Why it matters
AI coding agents often struggle with context management, leading to issues like context window decay and loss of important human input. SpecAnchor's new Harness Context Control aims to address these challenges by making Decision and Evidence contexts persistent and manageable, potentially leading to more reliable and efficient AI-driven development workflows. This could significantly reduce the need for agents to re-derive information and improve the quality of their output.

SpecAnchor, a tool designed to provide spec governance and an anti-decay layer for AI coding agents, has launched its v0.5.0-beta.1 release. This update redefines SpecAnchor as a "Harness Context Control plane," expanding its capabilities beyond just managing Spec Context to also include Decision and Evidence contexts.

The core motivation behind this release is to tackle the challenges associated with LLM context management, such as the degradation of long contexts and the loss of precision due to automatic compaction. Previously, SpecAnchor focused on organizing 'Spec Context' (Global, Module, Task contracts). However, human input during checkpoints often involves adding to specifications or clarifying existing ones, which was not persistently captured.

With v0.5.0-beta.1, SpecAnchor now makes these crucial signals persistent:

  • Decision Context: Checkpoint decisions are logged into Task Spec §5.2, including their status (active, superseded, withdrawn) and a pin flag. A hot/cold view automatically prunes decisions that are no longer relevant.
  • Evidence Context: Verification proofs are recorded in Task Spec §6.2, encompassing commands run, acceptance criteria mapping, unverified risks, manual checks, and rollback handles. Acceptance criteria are auto-pinned, and failed or unverified-risk entries remain 'hot'.

This release introduces a new schema (sdd-riper-one v2) with six new sections in the Task Spec template, along with enhanced configuration options in anchor.yaml for managing context control. New scripts (decision-filter.sh, evidence-filter.sh, specanchor-doctor.sh, specanchor-assemble.sh) and a new command (specanchor_handoff) have been added to support these features. A pre-commit hook is also included to enforce context-control linting.

Migration guidance is provided for existing projects, recommending either appending placeholder sections to legacy task specs or adjusting enforcement levels in anchor.yaml to avoid blocking the doctor linting process.

Sources

Written with AI assistance from the linked sources; every claim below was checked against them automatically. How we produce articles.

Claim check: 9/9 supported claims - 9 evidence links - 99% avg confidence
  • SpecAnchor v0.5.0-beta.1 introduces 'Harness Context Control' for AI coding agents.supported - github.com
  • The update expands context management to include Decision and Evidence contexts, in addition to existing Spec Context.supported - github.com
  • Decision Context logs checkpoint decisions into Task Spec §5.2 with status and pin flags, using a hot/cold filter for relevance.supported - github.com
  • Evidence Context records verification proofs in Task Spec §6.2, including commands, acceptance criteria, and risks, with auto-pinned acceptance criteria.supported - github.com
  • The release introduces a new `sdd-riper-one` v2 schema with six new sections in the Task Spec template.supported - github.com
  • New scripts and a `specanchor_handoff` command are included to support the new context control features.supported - github.com
  • A pre-commit hook is provided to enforce context-control linting.supported - github.com
  • The project has 5 stars on GitHub.supported - github.com
  • The project is written in Shell.supported - github.com

Caveats

  • GitHub star counts are dynamic and may change.
  • Single-source caution: verify critical details at the linked source.
Radar score 86/100 - how it was calculated
Reliability82
Freshness92
Novelty81
Technical77
Developer96
Ecosystem72
Confidence98
  • Reliability 82: GitHub metadata supports source trust
  • Freshness 92: Fresh GitHub activity
  • Novelty 81: Fresh GitHub release
  • Technical 77: Repository technical metadata
  • Developer 96: Developer tooling signals
  • Ecosystem 72: Fresh GitHub release
  • Confidence 98: Claims have reliable evidence
AI agentscoding agentscontext managementspec-driven developmentdeveloper toolsopen sourcegithub
Share
XLinkedInHacker News

Discussion

Loading comments...

Related articles

AI Coding - Oct 21, 2026HyperFrames: HTML to Video for AI AgentsHyperFrames is a TypeScript framework designed for AI agents to write HTML and render it into video. It leverages technologies like Puppeteer and FFmpeg for video generation.AI Coding - Oct 29, 2026Axocoatl v1.5.1: Unattended Coding Agent ExecutionAxocoatl v1.5.1 enables unattended execution of coding agents, supporting models like Claude Code, Codex, or local LLMs within a sandbox. It integrates host-side checks, a separate review model, and generates signed receipts with CI exit codes.AI Coding - Oct 29, 2026Tabularis: Open-Source SQL Workspace with AI Agent SupportTabularis is an open-source desktop SQL workspace featuring 3 built-in and 26 shipped database driver plugins. It includes an MCP server for AI agents like Claude, Cursor, and Devin, alongside SQL notebooks and visual EXPLAIN functionality.AI Coding - Oct 29, 2026Ataraxy Labs SEM: Semantic Version Control for Code AgentsAtaraxy Labs SEM offers semantic version control on top of Git, providing entity-level diffs, blame, and impact analysis. It supports 28 languages via tree-sitter and is specifically built for coding agents.AI Coding - Oct 25, 2026OctoCode: Structural Code Intelligence for AI AgentsOctoCode is a Rust binary providing structural code intelligence for AI agents, featuring semantic search, knowledge graphs, and an MCP server. It aims to give AI tools like Claude and Cursor a deeper understanding of codebases.AI Coding - Oct 24, 2026mu (μ) v0.2.0: AI Coding Agent with Fast JudgeThe mu (μ) project, version v0.2.0, introduces an AI coding agent designed to "think before it acts." It utilizes a small, fast judge for routine decisions, delegating complex tasks to a larger model, and is built upon the pi framework.