Parallel Project Threads Meet Unified Inline Edits and CLI Metrics

Friday is a multi-agent coordination and telemetry morning, so the useful work is how these four moves let you steer parallel agents and see what actually runs. Anthropic turns Projects into one chat that fans work into parallel cloud Claude Code threads, GitHub ships a single model for ghost text plus nearby and long-distance next edits, Copilot metrics finally show which CLI customizations get used, and Claude Code pins a skill sync plus a must-have gateway fix.

Treat today as a try-Projects, enable-NES, pull-metrics, and pin-v2.1.276 day. Join the Projects beta if you are on Pro or Max with empty project history, confirm next edit suggestions are on in current VS Code, request one Copilot 1-day usage report and scan the new totals_by_ arrays, and pin `@anthropic-ai/claude-code` to `v2.1.276` before you lean on account-synced skills behind a proxy.

Top Stories

Claude Code Projects redesigned: coordinator chat, parallel cloud threads
Practical dev impact: Starting Sep 17, Anthropic’s Projects beta in Claude Code treats a project as one conversation, so you set a goal and repos, Claude scopes the work, then a coordinator delegates parallel threads (each a Claude Code cloud session on its own branch and repo copy) while shared memory, files, and artifacts accumulate across the project. You steer from the main project chat or dive into a thread, and threads can fan out further with subagents, loops, and workflows, because overlap lands as normal merge conflicts. Beta access starts with select Claude Pro and Max subscribers who already use cloud sessions and have no existing web or desktop projects, then expands to more Pro/Max users over the coming week, then Team, Enterprise, Cowork, and regular Claude chats. Parallel threads burn usage faster, so check project-specific usage and set coordinator versus worker model and effort deliberately. Local tool and on-machine thread support is “very soon,” not today, so pilot on a multi-PR migration or latency goal rather than your hottest production hotfix until you see how merges and limits behave.

Copilot unified inline suggestions: one model for completions, NES, and long-distance edits
Practical dev impact: GitHub’s Sep 16 engineering post explains the paid VS Code Copilot inline path that unifies completions, next edit suggestions (NES), and long-distance NES behind a shared “diff patch” output format, so the model can choose the best edit end-to-end instead of client orchestration among three specialized models. Multi-patch responses cache follow-on edits for a faster tab-tab-tab flow, including cross-file patches, while online A/B testing of the shipping 2-in-1 candidate versus the three-model baseline reported no statistically significant regressions on key engagement metrics, about 10% lower time-to-shown suggestion, and about 61% fewer output tokens. The unified experience is available now for paid Copilot users in VS Code once next edit suggestions are enabled, so update VS Code, leave NES on, and watch whether ghost text still “wins” when a semantic rewrite would be better.

Copilot usage metrics API: agentic CLI skills, agents, MCP, slash commands, plugins
Practical dev impact: As of Sep 17, enterprise and organization Copilot usage reports (1-day and 28-day, including `day_totals`) expose agentic CLI customization activity, so platform owners can see which skills, custom agents, MCP servers, slash commands, and plugins are actually getting traffic, then fund the winners and retire dead weight. Top-five arrays such as `totals_by_skill`, `totals_by_custom_agent`, `totals_by_mcp`, `totals_by_slash_cmd`, and `totals_by_plugin` each carry `interaction_count`, plus distinct-use counts for the same categories. Recognized GitHub-provided names appear, while customer-defined skills, agents, MCP, and plugins roll up under `other` (slash commands under `custom`). MCP `interaction_count` increments on connect and reconnect attempts, not per tool call, and plugin skill invocations also appear in skill totals, so do not sum the two. The feature needs the Copilot usage metrics policy enabled and View Copilot Metrics permission, so pull one org 1-day report tomorrow morning and chart the new fields beside yesterday’s budget-request workflow.

Claude Code v2.1.275 and v2.1.276: claude.ai skill sync and gateway proxy fix
Practical dev impact: Pin past `v2.1.274` into `v2.1.275` (Sep 17) for product features, then immediately to `v2.1.276` (Sep 18) if you route through `ANTHROPIC_BASE_URL` proxies or gateways, because v2.1.276 fixes every request failing with `400 … Input tag ‘advisor_20260301’` (a v2.1.275 regression). In v2.1.275, terminal sessions signed in with a claude.ai account sync that account’s enabled skills and plugins (opt out with `syncClaudeAiSkills: false` or `syncClaudeAiPlugins: false`), gain a send-now interrupt (`Ctrl+Enter`, or `Ctrl+X` then `Ctrl+S`) that flushes the queue, show `/status` account confirmation on Claude apps gateway sign-in, warn when `otelHeadersHelper` fails at startup, and offer `/plugin install –marketplace ` to add a marketplace before install. Fleet operators who already pinned v2.1.274 for memory and MCP waits should move to v2.1.276 today so skill sync and gateway traffic do not fight each other.

Practical Impact Analysis

The through-line is turning agent sprawl into something you can steer and measure. Claude Projects put a coordinator over parallel cloud threads so multi-repo or multi-PR work stops living as five disconnected chats, while Copilot’s unified inline model collapses three suggestion paths into one so the editor stops picking the wrong specialized model first. On the ops side, Copilot’s new CLI customization metrics close the visibility gap that skills and MCP inventories opened, and Claude Code’s v2.1.276 pin keeps gateway fleets from eating a 400 on every turn after yesterday’s skill-sync release.

If you run Claude Pro or Max cloud sessions with a clean Projects slate, Friday is the day to try one coordinator-led goal (for example a deprecation across two repos) and watch merge conflicts and usage meters. If your team lives in VS Code Copilot, leave NES enabled and treat the 10%/61% numbers as a reason to re-evaluate muscle memory that always accepts the first ghost-text completion. Platform owners should add the new `totals_by_` fields to the same dashboard that already tracks budget requests, because unused custom agents are budget you cannot see. Anyone behind an Anthropic-compatible proxy should not linger on v2.1.275.

If you only do three things this morning: start one Projects beta thread set with a bounded goal, confirm NES is on after your latest VS Code update, and pin Claude Code to v2.1.276 (then pull one Copilot metrics day for the new CLI fields).

Tutorial

Pin Claude Code to v2.1.276 and verify claude.ai skill sync (or deliberately opt out). Use your existing Claude Code login. Keep secrets in the environment.

1. Install the `v2.1.276` pin and confirm `claude –version` (do not stop on `v2.1.275` if you use `ANTHROPIC_BASE_URL`).
2. Keep default skill and plugin sync, or write opt-outs in `~/.claude/settings.json` when a machine must stay isolated from account packs.
3. Start a signed-in session, confirm `/status` shows the expected account, and check that skills enabled on claude.ai appear in the slash menu (or that the opt-outs stuck).
4. Next, in VS Code enable next edit suggestions for paid Copilot, and for Copilot metrics run a 1-day org or enterprise report and inspect `totals_by_skill` / `totals_by_mcp`.

#!/usr/bin/env bash
set -euo pipefail

# Pin past v2.1.275 skill sync AND the v2.1.276 gateway/proxy advisor_20260301 fix.
npm install -g "@anthropic-ai/claude-code@2.1.276"
claude --version  # expect 2.1.276

mkdir -p ~/.claude
# Default on v2.1.275+: account skills/plugins sync into signed-in terminal sessions.
# Opt out only if you want terminal skills isolated from claude.ai:
cat > ~/.claude/settings.json <<'JSON'
{
  "syncClaudeAiSkills": true,
  "syncClaudeAiPlugins": true
}

... click "Show full code" below to expand

#!/usr/bin/env bash
set -euo pipefail

# Pin past v2.1.275 skill sync AND the v2.1.276 gateway/proxy advisor_20260301 fix.
npm install -g "@anthropic-ai/claude-code@2.1.276"
claude --version  # expect 2.1.276

mkdir -p ~/.claude
# Default on v2.1.275+: account skills/plugins sync into signed-in terminal sessions.
# Opt out only if you want terminal skills isolated from claude.ai:
cat > ~/.claude/settings.json <<'JSON'
{
  "syncClaudeAiSkills": true,
  "syncClaudeAiPlugins": true
}
JSON

# Behind ANTHROPIC_BASE_URL / corporate gateway: stay on v2.1.276+ after any v2.1.275 install.
# claude -p "List skills you can see and say whether account sync is on."

Recommended AI prompt

Copy this paragraph into ChatGPT, Claude, Gemini, Grok, or whatever you use.

You are my staff engineer for Claude Code Projects pilots, Copilot inline-suggestion rollout, Copilot CLI usage-metrics ops, and Claude Code fleet pins. Context for 2026-09-18: Anthropic (Sep 17) redesigned Projects in Claude Code so a project is one coordinator conversation that delegates parallel cloud threads (each a Claude Code cloud session on its own branch and repo copy) with shared memory, files, and artifacts; beta for select Pro/Max cloud users with no existing projects, then wider Pro/Max, then Team/Enterprise/Cowork/chat; local threads coming soon; parallel usage burns faster. GitHub/VS Code (Sep 16) shipped a unified Copilot inline suggestions model (completions + NES + long-distance NES) using a diff-patch format; 2-in-1 online flight showed ~10% faster time-to-shown and ~61% fewer output tokens vs three-model baseline with no key-metric regressions; available now for paid Copilot in VS Code with NES enabled. GitHub Changelog (Sep 17) added agentic CLI customization fields to Copilot usage metrics (totals_by_skill/custom_agent/mcp/slash_cmd/plugin plus distinct_* counts) in 1-day and 28-day org/enterprise reports; privacy redacts customer-defined names to other/custom; MCP counts connects; requires usage-metrics policy. Claude Code v2.1.275 syncs claude.ai skills/plugins into signed-in terminal sessions (opt out syncClaudeAiSkills/syncClaudeAiPlugins), adds Ctrl+Enter send-now, gateway account confirm in /status, otelHeadersHelper failure warning, /plugin install –marketplace; v2.1.276 fixes 400 advisor_20260301 when ANTHROPIC_BASE_URL points at a proxy/gateway. Ask which Claude plan, Projects access, VS Code/Copilot setup, metrics permission, and gateway topology we run. Then produce (1) a one-goal Projects pilot checklist, (2) a NES verification checklist for VS Code, (3) a metrics query plan for the new CLI fields, and (4) a v2.1.276 pin plus skill-sync policy note. Keep it concrete and copy-paste ready.

Sources: Anthropic (Projects redesigned) · The Verge · VS Code Blog · GitHub Changelog · GitHub Docs (Copilot usage metrics REST API) · Claude Code release v2.1.275 · Claude Code release v2.1.276