v0.3.3

Changelog

[0.3.3] - 2026-09-20

Added

• Multi-File Atomic Edits: Extended the edit tool with files: { path, search, replace, line_hint, edits }[] payload (with resilient per-item path normalization in edits). Implemented a 2-stage atomic commit with preflight validation, duplicate path detection, in-memory snapshot backups, and automatic rollback across all targets if any file fails patch application or syntax verification.
• Batch Multi-File read: Extended the read tool to accept paths: string[] (and resilient alias path: string[]), allowing the model to inspect multiple files in one turn while authorizing all targets in EpistemicGuard.
• Automatic Turn 1 Reproduction Diff Grounding: In before_agent_start, automatically detects unstaged/staged git diffs and injects a sanitized, clamped (max 25 lines, 2KB hard ceiling) reproduction status into initial context with session caching, eliminating redundant Turn 1 exploratory git status / git diff commands.

Changed

• Read Tool Budgeting & Anchor Guardrails: Bounded batch reading to 150 lines per file and a 24KB aggregate ceiling with explicit continuation markers ([truncated: <file>, returned lines 1-150 of <N>. Continue with read(path, offset=151)]). Added a 100-line cap on explicit anchors: true requests (MAX_ANCHOR_LINES = 100) to prevent models from generating expensive #HASH│ line hashes on large files.
• Targeted Reading Workflow Guidance: Updated AGENT_KERNEL_SYS_PROMPT.md to emphasize targeted sequential reads (read(path, offset, limit)) centered on search coordinates as the primary inspection workflow, reserving batch paths for short files or known regions.
• Search Specificity Guidance: Directed agents toward exact identifiers, error symbols, and test phrases, preventing noisy broad single-word searches from polluting conversation context.

Validated

• Verified 38 test suites passing (npm test) with 0 regressions.
• Multi-harness benchmarks on cx/gpt-5.6-luna:high across all 8 standard tasks demonstrate:
  • 100% Solve Rate (8 / 8 tasks solved).
  • Aggregate task time cut in half: 600s (10.0m) vs 1,223s (20.4m) (-50.9% time reduction, 2.04× faster).
  • Uncached input tokens reduced by -18.4%: 270k vs 331k (-60.8k tokens saved across the suite).
  • Tool calls reduced: 107 calls vs 112 calls.
  • Consistent speedup across all repositories: UFO (36s vs 119s), Fastify (47s vs 177s), p-limit (60s vs 176s), Ky (89s vs 253s), Zod (135s vs 212s), Hono (64s vs 96s), UUID (53s vs 64s), and Picomatch (116s vs 126s).

[0.3.2] - 2026-09-16

Added

• Statusline Retrieval Indicator & Extension Row Integration: Relocated the retrieval profile tag from Line 1 to Line 2 (extension statusline via ctx.ui.setStatus("retrieval", ...)) alongside external extensions (such as Ponytail), preserving individual extension ANSI colors.
• Profile Semantic Glyphs: Added distinct visual tier glyphs: 🌿 retrieval:bm25 (AST lexical), ◈ retrieval:hybrid-256d (Matryoshka hybrid), and 🧠 retrieval:dense-768d (Full semantic).
• Live Embedding Throughput & Stream Arrow Status: Added real-time chunk processing throughput (chunk/s) with a static Stream Arrow indicator during workspace indexing (🧠 retrieval:dense-768d ⇢ 45% (22/48 • 14.2 chunk/s)).
• Automatic Kernel Guidance Injection: Bundled AGENT_KERNEL_SYS_PROMPT.md is injected through Pi's before_agent_start lifecycle hook, enabled by default and configurable with [instructions] enabled = false.
• Benchmark Comparison Chart: Added the latest aggregate harness comparison graphic at static/benchmark-comparison.png for display in the project README.
• Six-Harness Cross-Agent Benchmark: Integrated comprehensive multi-harness comparison runner supporting Pi Kernel, Pi Vanilla, Claude Code, Codex CLI, OMP, and OpenCode across 8 real-world tasks.
• Passive Shield Architecture: Gated exploratory tools (ast_search, get_repo_map, lsp) behind enable_tools: false / PI_ENABLE_ALL_RETRIEVAL_TOOLS=1, keeping the active tool interface lean (read, edit, write, bash, and compact code_search) to eliminate per-turn schema overhead and tool distraction loops.
• Deterministic Delimiter Auto-Healing: Added patch-scoped lexical delimiter balancing and Tree-sitter AST repair (MISSING node detection) to automatically recover truncated closing braces/brackets (}), };, ]}) pre-write.
• Line-Hinted Range Disambiguation & Multi-Block Line Deltas: Added optional line_hint, start_line, and end_line parameters (supporting both camelCase and snake_case) to edit with sequential delta tracking to disambiguate identical search blocks without line-hash bloat.
• Capped Plain-Text Reads: Enforced 50KB / 2,000-line caps on broad reads with continuation hints (offset) and oversized line detection, recording precise exposure ranges in EpistemicGuard.
• Model Property Aliasing & Epistemic Confirmation: Supported model convention drift across parameter names (search/replace, old/new, old_text/new_text) and emitted concise positive confirmation strings (Successfully applied edit to <file>.) to prevent paranoid agent verification loops.

Changed

• Included AGENT_KERNEL_SYS_PROMPT.md in the published package so automatic guidance works after installation.
• Added lifecycle coverage for guidance injection, duplicate prevention, and the disabled configuration path.
• Tool Description Optimization: Completely redesigned tool descriptions for read and edit to explicitly declare clean plain-text defaults and precision parameter expectations, dropping 10k+ prompt inflation tokens on real benchmarks.
• Diagnostic Benchmark: Cut token consumption by 50.6% (762k vs 1.54M total turn tokens) and reduced tool calls by 24.8% (112 vs 149) compared to unguided Pi Vanilla on the 8-task Luna High diagnostic suite.
• Historical Benchmark: In the initial v0.3.1-to-v0.4.0 Passive Shield evaluation, token consumption dropped from 1,214,112 tokens down to 465,139 tokens (-61.7%) across 8 real-world repositories.
• Persistent Vector Profiles: Preserved dimension-specific binary caches (vectors-256d.bin, vectors-768d.bin) on disk across profile switches, preventing cache loss when toggling to BM25.
• On-Demand Embedder Loading: Removed premature model preloading on startup; weights load on-demand only when unindexed chunks are detected.

Fixed

• Startup Background Indexing Visibility: Fixed a bug where workspace indexing on session_start ran silently in the background without updating the UI statusline on cold or unindexed workspaces.
• Companion Symbol Token Inflation: Deduplicated companion symbols and clamped surrounding context lines during targeted symbol reads to eliminate token bloat.

Removed

• Removed the obsolete AGENT_KERNEL_PROPMPTS.md prompt document.