This is my Agent Harness configuration. Tell your agent to clone it to ~/.config/agents and symlink the parts you want into your harness.
Example setup (pi-coding-agent) git clone https://github.com/vedang/agents ~/.config/agentsmkdir -p ~/.pi/agentln -s ~/.config/agents/AGENTS.md ~/.pi/agent/AGENTS.mdln -s ~/.config/agents/APPEND_SYSTEM.md ~/.pi/agent/APPEND_SYSTEM.mdln -s ~/.config/agents/pi-settings.json ~/.pi/agent/settings.jsonln -s ~/.config/agents/agents ~/.pi/agent/agentsln -s ~/.config/agents/prompts ~/.pi/agent/promptsln -s ~/.config/agents/pi-extensions ~/.pi/agent/extensionsln -s ~/.config/agents/skills ~/.pi/agent/skillsThe claude/ directory holds the Claude Code surface of this
harness: CLAUDE.md (symlink to the shared AGENTS.md), settings.json
(Explanatory output style, notification hooks, statusline, permissions),
ported subagent role cards under claude/agents/ , and
skills under claude/skills/ (claude-specific skills plus
symlinks to the shared skills/).
See docs/install-claude.md for the symlink setup
into ~/.claude, the pi→Claude surface map, and what replaces each pi
package. Migration plans for my own pi extensions live in
docs/claude-compat-*.md .
Current Pi runtime profile Source of truth: pi-settings.json .
Setting Value Notes defaultProviderdeepseekPrimary provider. defaultModeldeepseek-v4-proDefault model for new sessions. defaultThinkingLevelhighModerate reasoning by default. steeringModeallApply steering globally. followUpModeone-at-a-timeSequential follow-up flow. hideThinkingBlockfalseShow thinking block. themedarkBase default; mac-system-theme.ts can override it in interactive macOS sessions. compaction.enabledtrueContext compaction on. compaction.reserveTokens16384Reserved token budget. compaction.keepRecentTokens20000Keep recent context before compacting.
zai-custom/zai-glm-4.7zai-custom/glm-5-turbozai-custom/glm-5.1openai-codex/gpt-5.5deepseek/deepseek-v4-flashdeepseek/deepseek-v4-proPackages loaded from pi-settings.json Package Source Purpose pi-askuserquestiongit:github.com/ghoseb/pi-askuserquestionStructured multi-choice questions via ask_user_question. pi-interactive-shellnpm:pi-interactive-shellOverlay, dispatch, and monitor workflows for interactive CLI agents. dafny-estimationgit:github.com/unravel-team/dafny-estimationFormal specification-driven project estimation skill. cavemangit:github.com/vedang/cavemanToken-efficient caveman response mode plus terse helper skills. pi-boomeranggit:github.com/vedang/pi-boomerangCollapse large task context into compact summaries. pi-custom-provider-zaigit:github.com/vedang/pi-custom-provider-zaiRegisters ZAI-hosted GLM models as custom provider models. pi-prompt-historygit:github.com/vedang/pi-prompt-historyPrompt-history search for past session inputs. pi-quizmegit:github.com/vedang/pi-quizmeQuiz-style recall and change-comprehension checks. pi-ralph-loopgit:github.com/vedang/pi-ralph-loopRalph-style autonomous planning loops via /ralph commands. pi-simplify-codegit:github.com/vedang/pi-simplify-codeRefactor/simplify changed code without changing behavior. pi-subagentsgit:github.com/vedang/pi-subagentsSubagent execution, parallel work, and chain orchestration. visual-explainergit:github.com/vedang/visual-explainerGenerate self-contained HTML visual explainers. pi-read-mapgit:github.com/Whamp/pi-read-mapBetter large-file reads with structure maps. chrome-cdp-skillgit:github.com/pasky/chrome-cdp-skillInteract with live local Chrome tabs through CDP. pi-autoresearchgit:github.com/davebcn87/pi-autoresearchAutonomous experiment loop with metrics tracking and confidence scoring.
File Description AGENTS.mdGlobal workflow rules: jj, planning folders, tests, commits, tagref. APPEND_SYSTEM.mdExtra system-prompt guidance layered into the harness. pi-settings.jsonPi runtime defaults, enabled models, and package list.
File Description context-builder.mdBuilds codebase context and a planner-ready meta-prompt from requirements. multimodal.mdReads PDFs, images, and other media for implementation details. plan-reviewer.mdCritiques plans for missing steps, bad assumptions, and risk. planner.mdTurns scoped context into a concrete implementation plan. researcher.mdRuns focused web research and produces a source-backed brief. reviewer.mdReviews implementation quality and can make small follow-up fixes. scout.mdFast codebase recon with compressed handoff context. worker.mdGeneral-purpose implementation subagent for scoped coding work.
Reusable prompt templates and operator instructions.
Checked-in local runtime extensions.
Checked-in local skills in this repo.
Skill Description docx/Create, edit, and analyze Word .docx documents. pdf/Extract, edit, merge, split, OCR, and otherwise process PDFs. pptx/Read, edit, and create PowerPoint .pptx decks. skill-creator/Create or update reusable agent skills. xlsx/Create, edit, and validate spreadsheet outputs.
Skills provided by installed packages These are available at runtime through packages in pi-settings.json, even though they are not stored under this repo's skills/ directory.
Skill Provided by Description cavemancavemanUltra-compressed communication mode for token efficiency. caveman-commitcavemanTerse Conventional Commits helper. caveman-helpcavemanQuick reference for caveman commands and modes. caveman-reviewcavemanOne-line compressed code review comments. caveman-compresscavemanCompress natural-language memory files while preserving technical content. dafny-estimationdafny-estimationRisk-adjusted project estimation using Dafny specs. visual-explainervisual-explainerHTML visual explainers for plans, systems, and diffs. chrome-cdpchrome-cdp-skillInspect and interact with live local Chrome tabs. interactive-shellpi-interactive-shellCheat sheet and workflow for launching TUI coding agents. codex-clipi-interactive-shellCodex CLI reference material. gpt-5-4-promptingpi-interactive-shellPrompting guidance for GPT-5.4. codex-5-3-promptingpi-interactive-shellPrompting guidance for Codex 5.3.