DetectLoops to break runaway agent cycles — a change that treats "agents that don't drift" as infrastructure, not application-layer responsibility. Meanwhile, Gemini CLI v0.66.0 ships with a banner explicitly warning developers it is "not yet optimized for Gemini 4," which — with Google having publicly targeted October 2026 for a Gemini 4 launch — is the clearest near-term model-release signal this week. Claude Code v2.1.312 continues the steady cadence of agentic infrastructure correctness work.dynamic_catalog=True is now the default — tools are appended rather than rewriting the cached prompt, which changes token usage patterns for any agent using dynamic tool sets. LLMReminder fires every 10 requests, not every request (down 10×). DetectLoops is new. Review before upgrading production agents.A quiet Tuesday on GitHub Trending — all repos with real weekly momentum are Claude/AI skills (LLM wrappers that don't qualify) or already-famous tools. Reaching into the watchlist for a pick that earns its slot today: with three separate coding-agent CLIs shipping bug fixes and behavioral changes this week (Claude Code, Gemini CLI, Pydantic AI), developers are staring at a lot of diffs.
brew install git-delta
cargo install git-delta
git config --global core.pager 'delta'
git config --global interactive.diffFilter 'delta --color-only'
git diff HEAD~1
git log -p --follow -- path/to/file.py
-s, --select interactive mode is removed from e2b template delete, e2b template publish, and e2b template unpublish. In the prior version these commands prompted you to pick a template interactively when none was specified; in v2.23.0 the template ID is a required argument. The unused -t, --project and --team options are also removed from those commands.e2b template delete, e2b template publish, or e2b template unpublish without an explicit template ID argument will now error rather than prompt.e2b template delete, e2b template publish, or e2b template unpublish calls that relied on interactive selection (e.g., calls with no template ID or that passed -s). Replace them with explicit template IDs. If you were using -t/--project or --team flags, those are gone — remove them. The positive change in this release is that e2b template list now pages correctly for large template sets (up to 1,000 in pretty output, unlimited in JSON via --json, configurable with -l 0 for no limit).e2b template delete, publish, or unpublish and relied on the interactive selector or the -t/--project flag.EffortModerate (audit and update CLI call sites; straightforward once you know which template IDs to pass).CodeMode now defaults to dynamic_catalog=True — revealed tools are appended rather than rewriting the cached prompt, changing token usage patterns for agents with dynamic tool sets; (2) LLMReminder fires every 10 requests by default, down from every request — a 10× reduction in prompt overhead for long-running agents; (3) Gemini Live reconnect replay is now passed to dial as a keyword-only replay argument; (4) Signed/encrypted history is replayed to providers with a matching canonical_system. New capabilities: DetectLoops capability interrupts agents caught in repeated-action cycles before they run up cost; IPythonCodeMode (alongside the Monty-based CodeMode) enables Jupyter/IPython kernel execution; PostgresSpendStore lets SpendLimits budgets be shared across processes or services through a Postgres database; ag-ui-protocol 1.0 support; OpenAI gpt-6.1-terra model support; Dataset.evaluate gains a task_timeout parameter.DetectLoops to automatically interrupt agents stuck in action loops.CodeMode, token usage will change after upgrading — appending revealed tools rather than rewriting the cached prompt is cheaper for most use cases but can alter caching behavior if your tool catalog is large. Test token counts on your representative workloads before upgrading production agents. LLMReminder change: If you had agents that depended on per-request reminders for correctness (e.g., strict format enforcement on every turn), you'll need to set LLMReminder(frequency=1) explicitly to restore the old behavior. DetectLoops: Add DetectLoops() to any long-running agent that could get stuck — it's a drop-in capability with no configuration required for default loop detection. PostgresSpendStore: If you run multiple agent processes that should share a cost budget (e.g., per-user or per-tenant spend limits), PostgresSpendStore lets you move that budget out of in-process memory and into a Postgres table — enabling true cross-process spend enforcement.CodeMode; you run long-running agents with LLMReminder; you build multi-process agent deployments with shared spend limits; you want automatic loop detection on autonomous agents.EffortModerate (behavioral defaults change — test token usage and LLMReminder behavior before upgrading production agents; new capabilities are opt-in).oidc.user_id_claim to the Claude apps gateway so organizations using Microsoft Entra ID (or any IdP that uses a claim other than sub) can configure their identifier field. Large file fix: Claude no longer overwrites a very large CLAUDE.md, rules file, or @-included file after reading only its start — write now prompts Claude to read the full file or edit in place. Skills routing fix: Per-directory skills and paths:-scoped skills are no longer silently skipped when the working directory changes via cd deeper inside a git repo. Sessions: Fixed a rare bug where the first reply after resuming a terminal session could be missing on the next resume; fixed sessions incorrectly showing as waiting for user input. Commands: Fixed an invalid command name in one plugin/command stopping ALL commands from loading (could leave claude -p with no output). Effort display: The spinner, OpenTelemetry effort values, and transcripts now show the actual effort level in force, not a stale value from earlier in the task. Telemetry: claude_code.safety_event is no longer sent as an OpenTelemetry event. Hooks: Prompt and agent hooks now run at most 24 at a time per session, with at most 12 additional hooks that have exceeded their timeout; a timeout under 1 second is treated as 1 second. Behavior: Bedrock, Vertex, Foundry, and Mantle use the shorter system prompt for unrecognized model names (set CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 to keep the full prompt). Usage limit messages now say "5-hour limit" instead of "session limit." Cloud: Added an effort picker to cloud routines, and a Bash editor with line numbers to cloud environment setup scripts.cd, the large-file overwrite bug, and a command-name crash that could leave claude -p silent — plus adds OIDC claim configuration for enterprise gateway.paths:-scoped skills and noticed they stopped applying after you cded deeper into a git repo, this is the fix — upgrade before your next session. Large-file write fix: The bug where Claude would overwrite a file after only reading the first chunk was silent and data-destroying — if you have a large CLAUDE.md or @-included doc, run git diff after this upgrade on any recent changes to verify nothing was corrupted. OIDC claim configuration: If you deploy the Claude apps gateway with Microsoft Entra ID or another IdP that uses oid or a custom claim as user identifier, set oidc.user_id_claim in your gateway config to get proper user attribution rather than falling back to sub. Hook concurrency: The new 24/12 hook concurrency limits mean hook-heavy setups won't spawn unbounded processes per session — if you have many concurrent hooks and were relying on unlimited parallelism, review the limit.paths:-scoped skills; you have large CLAUDE.md or @-included files; you deploy the Claude apps gateway with Entra ID; you have many concurrent hooks; you run claude -p programmatically.EffortQuick (drop-in update; check per-directory skills work on next session, verify large files).privateNetworkEndpoints opt-in for MCP OAuth flows, so OAuth redirects can target internal/private network URLs (useful for enterprise or locally-hosted MCP servers); a keys array in keybindings.json to support multiple key bindings for the same action; and a new banner warning that "Gemini CLI is not yet optimized for Gemini 4." Bug fixes include preserved terminal scrollback during viewport refreshes, disabled interactive Git credential prompts in safe-git-env, preserved ~/.gitconfig in shell environments, and a fix that validates the active binary after auto-update (preventing a false-success from PATH shadowing).privateNetworkEndpoints: true in your MCP config — this allows OAuth flows to complete using internal URLs that were previously unreachable. keybindings update: If you have a keybindings.json, the new keys: [...] array format replaces single-key binding and lets you assign multiple keys to one action.Nothing cleared the quality bar this period. arXiv (export.arxiv.org) DNS-blocked; huggingface.co DNS-blocked; simonwillison.net DNS-blocked; hnrss.org not checked.
ClaudeAgent and CodexAgent can now be team members and workflow steps, including inside Parallel, Loop, Condition, and Router. New session search endpoints (GET /sessions/filter-schema, POST /sessions/search) with nested AND/OR/NOT filters. github.com/agno-agi/agnoe2b template list now pages through large template sets: pretty output shows up to 1,000 templates, JSON output is unlimited, and -l 0 removes the cap. See Breaking Changes for the removed --select mode. github.com/e2b-dev/E2BCarried from prior digests. v1.0.0-beta.21 (Oct 23) was a documentation cleanup release; beta.20 (Oct 22) switched API clients to the OpenAI Responses endpoint. No GA date announced. Watch github.com/dagger/dagger/releases.
Google has publicly targeted late-October 2026 for Gemini 4 Pro. Gemini CLI v0.66.0 ships a developer notice that the CLI is not yet ready for it. No official announcement found as of this digest. Watch ai.google.dev and the Gemini CLI releases page.
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.