. Hooked into the Checkpoint Loop, the overwrite turned agent-check.sh RED. Our own free template's AGENTS.md passed clean. We have not run it against real Codex or Claude Code sessions, and it doesn't count your global ~/.codex/AGENTS.md or Codex fallback filenames. We haven't tested it on macOS either. The path check is a heuristic (backticked strings with a slash or a known file extension), so expect the odd false alarm.

🎮 GAME DEV CORNER

Epic AI MegaJam: AI is now optional, and SF registration opens Nov 1 PC Gamer reports Epic walked back its requirement to use AI in its own AI game jam, and Poptopic's read of the official rules says the extent of AI use won't raise or lower your score. Judges score the game itself: ambition and innovation, theme, fun, gameplay, visuals, level design and audio. The dates: registration for the in-person San Francisco build days (Nov 14–15) opens Nov 1 at 7 AM PST (9 AM CT) on Eventbrite, and space is limited. The theme drops Nov 11 at 11 AM PST (1 PM CT), the jam runs Nov 12–15, and the deadline is Nov 15 at 5:59 PM PST (7:59 PM CT). Epic's page says to get your coding agents and MCP working before kickoff, and to declare any prompts, scripts or assets you made before the jam. Our MegaJam guide has the rest. → Jam page · Poptopic on the rules

Tim Sweeney vs Steam's AI disclosures (Oct 24–25) Epic's CEO went after Valve's mandatory AI disclosures on X: "These AI statements are just as partisan as mandatory political essays or forcing developers to list their voting history to drive cancel campaigns and public struggle sessions." PC Gamer's Shaun Prescott made the case for Valve: the rule came in while the tech was "legally murky," and Steam released 18,433 games in 2024 to the Epic Games Store's 1,100. Whichever side you're on, keep a record of what AI did in your game. You'll need it for Steam, and MegaJam's Glass Box modifier requires a dev log. → PC Gamer

Creative Assembly on Shogun 2: Complete Edition (Oct 23) The remaster was pulled from sale days after launch. Total War VP Roger Collum: "This was not a Creative Assembly developed game," and "We took our eyes off its external development." He says CA doesn't use generative AI to create game assets and demands the same of external developers, and that the Complete Edition "is being investigated." His line on AI: "AI can support the creator. It shouldn't replace the creator." If you outsource, put your AI rules in the contract, because your name is still on the box. → CA's statement · PCGamesN

Strata: Unreal projects to browser links (Oct 22) Wonder Interactive unveiled Strata, an AI dev tool that turns Unreal Engine projects into playable browser links using WebAssembly and WebGPU, started from a single natural-language prompt. The pitch is playable demos and external playtests through a web link, without players installing a game client. It's live with sign-ups; we haven't tried it. → GamesBeat

Roblox Build: new creators, and the "AI slop" worry (Oct 26) Roblox's Vlad Loktev on its prompt-to-game Build tab: "Over 70 percent of people who are building in Build today have never even opened [Roblox] Studio, or created a game on the platform before." DT's Itai Cohen named the risk: "Do people actually invest the time and thought and effort into testing and validation?" Without that, he says, you get "the gaming version of AI slop." Building got cheap. Playtesting didn't. → GamesBeat

📦 Pack #001: Vibe Coder Game Jam Starter ($19)

MegaJam starts Nov 12, and AI is optional. If you do bring agents, bring a plan. Pack #001 has seven agent templates to take a game from idea to playable, fun and shippable: spec, AGENTS.md for game projects, playtest agent, "Is It Fun?" review (loop, hook and replay pull), juice pass, bug triage and ship checklist. Set it up before kickoff, and declare it like any other prep. → Get the pack ($19) · See what's inside

📦 Free: Template #001, The Checkpoint Loop. This week's workflow plugs into it. → Download free

Want to sponsor an issue? Hit reply. Hit reply and tell us how big your AGENTS.md is, and whether your agent ever ate it. We read everything. Agent Patch Weekly is researched and drafted with AI agents and reviewed before sending. Every link was checked on Oct 28, 2026.

← Previous: Issue #002 · All issues

Get the patch notes free →

Free, weekly, one-click unsubscribe.