Claude Daily — Signal Over Noise

🥇 Introducing Claude Opus 4.7

Fri, 17 Apr 2026 08:00:00 GMT

Anthropic released Claude Opus 4.7 — a direct upgrade to Opus 4.6 at the same price, with stronger coding, better multi-step agentic workflows, and significantly improved vision
SWE-bench Verified jumps from 80.8% → 87.6%; Cursor's internal coding benchmark goes from 58% → 70%; tool error rate in complex workflows drops by two-thirds
Vision gets a 3x resolution upgrade — images up to 3.75MP (up from 1.15MP) — so screenshots, diagrams, and dense documents are processed at much higher fidelity automatically, no API changes needed
Opus 4.7 is the most literally instruction-following Claude model yet — worth auditing prompts written for earlier models before migrating
Available today via Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry

🥈 v2.1.111

Fri, 17 Apr 2026 08:00:00 GMT

/ultrareview is new: comprehensive cloud-based code review using parallel multi-agent analysis — run with no args for your current branch, or /ultrareview <PR#> to review a specific GitHub PR
xhigh effort level is live for Opus 4.7 — sits between high and max, now tunable via an interactive /effort slider with arrow-key navigation
Auto mode is on by default for Max subscribers on Opus 4.7 — the --enable-auto-mode flag is no longer needed

🥉 [Release] anthropics/claude-agent-sdk-python: v0.1.60

Fri, 17 Apr 2026 08:00:00 GMT

list_subagents() and get_subagent_messages() are new session helpers for inspecting subagent message chains spawned during a session — no more manual transcript parsing to debug multi-agent workflows
OpenTelemetry distributed tracing added: install with pip install claude-agent-sdk[otel] to propagate W3C trace context from the SDK to the CLI subprocess, connecting spans end-to-end

🥇 v2.1.110

Thu, 16 Apr 2026 08:00:00 GMT

Run `/tui fullscreen` to switch to flicker-free rendering in the same conversation — finally fixes the flickering that made fullscreen hard to use in long sessions
Claude can now send mobile push notifications when Remote Control is enabled — your phone pings you when a long-running task finishes
The Write tool now tells Claude when you edit its proposed content in the IDE diff before accepting — Claude sees your manual edits and can reason about why you changed them
`/autocompact`, `/context`, `/exit`, and `/reload-plugins` now work from Remote Control (mobile/web) — full session control from your phone
`--resume`/`--continue` now resurrects unexpired scheduled tasks that were interrupted
Session recap is now enabled for Bedrock, Vertex, and Foundry users (previously only opt-in); opt out via `/config` or `CLAUDE_CODE_ENABLE_AWAY_SUMMARY=0`
Bash tool now enforces its documented maximum timeout instead of accepting arbitrarily large values

🥈 What's eating your Claude Code context window? I wrote a 500-line Python script to find out

Thu, 16 Apr 2026 08:00:00 GMT

`cc-healthcheck` is a single Python file, zero dependencies, that reads `~/.claude/` locally and tells you three things: what auto-loads into every session (CLAUDE.md chain, rules, skill frontmatter), whether your hooks are broken, and where last session's tokens actually went
Shows per-model token totals, cache hit ratio, and system-reminder injection counts — so you can see if your prompt cache is actually working
Catches common hook bugs: unquoted `|` operators, missing timeouts, and case-sensitivity traps that silently fail

🥉 Architectural drift with Agentic coding. Here's what I built to fix it.

Thu, 16 Apr 2026 08:00:00 GMT

`unkode` is a Claude Code skill that generates a `unkode.yaml` architecture map from your codebase — commit it to main as a versioned baseline
On each PR, it diffs the current module graph against that baseline and flags new cross-layer dependencies ("Auth now imports from Billing — is that intended?")
Solves a real agentic coding problem: Claude Code can add 14 imports across 8 files in one PR, and the diff is too big to catch the architectural implication by eyeballing it

🥇 v2.1.108

Wed, 15 Apr 2026 08:00:00 GMT

1-hour prompt cache TTL is now opt-in with `ENABLE_PROMPT_CACHING_1H` — previously API key users (and anyone who set `DISABLE_TELEMETRY`) were silently capped at the 5-minute TTL, burning extra cost on long sessions
New `/recap` command summarizes what happened since you stepped away so you can resume context instantly without re-reading the full transcript
Claude can now discover and self-invoke built-in slash commands like `/init`, `/review`, and `/security-review` via the Skill tool — opens the door to agent workflows that trigger standard operations without manual prompting
Error messages are now meaningfully differentiated: rate limits vs plan limits are separate, 5xx/529 errors link directly to status.claude.com, and unknown slash commands suggest the nearest valid match
`/resume` defaults to sessions in your current directory; Ctrl+A expands to all projects

🥈 YAML vs Markdown vs JSON vs TOON: Which Format Is Most Efficient for the Claude API

Wed, 15 Apr 2026 08:00:00 GMT

The same 200-product catalog costs 15,879 tokens in JSON vs 6,088 in TOON — a 62% difference that hits your API bill every call
This is the first public benchmark run on Claude specifically (Haiku 4.5, Sonnet 4.6, Opus 4.6): 450 API calls, 120 data files, 8 real-world scenarios, Anthropic's production tokenizer
Practical rule: prefer Markdown or TOON over JSON whenever you're passing structured data as context — not just in prompts, but in tool responses and retrieved documents

🥉 Context budget optimization: how to design MCP tools that don't waste tokens

Wed, 15 Apr 2026 08:00:00 GMT

Most MCP tools return raw API JSON (4,000+ tokens) when Claude only needs a pre-digested summary — a 10-100x token waste per call
Design MCP responses to include what changed, what it means, and what to do next — not raw data dumps

🥇 v2.1.105

Tue, 14 Apr 2026 08:00:00 GMT

New PreCompact hook: a hook can now block compaction entirely by exiting with code 2 or returning `{"decision":"block"}` — useful if you want to prevent Claude from summarizing context mid-session
Background plugin monitors now auto-arm at session start or on skill invoke via a top-level `monitors` manifest key — no more manual setup
`/proactive` is now an alias for `/loop`, making the recurring-task mode easier to remember
`WebFetch` now strips `<style>` and `<script>` contents from fetched pages so CSS-heavy sites stop exhausting the content budget before reaching actual text
`/doctor` gets a facelift with status icons and an `f` key shortcut that has Claude fix reported issues automatically
15+ bug fixes: images on queued messages no longer dropped, blank screen on wrapped prompt input fixed, ASCII art no longer has leading whitespace stripped, and one-shot scheduled tasks no longer re-fire repeatedly

🥈 Exploring the new `servo` crate

Tue, 14 Apr 2026 08:00:00 GMT

Simon Willison tasked Claude Code with exploring the brand-new `servo` 0.1.0 crate (the Servo browser engine, now embeddable as a Rust library — released yesterday)
Claude Code built `servo-shot`, a working CLI tool that takes real screenshots of web pages using the Servo engine; the result accurately renders Hacker News
It also determined that compiling Servo to WebAssembly isn't feasible (threads + SpiderMonkey) and built an alternative playground using `html5ever` compiled to Wasm instead

🥉 I Built a Free MCP Server for European Financial Data

Tue, 14 Apr 2026 08:00:00 GMT

Free, open MCP server with 6 tools that pull ECB and Eurostat data into Claude as clean flat JSON — no more hand-parsing SDMX-XML
Covers: ECB deposit/refi/marginal lending rates, EUR exchange rates (30+ currencies), EU inflation by country (HICP), GDP by country (quarterly), and more
Install in 30 seconds; all sources are official and free (ECB Statistical Data Warehouse, Eurostat)

🥇 Claude Code v2.1.101 — /team-onboarding, OS CA cert trust, command injection fix, memory leak, 20+ fixes

Sat, 11 Apr 2026 08:00:00 GMT

New `/team-onboarding` command: generates a teammate ramp-up guide from your own local Claude Code usage history — useful for async onboarding without writing docs manually
OS CA certificate store is now trusted by default, so enterprise TLS proxies work without extra setup; opt out with `CLAUDE_CODE_CERT_STORE=bundled` if you need the old behavior
Security fix: command injection vulnerability in the POSIX `which` fallback used by LSP binary detection — patched
Memory leak fix: long sessions were retaining dozens of historical copies of the full message list in the virtual scroller — now cleaned up
`/ultraplan` and remote-session features no longer require web setup first — they auto-create a default cloud environment
Multiple `--resume` fixes: context loss on large sessions, crash when a persisted Edit/Write result was missing `file_path`, and cross-subagent chain bridging into the wrong conversation

🥈 Stop Putting Best Practices in Skills

Sat, 11 Apr 2026 08:00:00 GMT

Skills in Claude Code only get invoked 6–66% of the time; CLAUDE.md is always in context — so anything you want Claude to *always* do should live in CLAUDE.md, not a skill
The author ran 51 multi-turn evals across 4 configurations, replicated Vercel's single-shot experiment in realistic sessions, and read Claude Code's source to confirm: skills and CLAUDE.md are both just prompts, but skills depend on a chain of decisions that frequently fails
Rule of thumb: CLAUDE.md for guidelines, coding standards, and non-negotiable behavior; skills for on-demand recipes ('run this audit', 'generate this template', 'open a PR')

🥉 Most of your Claude Code agents don't need Sonnet

Sat, 11 Apr 2026 08:00:00 GMT

Haiku costs $0.25/1M input tokens vs Sonnet's $3/1M — a 12x difference — and for mechanical tasks (commit messages, code review, docs, CI config, test running) Haiku is plenty
The author runs 50 agent calls/day; only 8 hit Sonnet — the remaining 68% run Haiku, with 2 local Ollama models at zero API cost
Sonnet stays on tasks where wrong answers are expensive: planning, multi-file debugging, security review, complex implementation

🥇 The Advisor Strategy: pair Opus as a mid-task advisor with Sonnet or Haiku as executor — now in beta

Fri, 10 Apr 2026 08:00:00 GMT

New Claude Platform API feature in beta: designate Opus as an 'advisor' that your Sonnet or Haiku agent consults when it hits a hard mid-task decision — Opus returns a plan, the executor keeps running, all inside one API request
Anthropic's evals: Sonnet + Opus advisor beats plain Sonnet by 2.7 percentage points on SWE-bench Multilingual while costing 11.9% less per task — near-Opus quality at near-Sonnet prices
SDK support shipped simultaneously across all five languages: Python v0.93.0, TypeScript v0.87.0, Go v1.34.0, Java v2.24.0, C# v12.14.0 — the 'beta advisor tool' feature is in each one
Full write-up at claude.com/blog/the-advisor-strategy

🥈 Claude Code v2.1.98 — three Bash permission security fixes + Vertex AI setup wizard, Monitor tool, subprocess sandboxing

Fri, 10 Apr 2026 08:00:00 GMT

Security fix: a backslash-escaped flag in a Bash command could be auto-allowed as read-only and escalate to arbitrary code execution — patched
Security fix: compound Bash commands (e.g. `cmd1 && cmd2`) were bypassing forced permission prompts in auto and bypass-permissions modes — patched
Security fix: redirects to `/dev/tcp/...` or `/dev/udp/...` were silently auto-allowed instead of prompting — patched
New: interactive Google Vertex AI setup wizard on the login screen — guides you through GCP auth, project/region config, credential check, and model pinning with no manual config files
New: `CLAUDE_CODE_PERFORCE_MODE` — Edit/Write/Notebook ops fail fast on read-only Perforce files with a `p4 edit` hint instead of silently overwriting
New: Monitor tool for streaming events from long-running background scripts

🥇 Introducing Claude Managed Agents, now in public beta

Thu, 09 Apr 2026 08:00:00 GMT

Anthropic just launched Managed Agents: you define what your agent does, and Anthropic runs it on their infrastructure — no devops, no agent loop babysitting
Before this, shipping a production agent meant months of work on infrastructure, state management, and permissioning — now it's supposed to be days
All five major Anthropic SDKs (Python, TypeScript, Java, Go, C#) shipped same-day support for the Managed Agents API, so you can start coding against it right now
Notion, Sentry, Rakuten, Asana, and Vibecode are early customers — Notion is running dozens of tasks in parallel directly inside its workspace
Multi-agent coordination is in research preview (request access); full blog at claude.com/blog/claude-managed-agents

🥈 Claude Code v2.1.97 — focus view, permission hardening, MCP memory leak fix, 20+ fixes

Thu, 09 Apr 2026 08:00:00 GMT

New `Ctrl+O` focus view in NO_FLICKER mode: shows just your prompt, a one-line tool summary with diff stats, and the final response — great for staying focused
MCP HTTP/SSE connections were leaking ~50 MB/hour of memory when servers reconnect — now fixed
`--dangerously-skip-permissions` was silently being downgraded when you approved a write to a protected path — fixed
`settings.json` was being silently ignored if you had a permission rule named `toString` or another JS prototype property — fixed
`/resume` gets a pile of fixes: `--resume <name>` is now editable, diffs survive resume on large files, messages typed while Claude works are now persisted

🥉 Claude Code GitHub Action v1.0 — Generally Available, simplified API, breaking changes

Thu, 09 Apr 2026 08:00:00 GMT

The Claude Code GitHub Action is now v1.0 GA — if you're still on v0.x, you need to migrate
Old: `mode`, `direct_prompt`, `override_prompt`, `custom_instructions`, `model`, `allowed_tools` inputs. New: one `prompt` input and `claude_args` for everything else
Auto mode detection: no more manually setting `mode` — the action figures out interactive vs automation on its own
New example workflows published: automated PR reviews, CI failure fixes, issue triage, doc generation, security scanning

🥇 claude-mythos-preview appears simultaneously in all Anthropic SDKs

Wed, 08 Apr 2026 08:00:00 GMT

All five major Anthropic SDKs (Python, TypeScript, Go, Java, C#) shipped a new release on the same day, and they all have exactly one thing in common: `claude-mythos-preview` is now a supported model ID
You can start passing `model="claude-mythos-preview"` in your API calls right now — the SDK handles it, though actual access depends on Anthropic granting your account
This is the standard pattern Anthropic uses just before a model becomes broadly available: SDK support lands first, public announcement follows
The synchronized multi-language rollout suggests this isn't a leak — it's a coordinated release prep
Related context: a Thomas Friedman NYT piece and Anthropic's "Glasswing" page are already circulating with details about Mythos' capabilities, including finding vulnerabilities in major software systems

🥈 Claude Code v2.1.94 — effort defaults back to high for API users, 20+ fixes

Wed, 08 Apr 2026 08:00:00 GMT

Default effort level is now `high` again for API-key, Bedrock/Vertex/Foundry, Team, and Enterprise users — it was silently changed to `medium` a while back, this reverses that; control it with `/effort`
Bedrock Mantle is now supported via `CLAUDE_CODE_USE_MANTLE=1`
Fixed: agents stuck silently after 429 rate-limit (error now surfaces immediately)
Fixed: Console login on macOS silently failing with "Not logged in" when keychain is locked — `claude doctor` now diagnoses it
Fixed: CJK and multibyte text corruption when chunk boundaries split a UTF-8 sequence in stream-json mode
`--resume` now works across worktrees of the same repo directly, no more `cd` command required

🥉 Amazon Bedrock Mantle client lands across all Anthropic SDKs

Wed, 08 Apr 2026 08:00:00 GMT

A new `AnthropicBedrockMantle` client is now available in Python (v0.91.0), TypeScript (v0.85.0), Go (v1.32.0), and C# Bedrock (v0.2.0) — also in Claude Code itself via `CLAUDE_CODE_USE_MANTLE=1`
Mantle is Amazon's unified inference layer for Bedrock — this client lets you talk to Claude through Mantle's routing without managing region-specific endpoint strings manually

🥇 Boris Cherny explains some recent changes to Claude Code

Tue, 07 Apr 2026 08:00:00 GMT

Boris Cherny (Claude Code team lead) clarified why Claude Code "thinking" seems to have dropped: the `redact-thinking-2026-02-12` header hides thinking from the UI but doesn't reduce it — Claude still thinks just as hard, you just don't see the transcript
The actual thinking depth change happened Feb 9 when Opus 4.6 launched: it switched from a fixed thinking budget to *adaptive* thinking, where Claude decides how long to think. Turn this off with `CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING` env var if you want the old behavior
On March 3 they also changed the default effort from 100 to 85 (the sweet spot they found for speed vs. intelligence). To get full effort back: type `/effort` in Claude Code or add `"effort": "high"` to settings.json — it sticks across sessions
If you want thinking shown in the UI again: add `"showThinkingSummaries": true` to settings.json
Key takeaway: post GitHub issues, not Reddit, if you want Anthropic to see bug reports

🥈 Claude Code v2.1.92 introduces Ultraplan — draft plans in the cloud, review in your browser, execute anywhere

Tue, 07 Apr 2026 08:00:00 GMT

Claude Code v2.1.92 ships `/ultraplan` (beta): run it in your terminal, get a structured plan back as a shareable web page at claude.ai/code with inline commenting
You can review and annotate the plan in the browser, then either execute it remotely in the cloud or send it back to your local CLI to run
This is the first step toward cloud-first Claude Code workflows — the terminal stays the power-user entry point but planning and review can now happen anywhere

🥉 Claude Code can now submit your app to App Store Connect and help you pass review

Tue, 07 Apr 2026 08:00:00 GMT

Blitz is an open-source macOS app (Apache 2.0) that adds MCP servers so Claude Code can control App Store Connect: metadata, screenshots, localization, build submission, TestFlight groups — all from the terminal
Includes a built-in terminal with Claude Code so agents can build, test, and submit without leaving one window; runs fully locally, MCP server is localhost-only
Repo: github.com/blitzdotdev/blitz-mac — there's a demo of an agent submitting an end-to-end App Store review

🥇 Claude Code GitHub Action v1.0 — Now Generally Available

Mon, 06 Apr 2026 08:00:00 GMT

The Claude Code GitHub Action hit v1.0 GA — you can now wire Claude directly into any GitHub workflow to do automated PR reviews, fix failing CI, triage issues, generate docs, and run security scans
Configuration got a major cleanup: one unified `prompt` input replaces the old `mode`/`direct_prompt`/`override_prompt` trio, and all CLI options go through `claude_args` instead of scattered fields
The action now auto-detects whether to run in interactive mode (responding to @claude mentions) or automation mode (running on triggers like PR opened), so no more manual `mode:` configuration
AWS Bedrock and Google Vertex AI are now fully supported alongside the default Anthropic API
Breaking changes from v0.x: if you already use the action, read the migration guide at github.com/anthropics/claude-code-action/blob/main/docs/migration-guide.md before upgrading

🥈 71.5x token reduction by compiling your raw folder into a knowledge graph instead of reading files. Built from Karpathy's workflow

Mon, 06 Apr 2026 08:00:00 GMT

Instead of reloading raw files every Claude Code session, `graphify` compiles your whole folder — code (13 languages via AST), PDFs, images, markdown — into a structured wiki once, then answers questions from the graph
Install with `pip install graphify && graphify install`, then call `/graphify ./raw` inside Claude Code — it works as a native Claude Code skill
Every relationship is tagged EXTRACTED, INFERRED, or AMBIGUOUS, so you know exactly what came from source vs. what the model reasoned
Tested at 71.5x fewer tokens per query vs. reading raw files cold; drop new content in and `--update` merges it into the existing graph

🥉 After months with Claude Code, the biggest time sink isn't bugs — it's silent fake success

Mon, 06 Apr 2026 08:00:00 GMT

Claude Code is optimized to produce "working" output — so when it can't get auth working, it quietly inserts a try/catch returning sample data, and you won't notice until three days later
Add an "Error Handling Philosophy: Fail Loud, Never Fake" section to your CLAUDE.md: prefer visible failures over silent fallbacks, never substitute placeholder data, and always disclose degraded mode
The priority ladder: works correctly → disclosed fallback (with a banner/log) → clear error message → silent degradation (never acceptable)

🥇 I gave Claude Code a 285-line operating manual and 5 lifecycle hooks inside my Obsidian vault. After a month, it basically runs my work documentation for me.

Sun, 05 Apr 2026 08:00:00 GMT

You can wire up Claude Code's session hooks (SessionStart, UserPromptSubmit, PostToolUse, PreCompact, Stop) in `.claude/settings.json` to fully automate work documentation in an Obsidian vault
SessionStart injects your active projects, recent git changes, and open tasks so Claude already knows your context — no more "let me catch you up" at the start of every session
UserPromptSubmit classifies your messages before Claude responds: say "just had a 1:1 with Sarah, she wants error monitoring before release" and it automatically creates the 1:1 note, updates Sarah's person file, logs the decision, and adds the win to your brag doc
PostToolUse validates every .md Claude writes (frontmatter, wikilinks, correct folder) before mistakes compound; Stop auto-archives completed projects and checks for orphan notes
The 285-line CLAUDE.md is the keystone — consistent note schema is what makes synthesis work, not clever prompting

🥈 I got Claude Code to stop burning 40K tokens by just figuring out my codebase

Sun, 05 Apr 2026 08:00:00 GMT

Most Claude Code sessions waste 25K–66K tokens just re-exploring your codebase at the start of every conversation; codesight fixes this in one command
Run `npx codesight scan` and it writes compact context files into `.codesight/` — architecture overview, all API routes, DB schema, UI components, env vars, and an import graph with "blast radius" analysis
Add one line to CLAUDE.md pointing Claude at `.codesight/CODESIGHT.md` and future sessions start from a pre-computed map instead of rediscovering everything (~12x fewer orientation tokens on real production codebases)

🥉 Senior engineer best practice for scaling yourself with Claude Code

Sun, 05 Apr 2026 08:00:00 GMT

Git worktrees let you run 4–8 Claude Code agents in parallel on the same repo, each on its own branch — if Claude Code made you 10x faster, worktrees multiply that again depending on how many agents you can manage at once
Without a structured planning step first (like the compound-engineering-plugin's multi-agent brainstorm → plan → execute → review flow), parallel worktrees create collisions and confusion rather than speed

🥇 v2.1.92

Sat, 04 Apr 2026 08:00:00 GMT

New interactive Bedrock setup wizard walks you through AWS auth, region config, credential verification, and model pinning right from the login screen — no more manual config fumbling
`/cost` now shows a per-model and cache-hit breakdown for subscription users, so you can actually see which models are eating your quota
Pro users get a footer hint when prompt cache expires mid-session, telling you roughly how many tokens the next message will burn uncached
Write tool diff computation is 60% faster on files containing tabs, ampersands, or dollar signs
`/release-notes` is now an interactive version picker instead of a wall of text
13 bug fixes including: subagent spawning failures after tmux window kills, Stop hook failures with fast models, streaming JSON encoding bugs, phantom feedback submissions from autopilot keypresses, and a duplicate-message scroll glitch in iTerm2/Ghostty

🥈 Using third-party harnesses with your Claude subscriptions

Sat, 04 Apr 2026 08:00:00 GMT

Starting April 4 at 12pm PT, Claude subscriptions no longer cover usage on third-party tools like OpenClaw that authenticate via Claude login
Affected users get a one-time credit equal to their monthly plan cost, plus access to discounted usage bundles
To keep using these tools: switch to a Claude API key, or buy the new extra usage bundles
Agent SDK is unaffected for now

🥉 I reverse-engineered why Claude Code burns through your usage so fast. 7 bugs that stack on top of each other — and the worst one activates when Extra Usage kicks in

Sat, 04 Apr 2026 08:00:00 GMT

The native installer binary silently breaks prompt caching on every turn — switching to the npm version (`npm install -g @anthropic-ai/claude-code`) fixes it immediately
When Extra Usage kicks in, the client quietly downgrades your cache TTL from 1-hour to 5-minute, causing cache misses to spike exactly when you can least afford them
Three of the seven bugs (session resume dropping attachments, autocompact infinite loop, and the cache downgrade) were present for weeks and have since been patched in v2.1.89–v2.1.91

🥇 v2.1.91

Fri, 03 Apr 2026 08:00:00 GMT

MCP tool results can now be up to 500K characters (up from ~50K) by setting `_meta["anthropic/maxResultSizeChars"]` on the tool annotation — no more silent truncation of large DB schemas or file dumps
Plugins can now ship executables under `bin/` and invoke them as bare commands from the Bash tool, making plugin-packaged CLIs a first-class citizen
Multi-line prompts in `claude-cli://open?q=` deep links now work — encoded newlines (`%0A`) are no longer rejected
A bug where `--resume` could silently lose conversation history when async transcript writes fail is fixed
Plan mode no longer loses track of the plan file after a container restart in remote sessions, which was causing spurious permission prompts and empty plan-approval modals
The Edit tool now uses shorter `old_string` anchors, reducing output token count on every file edit

🥈 Claude Code GitHub Action v1.0 — Now Generally Available

Fri, 03 Apr 2026 08:00:00 GMT

The Claude Code GitHub Action hits v1.0 GA — the action that lets you trigger Claude Code on PRs, issues, and CI events with `@claude` mentions or automation rules
Configuration is drastically simplified: `mode`, `direct_prompt`, `override_prompt`, and `custom_instructions` are all replaced by a single `prompt` input and a `claude_args` passthrough for CLI flags
Now supports AWS Bedrock and Google Vertex AI natively, so teams with cloud-provider constraints can use it without routing through the Anthropic API
Breaking changes from v0.x — check the migration guide before upgrading

🥉 Follow-up on usage limits

Fri, 03 Apr 2026 08:00:00 GMT

Anthropic's official account explains what's been driving faster usage burn: peak-hour limits tightened and 1M-context sessions got larger, not a billing bug
Key tips: use Sonnet 4.6 instead of Opus on Pro (Opus burns ~2× as fast), turn off extended thinking when you don't need it, start fresh sessions rather than resuming idle ones, and set `CLAUDE_CODE_AUTO_COMPACT_WINDOW=200000` to cap context growth

🥇 v2.1.90

Thu, 02 Apr 2026 08:00:00 GMT

Claude Code v2.1.90 ships `/powerup` — an interactive, animated tutorial system that teaches you Claude Code features in-context; great onboarding shortcut if you've been skipping the docs
The long-running `--resume` cache bug is fixed: since v2.1.69, resuming a session with MCP servers, deferred tools, or custom agents triggered a full prompt-cache miss on the first request, quietly burning up to 11.5× more tokens than expected every single time
Auto mode now correctly honors explicit user boundaries ('don't push', 'wait for X before Y') — before this fix those instructions were silently ignored even when the action was otherwise allowed
Three quadratic-time performance regressions fixed in one release: SSE transport large-frame handling, SDK session transcript writes, and per-turn MCP tool schema lookups are all now linear
PowerShell security hardened: four distinct bypass techniques (background job `&`, `-ErrorAction Break` debugger hang, archive-extraction TOCTOU, parse-fail deny-rule degradation) were all patched simultaneously

🥈 anthropic-sdk-python v0.88.0 / anthropic-sdk-typescript v0.82.0

Thu, 02 Apr 2026 08:00:00 GMT

Both the Python and TypeScript SDKs now return `structured stop_details` on message responses — instead of inferring why a generation stopped, you get a typed object telling you exactly whether it was `tool_use`, `end_turn`, `max_tokens`, or another reason
Bedrock now supports API key authentication in both SDKs, removing the AWS credential ceremony for teams that want simpler Bedrock access
A new `aws-sdk` package lands in the TypeScript SDK (v0.2.0) for first-class AWS-native integration

🥉 I investigated Claude Code's --resume cache bug. Here's what was actually happening

Thu, 02 Apr 2026 08:00:00 GMT

A developer traced the v2.1.69–v2.1.90 `--resume` regression to its root cause: session state was silently dropping `deferred_tools_delta` and `mcp_instructions_delta` records on save, so resume reconstructed a different cache key and triggered a full miss every time
Real cost numbers: ~11.5× higher token spend on the first resumed request; one heavy MCP user estimated 12.6M extra tokens burned and ~$43.56 in wasted API value over 27 days — with a broader community midpoint around $285K

🥇 v2.1.89

Wed, 01 Apr 2026 08:00:00 GMT

Claude Code v2.1.89 lands with 20+ changes — the headline is a new `"defer"` value for PreToolUse hooks: headless sessions can now pause mid-task at any tool call and resume later with `-p --resume`, so long automation pipelines can checkpoint instead of running all-at-once or stopping cold
`CLAUDE_CODE_NO_FLICKER=1` is a new env var for flicker-free alt-screen rendering — if you're running Claude Code in tmux or over SSH and seeing visual glitches, set this and they go away
Auto mode now shows denied commands in a `/permissions` → Recent tab where you can retry with `r` — previously blocked actions were invisible; now you can see and re-evaluate them
Windows users get two critical fixes: Edit/Write were doubling CRLF line endings and stripping Markdown hard line breaks (two trailing spaces), silently corrupting files — update now if you're on Windows
`/stats` now correctly counts subagent token usage and preserves data past 30 days — your token history was likely undercounting the whole time

🥈 Claude Code GitHub Action v1.0 — Now Generally Available

Wed, 01 Apr 2026 08:00:00 GMT

The Claude Code GitHub Action hits v1.0 GA — the old `mode`, `direct_prompt`, and `override_prompt` inputs are replaced with a single `prompt` input, and all CLI options now flow through `claude_args`, making setup dramatically simpler
The action now auto-detects interactive vs automation mode (no more manual configuration), and adds native AWS Bedrock and GCP Vertex AI support if you want to route away from the Anthropic API directly
Breaking changes from v0.x — review the migration guide before upgrading production workflows

🥉 claude-agent-sdk-typescript: v0.2.89

Wed, 01 Apr 2026 08:00:00 GMT

New `startup()` function pre-warms the CLI subprocess before your first `query()` call — pay the cold-start cost upfront and get ~20x faster first responses when you control the timing
New `listSubagents()` and `getSubagentMessages()` functions expose what subagents did inside a session — finally lets you inspect multi-agent execution history programmatically

🥇 Computer use is now in Claude Code

Tue, 31 Mar 2026 08:00:00 GMT

Claude Code can now open your actual apps, click through the UI, and test what it built — all directly from the CLI, without you touching the keyboard
Works on anything you can open on a Mac: a compiled SwiftUI app, a local Electron build, or any GUI tool that doesn't have a CLI — so Claude can visually verify its own output
This is a research preview, available now on Pro and Max plans on macOS. Enable it with /mcp in Claude Code
Docs at code.claude.com/docs/en/computer-use — this is the first time Claude can close the loop from 'write code' to 'run it and see if it looks right'

🥈 v2.1.88

Tue, 31 Mar 2026 08:00:00 GMT

Adds a new PermissionDenied hook that fires when auto mode blocks an action — return {retry: true} to let the model try a different approach instead of just stopping
Fixes a StructuredOutput schema cache bug that was causing ~50% failure rates in workflows with multiple schemas — a likely silent source of mysterious failures in agent pipelines
Fixes prompt cache misses in long sessions caused by tool schema bytes changing mid-session (a known cost multiplier)
Fixes the misleading 'Rate limit reached' error that was actually an entitlement error — now shows the real message with actionable hints

🥉 PSA: Claude Code has two cache bugs that can silently 10-20x your API costs — here's the root cause and workarounds

Tue, 31 Mar 2026 08:00:00 GMT

Bug 1: The standalone Claude Code binary (from claude.ai/install.sh or npm install -g) has a native string replacement that can break your prompt cache if your conversation ever discusses Claude Code's own billing internals — workaround: run via npx @anthropic-ai/claude-code instead
Bug 2: Every --resume since v2.1.69 causes a full cache miss on the entire conversation history — only the system prompt stays cached, so a long session resumed once could cost $3-10 extra in wasted input tokens

🥇 Bringing Code Review to Claude Code

Mon, 30 Mar 2026 08:00:00 GMT

Claude Code now has a built-in code review workflow — ask Claude to review your changes the same way you'd request a teammate review
This moves Claude Code from 'write code for me' toward 'be my full development partner', covering the review step that was previously manual
Official Anthropic blog post — code review is one of the most common developer workflows, and having it native to Claude Code removes the context-switch to a separate reviewer
Pairs naturally with the Claude Code GitHub Action v1.0 (released last week) for automated PR review pipelines

🥈 Why the 1M context window burns through limits faster and what to do about it

Mon, 30 Mar 2026 08:00:00 GMT

Every message you send re-sends your *entire* conversation to the API — message 50 includes all 49 prior turns before Claude starts on your new one. Without caching, a 100-turn Opus session would cost $50-100 in input tokens alone
Anthropic caches aggressively (90% off for cache hits). One measured session hit 96.39% cache hit rate: 47M tokens sent, only 1.6M needed real compute
The real cost driver is cache *busts* caused by the 5-minute TTL. A 6-minute coffee break on a 500K-token conversation costs ~$3.13 just in cache-write fees (billed at 125% of normal input rate)
Common accidental cache busters: timestamps or dynamic content in your system prompt, switching models mid-session, adding/removing MCP tools (tool definitions are part of the cached prefix)
Fix: keep system prompts fully static, don't swap models mid-session, batch MCP tool changes to the start, avoid pauses longer than 5 minutes when your context is large

🥉 MEX: structured context scaffold for Claude Code with drift detection

Mon, 30 Mar 2026 08:00:00 GMT

MEX replaces one monolithic context file with a routing table in `.mex/` — Claude loads only the context relevant to the current task (working on auth? loads `context/architecture.md`, writing new code? loads `context/conventions.md`)
A CLI runs 8 zero-token, zero-AI drift checkers: finds referenced file paths that no longer exist, npm scripts your docs mention that were deleted, dependency version conflicts, scaffold files not updated in 50+ commits
When drift is found, `mex sync` builds a targeted repair prompt and fires Claude Code on only the broken files

🥇 [Release] anthropics/claude-code-action: Claude Code GitHub Action v1.0

Sun, 29 Mar 2026 08:00:00 GMT

Claude Code GitHub Action just hit v1.0 GA — it's now production-ready with a cleaner, unified interface
The old `mode`, `direct_prompt`, `override_prompt` inputs are gone — everything goes through a single `prompt` input now
All Claude Code CLI options are now passable via `claude_args`, giving you direct access to custom system prompts, tool configs, and model selection
AWS Bedrock and Google Vertex AI compatibility improved
New example workflows bundled: automated PR review, CI failure fixes, issue triage, doc generation, security scanning
Breaking change from v0.x — if you're on v0, check the migration guide before upgrading

🥈 [Release] anthropics/claude-agent-sdk-python: v0.1.52

Sun, 29 Mar 2026 08:00:00 GMT

`get_context_usage()` lands in the Python SDK — finally see exactly what's eating your context window, by category (parity with TypeScript SDK)
`@tool` decorator now supports `typing.Annotated` for per-parameter descriptions, making auto-generated JSON schemas much more readable
Critical bug fixed: `connect(prompt="...")` was silently dropping the string prompt, causing `receive_messages()` to hang forever — upgrade if you've seen mysterious hangs
In-flight hook callbacks are now properly cancelled when the CLI abandons them

🥉 My 10 Pro Tips for Claude Code users

Sun, 29 Mar 2026 08:00:00 GMT

`/effort high` + `ultrathink` in your prompt forces full extended thinking with a 31,999-token budget — the most tokens you can give it
`/fast` does NOT change the model — it's the same Opus with faster output; pair with `/effort medium` for best speed/quality balance
Exit code 2 in a PostToolUse/PreToolUse hook silently blocks actions and forces a rewind — powerful for building guardrails

🥇 v2.1.86

Sat, 28 Mar 2026 08:00:00 GMT

A new `X-Claude-Code-Session-Id` header is added to every API request so proxies and observability tools can group requests by session without parsing the body
`.jj` and `.sl` directories (Jujutsu and Sapling VCS) are now excluded from Grep and file autocomplete — no more accidentally descending into VCS metadata
`--resume` no longer fails with 'tool_use ids were found without tool_result blocks' on sessions created before v2.1.85 — safe to resume old sessions again
Write/Edit/Read now work on files outside the project root (e.g., `~/.claude/CLAUDE.md`) when conditional skills or rules are configured — this was a quiet but painful regression
Official marketplace plugin scripts no longer fail with 'Permission denied' on macOS/Linux (broken since v2.1.83)
`--bare` mode now correctly passes through MCP tools in interactive sessions and no longer silently drops messages queued mid-turn
Prompt cache hit rate for Bedrock, Vertex, and Foundry users improved by removing dynamic content from tool descriptions
Token overhead for `@`-mentioned files reduced — raw string content is no longer JSON-escaped before being sent
Memory filenames in the 'Saved N memories' notice now open on click
Fixed masked OAuth input leaking token start characters when wrapping on narrow terminals

🥈 [Release] anthropics/claude-agent-sdk-python: v0.1.51

Sat, 28 Mar 2026 08:00:00 GMT

`fork_session()` and `delete_session()` land in the Python SDK — you can now branch a session mid-conversation and clean up sessions programmatically
`task_budget` option added for token budget management — tell the agent upfront how many tokens it has so it paces itself instead of blowing the budget mid-task
`AgentDefinition` now accepts `disallowedTools`, `maxTurns`, and `initialPrompt` — you can bake constraints and a starter prompt directly into the agent definition
`--system-prompt-file` CLI flag now available as `SystemPromptFile` in the SDK
Python 3.10 users: `NotRequired` TypedDict fields now work correctly via `typing_extensions`

🥉 [Release] anthropics/claude-agent-sdk-typescript: v0.2.86

Sat, 28 Mar 2026 08:00:00 GMT

New `getContextUsage()` control method returns a breakdown of context window usage by category — finally a way to see exactly what's eating your context budget
`session_id` is now optional in `SDKUserMessage` — the SDK assigns one automatically if you don't provide it, removing a common footgun for new SDK users

🥇 v2.1.85

Fri, 27 Mar 2026 08:00:00 GMT

Hooks just got a lot smarter: you can now add an `if` field using permission rule syntax (e.g., `Bash(git *)`) so hooks only fire on the tool calls you actually care about — no more spawning a process for every single command
PreToolUse hooks can now answer an `AskUserQuestion` by returning `updatedInput` alongside `permissionDecision: "allow"` — this is the missing piece for headless CI integrations that need to inject answers without a human at the keyboard
The `/compact` command no longer crashes when the conversation is so large that the compact request itself exceeds the context limit — it quietly handles the overflow
Timestamp markers now appear in transcripts whenever `/loop` or `CronCreate` tasks fire — finally you can see exactly when a scheduled task ran in a long session
Deep link queries (`claude-cli://open?q=…`) now support up to 5,000 characters, with a scroll-to-review warning for anything long
Plugins blocked by organization policy are now hidden from the marketplace entirely, not just disabled
Fixed `deniedMcpServers` not actually blocking claude.ai MCP servers — the setting now works as documented
Fixed `switch_display` returning "not available in this session" on multi-monitor setups in computer-use
Fixed a memory leak in remote sessions when a streaming response is interrupted
Fixed ECONNRESET errors during edge connection churn by using a fresh TCP connection on retry

🥈 Exclusive: Anthropic acknowledges testing new AI model representing 'step change' in capabilities, after accidental data leak reveals its existence

Fri, 27 Mar 2026 08:00:00 GMT

Anthropic accidentally leaked the existence of a new model called "Mythos" through a data leak before any official announcement
After the leak, Anthropic confirmed they are actively testing Mythos and described it as a "step change" in capabilities — not an incremental update
No public release date or pricing yet — this is purely a confirmation that it exists and is in testing

🥉 [Release] anthropics/claude-agent-sdk-typescript: v0.2.85

Fri, 27 Mar 2026 08:00:00 GMT

New `reloadPlugins()` method lets SDK-based integrations hot-reload plugins and get back refreshed commands, agents, and MCP server status without restarting
Fixed a bug where PreToolUse hooks returning `permissionDecision: "ask"` were silently ignored in SDK mode — they now correctly prompt as expected

🥇 v2.1.84

Thu, 26 Mar 2026 08:00:00 GMT

Windows users get a PowerShell tool as an opt-in preview — finally a native shell experience instead of relying on bash workarounds
If you've ever come back to a stale Claude Code session hours later and burned tokens re-caching context, a new idle-return prompt now nudges you to /clear after 75+ minutes away
New TaskCreated hook fires when a task spins up via TaskCreate — useful for logging, notifications, or triggering side effects in multi-agent workflows
WorktreeCreate HTTP hooks can now return the created worktree path, making it easier to integrate with external CI systems
MCP tool descriptions are now capped at 2KB, preventing OpenAPI-generated servers from stuffing your context with thousands of tokens of schema noise
Token counts ≥1M now display as '1.5m' instead of '1512.6k' — small detail, big improvement for readability in long agentic sessions

🥈 Claude Code GitHub Action v1.0 — Now Generally Available

Thu, 26 Mar 2026 08:00:00 GMT

The official Claude Code GitHub Action hits v1.0 GA — it's now the stable, supported way to wire Claude into your CI/CD pipelines
Mode detection is now automatic: no more manually setting 'interactive' vs 'automation'. The action figures it out from your setup
All prompts go through a single 'prompt' input now, and CLI options through 'claude_args' — simpler, more consistent with how you'd use the CLI directly
Ready-made examples cover automated PR review, CI failure fixes, issue triage, doc generation, and security scanning
Breaking changes from v0.x: if you're already using this action, check the migration guide before updating

🥉 [Release] anthropics/claude-agent-sdk-typescript: v0.2.84

Thu, 26 Mar 2026 08:00:00 GMT

New 'taskBudget' option lets you tell the model upfront how many tokens it has for a task, so it paces tool use instead of going all-in on exploration and hitting limits mid-task
'enableChannel()' method and 'capabilities' field on McpServerStatus give you SDK-level control over which MCP channels are active

🥇 Auto mode for Claude Code

Wed, 25 Mar 2026 08:00:00 GMT

Claude Code's biggest daily friction — approving every single file write and bash command — now has a middle path between babysitting and full danger mode
Auto mode uses a classifier that checks each tool call before it runs: safe actions proceed automatically, risky ones get blocked and Claude finds a different approach instead
You no longer need --dangerously-skip-permissions to get flow state — auto mode gives you speed without removing all guardrails
Currently research preview on Team plan. Enterprise and API access rolling out in the coming days
Anthropic recommends running it in isolated environments — the classifier reduces risk but doesn't eliminate it

🥈 Claude Code with --dangerously-skip-permissions is a real attack surface. Lasso published research + an open-source defender worth knowing about.

Wed, 25 Mar 2026 08:00:00 GMT

When Claude reads files, fetches web pages, or gets output from MCP servers, it can't reliably tell your instructions apart from malicious ones hidden in that content
If you're running --dangerously-skip-permissions and Claude clones a repo with a poisoned README, it might just follow the embedded instructions with full system access
Attack vectors include hidden instructions in code comments, malicious web pages Claude fetches, edited Notion/GitHub/Slack pages through MCP connectors, and Base64/homoglyph-encoded payloads
Lasso Security released an open-source PostToolUse hook that scans tool outputs against 50+ detection patterns — warns rather than hard-blocks, takes ~5 minutes to set up

🥉 Agent Flow: A beautiful way to visualize what Claude Code does

Wed, 25 Mar 2026 08:00:00 GMT

Claude Code is a black box — you see what it built, not how it decided to build it. Agent Flow is a VS Code extension that makes the agent's decisions visible in real time
See live tool calls with timing and token costs, watch subagents spawn and coordinate, click into any step to inspect the full transcript
File attention heatmaps show which files Claude is reading and writing most — useful for catching unnecessary context reads that drain your usage budget

🥇 Put Claude to work on your computer

Tue, 24 Mar 2026 08:00:00 GMT

Claude can now physically use your Mac — open apps, navigate browsers, fill in spreadsheets — just like a person sitting at your desk
It uses your connected integrations (Slack, Calendar, etc.) first; when no connector exists, it asks permission to take over your screen directly
The key trick: assign a task from your phone, put it down, and come back to finished work. Claude picks up the conversation exactly where it left off
You can set recurring instructions once — "scan my email every morning" or "pull a report every Friday" — and Claude handles it on schedule
Currently research preview: Pro and Max plans, macOS only. Update your desktop app and pair with mobile to try it

🥈 [Release] anthropics/claude-code-action: Claude Code GitHub Action v1.0

Tue, 24 Mar 2026 08:00:00 GMT

The Claude Code GitHub Action hit v1.0 GA — the API got a complete overhaul and it's cleaner: one `prompt` input instead of three separate modes, all CLI options through `claude_args`
Automatic mode detection — you no longer need to manually specify interactive vs. automation mode; the action figures it out
Breaking change: `mode`, `direct_prompt`, `custom_instructions`, `model`, `allowed_tools` are all gone. If you're on v0.x, read the migration guide before upgrading
New use case examples are bundled: automated PR review, CI failure auto-fix, issue triage, doc generation, security scanning

🥉 I made Claude aware of my entire git history

Tue, 24 Mar 2026 08:00:00 GMT

claudememory indexes your entire git history into a local vector database and exposes it to Claude Code via MCP tools — one `pip install claudememory` away
Claude can now answer "why was X removed?", see the full commit history of a file before editing it, and look up past bug fixes near code it's about to touch
The concrete payoff: Claude stopped re-introducing bugs that were already fixed, because it now checks the fix history before adding new code in that area

🥇 Orchestration -- the exact prompts I use to get 3-4 hour agentic runs

Mon, 23 Mar 2026 08:00:00 GMT

The entire system is a 25-line PLAN.md file that tells Claude *how* to plan AND *how* to implement — not a complex multi-agent framework, something anyone can write themselves
Planning phase (~30 min): Claude asks a few questions upfront, then generates a milestone-specific plan file that embeds its own implementation instructions
Execution phase (2-4 hours): a fresh Claude agent reads that plan file and runs fully autonomously — implementation, validation, and self-correction all baked into the file's instructions
Built-in code review: 4 separate second-opinion calls to a different agent (Claude/Codex) covering KISS, style, correctness, and goal fulfillment — Claude *must* address objections before it ships
Author links to the actual PLAN.md and a real PLAN-M3.md on GitHub so you can copy the format directly into your own projects

🥈 Hook-Based Context Injection for Coding Agents

Mon, 23 Mar 2026 08:00:00 GMT

Instead of dumping everything into CLAUDE.md at session start (where it gets buried 20 minutes later), inject only the *relevant* conventions the moment Claude touches a specific file — via PreToolUse hooks
A billing file gets billing-specific docs injected; a frontend component gets component conventions — all-match routing, general first, domain-specific last, landing at the recency end of the context window
PostToolUse runs grep-based architecture checks that block basic violations (wrong logger, fetch outside hooks, etc.) before Claude can move on
Tested on 15-file context decay runs with both Haiku and Sonnet: 108/108, zero degradation from file 1 to file 15

🥉 PSA for heavy daily use Claude Code users: give yourself a gift and get 'claude-devtools'

Mon, 23 Mar 2026 08:00:00 GMT

claude-devtools reads the log files Claude Code already writes to ~/.claude/ and turns them into something readable — no wrapper, no changed behavior, just visibility
Token breakdown across 7 categories (CLAUDE.md files, tool inputs/outputs, thinking tokens, skill activations, user text, etc.) — finally see exactly what's eating your context window
Full subagent tree: when Claude spawns sub-agents, you see each one's prompts, tool calls, cost, and duration as a nested tree — including agents that spawn more agents

🥇 Bringing Code Review to Claude Code

Sun, 22 Mar 2026 08:00:00 GMT

Claude Code now has a built-in Code Review feature — get structured feedback on your code without leaving the terminal
Ask Claude Code to review any file, a diff, or your full recent changes; it flags issues, suggests improvements, and explains its reasoning inline
Closes the loop: write code → get review → fix → commit, all inside one session — no context switch to a separate chat window
Especially useful for solo devs who don't have peers to review PRs, or teams that want a fast AI first-pass before human review
Ships as a native Claude Code capability, not a third-party add-on

🥈 Advancing Claude for Excel and PowerPoint

Sun, 22 Mar 2026 08:00:00 GMT

Anthropic has shipped updates to how Claude works inside Microsoft Excel and PowerPoint
Improvements cover complex formula handling, richer data analysis in spreadsheets, and better slide generation in PowerPoint
If you're on Microsoft 365, Claude's capabilities inside those apps just got meaningfully deeper — no extra setup needed

🥉 How I got Claude Code to maintain its own documentation (and stop breaking production)

Sun, 22 Mar 2026 08:00:00 GMT

Solo dev with 135 DB tables and 60+ API routes kept having Claude Code edit the wrong file, use the wrong table name, or invent column names that don't exist — costing 15-30 min debug sessions each time
Fix: two living markdown docs that Claude reads before every task and updates after — a 2,000+ line auto-generated ROUTE_REFERENCE.md and a hand-curated SYSTEM_DOCS.md
To bootstrap: point Claude Code at your routes file, server mounts, and sidebar nav → one read-only scan generates the skeleton in ~7 minutes

🥇 Claude Code v2.1.81

Sat, 21 Mar 2026 08:00:00 GMT

New `--bare` flag makes Claude Code safe for scripted/automation pipelines: it skips hooks, LSP, and plugin sync so `-p` calls run cleanly without interactive interruptions — requires an API key (no OAuth/keychain in bare mode)
Multi-session OAuth fix: if you're running several Claude Code windows at once, one session refreshing its token no longer forces all the others to re-authenticate
`--channels` permission relay expanded: channel servers that declare the permission capability can now forward tool approval prompts to your phone (approve/deny file writes, bash commands, etc., without being in the terminal)
Voice mode gets two reliability fixes: silent retry failures now show the real error instead of "check your network", and audio recovers when the server silently drops the WebSocket connection
MCP read/search tool calls now collapse into a single "Queried {server}" line (expand with Ctrl+O), cutting transcript clutter; plugin hooks no longer block prompt submission if the plugin directory is deleted mid-session

🥈 Claude Code GitHub Action v1.0 — Now Generally Available

Sat, 21 Mar 2026 08:00:00 GMT

The Claude Code GitHub Action officially hits v1.0 GA with a cleaned-up API: the old `mode`, `direct_prompt`, `override_prompt`, `custom_instructions`, `allowed_tools`, and `mcp_config` inputs are gone — everything now flows through a single `prompt` input and `claude_args` (matching the CLI format)
Automatic mode detection means you no longer configure interactive vs. automation mode manually — the action figures it out from your setup
Breaking changes from v0.x require migration — check the migration guide before upgrading. New example workflows cover automated PR reviews, CI failure auto-fix, issue triage, documentation generation, and security scanning

🥉 claude-agent-sdk-python v0.1.49

Sat, 21 Mar 2026 08:00:00 GMT

`AgentDefinition` now accepts `skills`, `memory`, and `mcpServers` fields — you can define agents with full Claude Code infrastructure (skill files, memory, MCP servers) directly in Python code rather than wiring it separately
Per-turn token usage now preserved on `AssistantMessage`, new typed `RateLimitEvent`, and programmatic `tag_session()` / `rename_session()` — better observability and organization for multi-session agent runs

🥇 Claude Code v2.1.80

Fri, 20 Mar 2026 08:00:00 GMT

The big headline: `--channels` (research preview) lets MCP servers push messages directly into your Claude Code session — meaning you can control Claude Code from Telegram or Discord, right from your phone
Rate limit visibility is finally here: a new `rate_limits` field in statusline scripts shows your Claude.ai usage percentage and reset times for both 5-hour and 7-day windows — no more guessing when you'll run out
You can now set `effort` in skill/slash command frontmatter to override the model effort level when that command runs — useful for making lightweight commands cheaper without touching global settings
Memory usage drops by ~80MB on large repositories (250k+ files) — if you work on huge codebases this is a noticeable improvement on startup
Bug fixes that matter: `--resume` no longer drops parallel tool results (sessions with parallel tool calls now restore correctly), voice mode WebSocket failures from Cloudflare bot detection are fixed, and plugin installs are now a single `/plugin install` command instead of a two-step flow

🥈 From Zero to Fleet: The Claude Code Progression Ladder

Fri, 20 Mar 2026 08:00:00 GMT

Someone who built a 668,000-line platform with autonomous Claude agents maps out 5 distinct levels of Claude Code mastery: raw prompting → CLAUDE.md → Skills → Hooks → Orchestration
The key insight: you don't level up by deciding to — you get pushed up when something breaks. The fix is always more infrastructure, not more effort
CLAUDE.md has a real ceiling at ~100 lines (theirs crept to 190 with 40% redundancy); Skills are the right place for deep expertise — they have 40 skills totaling 10,800 lines that cost zero tokens when not in use
Level 5 (Orchestration) means parallel agents in isolated worktrees, 198 agents across 109 waves — don't try to skip to it before having solid hooks in place

🥉 How I use Haiku as a gatekeeper before Sonnet to save ~80% on API costs

Fri, 20 Mar 2026 08:00:00 GMT

Simple two-stage pipeline: send everything to Haiku first with a yes/no prompt ("does this contain a real complaint/need?"), then only pass the ~15% that pass to Sonnet for the real work
Result: running Sonnet on 15% of input instead of 100% — the cost difference at scale is massive, and Haiku is surprisingly good at the gate job with few false negatives

🥇 Claude Code GitHub Action v1.0 — Now Generally Available

Thu, 19 Mar 2026 08:00:00 GMT

The Claude Code GitHub Action just hit v1.0 and is now officially stable — you can automate AI-assisted code review, CI failure diagnosis, issue triage, and docs generation directly in your GitHub workflows
Big simplification: instead of juggling separate `mode`, `direct_prompt`, `model`, and `allowed_tools` inputs, everything now flows through a single `prompt` input and a `claude_args` passthrough for any CLI option
The action now intelligently detects whether to run in interactive mode (responding to @claude mentions in PRs) or automation mode (triggered by CI events) — no more manual configuration
Breaking changes from v0.x: several old inputs (`mode`, `direct_prompt`, `override_prompt`, `trigger_phrase`) are gone — check the migration guide before upgrading
Now works with AWS Bedrock and Google Vertex AI, and you get direct access to every Claude Code CLI flag through `claude_args`

🥈 Claude Code v2.1.79

Thu, 19 Mar 2026 08:00:00 GMT

New `/remote-control` command in VSCode lets you bridge your local Claude Code session to claude.ai/code so you can pick it back up from a browser or phone mid-task
Session tabs in VSCode now get AI-generated titles based on your first message — no more mystery sessions
Added `--console` flag to `claude auth login` so you can authenticate against the Anthropic Console (API billing) directly from the CLI
Startup memory usage drops by ~18MB across all scenarios, and non-streaming API calls now have a 2-minute timeout so sessions won't hang indefinitely
Bug fixes: `claude -p` no longer hangs when spawned as a Python subprocess, Ctrl+C works again in print mode, and enterprise users can retry on rate-limit (429) errors

🥉 Claude Code Hooks — all 23 explained and implemented

Thu, 19 Mar 2026 08:00:00 GMT

Someone built a complete open-source reference repo that implements all 23 Claude Code hooks, with a companion YouTube video walking through every use case — hooks are one of the biggest differentiators between Claude Code and other CLI coding agents
Hooks let you run shell scripts automatically at specific points in Claude's workflow (before tool calls, after commits, when sessions end, etc.) — think of them as programmable guardrails and automation triggers that other AI CLIs don't have

🥇 The CLAUDE.md rules that actually change Claude's behavior

Wed, 18 Mar 2026 08:00:00 GMT

CLAUDE.md files work better as hard rules ("never do X") instead of soft guidelines ("prefer to do X"), because Claude will skip soft suggestions when tasks feel urgent
The most effective rules are binary and specific—"TypeScript only, never .js files" works way better than "prefer TypeScript when possible"
Three game-changing rules: (1) only modify files directly needed for the task, don't refactor random stuff; (2) never claim work is done without showing actual test output; (3) don't ask for permission or clarification unless genuinely stuck
Keep CLAUDE.md short (under 50 lines) because it loads into every message and burns tokens—only include rules you'd actually enforce
Add a status file (like tasks/current-task.md) that persists across context resets so Claude doesn't lose track of what it was doing

🥈 I use Claude Code to research Reddit before writing code — here's the MCP server I built for it (470 stars)

Wed, 18 Mar 2026 08:00:00 GMT

Someone built a Reddit search tool (called reddit-mcp-buddy) that works with Claude Code, letting you ask AI questions like "what do developers complain about with this library?" and get real Reddit discussions instead of searching manually—it's got 470 stars and 76K downloads.
The main uses are: researching libraries before you pick one, finding solutions to weird bugs people actually encountered, checking out complaints about competing products, and staying updated on trends—all without leaving your coding window.
It's better than a regular web search because Claude gets clean, structured data from Reddit (not messy HTML), results get cached so repeated questions don't cost API calls, and it's focused on just Reddit instead of giving you a browser to figure out yourself.

🥉 Claude Code .claude directory: what goes where

Wed, 18 Mar 2026 08:00:00 GMT

Claude Code looks for a `.claude/` folder in your project where you set up permissions, modes, and integrations — `settings.json` is the main config file that controls whether Claude asks permission before running commands and lets you connect external tools like databases.
You can also add a `settings.local.json` for personal overrides (keep it out of git), a `skills/` folder for custom abilities, plus a `CLAUDE.md` file for rules and a `.claudeignore` file to keep sensitive files out of Claude's context.

🥇 v2.1.77

Tue, 17 Mar 2026 08:00:00 GMT

Claude's big language models can now output way more text at once — up to 64k tokens normally, and 128k tokens max (that's like writing a whole short story in one response)
A bunch of permission and security fixes so that "Always Allow" rules work properly for complex commands, and deny rules actually get respected instead of being bypassed
Quality-of-life improvements for the code editor like fixed copy/paste, better vim key support, and `/copy N` command so you can grab earlier responses instead of just the latest one
Various technical fixes for long-running sessions (memory leaks), file handling (line endings), and terminal display issues (colors in tmux, hyperlinks opening twice, etc.)
Fixed a critical bug where Claude Desktop was accidentally using the wrong API credentials, and several edge cases that could crash or hang the app

🥈 Self-hosted AI ops on Mac mini: Claude Max + Discord + MCP at $0/month extra

Tue, 17 Mar 2026 08:00:00 GMT

Someone built a personal AI assistant that runs 24/7 on a Mac mini for basically free, using Claude AI through Discord as the interface—it's been running for 2 months with 99.7% uptime and handles 30+ scheduled tasks plus reactive chat all at once
The main challenge was Claude's context window (memory) filling up too fast, so they created a compression system that cuts through 98% of useless output before it reaches Claude, turning 315 KB of noise into just 5.4 KB of useful info, letting sessions run for hours instead of minutes
The whole thing is self-healing with multiple backup recovery layers, so when things break (which happened 14 times), the AI can diagnose and fix itself 64% of the time—plus they open-sourced the code so anyone can set up something similar with no monthly cloud costs

🥉 I turned my idle Claude Max subscription into a 24/7 AI company — $0 extra

Tue, 17 Mar 2026 08:00:00 GMT

Someone realized their Claude Max subscription was sitting idle most of the day, so they built a 24/7 system that runs 11 specialized AI teams handling everything from security audits to market alerts—all for free by using Claude's headless command-line tool that's already included in the subscription.
The secret sauce is a custom "smart filter" that compresses data flowing to Claude by 98% (shrinking 315 KB down to 5.4 KB), which lets Claude keep working on the same task for 3+ hours straight instead of losing context after 30 minutes, and the whole system runs with 99.7% uptime and automatically fixes itself when things break.