v0.4.2 - Agent-visible hook warnings

Changes

Fix: Pi hook warnings now visible to agent (#2)

Previously, the Pi hook used ctx.ui.notify() which only showed notifications to the human user in the TUI. The agent never saw these warnings.

Fix: Changed from tool_call with ctx.ui.notify() to tool_result event that modifies bash output content directly. The warning is now appended to git add output so the agent sees it.

Claude Code: Switch to prompt-based hook

Changed from PostToolUse command hook to PreToolUse with type: 'prompt'. This is a cleaner integration - Claude evaluates the prompt and includes the reminder in its response.

Consistent warning message

All hooks now show the same message with a features.json reminder:

⚠️ Long-task-harness: .long-task-harness/long-task-progress.md is not staged.
   If relevant, also update features.json with completed/failed features.
   Update these files before committing to maintain session continuity.

Removed

• droid_install_hooks.py and droid_precommit_check.py - Droid hooks didn't work as expected
• claude_code_precommit_check.py - replaced by prompt-based hook

v0.4.1

Added

• Multi-agent hooks: Claude Code, Factory Droid, and Pi (session reminders, pre-commit warnings, Pi session title tracking).
• Status line (status_line.py) and Ralph-style iteration loops (iteration_loop.py).
• Declarative rules system (check_rules.py) and git_add.py wrapper for staging-time checks.
• Optional repo-local git pre-commit hook installer for unsupported agents (Codex/Cursor/etc.) via precommit_install_hook.py with shared precommit_check.py.

Docs

• Updated README/SKILL instructions and harness tracking for v0.4.x.

v0.3.0

Cleaner layout, better workflow, dogfooding

What's New

🗂️ Cleaner File Layout
All harness files now live in .long-task-harness/ directory, keeping your project root clean.

⚡ Improved Commit Flow
Pre-commit hook now warns instead of blocking. Suggests a follow-up commit for progress updates rather than amending (preserving commit hashes for session references).

🏷️ Clearer Script Naming
Claude Code-specific scripts now prefixed with claude_code_ to clarify they're optional and only for Claude Code users.

📝 Surprise Logging Pattern
New guidance for logging surprises in session entries—things that indicate model uncertainty and contain information-dense context for future sessions.

🐕 Dogfooding
The skill now tracks its own development using itself.

Migration from v0.2.x

If upgrading an existing project:

1. Move your files: mkdir .long-task-harness && mv long-task-progress.md features.json .long-task-harness/
2. Update any custom scripts referencing the old paths

v0.2.1 - Multi-Agent Support

What's New

Multi-Agent Support

The harness now works with any agent that can read markdown files:

• Claude Code: Session hooks (SessionStart, PreToolUse)
• Cursor, Codex, Droid, Pi: AGENTS.md instructions

Agent-Agnostic Naming

• Renamed claude-progress.md → long-task-progress.md

Simplified Configuration

• Agent-driven setup: on first invocation, the agent checks if AGENTS.md/CLAUDE.md has the harness snippet and offers to add it
• Dynamic path insertion: agent uses its actual skill path, no hardcoding needed
• Removed noisy Cursor hooks (afterFileEdit fired too often)

Concise SKILL.md

• Rewrote SKILL.md following skill-creator principles
• Progressive disclosure, less verbose
• Focus on what agents need to know

Migration from v0.2.0

1. Rename claude-progress.md to long-task-progress.md in your projects
2. Update any CLAUDE.md/AGENTS.md snippets to reference the new filename

long-task-harness 0.2.0

v0.2.0 brings improved history management and better tooling for multi-session projects. It follows patterns from Anthropic's Effective Harnesses for Long-Running Agents and HumanLayer's ACE-FCA.

New Features

• Bidirectional Linking (v2 templates): Sessions reference features; features track their session history
• Targeted History Reading: --session N and --list flags for read_progress.py
• Feature Filtering: --feature ID and --json flags for read_features.py
• Git Metadata Tooling: session_metadata.py generates commit info for session entries
• Enhanced Pre-commit Hook: Outputs session metadata when blocking commits
• History Research Protocol: Guidance for using subagents as "scouts" to find relevant history

Changes

• Flattened repo structure (removed .claude-plugin/, skill now at root)
• Simpler installation: gh repo clone tmustier/long-task-harness ~/.claude/skills/long-task-harness
• Generalized language from "Claude Code" to "filesystem-based agents"
• Follows agentskills.io specification

long-task-harness 0.1.0

AI agents lose context across sessions. Each new context window starts fresh, creating a "shift change" problem where continuity is lost. This is a lightweight session continuity plugin (agent skill) to handle agent "shift changes" for long-running tasks.

v0.1.0 is optimised for Claude Opus 4.5. It codifies the patterns and practices from Anthropic's engineering article: Effective Harnesses for Long-Running Agents. Key features:

• Progress Documentation: Maintain claude-progress.md files documenting work history
• Feature List Strategy: Track testable features in structured JSON format
• Session Startup Protocol: Consistent initialization steps for each new session
• Context-Efficient Scripts: Load only recent sessions (~78% context reduction)
• Session Hooks: Auto-remind Claude to invoke the skill on startup