ultimate-pi 0.1.2 → 0.1.4

package/vault/wiki/index.md

@@ -0,0 +1,437 @@
+---
+type: index
+status: active
+created: 2026-04-28
+updated: 2026-05-05
+tags: [meta, index, catalog]
+
+---
+
+# Codebase Map & Architecture Wiki
+
+## Overview
+This wiki maps the codebase architecture, tracks key software design decisions, and serves as the persistent memory layer for the ultimate-pi agentic harness. See [[harness-implementation-plan]] for the authoritative master plan (updated May 2026: skill-first v2 architecture).
+
+## Harness Pipeline
+
+| Layer | Module | Summary |
+|-------|--------|---------|
+| — | [[harness]] | 8-layer mandatory pipeline with drift monitor + sentrux structural gate + cross-cutting tool enhancements |
+| — | [[harness-implementation-plan]] | Master build plan (Skill-First v2): phases mapped to skills vs code, unified token budget, sentrux P44 integration |
+| L1 | [[spec-hardening]] | Block execution until ambiguities resolved |
+| L2 | [[structured-planning]] | Machine-readable task DAG + sprint contracts |
+| L2.5 | [[drift-detection-unified]] | Runtime drift monitor: 3 paradigms (tool-call, spec, implementation) |
+| L3 | [[grounding-checkpoints]] | Smallest verifiable change + spec-drift detection |
+| L3 | [[think-in-code-enforcement]] | Mandatory code-over-data paradigm at tool layer |
+| L4 | [[adversarial-verification]] | Critic agents attack with hard-threshold pass/fail |
+| L5 | [[automated-observability]] | Instrumentation is definition-of-done |
+| L6 | [[persistent-memory]] | claude-obsidian wiki as knowledge base |
+| L7 | [[schema-orchestration]] | Archon workflow DAG, enforces wiki contract |
+| L8 | [[wiki-query-interface]] | LLM-native search via claude-obsidian skills |
+
+## Pipeline Integration
+
+| Page | Summary |
+|------|---------|
+| [[harness-wiki-skill-mapping]] | Skill-to-layer contract: which wiki skill fires when |
+| [[harness-wiki-pipeline]] | Read-first/write-after data flow between harness and wiki |
+
+## Decisions
+
+| Decision | Status | Summary |
+|----------|--------|---------|
+| [[adr-008]] | active | Spec-Only Black-Box QA |
+| [[adr-009]] | active | claude-obsidian Mode B for Persistent Memory |
+| [[adr-010]] | active | Agentic Harness ↔ Wiki Tight-Coupling Contract |
+| [[adr-011]] | accepted | Multi-Agent Consensus Debate with Selective Routing (iMAD) |
+| [[adr-017]] | **superseded** | ~~Harness Project Structure — Event bus removed, pi built-in replaces custom wiring~~ |
+| [[2026-04-30-pi-lean-ctx-native]] | active | Adopt pi-lean-ctx native package, drop custom extension |
+| [[colocate-wiki]] | active | Co-locating Wiki with Codebase |
+
+## Formal Models & Control Frameworks
+
+| Framework | Summary |
+|-----------|---------|
+| [[harness-control-frameworks]] | Unified view: H-Formalism + Feedforward-Feedback + Generator-Evaluator |
+| [[harness-h-formalism]] | H = (E, T, C, S, L, V) — 6-component formal model |
+| [[feedforward-feedback-harness]] | Guides + Sensors, Computational + Inferential |
+| [[generator-evaluator-architecture]] | GAN-inspired separation of generator and evaluator agents |
+| [[model-adaptive-harness]] | Harness generates provider-native prompts (v2 redesign, May 2026) |
+| [[provider-native-prompting]] | Generate prompts optimized for each provider's official conventions |
+| [[Prompt Renderer]] | Build-time compilation: base spec → per-model prompts via pluggable renderers |
+| [[Build-Time Prompt Compilation]] | Compile prompts at build time, ship as static JSON in npm — zero runtime cost |
+| [[harness-configuration-layers]] | Four-layer config: Signal, Gate, Channel, Completion — now provider-native (v2) |
+| [[self-evolving-harness]] | Auto-synthesis and meta-optimization of harness code |
+
+## Concepts — Legendary Engineering Patterns
+
+| Concept | Summary |
+|---------|---------|
+| [[legendary-engineering-patterns-harness]] | 10 patterns from Torvalds/Thompson/Ritchie/Stroustrup/Hejlsberg/van Rossum mapped to harness layers |
+
+## Concepts — Execution & Drift
+
+| Concept | Summary |
+|---------|---------|
+| [[drift-detection-unified]] | Three complementary drift paradigms at L2.5, L3, L4 |
+| [[context-drift-in-agents]] | Progressive degradation of agent behavior over extended interactions |
+| [[meta-agent-context-pruning]] | Detect stuck → prune dead context → restart agent (novel synthesis) |
+| [[agent-loop-detection-patterns]] | Production patterns: repetition, ping-pong, retry-without-progress |
+| [[guardian-agent-pattern]] | Pre-execution and post-execution safety validation |
+| [[codex-harness-innovations]] | Codex innovations: 10 patterns cataloged from open-source Rust agent, first-principles analysis |
+| [[codebase-intelligence-harness-integration]] | 7-point integration map for fallow codebase intelligence in harness pipeline |
+| [[codebase-intelligence-ecosystem-comparison]] | Cross-language gap analysis: no ecosystem has fallow-equivalent |
+| [[cursor-harness-innovations]] | 10 Cursor innovations cataloged with first-principles analysis for harness integration |
+| [[context-anxiety]] | Models rush to finish as context window fills |
+| [[verification-drift-detection]] | Stub — see [[grounding-checkpoints]] |
+| [[fork-safe-spec-storage]] | Fork isolation: gitignored cache + `harness init` bootstrap |
+| [[content-addressed-spec-identity]] | Content-hash spec identity + `harness migrate` transfer-on-merge |
+| [[harness-engineering-first-principles]] | Synthesized 12 first principles from Fowler, OpenAI, LangChain, Augment |
+| [[skill-first-architecture]] | **UPDATED (May 2026)**: Harness layers as markdown skills — only drift monitor remains as code. Event bus handled by pi built-in. Zero-compile iteration. |
+| [[agent-skills-pattern]] | Progressive disclosure: skills loaded on-demand to prevent context rot |
+| [[policy-engine-pattern]] | Pre-execution gates: deterministic constraints vs probabilistic compliance |
+| [[gemini-cli-architecture]] | Gemini CLI SOTA: 15 harness innovations, 2-package architecture, weekly releases |
+
+## Concepts — Chunking & Embeddings
+
+| Concept | Summary |
+|---------|---------|
+| [[AST-Aware Code Chunking]] | Split code at AST boundaries, not character limits; +4.3 Recall@5 |
+| [[Contextualized Text Embedding]] | Prepend scope/signatures/imports before embedding raw code |
+| [[Late Chunking vs Early Chunking]] | Embed full doc then pool vs embed chunks separately; contextual retrieval is sweet spot |
+
+## Concepts — Context & Search
+
+| Concept | Summary |
+|---------|---------|
+| [[think-in-code]] | Paradigm: write code to process data, not read raw data |
+| [[agentic-harness-context-enforcement]] | 5-layer enforcement for context-efficient behavior |
+| [[context-mode]] | FTS5 sandbox architecture for tool output |
+| [[lean-ctx]] | Context Runtime with AST compression and 90+ shell patterns |
+| [[context-continuity]] | Session state preservation across context compaction |
+| [[ast-compression]] | Tree-sitter signature extraction, 60-95% reduction |
+| [[shell-pattern-compression]] | 90+ command patterns (git, npm, cargo, docker) auto-compressed |
+| [[fts5-sandbox]] | context-mode intercept-and-sandbox architecture |
+| [[ast-truncation]] | Stubbing function bodies at AST level for token-efficient reading |
+| [[fuzzy-edit-matching]] | Diff algorithm tolerating formatting drift |
+| [[inline-post-edit-validation]] | Syntax-only post-edit compiler/parser check (<2s) |
+| [[hybrid-code-search]] | BM25 + embeddings + RRF fusion |
+| [[agent-search-enforcement]] | 3-layer defense to force semantic search over grep/cat |
+| [[ts-execution-layer]] | Replace flat tool calling with typed TypeScript API + sandboxed runtime (3-4x context reduction) |
+| [[mcp-tool-routing]] | MCP protocol for semantic search as first-class agent tool |
+| [[codebase-to-context-ingestion]] | Converting entire codebases into structured LLM context |
+| [[memory-system-of-record-vs-ephemeral-cache]] | Canonical wiki memory + optional local fast cache layering |
+
+## Concepts — Orchestration & Harness
+
+| Concept | Summary |
+|---------|---------|
+| [[agentic-orchestration-pipeline]] | Three patterns: subagent delegation, team dispatch, sequential chaining |
+| [[agent-harness-architecture]] | Agent = Model + Harness. Scaffolding + Harness phases. Feedforward guides + Feedback sensors |
+| [[multi-agent-specialization]] | Specialization by role, model, and tool set. Team composition via YAML config |
+| [[context-engineering]] | Staged compaction, dual-memory, event-driven reminders, lazy tool discovery, prompt caching |
+| [[safety-defense-in-depth]] | Five-layer architecture: prompt → schema → approval → validation → hooks |
+
+## Concepts — Agent Skills & Methodology
+
+| Concept | Summary |
+|---------|---------|
+| [[pi-vscode-extension-landscape]] | Three extension patterns for Pi in VS Code: official bridge, LM provider, community full UI |
+| [[vcc-conversation-compaction-for-pi]] | VCC in Pi means deterministic conversation compaction + recall (`pi-vcc`), not VS Code UI |
+| [[deterministic-session-compaction]] | **UPDATED**: No-LLM compaction pattern: pi-vcc, Distill, MemoSift, Codex DSC. Now includes context folding comparison and 3-layer token management architecture |
+| [[context-folding]] | **NEW**: RL-learned branch/return sub-trajectories. 10x context reduction (32K vs 327K). arXiv 2510.11967 (ByteDance/CMU/Stanford). Now in Anthropic's API |
+| [[superpowers-methodology]] | Hard-gate discipline framework: brainstorm→plan→TDD→subagent→review pipeline |
+| [[agent-skills-ecosystem]] | 490K+ SKILL.md marketplace, open standard, cross-agent, security risks |
+
+## Concepts — Agent Architecture
+
+| Concept | Summary |
+|---------|---------|
+| [[Multi-Agent AI Coding Architecture]] | Universal pattern: Planner→Architect→Coder/Generator→Evaluator with structured handoffs |
+| [[Context-Aware System Reminders]] | Event-driven injection of behavioral guidance at decision points to counteract attention-decay |
+| [[Context Engine (AI Coding)]] | Semantic search engine providing deep codebase understanding |
+| [[Semantic Codebase Indexing]] | Converting code to vector embeddings for similarity search |
+| [[Dual-Model Agent Architecture]] | Fast model for iteration + deliberative model for selection |
+| [[Prompt Enhancement]] | Pre-processing queries with retrieved codebase context |
+| [[Majority Vote Ensembling]] | Generate N solutions, LLM selects best |
+| [[Contractor vs Employee AI Model]] | Context makes the difference, not intelligence |
+| [[agentic-harness]] | Central execution pipeline concept |
+| [[consensus-debate]] | Multi-agent dialectical debate protocol for harness decisions |
+| [[lifecycle-hooks]] | Tool-level deterministic hooks (30+ events, exit-code semantics, 100% compliance) |
+| [[structured-compaction]] | Five-layer compaction with forked subagent, ~85% context reduction |
+| [[subagent-worktree-isolation]] | Fresh context windows + isolated git worktrees per subagent |
+| [[selective-debate-routing]] | iMAD: trigger debate only when beneficial (92% token savings) |
+| [[agent-codebase-interface]] | Designing tool interfaces for AI agents, not humans |
+| [[progressive-disclosure-agents]] | Layered codebase information (L0-L3) within token budgets |
+| [[repo-map-ranking]] | Graph centrality algorithm for selecting important symbols |
+| [[execution-feedback-loop]] | Agent debug-test-iterate cycle via structured test output |
+| [[model-routing-agents]] | Dispatching exploration to cheap models, code gen to frontier |
+| [[pi-messenger-analysis]] | What we adopt/strip from pi-messenger for consensus transport |
+
+## Concepts — Antigravity Innovations
+
+| Concept | Summary |
+|---------|--------|
+| [[antigravity-agent-first-architecture]] | Two-view control plane: Editor View + Manager View for multi-agent orchestration |
+| [[agent-artifacts-verifiable-deliverables]] | Trust via human-reviewable deliverables instead of raw tool logs |
+| [[agent-browser-browser-automation]] | Vercel Labs agent-browser (31.4K stars, Apache 2.0, Rust-native) — browser automation CLI for AI agents. Snapshot + refs, annotated screenshots, structured diff. Replaces browser-harness for P30 May 2026. |
+| [[browser-subagent-visual-verification]] | Headless browser agent that visually verifies UI changes (now uses Vercel Labs agent-browser — May 2026) |
+
+## Modules — Other
+
+| Module | Summary |
+|--------|---------|
+| [[skills]] | Agent capability plugins |
+| [[extensions]] | Programmatic hooks |
+| [[bench]] | Evaluation tools (terminal-bench) |
+
+## Entities — Projects & Tools
+
+| Entity | Summary |
+|--------|---------|
+| [[gsd-get-shit-done]] | **NEW**: 60K ⭐ meta-prompting system. Discuss→plan→execute→verify→ship pipeline. Downstream from our harness. |
+| [[pi-coding-agent]] | Open-source terminal coding agent by Mario Zechner. TypeScript extension API. Our harness platform |
+| [[opendev]] | Open-source terminal agent. First comprehensive architecture paper (arXiv:2603.05344v1) |
+
+## Entities — People
+
+| Entity | Summary |
+|--------|---------|
+| [[jesse-vincent]] | **NEW**: Creator of Superpowers. Request Tracker, K-9 Mail, Prime Radiant founder |
+| [[disler-indydevdan]] | Creator of pi-vs-claude-code. 15+ Pi extensions. Reference implementation for orchestration |
+
+## Entities — Companies
+
+| Entity | Summary |
+|--------|---------|
+| [[Lovable (company)]] | Full-stack AI dev platform (formerly GPT Engineer). SOC 2, ISO 27001, multi-agent pipeline |
+| [[Bolt.new (StackBlitz)]] | Browser-based AI web dev. WebContainers + Claude. 0→$4M ARR in 4 weeks. Open source |
+| [[Rocket.new]] | "Vibe Solutioning" platform: strategy + building + competitive intelligence. $15M seed, 1.5M users |
+| [[Emergent Labs]] | YC S24. Autonomous coding agents replacing traditional software development |
+| [[Linus Torvalds]] | Linux kernel, Git, chain-of-trust, don't break userspace |
+| [[Ken Thompson]] | Unix co-creator, extreme leverage from deep understanding |
+| [[Dennis Ritchie]] | Unix co-creator, C language, K&R, economy from constraints |
+| [[Anders Hejlsberg]] | Turbo Pascal, Delphi, C#, TypeScript — fast feedback, behavioral compatibility |
+| [[Guido van Rossum]] | Python — pragmatism over perfection, human control over architecture |
+| [[Bjarne Stroustrup]] | C++ — evolutionary design, compatibility over purity, static typing |
+| [[Augment Code]] | AI coding platform with Context Engine, #1 SWE-bench Pro |
+| [[lean-ctx]] | Context Runtime for AI Agents — adopted via pi-lean-ctx native |
+| [[ck-tool]] | Hybrid code search tool, Rust, 1.6k ⭐, MCP-native |
+| [[vgrep-tool]] | Vector embedding search engine, Rust, 144 ⭐ |
+| [[autodev-codebase]] | TypeScript MCP server with call graphs + Ollama embeddings |
+| [[ops-codegraph-tool]] | Python dependency graph engine + semantic search |
+| [[codesearch]] | Rust MCP server for Claude Code/OpenCode |
+| [[javascript-runtimes]] | Node.js (stable, mature), Deno (secure, tooling-rich), Bun (fast, drop-in replacement) |
+| [[vitest]] | Vite-native test framework, Jest-compatible, v4.1.5, Oxc-powered TypeScript |
+
+## Concepts — Code Quality & Architecture Governance
+
+| Concept | Summary |
+|---------|---------|
+| [[Quality Signal (sentrux)]] | Geometric mean of 5 root cause metrics (0-10,000 score), Nash Social Welfare theorem |
+| [[Five Root Cause Metrics (sentrux)]] | modularity, acyclicity, depth, equality, redundancy — 3 edge + 2 node graph properties |
+| [[sentrux Rules Engine]] | TOML-based architectural constraint system with layers, boundaries, CI enforcement |
+| [[sentrux MCP Integration]] | 9-tool MCP server for AI agent feedback loops: scan, session_start/end, check_rules, etc. |
+
+## Concepts — TypeScript & Code Organization
+
+| Concept | Summary |
+|---------|--------|
+| [[typescript-strict-mode]] | `strict: true` enables 8+ compiler checks that prevent null reference bugs |
+| [[barrel-files]] | Re-export files: useful for libraries, harmful for app code (circular imports, 68% module bloat) |
+| [[monorepo-architecture]] | Single-repo multi-package: built-package vs internal-packages, Turborepo, ESM rules |
+| [[result-monad-error-handling]] | Functional pattern: `Result<Ok, Err>` with map/flatMap/match — errors as values |
+
+## Resolved Questions
+
+| Resolution | Resolves |
+|-----------|---------|
+| [[resolved-context-pruning-inplace-vs-restart]] | In-place vs restart, cache interaction, chain-of-thought, meta-agent regress |
+| [[resolved-treesitter-dynamic-languages]] | Tree-sitter + dynamic languages, 3-layer solution (syntax→types→runtime) |
+| [[resolved-imad-debate-gating-transfer]] | iMAD QA→code review transfer, selective routing, model-specific classifiers |
+| [[resolved-small-model-meta-agents]] | Haiku/Flash for drift detection, cost analysis, infinite regress resolved |
+| [[resolved-mcp-tool-preference]] | MCP priority system (none exists), 3-layer enforcement, false positive rates |
+| [[resolved-context-window-economics]] | Token allocation model, monorepo handling, caching, Gitingest questions |
+
+## Research Syntheses
+
+| Synthesis | Summary |
+|-----------|---------|
+| [[Research: pi-vcc]] | **UPDATED**: Pi compaction ecosystem grew 4→7 extensions. Anthropic launched official compaction API. Context Folding (arXiv) achieves 10x reduction. pi-vcc still only zero-LLM option. 80K token accuracy cliff validated |
+| [[Research: vcc extension for pi coding agent]] | **NEW**: Topic split into two meanings: VS Code extension ecosystem vs literal `pi-vcc` compaction package |
+| [[Research: how GSD fits into our coding harness setup]] | **NEW**: GSD is downstream (builds apps); our harness is upstream (controls agent behavior). Complementary. 3 adoptable patterns, 3 integration levels, 4 open questions. |
+| [[Research: pi-vs-claude-code Agentic Orchestration Pipeline]] | **NEW**: Three orchestration patterns, harness architecture, safety layers, implementation path |
+| [[Research: Augment Code Context Engine]] | Context Engine architecture, benchmarks, integration plan for harness |
+| [[Research: Model-Adaptive Agent Harness Design]] | Four-layer harness must be specialized per model |
+| [[Research: context-mode vs lean-ctx]] | context-mode vs lean-ctx + Think in Code enforcement |
+| [[Research: semantic code search tools]] | Self-hosted semantic code search — ck recommended |
+| [[research-agent-first-codebase-exploration]] | Rethinking OSS strategies for AI agents (ACI design) |
+| [[research-wozcode-token-reduction]] | WOZCODE architecture analysis and harness integration |
+| [[research-gitingest-gitreverse-integration]] | Gitingest adopted, GitReverse skipped |
+| [[Research: Meta-Agent Context Drift Detection]] | Novel synthesis: detect → prune → restart |
+| [[research-agentic-coding-harness-latest-papers]] | 5 pipeline improvements, 3 future phases, formal H-model |
+| [[Research: GitHub Issues as Harness Spec Storage]] | GitHub Issues as cloud-persistent spec storage with sub-issues + dependencies |
+
+| [[Research: Codex State-of-the-Art Harness Improvements]] | Codex open-source architecture: 7 validations, 5 new gaps, 3 novel patterns, 5 new phases |
+| [[Research: Automating Software Engineering - Lovable, Bolt, Emergent, Rocket]] | Platform architectures (Lovable, Bolt, Rocket, Emergent) + first-principles harness patterns from OpenAI/Anthropic/OpenDev |
+| [[Research: Fallow Codebase Intelligence Harness Integration]] | Fallow: 7 harness integration points, cross-ecosystem gap analysis, P44 phases |
+| [[Research: Claude Code State-of-the-Art Harness Improvements]] | Claude Code architecture: 6 gaps, 5 new phases, 4 first principles |
+| [[Research: cursor.sh Harness Innovations]] | Cursor production harness: 5 validations, 4 new gaps, first-principles lessons |
+| [[Research: Google Antigravity Harness Integration]] | Google Antigravity SOTA: artifacts, browser subagent, agent-first architecture, 3 new phases |
+| [[Research: Gemini CLI SOTA Harness Integration]] | Gemini CLI SOTA: 15 innovations mapped, 7 integration priorities from first principles |
+
+| [[Research: TypeScript Execution Layer for Agent Tool Calling]] | TS execution layer: CodeAct → Cloudflare Code Mode → Executor convergence, 3-4x context reduction, P43 phase |
+| [[Research: executor.sh Harness Integration]] | executor.sh: integration layer scope, 3 new P43 sub-phases (catalog, discovery, policy), build vs integrate decision |
+| [[Research: Prompt Renderer for Multi-Model Agent Harness]] | Build-time prompt renderer: per-model compilation, caching, variable system, npm distribution |
+| [[Research: TypeScript Best Practices and Codebase Structure]] | Strict mode, runtimes, barrel files, monorepos, folder structure, error handling, testing |
+| [[Research: Superpowers Skill for Agentic Coding Agents]] | **NEW**: Superpowers methodology, hard-gate enforcement, harness integration, ecosystem analysis |
+| [[Research: Skill-First Harness Architecture]] | **UPDATED (May 2026)**: Rethought MVP + harness plans from first principles. Skills as atomic unit. 3 code files vs 15. Event bus removed. |
+| [[Research: Engineering Workflows of Legendary Programmers and AI Harness Mapping]] | **NEW (May 2026)**: 10 patterns from 6 legendary programmers mapped to harness layers |
+| [[Research: claude-mem over Obsidian for Harness Layer]] | Replacement assessment: keep Obsidian as source of truth, optional claude-mem as cache |
+| [[Research: how claude-mem fits into our workflow. and whether it should replace obsidian in the codebase. no hard feelings about previous actions, rethink from first principles always]] | First-principles reassessment: keep Obsidian canonical, use claude-mem only as non-authoritative acceleration cache |
+| [[Research: claude-mem over obsidian wiki as the knowledge base for our agentic harness pipeline. think from first principles. does this replace or complement our current setup? no hard feelings about previous decisions. gimme accurate points]] | First-principles verdict: complement, not replace. Keep wiki canonical; use claude-mem only as fast cache |
+
+## Sources (External Research)
+
+| Source | Type | Summary |
+|--------|------|---------|
+| [[distill-deterministic-context-compression]] | github-repo | **NEW**: Siddhant-K-code/distill (143 stars), 4-layer deterministic context compression, ~12ms, different layer than pi-vcc |
+| [[codex-dsc-rfc-8573]] | github-issue | **NEW**: Codex DSC RFC — deterministic checkpoint proposal, validates pi-vcc pattern, closed as not_planned |
+| [[pi-compaction-extensions-ecosystem]] | ecosystem-analysis | **UPDATED**: 7 Pi compaction extensions across 3 layers: prevention (rtk-optimizer), mid-session (context-prune), boundary (vcc, omni-compact, model-aware, custom, agentic) |
+| [[pi-rtk-optimizer-github-repo]] | github-repo | **NEW**: Upstream token reduction via command rewriting + output compaction, v0.5.3, deterministic, MIT |
+| [[pi-omni-compact-github-repo]] | github-repo | **NEW**: Large-context model (1M+) subprocess compaction, v0.1.2, opposing philosophy to pi-vcc |
+| [[pi-context-prune-github-repo]] | github-repo | **NEW**: Tool-call batch summarization with original preservation + prefix caching, 5 trigger modes |
+| [[anthropic-compaction-api]] | official-docs | **NEW**: Anthropic server-side compaction API (beta Jan 2026). Claude Opus 4.7/4.6, Sonnet 4.6. LLM-based |
+| [[context-folding-paper]] | paper | **NEW**: arXiv 2510.11967 (ByteDance/CMU/Stanford). 10x context reduction via RL-trained branch/return. 62% BrowseComp-Plus, 58% SWE-Bench at 32K tokens |
+| [[pi-mono-compaction-docs]] | official-docs | Core Pi compaction docs (`/compact`, `/autocompact`, thresholds). Baseline for evaluating `pi-vcc` |
+| [[pi-vscode-marketplace]] | marketplace-listing | **NEW**: Official Pi VS Code extension (`pi0.pi-vscode`), 1,348 installs, IDE bridge + `@pi` chat participant |
+| [[pi-vscode-model-provider-marketplace]] | marketplace-listing | **NEW**: Pi as VS Code LM provider (`vscode.lm.*`) for Copilot Chat + model picker |
+| [[vscode-pi-community-extension]] | marketplace-listing | **NEW**: Community unofficial extension (`cdervis.vscode-pi`), 352 installs, RPC sidebar/chat UX |
+| [[pi-vcc-github-repo]] | github-repo | **UPDATED**: `pi-vcc` deterministic compaction/recall package for Pi sessions, v0.3.12, includes `overrideDefaultCompaction` route control and refreshed install metrics |
+| [[gsd-github-repo]] | github-repo | **NEW**: 60K ⭐ meta-prompting/spec-driven dev system for Claude Code + 15 runtimes. MIT. v1.40.0. |
+| [[gsd-codecentric-deep-dive]] | engineering-blog | **NEW**: Deep architecture analysis of GSD's slash commands, agents, hooks, state persistence |
+| [[gsd-hn-discussion]] | discussion-thread | **NEW**: 473-pt HN discussion: community consensus, comparisons, limitations |
+| [[superpowers-github-repo]] | github-repo | **NEW**: 179K ⭐ agentic skills framework, 14+ composable skills, hard-gate enforcement |
+| [[superpowers-release-blog]] | blog-post | **NEW**: Original release announcement, persuasion principles, TDD for skills |
+| [[superpowers-termdock-analysis]] | article | **NEW**: Third-party deep dive, skill-by-skill analysis, philosophy breakdown |
+| [[sources/disler-pi-vs-claude-code]] | github-repo | Pi extension showcase: 15+ extensions for orchestration, safety, UI, cross-agent integration |
+| [[sources/opendev-arxiv-2603.05344v1]] | academic-paper | OpenDev architecture: compound AI system, ACC, system reminders, 5-layer safety (Mar 2026) |
+| [[sources/martin-fowler-harness-engineering]] | engineering-blog | Harness = Guides + Sensors + Steering Loop. Computational vs Inferential controls (Apr 2026) |
+| [[sources/mindstudio-four-agent-types]] | article | Four agent types: Coding Harness, Dark Factory, Auto Research, Orchestration (2026) |
+| [[sources/anthropic-effective-harnesses]] | engineering-blog | Authoritative harness definition: runtime framework for tool dispatch, context, safety (2025) |
+| [[Source: Lovable Architecture & Clone Analysis]] | blog | Multi-agent architecture: Planner→Architect→Coder, Pydantic-typed handoffs, LangGraph orchestration |
+| [[Source: Bolt.new Architecture & Case Study]] | case-study | WebContainers + Claude, 0→$4M ARR in 4 weeks, Remix frontend, open source |
+| [[Source: Rocket.new — Vibe Solutioning Platform]] | news | Strategy→Build→Intelligence platform, $15M seed, 1.5M users, "vibe McKinsey" |
+| [[Source: OpenAI Harness Engineering — 0 Lines of Human Code]] | engineering-blog | Codex-built product: 0 human-written lines, ~1M lines code, progressive disclosure, architecture enforcement |
+| [[Source: OpenDev — Building AI Coding Agents for the Terminal]] | paper | Compound AI system, adaptive compaction, system reminders, 5-layer safety, MCP lazy discovery |
+| [[linux-kernel-coding-workflow]] | official-documentation | Linux kernel coding style + dev process (Torvalds, 2026) |
+| [[unix-philosophy]] | encyclopedia | Unix philosophy (Thompson, Ritchie, McIlroy, Raymond) |
+| [[birth-of-unix-kernighan-interview]] | interview-podcast | Bell Labs Unix creation (Kernighan, 2020) |
+| [[hejlsberg-7-learnings]] | interview-blog | 7 engineering lessons (Hejlsberg, GitHub Blog 2026) |
+| [[guido-python-design-philosophy]] | interview-blog | Python design philosophy (van Rossum, 2009/2025) |
+| [[Source: Google Gemini CLI Architecture Docs]] | official-docs | Architecture: 2 packages (cli+core), ReAct loop, tool system |
+| [[Source: Google Blog - Gemini CLI Announcement]] | official-announcement | Launch: free tier, 1M token window, MCP, Google Search grounding |
+| [[Source: Render AI Coding Agents Benchmark 2025]] | benchmark-report | Independent: Cursor 8/10, Gemini 6.8/10, strengths/weaknesses |
+| [[Source: Martin Fowler - Harness Engineering]] | engineering-blog | Canonical: feedforward/feedback, computational/inferential, steering loop |
+| [[Source: LangChain - Anatomy of Agent Harness]] | engineering-blog | Agent=Model+Harness, primitives derivation, model-harness co-evolution |
+| [[Source: OpenAI Harness Engineering Five Principles]] | engineering-blog | 5 principles from 1M-line agent-built codebase |
+| [[Source: Augment - Harness Engineering for AI Coding Agents]] | engineering-blog | PEV loop, constraint layers, measurement metrics |
+| [[Source: Gemini CLI Changelogs]] | official-changelog | Feature evolution: v0.4–v0.40 across 40+ weekly releases |
+| [[Augment Context Engine Official]] | product-page | Context Engine features, benchmarks, team impact |
+| [[Augment SWE-bench Agent GitHub]] | github-repo | #1 open-source SWE-bench agent, dual-model architecture |
+| [[Augment SWE-bench Pro Blog]] | blog-post | #1 SWE-bench Pro at 51.80%, same-model comparison |
+| [[Augment Code WorkOS ERC 2025]] | conference-recap | Context Engine as prompt enhancer, live demo |
+| [[Augment Code Codacy AI Giants]] | podcast-recap | Engineering interview: context strategy, pricing, onboarding |
+| [[Augment Code MCP SiliconAngle]] | news-article | MCP launch, 30-80% improvement as context provider |
+| [[Auggie Context MCP Server]] | github-repo | Community MCP wrapper for Augment context engine |
+| [[meng2026-agent-harness-survey]] | paper | H=(E,T,C,S,L,V), 110+ papers, 23 systems |
+| [[anthropic2026-harness-design]] | blog | GAN-inspired generator-evaluator architecture |
+| [[bockeler2026-harness-engineering]] | blog | Feedforward/feedback controls (Martin Fowler) |
+| [[forgecode-gpt5-agent-improvements]] | blog | Model-adaptive harness findings (GPT 5.4 vs Opus 4.6) |
+| [[openai-prompt-guidance]] | official-docs | OpenAI prompt guidance: GPT-5.5 through GPT-4.1 |
+| [[anthropic-prompt-best-practices]] | official-docs | Anthropic prompt engineering: Claude Opus 4.7 through Haiku 4.5 |
+| [[gemini-3-prompting-guide]] | official-docs | Google Gemini 3 prompting guide on Vertex AI |
+| [[lou2026-autoharness]] | paper | AutoHarness: LLM synthesizes harness from feedback |
+| [[lee2026-meta-harness]] | paper | Meta-Harness: outer-loop harness optimization |
+| [[fan2025-imad]] | paper | iMAD: selective debate routing, 92% token savings |
+| [[ironclaw-drift-monitor]] | github-issue | Rule-based stuck-pattern detection (nearai/ironclaw #1634) |
+| [[langsight-loop-detection]] | blog | Production loop detection with argument hashing |
+| [[agent-drift-academic-paper]] | paper | Agent Drift: ASI framework, 42% task success reduction |
+| [[vectara-guardian-agents]] | blog | Guardian Agents benchmark (~900 scenarios, 6 domains) |
+| [[wozcode]] | product-doc | WOZCODE token-reduction architecture (25-55% savings) |
+| [[ck-semantic-search]] | docs | ck hybrid code search documentation |
+| [[vgrep-semantic-search]] | readme | vgrep vector embedding search |
+| [[cast-code-chunking-paper]] | paper | cAST: AST-aware code chunking for RAG, +4.3 Recall@5 (June 2025) |
+| [[vectara-chunking-vs-embedding-naacl2025]] | paper | Chunking strategy ≥ embedding model for retrieval quality, NAACL 2025 |
+| [[coir-code-retrieval-benchmark]] | paper | CoIR: standard benchmark for code retrieval, ACL 2025, pip-installable |
+| [[code-chunk-library-supermemory]] | tool | Production AST-aware code chunking library (npm), tree-sitter based |
+| [[embedding-models-benchmark-supermemory-2025]] | benchmark | MiniLM vs BGE vs E5 vs Nomic on BEIR TREC-COVID (June 2025) |
+| [[context-mode-website]] | website | context-mode.com landing page |
+| [[leanctx-website]] | website | leanctx.com landing page |
+| [[think-in-code-blog]] | blog | "Think in Code" by B. Mert Köseoğlu |
+| [[gitingest]] | tool | Codebase-to-structured-text for LLM context |
+| [[gitreverse]] | tool | Repo-to-synthetic-prompt (not adopted) |
+| [[github-sub-issues-docs]] | documentation | GitHub official docs on sub-issues (parent-child hierarchies) |
+| [[github-issue-dependencies-docs]] | documentation | GitHub official docs on issue dependencies (blocked-by/blocking) |
+| [[gh-sub-issue-extension]] | github-repo | Community gh CLI extension for sub-issue management |
+| [[gh-cli-sub-issue-rfc]] | github-issue | Official gh CLI feature request for native sub-issue support |
+| [[github-fork-issues-discussion]] | github-discussion | Fork issues enablement evolution (Jun-Dec 2025) |
+| [[aider-repomap-tree-sitter]] | blog | Tree-sitter + graph ranking for LLM code context |
+| [[swe-agent-aci]] | paper | Agent-Computer Interfaces concept (2024) |
+| [[swe-bench]] | paper | Real-world software engineering benchmark |
+| [[openhands-platform]] | paper | Open platform for AI developer agents |
+| [[oss-guide-codebase-exploration]] | blog | Human-centric OSS codebase exploration guide |
+| [[cursor-shadow-workspace-2024]] | engineering-blog | Shadow Workspace: hidden Electron window LSP pre-validation |
+| [[cursor-agent-best-practices-2026]] | engineering-blog | Agent best practices: Plan Mode, hooks, skills, context management |
+| [[cursor-harness-april-2026]] | engineering-blog | Harness evolution: dynamic context, Keep Rate, error classification |
+| [[cursor-shipped-coding-agent-2026]] | engineering-blog | Composer system architecture, MoE, latency, sandboxing |
+| [[cursor-instant-apply-2024]] | engineering-blog | Speculative edits: 1000 tok/s, fast apply model, diff problem |
+| [[cursor-fork-29b-2025]] | analysis | VS Code fork architecture, vertical agent thesis |
+| [[codex-open-source-agent-2026]] | github-repo | Codex CLI: 79.2K stars, Rust, sandboxing, hooks, MCP, skills |
+| [[claude-code-architecture-vila-lab-2026]] | academic-paper | VILA-Lab arxiv 2604.14228: comprehensive Claude Code architecture analysis |
+| [[claude-code-architecture-qubytes-2026]] | blog | Qubytes: five-layer architecture breakdown (agent loop, permissions, tools, state, compaction) |
+| [[claude-code-architecture-karaxai-2026]] | blog | KaraxAI: full stack walkthrough — CLAUDE.md, skills, plugins, MCP, hooks |
+| [[codeact-apple-2024]] | paper | CodeAct: 20% higher success rate, 30% fewer turns (ICML 2024) |
+| [[cloudflare-codemode]] | official-docs | Cloudflare Code Mode: TypeScript execution layer, Worker sandbox |
+| [[executor-rhyssullivan]] | github-repo | Executor: local-first TS runtime, 1.3K stars, cross-agent tool catalog |
+| [[colinmcnamara-context-optimization-codemode]] | blog | Context optimization: sub-agents vs TypeScript interfaces |
+| [[fallow-rs-codebase-intelligence]] | github-repo | Fallow: Rust-native TS/JS codebase intelligence, 1.7K stars, MIT. Dead code, duplication, complexity, boundaries, MCP server |
+| [[google-antigravity-official-blog]] | engineering-blog | Google Antigravity: agent-first IDE, Manager View, Artifacts, Skills |
+| [[google-antigravity-wikipedia]] | encyclopedia | Antigravity: VS Code fork, Gemini 3.1, Windsurf $2.4B acquisition |
+| [[cursor-vs-antigravity-2026]] | comparison | Cursor v2.6 vs Antigravity v1.20: centaur vs manager, benchmarks |
+| [[Source: Build-Time Prompt Compilation Architecture]] | architecture-analysis | Real tools: DIY pipeline (js-yaml + PromptWeaver + per-model renderers), replaces fabricated PromptKit PackC |
+| [[Source: SwirlAI Agent Skills Progressive Disclosure]] | newsletter | Three-tier progressive disclosure, ecosystem adoption speed, skills as system design pattern (Mar 2026) |
+| [[Source: Claude API Agent Skills Overview]] | official-docs | Filesystem-based skill architecture, three loading levels, security model |
+| [[Source: Blake Crosley Agent Architecture Guide]] | engineering-blog | Complete harness pattern: hooks, skills, subagents, multi-agent orchestration, production results (Apr 2026) |
+| [[Source: Vercel Labs agent-browser]] | official-repo | agent-browser: 31.4K stars, Apache 2.0, Rust-native browser automation CLI for AI agents — snapshot + refs, annotated screenshots, structured diff. Replaces browser-harness for P30. |
+| [[Source: browser-harness CDP Harness]] | official-repo | browser-harness: 9.4K stars, MIT, thin CDP harness — self-healing LLM-to-Chrome bridge. SUPERSEDED by agent-browser for P30 (May 2026). |
+| [[Source: AgentBus Jinja2 Prompt Pipelines]] | engineering-blog | Jinja2 templating: inheritance, conditionals, loops, pipeline runner |
+| [[Source: TianPan Prompt Caching Architecture]] | engineering-blog | Multi-tier caching: semantic→prefix→full, 60-90% savings, cache boundary control |
+| [[Source: Arxiv — Don't Break the Cache]] | academic-paper | PwC evaluation: 41-80% cost reduction, system-prompt-only caching optimal |
+| [[sentrux-github-repo]] | github-repo | sentrux: 1.9k stars, 52 languages, MCP server, 5 metrics, MIT |
+| [[sentrux-dev-landing]] | website | sentrux.dev: landing page, cybernetic feedback loop, pricing |
+| [[sentrux-docs-quality-signal]] | documentation | Quality Signal formula, geometric mean justification, Nash theorem |
+| [[sentrux-docs-root-cause-metrics]] | documentation | 5 metrics mathematical definitions, graph-theoretic completeness |
+| [[sentrux-docs-rules-engine]] | documentation | .sentrux/rules.toml constraint system with layers and boundaries |
+| [[sentrux-docs-pro-architecture]] | documentation | Pro tier: runtime dylib, Ed25519 licenses, $15/mo, per-user watermarking |
+| [[ts-strict-mode-rishikc]] | article | TypeScript strict mode guide: 8+ sub-flags, migration strategy, ESLint pairing |
+| [[ts-runtimes-comparison-betterstack]] | article | Node.js vs Deno vs Bun: benchmarks, tooling, security, TS support (2026) |
+| [[barrel-files-tkdodo]] | article | Barrel files cause 68% module bloat, circular imports — stop using in app code |
+| [[ts-monorepo-koerselman]] | article | Monorepo patterns: built-package vs internal-packages, Turborepo, ESM, IDE setup |
+| [[vitest-official]] | official-docs | Vite-native test framework, Jest-compatible, v4.1.5 |
+| [[ts-folder-structure-mingyang]] | article | Production-grade Node.js/TS folder structure: Clean Architecture approach |
+| [[ts-best-practices-2025-devto]] | article | TS best practices 2025: type safety, advanced types, linting, tooling |
+| [[ts-result-error-handling-kkalamarski]] | article | Result monad pattern for declarative error handling in TypeScript |
+
+## Questions
+
+| Question | Answer Summary |
+|----------|---------------|
+| [[mvp-implementation-blueprint]] | **UPDATED (May 2026)**: Skill-First v2. 3 TS files + 12 skill files. 19 files total, ~8 weeks. |
+| [[how-to-enable-semantic-code-search-now]] | **IMPLEMENTED (L1+L2)**: ck installed, index built, skill created, grep tool intercepted via pi extension, file watcher auto-reindexes |
+
+## Flows
+
+| Flow | Summary |
+|------|---------|
+| [[harness-wiki-pipeline]] | Read-first/write-after data flow between harness and wiki |