---
type: "post"
id: "phosphor-v0-7-1-improves-performance-and-windows-63e83b"
slug: "phosphor-v0-7-1-improves-performance-and-windows-63e83b"
title: "Phosphor v0.7.1 improves performance and Windows sandbox stability"
description: "Phosphor v0.7.1 changes how run context reaches the model so provider prompt caches can remain warm across agent turns, while fixing several Windows sandbox stalls."
retrieval_nugget: "Phosphor v0.7.1 changes how run context reaches the model so provider prompt caches can remain warm across agent turns, while fixing several Windows sandbox stalls. For teams using the coding agent, it is a runtime reliability release: it aims to preserve ordinary git workflows and bounded permission handling without weakening confinement."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["sandbox"]
entities: ["Phosphor","hackafterdark","golang.org/x/net"]
related_patterns: ["agent-security-moves-to-runtime-boundaries"]
novelty_label: "new_fact"
what_changed: "Phosphor v0.7.1 adds warm provider prompt caches across agent turns and fixes Windows sandbox startup and shutdown stalls."
prior_refs: ["https://newruntime.com/posts/phosphor-coding-agent-v0-7-0-adds-kernel-048069/"]
source_url: "https://github.com/hackafterdark/phosphor/releases/tag/v0.7.1"
source_urls: ["https://github.com/hackafterdark/phosphor/releases/tag/v0.7.1"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:phosphor-v0-7-1-improves-performance-and-windows-63e83b"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/phosphor-v0-7-1-improves-performance-and-windows-63e83b/","markdown":"https://newruntime.com/posts/phosphor-v0-7-1-improves-performance-and-windows-63e83b.md","json":"https://newruntime.com/posts/phosphor-v0-7-1-improves-performance-and-windows-63e83b.json"}
markdown_url: "https://newruntime.com/posts/phosphor-v0-7-1-improves-performance-and-windows-63e83b.md"
---

# Phosphor v0.7.1 improves performance and Windows sandbox stability

## Retrieval answer

Phosphor v0.7.1 changes how run context reaches the model so provider prompt caches can remain warm across agent turns, while fixing several Windows sandbox stalls. For teams using the coding agent, it is a runtime reliability release: it aims to preserve ordinary git workflows and bounded permission handling without weakening confinement.

## Summary
Phosphor v0.7.1 changes how run context reaches the model so provider prompt caches can remain warm across agent turns, while fixing several Windows sandbox stalls. For teams using the coding agent, it is a runtime reliability release: it aims to preserve ordinary git workflows and bounded permission handling without weakening confinement.

## What happened
Phosphor’s GitHub release notes describe v0.7.1 as a performance and stability release. The project says dynamic run state—including the todo list, active goal and post-turn reminders—was previously rebuilt into the system prompt on every agent-loop iteration. A todo update therefore changed the prompt prefix and invalidated provider prefix caches. In v0.7.1, that state is sent once in the run’s user message as a `<run_context>` block, leaving the wire prefix byte-identical across those updates. The release also orders MCP server instructions deterministically, intended to stop tooling from unnecessarily changing the prompt and disturbing the cache. When a model drifts, Phosphor now re-anchors it on its open todo through tool results instead of rewriting the system prompt.

The release notes also describe Windows sandbox fixes. Container ACL grants and revokes on deep workspace or toolchain-cache trees could block process startup or exit for minutes. Those operations now have explicit deadlines; if a grant exceeds its deadline, launch fails closed, and leftover revoked access-control entries are cleaned up at the next launch. A second Phosphor instance using the same data directory now fails immediately and reports the owner’s process ID instead of waiting through a 30-second SQLite busy timeout. Short-lived read-only commands can still inspect a database while a TUI is using it.

Other changes restore unlink and rename operations in writable sandbox roots so `git add`, `git commit` and `git status` work inside the sandbox, while `.git` metadata remains fenced from direct agent writes. Confined child processes launch with `CREATE_NO_WINDOW`, preventing early failures such as `ssh.exe` from printing raw text over the terminal UI. The release also fixes delivery of recovery instructions and todo snapshots on continuation and queued turns, plus a race in a Windows CI test. For security maintenance, Phosphor upgraded `golang.org/x/net` to v0.62.0 to address CVE-2026-97045 and CVE-2026-78661. The release notes say remaining code-scanning and Dependabot alerts were reviewed and dismissed where vulnerable paths were unreachable in this codebase.

## What the archive adds
New Runtime covered Phosphor v0.7.0 on [November 1, 2026](https://newruntime.com/posts/phosphor-coding-agent-v0-7-0-adds-kernel-048069/), when the coding agent added kernel-enforced sandboxes for agent-run shell commands and MCP stdio servers across Linux, macOS and Windows. Version 0.7.1 is a new release in that story: it concentrates on cache stability, Windows behavior and dependency maintenance rather than introducing a new sandboxing boundary.

## Trend signal
This release reinforces [agent security moves to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/). Phosphor’s approach is not only to ask an agent to respect a repository boundary: it keeps `.git` metadata outside direct agent writes while allowing required file operations in writable sandbox roots. Explicit ACL deadlines, cleanup of residual revoked entries and fail-closed startup behavior show that a runtime boundary must remain reliable under failure, not merely restrictive in the happy path.

## Why now
The release follows Phosphor v0.7.0’s introduction of kernel-enforced sandboxing. That makes the timing legible: once confined commands become part of normal agent execution, permission operations, git compatibility and prompt-cache behavior are no longer peripheral implementation details. They determine whether the boundary can remain enabled during everyday, longer-running work.

## Why it matters
A sandbox only helps a development team when it constrains an agent without breaking routine operations or making runs unpredictably slow. The writable-root changes aim to preserve common git actions, while `.git` fencing keeps a more sensitive part of the working tree outside direct agent writes. The Windows changes similarly turn a potentially multi-minute permission stall into an explicit, bounded outcome: a delayed grant fails closed rather than allowing an unclear startup state.

The cache change matters for long-running loops whose todo state changes often. By moving that changing state out of the system-prompt prefix, Phosphor is attempting to let providers reuse cached prompt work across turns. That is an architectural choice teams can recognize in their own harnesses: separate durable instructions from mutable run state where the provider’s cache semantics make that worthwhile. It does not, by itself, establish a particular speedup or cost reduction.

For evaluators, the significant signal is maintenance at the boundary between security, model context and developer ergonomics. A confinement layer that prevents basic source-control operations will be disabled; one that hides slow permission handling behind indefinite waits will be difficult to trust in automation. This release addresses those adoption risks, even though it does not add a new headline feature.

## Limits and uncertainty
The release notes do not publish cache-hit rates, latency reductions, token-cost savings or benchmarks for the new prompt layout. They also do not quantify the frequency of Windows startup and shutdown stalls, identify all affected providers, or specify compatibility constraints for `<run_context>` handling. The two dependency CVEs are named, but the notes do not describe exploitation status or broader impact beyond the stated fixes. These are project-reported changes; independent testing of performance, Windows behavior and security properties is not established here.

## Skill to build now
Audit one agent loop this week for mutable state that is repeatedly merged into a stable instruction prefix. Separate durable policy, tool guidance and role instructions from changing task state such as a todo list, active objective or recovery note. If the model provider supports prefix caching, measure cache reuse before and after the change. Keep the mutable state structured and delivered through an explicit runtime channel, then verify that recovery and queued-turn behavior still reaches the model in the intended order.

## Reader takeaway
Phosphor v0.7.1 is a reliability and hardening release for an existing sandboxed coding agent, not a new general capability. Teams already using it should review the release against Windows and long-run workflows, especially where sandboxed git operations, permission handling and provider prompt caching affect whether the runtime remains practical to use.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Sandbox](https://newruntime.com/topics/sandbox.md): Explore the sandbox topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
