The AI Toolchain — issue 004, August 22, 2026

Agent API presets automatically assign stable prompt cache keys, letting independent requests sharing the same preset reuse the cached system prompt and tool definitions prefix, reducing costs by ~5% with no request changes required. An explicit prompt_cache_key field can still be set per-request to override the preset default when finer cache control is needed.

The Python SDK v1.0 requires Python 3.10 or later and moves the HTTP layer from httpx to httpx2. Custom http_client, Timeout, and transport objects must now be built from httpx2, though the DefaultHttpxClient helpers are unchanged; users who patch httpx for tracing or mocking can preserve compatibility by calling httpx2.alias_httpx() before importing anthropic.

import httpx2
httpx2.alias_httpx()

import anthropic
client = anthropic.Anthropic()

On the async client, .with_raw_response results now require await response.parse() instead of the previous synchronous .parse() call; calling it synchronously no longer works.

import anthropic
import asyncio

async def main():
    client = anthropic.AsyncAnthropic()
    raw = await client.messages.with_raw_response.create(
        model='claude-opus-4-5',
        max_tokens=256,
        messages=[{'role': 'user', 'content': 'Hello'}]
    )
    message = await raw.parse()
    print(message.content)

asyncio.run(main())

Ships a full audio FX rack in Studio (staged rollout, now enabled for all users alongside track groups, mute, and solo) including preset folding, per-preset title design, signal-path view, frequency ruler, and family/tint lettering; a Tone module (multi-band EQ) on faders in Studio and at the Core preset-node level; a fifth FX worklet for granular pitch-shifting unlockable through character presets; an audio FX preset catalogue with one-click application; FX tails that decay naturally past the clip boundary; and the Even Out Levels levelling script in Core for automatic level normalisation. Automation lanes let per-parameter keyframed envelopes be copied/pasted, retimed, edge-stretched, shaped/simplified, time-range selected, and edited via curve/snap/direct value entry, with a shared automation-lane row per track and live value display at the playhead. Track groups add group rows to the timeline with volume control, a level meter, split disclosure, and mute-group and solo ('hear-only-this', preview-only) controls, rendering grouped audio through a summed, FX-processed bus in both preview and export. The /hyperframes-audio skill adds agent-driven audio workflows with a waveform cache keyed by file.

Adds HF_DE_PARALLEL_ROUTER environment variable to opt in (=true) or opt out (=false) of the parallel drawElement capture path, default-on for a 5% canary ramp behind a per-install circuit breaker that automatically disables the feature if a render fallback occurs. The router's engagement threshold was lowered from 2000 to 700 frames to reduce latency on mid-length renders. Circuit-breaker state now survives CLI config wipes, and parallel-worker init telemetry in the Producer surfaces the render band's motion axis for short-comp routing decisions; a default-off flag also enables single-worker inversion for short compositions under an element ceiling as groundwork for a fast-path render route.

Catalog pages now play the real composition in the player instead of a pre-recorded video; the sidebar groups 358 items under eight collapsible sections organized by what you want to make; items with variables expose an editable variables panel whose preview rebuilds live as you change values — matching exactly the accent, size, or other variable selected — with shareable links carrying those values; the install command copies the item with variables attached; reinstalling a catalog item no longer overwrites local edits; and provider cost tier for media-use items is now derived from the registry.

Adds device authorization login (RFC 8628) to the CLI, enabling sign-in from attended SSH and remote-terminal sessions without API keys or loopback callbacks. The fail-closed flow validates tokens before persisting them, and leaves existing browser login unchanged.

Adds outputDynamicRange: auto | hdr | sdr field to the Producer API for explicit control over output dynamic range, while temporarily accepting the legacy hdrMode field for rolling migrations.

Adds BRIEF.md output from the agent's intent interview, so a later session can resume without re-asking run-shape questions, and designates frame.md as the canonical design spec, replacing design.md references throughout the prompting documentation. Adds a seven-level novice-to-capstone Prompt Guide arc at /prompting, with 18 verified prompt-to-render examples and a 62-second capstone film dissected chapter by chapter; a motion grammar reference covering eight rules, each shown as an A/B render built from the same composition; a capstone chapter dissecting 'The Timeline' — one continuous camera across nine regions, rendered twice from one template; new chapters covering storyboards, porting from Remotion, and recreating a reference video from text alone; and an appendix documenting known traps: layout waivers, the contrast-gate side effect, fromTo back-render, and round-linecap dots.

Adds the anchored-connector lint rule with source-traceable visuals doctrine to the Skills system, plus telemetry measuring which lint rules fire, their cost, and which fail to converge. Adds the data-layout-allow-caption-zone waiver letting narrow caption-zone layouts bypass the layout constraint; the opt-in proseCoverageFloor rule enforcing a minimum prose-coverage threshold; content_overlap dense motion re-sampling and off_pivot_rotation hub-referenced layout checks; and rotation_pivot_drift, surfacing off-center rotation pivots during layout audits before render. Lint findings now surface inside Studio alongside CLI output.

Introduces distributed plan protocol v2 in the Producer — versioned, integrity-checked, content-addressed artifacts — retaining Plan v1 compatibility, with AWS Lambda worker support extended to protocol v2. Adds direct Plan v2 publishing to S3 from AWS Lambda and to GCS from GCP Cloud Run, streaming content-addressed artifacts without an intermediate staging step.

Distributed cloud renders now default to Plan v2 transport. AWS Lambda or GCP Cloud Run infrastructure must be redeployed from the same package version before upgrading SDK callers, or pass planProtocol: "v1" explicitly to keep prior behavior; v0.8.0 is the supported migration boundary.

Adds local, on-device semantic (meaning-based) catalog search via the CLI, so blocks can be found by concept rather than exact keyword. Catalog search now ranks results by word position and term rarity, improving accuracy for name and rare-word queries. Search and browse fall back to the last on-disk copy when the registry is unreachable, preventing a network timeout from wiping the local catalog, and when a search finds no match the CLI outputs the command to report the gap to the registry.

Enables editing and styling text directly in the Studio preview panel, including applying a style to a partial run of characters (sub-selection styling) and displaying every colour present in a mixed text selection simultaneously in the colour swatch. Rich text is sanitized on ingestion into a composition in the Core pipeline.

Adds normalize-audio CLI command to match one clip's loudness level to another.

Makes caption-editorial-emphasis data-driven with an emphasis heuristic, caption-highlight data-driven with automatic grouping, caption-emoji-pop data-driven with a generic emoji lexicon, caption-pill-karaoke data-driven via the caption-data runtime, and caption-weight-shift data-driven via the caption-data runtime. Caption template runtimes are now encapsulated in IIFEs for isolation.

Producer now sniffs HTML payloads before invoking ffprobe in resolveMediaDuration, preventing non-media HTML sources from being misidentified as media files.

Extends drawElement fast capture (~2x speed) to Windows hosts with a hardware GPU, previously macOS-only; Linux and software-GPU hosts are unchanged.

Adds RenderPerfSummary telemetry for host/render performance in the Producer. Emits the canary decision reason alongside cohort assignment on Core and CLI telemetry events, extending the same reason emission to Studio events so forced overrides are distinguishable from ordinary cohort rolls; classifies identity persistence on every CLI telemetry event, indicating whether an install's anonymous ID will survive to the next run; adds on_battery and low_power_mode fields to render telemetry for per-machine diagnosis; makes CLI telemetry opt-out durable across invocations; and embeds invisible, unsigned renderer and version metadata into encoded output files for diagnostics.

Makes creator media edits render-safe, ensuring edits made in creator mode are compatible with the final render pipeline. Adds copyable recipes in Creator Skills for cuts, trims, zooms, masks, transitions, and placed audio, giving agents and users reusable edit patterns.

New external conflict recovery UI in Studio detects when a project file has been edited outside HyperFrames and lets users resolve the conflict without losing either version, coordinating and preserving external file changes to prevent silent overwrites. Logical timeline navigation replaces raw frame-position movement in Studio, making keyboard control of the timeline more predictable.

Adds track-timeline performance monitoring in Studio. Adds bulk-edit easing for merged keyframes, enabling uniform easing changes across multiple selected keyframes at once. Adds expandable keyframe property lanes, track headers, precise keyframe retiming interactions, and nested-composition timeline support.

Adds media treatment capabilities to the Core, agent-first media treatment tools to the CLI, a media treatment inspector to Studio, media treatment overlays to the Registry catalog, and renders media treatments deterministically in the Runtime.

Adds morphological_graphs() (plural) for batch construction of morphological graphs across multiple areas in a single call, with improved error handling for GeometryCollection inputs, and expands options on the existing morphological_graph() (singular) for finer control over generation.

import city2graph

graphs = city2graph.morphological_graphs(buildings_gdf, tessellation_gdf)

A new /connect menu manages provider connections directly from the CLI, adding beta support for Bring Your Own Key (BYOK) providers so teams can supply their own API keys for AI backends, plus a local-only mode that keeps BYOK credentials and traffic off external networks.

Adds a Skills section in Settings that lists installed skills grouped by project, global, or bundled source with token cost shown for each skill. Supports installing skills from a GitHub repository with per-skill selection, save-location choice, and a native confirmation dialog before installation, and lets users enable, disable, and remove managed skills from the Desktop app while keeping skill state in sync with the CLI.

Adds per-feature model selection in Settings for compaction, session titles, command explanations, image vision, and other feature tasks. Feature-model choices now show plan availability and keep saved model overrides visible, and compact mode names and descriptions match the CLI with the selected compaction model shown in Config.

Adds keyboard shortcuts and a right-click menu option to copy selected terminal text, terminal text size control from Settings with zoom shortcuts when the terminal is focused, and pointer and keyboard controls to resize the terminal drawer with drag-down to close it.

Adds Qwen 3.8 27B as an available model option, then adds low, medium, and xhigh reasoning effort levels for the Qwen 3 8B and 27B models.

The /status command now reports the live session's active model and current reasoning effort.

The CLI adds a workspaces get <workspace-id> subcommand to look up a workspace by ID, set-wake <host-id> <command> and wake <host-id> subcommands that register and trigger a configurable per-host wake command before connecting, and single-host reads with explicit host targeting in the CLI and host-service.

$ superset workspaces get <workspace-id>

$ superset set-wake <host-id> "<wake-command>"
superset wake <host-id>

$ superset workspaces get <workspace-id>

$ superset set-wake <host-id> 'ssh myhost sudo systemctl start superset-host'

$ superset wake <host-id>

$ superset workspaces get <workspace-id>

$ superset set-wake <host-id> "wakeonlan AA:BB:CC:DD:EE:FF"

$ superset wake <host-id>

$ superset workspaces get <workspace-id>

Introduces a trigger editor in the desktop and tRPC layer for building and managing automation triggers, fires automations from GitHub events by recording GitHub deliveries as automation events via the API, adds automation trigger and event tables to the database, dispatches automations from triggers via the API and tRPC, saves triggers as a set alongside the automation, adds identityProvider and automationId scope and triggerId narrowing on the shared dispatcher, expires raw webhook payloads after 14 days, introduces one identity table for every integration, and adds enum values for every planned trigger provider up front in the database.

Adds a unified MCP-install page at /mcp-install for connecting AI agents to Superset, a /mcp alias, a docs MCP server with agent_auth identity types and a reinstated OpenAPI spec, MCP server instructions and destructive/read-only tool annotations for MCP v2, and MCP discovery, markdown twin pages, and agent auth metadata as agent-readiness surfaces across the API and docs.

Introduces cloud workspaces backed by Blaxel sandboxes across the desktop, tRPC, and host-service layers, adds pre-baked repo and schema to sandbox provisioning so the sandbox can start itself, and brings cloud workspace management to the mobile app — list, open, terminal access, provisioning, and actions — with creation via tRPC and optimistic 'creating'/'failed' interstitial states.

The v2 sidebar gains drag-and-drop pinning with collapsible Pinned and Sessions sections, reorderable sidebar groups anywhere in the project list, configurable folder-link clicks with an open-in-Finder binding, separated tasks and pull requests sections, a ChatHistorySidebar message rail, safe-triangle hover intent on hover cards, bulk workspace actions, right-click import of untracked worktrees from the context menu, workspace pinning on the v2 dashboard sidebar, click-to-open-file/cmd-click-to-open-diff behavior, a cleaner denser restyle, and a replacement of the sidebar activity strip with port/agent chips and bulk-close hover cards.

Moves several hot paths off the Node event loop: adds an offLoop() worker-procedure resolver to host-service, ports branchPrefix.gitInfo off the event loop, moves workspace deletion off the event loop while removing status checks that blocked bulk delete, moves hot git reads off the event loops in host-service and desktop with ratchet enforcement, limits diff-stat queries to only the active sidebar workspace item, and derives domain constants from NEXT_PUBLIC_ROOT_DOMAIN for the boid.so cutover.

Adds a superset settings CLI command to manage desktop app settings and theme from the command line, with live-reload when changes are applied.

$ superset settings

Adds a top-bar ports dropdown as an alternative layout for accessing forwarded ports, adds collapsible per-workspace groups within that dropdown, and improves port-scanner performance with a shared ps per scan, lsof -a, and idle-session decay for faster local port discovery.

Adds a Superset orchestration skill and shared 'decide' and 'redesign' skills to the agents layer, ships superset:* plugin skills to all agents with a private feedback pipeline, and adds a browser design mode that lets users click any page element and send it to an agent, computer control via Cua Driver, and an engine-generic browsing skill with a Browser Use 3.0 option.

The MCP package is renamed from packages/mcp-v2 to packages/mcp, with all user-visible surfaces de-versioned; any references to mcp-v2 paths or versioned surface names will break.

Adds a Sydney (syd) region to the relay fleet for lower-latency global coverage, relay WebSocket outage telemetry with version-stamped desktop events, typed WebSocket close codes with Sentry exception capture in the relay layer, and serves host presence live from the relay Durable Object, dropping database presence writes for lower latency.

Replaces Electric sync with plain tRPC + React Query for data fetching on desktop and mobile; the EXPO_PUBLIC_ELECTRIC_URL environment variable is removed.

Adds previous_permission_mode and next_permission_mode actions to tui.keymap.chat in config.toml, letting users bind keys to cycle through available built-in permission modes in the TUI.

[tui.keymap.chat]
next_permission_mode = "]"
previous_permission_mode = "["

Adds existingCall transport to thread/realtime/start, accepting a client-provided callId and optional realtimeSessionId so clients can attach Codex to a pre-negotiated realtime call.

Allows codex agents to accept session-configuration options — model, approval policy, sandbox mode, web search, working directory, and config overrides — when opening the agents dashboard.

New /claude-api upgrade subcommand migrates Python projects from the anthropic 0.x SDK to 1.x, including updating timeout usage from httpx.Timeout to anthropic.Timeout.

Plugins synced from claude.ai now appear as name@synced in cloud sessions and can be managed with claude plugin enable <name>@synced / claude plugin disable <name>@synced; a synced plugin never overrides a locally installed plugin with the same name.

$ claude plugin enable my-plugin@synced

Windows now supports cross-session messaging, letting Claude Code sessions message each other with SendMessage and discover each other via ListAgents, matching existing macOS/Linux behavior. ListAgents and /list-agents now also include live teammates, not just subagents and other local sessions.

Cost estimates shown via /cost, the status line, and --max-budget-usd now factor in the 1.1× US-only-inference premium charged for data-residency workspaces.

CLAUDE_CODE_RETRY_WATCHDOG persistent retry mode now fails immediately on organization spend-limit and out-of-credits errors instead of retrying indefinitely.

The automation_script, web_scrape, and cache tools are removed from the computercontroller extension.

Goose now supports pre-registered OAuth clients for streamable_http extensions, enabling MCP extensions that use pre-configured OAuth credentials without requiring interactive login.

Introduces file-store for the DeepSeek Files API with a MAX_CHAT_IMAGE_BYTES cap of 32 MiB, upload deduplication, expiry-based invalidation, and quota-recovery batch cleanup.

Stores a deterministic canonical image encoding, versioned as REQUEST_IMAGE_TRANSFORM_VERSION = 'request-image-v4', with preferred quality levels of 85 and 80, enabling stable cache and upload-index identity across requests.

Unifies the 'master' and Files request image pipelines via offloadRequestImagesWithPolicy, consolidating how inline base64 and Files-API-referenced images are prepared for chat requests, and expands the source upload envelope to accept a broader set of image sources through this unified path.

New devin.agentWindow.location setting splits the Agent Command Center into a separate window from the editor. The Command Center now follows whichever space is selected instead of being tied to one folder, converts an editor window in place with no reload or save prompt, and supports any number of agent windows running side by side.

post_setup_worktree hooks now run for worktrees created from a Devin Local session, copying .env files and other untracked setup before the session starts.

On Windows, Install Devin CLI now writes a shim to the bundled CLI so updating Devin Desktop also updates the devin command.

Live shell output now streams into the session while a command runs; finished shell rows expand to show the full command and its complete output.

Introduced in v0.7.0 as a command for compact, scored task-neighborhood retrieval, backed by a deterministic JSONL agent protocol with explicit detail levels, scored selection reasons, stable ordering, node quotas, and a hard estimated-output-token ceiling. In v0.7.2, mex graph scope now defaults to bounded source-backed retrieval, returning meta, source, flow, and summary JSONL records prioritized by relevance and execution-path confidence instead of a minimal manifest; adds evidence-aware JSONL protocol v3 records for source ranges, directed execution flows, summaries, omissions, and trustworthy fallback guidance; and scope budgets now adapt to repository size, enforcing per-response file, node, flow, and source ceilings while distinguishing mandatory evidence from optional truncation.

$ mex graph scope "how does the authentication middleware handle token expiry"

$ mex graph scope <node-id>

Adds a twelfth drift checker for grounds_to code-node grounding, covering drift, gone, ambiguous, and durable moved-node repair behavior. Adds checkFrontmatterCompleteness, flagging context/*.md and patterns/*.md files missing recommended frontmatter fields (name, description, last_updated). Adds checkStalePatterns, flagging pattern files with no inbound reference from ROUTER.md or any context/*.md file. Improves checkToolConfigSync to skip user-authored files lacking the scaffold marker and to identify the actual managed config file that drifted rather than always blaming the first file in the list.

Adds mex graph, mex graph query, mex graph get, and mex impact commands for graph building, structural lookup, targeted source expansion, and blast-radius analysis.

$ mex graph && mex impact <node-id>

Adds compiler-backed TypeScript extraction via the TypeScript compiler API, resolving calls, imports, inheritance, containment, callback flow, and declaration-aware source regions. TypeScript 5.9.3 is now an exact runtime dependency, so projects on a different TypeScript version must pin or align to 5.9.3 for graph construction to work.

Builds a deterministic local SQLite code graph (stored in .mex/graph.db) for TypeScript, TSX, JavaScript, JSX, Python, and Rust, with cross-file resolution, body hashes, MinHash fingerprints, and LSH reconciliation. This graph engine requires Node.js 22.5 or later, raised from the previous minimum.

Adds viz_node_limit=<int> key to .graphifyrc — graphify hook install reads it and bakes the value into generated git hooks as a ${GRAPHIFY_VIZ_NODE_LIMIT:-<n>} default, so a per-run env var still overrides the committed limit. graphify hook status now reports the baked viz_node_limit from .graphifyrc and degrades gracefully on a malformed .graphifyrc.

graphify extract --no-dedup skips fuzzy near-duplicate node merging during build and incremental merge runs, while preserving exact-id uniqueness and refusing surprising node drops via a shrink guard.

$ graphify extract --no-dedup .

Adds Common Lisp extraction for .lisp, .cl, .lsp, and .asd files via tree-sitter-commonlisp (optional [commonlisp] extra), covering packages, classes, functions, methods, generics, macros, variable definers, same-file calls, and cross-file opened/:used package resolution.