Memory Warnings Meet Dev Container Agents and Budget Escalations

Thursday is an agent-runtime and review morning, so the useful work is how these four moves keep long sessions alive under real machine, environment, and budget pressure. Anthropic ships a Claude Code pin that surfaces memory pressure and bounds MCP startup waits, VS Code puts agent sessions inside your project’s Dev Container with a more portable Codex harness, GitHub turns budget lockouts into an approve-or-deny workflow, and Zed opens a public beta that keeps humans and agents in one continuous thread instead of a giant PR diff.

Treat today as a pin, containerize, escalate, and try-Delta day. Upgrade Claude Code to v2.1.274 and set an MCP startup wait for headless jobs, enable Agents-window Dev Container runs on one repo that already has a `.devcontainer`, walk one Copilot budget request through Approve and increase, and open one Delta thread with a teammate on a small change before you rewrite your review process.

Top Stories

Claude Code v2.1.274: memory warnings, MCP startup waits, stuck-session fixes
Practical dev impact: Pin `@anthropic-ai/claude-code` to `v2.1.274` so long sessions get a visible warning when memory usage is critical (with steps to free memory or restart safely), and so the first non-interactive turn can bound how long it waits for connecting MCP servers via `CLAUDE_CODE_MCP_STARTUP_WAIT_MS` (`0` means do not wait). The same release stops sessions from endlessly retrying unexpected `tool_use_id` 400 errors by self-healing corrupted transcripts when possible, or ending the loop with a clear error and a `/rewind` hint. Operators also get OpenTelemetry hooks that matter in fleets: an `effort` attribute on the `claude_code.llm_request` span (aligned with the `api_request` event), plus `claude_code.managed_settings_resolved` for managed-settings sources (redacted details with `OTEL_LOG_MANAGED_SETTINGS=1`). Streamable HTTP MCP tool calls now honor a longer per-server `timeout` instead of dying around five minutes, and HTTP MCP servers that only speak legacy HTTP+SSE can connect after a first-request 4xx. Start with the version pin, one forced low-memory observation if you can, and one `-p` run that proves MCP wait behavior under a short `CLAUDE_CODE_MCP_STARTUP_WAIT_MS`.

VS Code v1.138: Dev Container agents and portable Codex sessions
Practical dev impact: Upgrade to VS Code `v1.138` (Sep 16) when you want Agents-window work to use the same toolchain as the repo instead of whatever happens to be installed on the laptop. With `chat.agentHost.devContainer.enabled`, local folders that already have a supported Dev Container config can start an agent session inside that container (Docker required), so builds and tests match project dependencies. The same release expands Codex in the agent host: pick GitHub Copilot or ChatGPT subscription models from the picker without losing the conversation, continue the same Codex session between the ChatGPT app and VS Code, reuse ChatGPT computer-use setup when present, and let Codex call VS Code built-in, extension, and MCP tools. Automations are enabled by default and can be exported or imported for sharing. Preview cleanup can nudge Mark as Done when a session’s PRs merge, with optional auto-mark and auto-delete grace periods left off by default. Roll out on one containerized repo first, then decide whether Copilot-backed or ChatGPT-backed Codex is the default for your team.

Copilot budget increase requests are generally available
Practical dev impact: When a member exhausts their Copilot AI credits under usage-based billing, they can now request more budget at the moment they are blocked, and organization owners, enterprise owners, or billing managers can approve, adjust, or deny that request without leaving settings. Requests route to whichever account pays for the budget (organization-owned requests land in org settings; enterprise-owned in enterprise settings). Review pending items under Requests from members, set the new amount, and choose Approve and increase so the member’s budget updates and credit-consuming Copilot features restore immediately. The feature ships for GitHub Copilot Business and GitHub Copilot Enterprise on usage-based billing. Pair it with your existing budget caps so a hard stop becomes a ticketed escalation instead of a silent productivity cliff.

Zed Delta public beta: agent threads instead of pull-request review
Practical dev impact: If agent-authored diffs are drowning your PR queue, try Zed’s Delta public beta (Sep 16): a multiplayer environment where you invite teammates into the same agent thread, share worktrees, ask the original agent why a change landed, and open a review subthread with an isolated copy of those worktrees so reviewers can explore or fix without breaking the parent thread. Delta is built on DeltaDB (Git-compatible deltas between commits plus human and agent messages), ships clients for macOS, Linux, and Windows plus web and mobile browser access, and stays free during the public beta. You do not have to abandon GitHub: Zed is keeping `zed-industries/zed` on GitHub for community PRs while encouraging Delta threads alongside them. Run one end-to-end fix-review-land flow on a non-critical change before you turn off PRs anywhere that still needs external contributors.

Practical Impact Analysis

The through-line is keeping long agent work alive and reviewable under real machine, environment, and budget constraints. Claude Code v2.1.274 and VS Code v1.138 both tighten the runtime: one warns when memory is critical and stops infinite `tool_use_id` retries while bounding MCP waits for `-p` and SDK jobs, while the other runs the agent inside the project’s Dev Container and lets Codex sessions move between Copilot and ChatGPT without a cold start. If your fleet still floats `latest` on Claude Code, or still runs Agents-window work against host toolchains that do not match CI, Thursday is the day to pin v2.1.274 and try one Dev Container agent session on a repo that already has the config.

Copilot’s budget request GA closes the human loop that hard usage caps opened. Budgets still matter, but a member who burns through AI credits mid-incident can escalate inside GitHub instead of pinging finance on Slack. Wire the Requests from members queue into the same on-call or platform rotation that already owns Copilot policy, and document how much you will approve without a second review.

Zed Delta is the collaboration bet. It does not replace Git storage, and it will not fit every open-source or compliance workflow on day one, but it directly attacks the “agent wrote 2,000 lines, please review the PR” failure mode by keeping decision context with the change. Pilot it beside GitHub PRs rather than as a big-bang cutover.

If you only do three things this morning: pin Claude Code to v2.1.274 and set an MCP startup wait for headless runs, enable Dev Container agent sessions on one prepared repo in VS Code v1.138, and dry-run one Copilot budget Approve and increase (then optionally open one Delta review thread on a small change).

Tutorial

Pin Claude Code to v2.1.274 and bound MCP startup wait for a non-interactive run. Export whatever auth your org already uses for Claude Code. Keep secrets in the environment.

1. Install the `v2.1.274` pin and confirm `claude –version`.
2. Set a short `CLAUDE_CODE_MCP_STARTUP_WAIT_MS` so the first `-p` turn fails fast when servers are down (`0` means do not wait).
3. Run one non-interactive prompt and confirm a missing MCP server no longer stalls past your wait.
4. Tell operators `/rewind` is the escape hatch when a transcript corruption error ends a retry loop, then next open a Dev Container Agents-window session in VS Code v1.138 and practice one Copilot budget Approve and increase in a non-prod org.

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

# Pin past v2.1.274 memory warning, MCP wait, and stuck tool_use_id fixes.
npm install -g "@anthropic-ai/claude-code@2.1.274"
claude --version  # expect 2.1.274

# Bound how long the first non-interactive turn waits for MCP connect.
# 0 = do not wait; raise for CI that needs servers ready before turn one.
export CLAUDE_CODE_MCP_STARTUP_WAIT_MS=3000

# Optional fleet telemetry (managed settings redaction).
# export OTEL_LOG_MANAGED_SETTINGS=1

claude -p "Reply with ok and the MCP servers you can see."

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

# Pin past v2.1.274 memory warning, MCP wait, and stuck tool_use_id fixes.
npm install -g "@anthropic-ai/claude-code@2.1.274"
claude --version  # expect 2.1.274

# Bound how long the first non-interactive turn waits for MCP connect.
# 0 = do not wait; raise for CI that needs servers ready before turn one.
export CLAUDE_CODE_MCP_STARTUP_WAIT_MS=3000

# Optional fleet telemetry (managed settings redaction).
# export OTEL_LOG_MANAGED_SETTINGS=1

claude -p "Reply with ok and the MCP servers you can see."

Recommended AI prompt

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

You are my staff engineer for Claude Code fleet pins, VS Code Agents-window Dev Containers, Copilot usage-budget ops, and Zed Delta review pilots. Context for 2026-09-17: Claude Code v2.1.274 (Sep 17) adds a visible critical-memory warning, CLAUDE_CODE_MCP_STARTUP_WAIT_MS to bound the first non-interactive turn’s wait for connecting MCP servers (0 = do not wait), OpenTelemetry effort on claude_code.llm_request plus claude_code.managed_settings_resolved (redact with OTEL_LOG_MANAGED_SETTINGS=1), stuck tool_use_id retry self-heal or clear error with /rewind, Streamable HTTP MCP timeout honors per-server timeout, and legacy HTTP+SSE MCP connect after first-request 4xx. VS Code v1.138 (Sep 16) adds Agents-window sessions in local Dev Containers via chat.agentHost.devContainer.enabled (Docker required), expands Codex in the agent host with Copilot or ChatGPT subscription switching without losing the conversation, cross-app Codex continuation with the ChatGPT app, optional computer use, and VS Code/MCP tools; automations on by default with export/import; preview session cleanup when PRs merge. GitHub Copilot (Sep 16) makes budget increase requests GA for Business and Enterprise on usage-based billing: members request more AI credits when blocked; org/enterprise owners or billing managers review Requests from members, set an amount, and Approve and increase to restore access immediately; requests route to the paying account. Zed Delta public beta (Sep 16) is a multiplayer agent coding and review environment with shared threads, review subthreads and isolated worktrees, DeltaDB on top of Git, macOS/Linux/Windows/web/mobile clients, free during beta. Ask which Claude Code pin, MCP topology, VS Code Agents/Dev Container setup, Copilot plan/billing account, and review workflow we run. Then produce (1) a v2.1.274 pin plus MCP wait and /rewind checklist, (2) a one-repo Dev Container agent trial plan for v1.138, (3) a budget-request approve matrix for platform on-call, and (4) a Delta-alongside-PR pilot outline. Keep it concrete and copy-paste ready.

Sources

1. Claude Code v2.1.274 release
2. Claude Code CHANGELOG
3. VS Code 1.138 release notes
4. Copilot budget increase requests are generally available
5. Managing Copilot budget requests
6. Zed: Replace PRs with Delta (public beta)