cliff enable is the entire integrationuv tool install cliffcompaction
pip install cliffcompaction
cliff enable # sets environment variables; Claude Code now routes through the proxy automatically
cliff status # verify the proxy is running
cliff watch # live view of compaction operations as your session runs
For Codex, add a provider block to ~/.codex/config.toml pointing to the proxy port (see README).
--system-prompt flag is being deprecated; wrappers and CI pipelines must switch to --system-prompt-file pointing to a file with the system prompt content.claude --system-prompt "..." on a self-hosted runner will need to be updated to use --system-prompt-file /path/to/system-prompt.txt.--system-prompt as a direct CLI argument to Claude Code. Create a file with the same system prompt text and update the flag to --system-prompt-file <path>. The motivation is privacy: inline flags can appear in process listings and shell history, while a file can be mode-restricted. No changes needed for interactive or SDK-invoked sessions — this only affects self-hosted runner invocations that explicitly pass a system prompt via CLI flag.--system-prompt in your invocation scripts.EffortQuick — create a file with your system prompt content; replace the flag name; update any scripts that construct the CLI invocation dynamically.name, description, and parameters fields — no code injection required. If your agent discovers tools from third-party MCP servers (via a registry, marketplace, or user-supplied URL), the metadata in those tool definitions can be crafted to influence which tool your agent calls. (2) The four attack scenarios are concrete. Information Exfiltration (tool returns steer agent to leak context), Environment Integrity Compromise (tool returns trigger unintended writes/commands), Reasoning Derailment (tool returns corrupt agent chain-of-thought), and Cognitive Denial of Service (optimized metadata causes 32.4× token overconsumption). All four succeed at rates above 60% in the paper's evaluation. (3) What to do now: Treat third-party MCP server metadata with the same distrust you apply to user inputs. For production agents: pin your MCP tool catalog (the mcp-client-2026-09-15 beta from yesterday's SDK release does exactly this); vet metadata from external MCP servers before allowing agent access; use claude plugin validate (added in v2.1.281 today) which now checks for silently dropped entries and insecure URLs; consider filtering tools to an explicit allowlist rather than passing the full server catalog. (4) Perplexity's Bumblebee (featured Sep 14) already scans your local MCP servers for known-compromised packages — run it before expanding your MCP server catalog.cliff enable sets environment variables so Claude Code routes through it automatically. No code changes to your agent or hooks needed. (2) The key technique is the "compaction cliff" insight: standard compaction rewrites history into a summary, which creates a cache miss on every compacted session and degrades performance as the summary diverges from the original reasoning. CliffCompaction avoids rewrites by only keeping verbatim fragments (ordered by a learned importance scorer) — the remaining kept context is unchanged, so prompt caching still hits on it. (3) Both Anthropic Messages API and OpenAI Chat Completions API are supported — configure once, works with Claude Code (automatic via cliff enable) and Codex CLI (manual config block in ~/.codex/config.toml). (4) The paper also benchmarks on KernelBench where the compaction savings make additional test-time scaling compute-efficient. If you run multi-attempt or best-of-N agent loops, CliffCompaction multiplies the effective budget. (5) Caveat: 19 stars, day-one software — treat as experimental for production pipelines. The MIT license and transparent proxy design make it low-risk to try in dev/staging before committing.cliff enable); Moderate for Codex (manual config); evaluate on your own workloads before trusting the paper's numbers as universal.claude plugin validate, MCP URL-mode elicitation for browser-based flows, and a new "attribution": false setting to suppress commit/PR attribution. Dozens of crash and session-stability fixes.--max-turns being ignored on retry loops, and fixes five classes of stream-corruption issues that caused truncated or silently-lost responses behind proxies.guardrail: {id, version} to your Bedrock upstream in gateway.json — every request routes through the guardrail automatically, without app-level changes. (2) Cross-account IAM is now supported. Add assume_role to your Bedrock upstream to use an IAM role from a different AWS account — useful for multi-account enterprise setups where the Claude Code operator account differs from the Bedrock account. (3) claude plugin validate now checks MCP servers. It reports silently dropped entries (entries your MCP server is advertising but Claude Code is ignoring) and insecure URLs (plain HTTP MCP endpoints) — run this after adding any new MCP server. (4) Proxy/gateway stream corruption fixed. If your Claude Code deployment sits behind an enterprise proxy or gateway, four classes of stream-truncation bugs were fixed: responses cut short showing as complete, "Content block not found" errors, lost stop reasons, and fast mode 0-second retry loops. Upgrade if you've seen inexplicable silent truncations. (5) Sessions with pending permission prompts no longer corrupt on restart. Previously, resuming a session after a restart that occurred during a permission prompt broke prompt cache. Fixed. (6) Dangerous rm commands now time out to deny after 2 minutes instead of waiting indefinitely — reduces blast radius of unreviewed recursive deletes in auto mode.UserPromptSubmit and TaskStart hooks can now inject context again (this was broken in v4.1.15–v4.1.19) — if your Cline workflow relies on hooks to prepend context to tasks, upgrade immediately.UserPromptSubmit or TaskStart hooks; you run Cline with Claude Opus 5.5 and complex reasoning tasks.EffortQuick — update Cline extension to v4.1.20.mcp-client-2026-09-15, added in yesterday's SDK release) becomes more important as the catalog grows — pin your tool list for long-running agents to prevent unexpected catalog changes mid-session.BenchAlign coding (September 2026): Claude Opus 5.5 leads with 87.1 (Fable 5.1: 81.4, Fable 5: 74.6). BenchAlign evaluates coding competency across professional alignment scenarios. Source: benchlm.ai/coding, September 2026 update
SWE-bench Pro (September 2026): Muse Spark 1.1 moves to the top of the Pro leaderboard, displacing GPT-5.4 (previously public-set leader through September, now 4th on commercial code at 43.4%). Claude Opus 4.5 shows a notable 22.5-point gap between public (45.9%) and commercial code (23.4%) benchmarks — the largest such split in the current leaderboard, suggesting public benchmark saturation. Source: morphllm.com/swe-bench-pro, September 2026
Nothing cleared the quality bar this period. No HN threads with score >200 and concrete technical depth on new Sep 23 items were confirmed. Nathan Lambert, Eugene Yan, and Sebastian Raschka had no Sep 23 posts confirmed; Simon Willison (egress-blocked) likely covered A2M and CliffCompaction but primary source articles were used directly.
claude-opus-5-5, gpt-6-sol, and gpt-6-luna model support; genai-prices bumped to 0.1.8; fixes UIMessage.id and AG-UI Message.id preservation across load/dump. [github.com/pydantic/pydantic-ai/releases/tag/v2.48.0]deploy command accepts agent_id and environment args; new --image-uri flag for self-hosted deploys; AnyIO, httpx2, cryptography dependency bumps. [github.com/langchain-ai/langgraph/releases]Anthropic confirmed both follow Opus 5.5 within weeks of the September 22 release. No pricing or specific dates announced. [anthropic.com/claude-opus-5-5]
Migrate to antigravity-preview-09-2026, rename snake_case built-in tool parameters to PascalCase, and update file-edit handling. See September 18 digest Breaking Changes for the full migration guide. [ai.google.dev/gemini-api/docs/changelog]
Standard pricing: $10/$1 (cached)/$50 per MTok. Long context (>272K tokens): $20/$75 per request. Free billing period ends October 4. [platform.openai.com/docs/changelog]
Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.