You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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.
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.
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.