pawan-core 0.5.16

Pawan (पवन) — Core library: agent, tools, config, healing
Documentation

pawan-core v0.5.16

Core library for the Pawan CLI coding agent. Contains the agent engine, tool system, configuration, and healing/recovery logic.

Unreleased

What's New in v0.5.16

  • Version alignment — published with the workspace release; RMUX session-list card rendering lives in the pawan CLI crate.

What's New in v0.5.15

  • Version alignment — published with the workspace release; RMUX pane-list card rendering lives in the pawan CLI crate.

What's New in v0.5.14

  • RMUX pane discovery — rmux tool now exposes list_panes with optional session/title/command/cwd/running filters for active-pane inventory.

What's New in v0.5.13

  • RMUX session discovery — rmux tool now exposes list_sessions for active-session inventory before pane operations.

What's New in v0.5.12

  • Version alignment — published with the workspace release; RMUX snapshot-card rendering lives in the pawan CLI crate.

What's New in v0.5.11

  • Live RMUX verification — ignored integration test covers ensure_session → wait_for_text → snapshot → kill_session when PAWAN_RMUX_LIVE=1 and the rmux binary are available.
  • RMUX cleanup — kill_session action and /rmux kill <session> prompt routing support explicit teardown.
  • RMUX validation — missing sessions and partial terminal sizes are rejected before daemon startup; connection errors mention installation/PATH/daemon checks.

Live RMUX test

PAWAN_RMUX_LIVE=1 cargo test -p pawan-core --test rmux_live -- --ignored

The test starts or connects to RMUX, creates a short-lived session, waits for a marker, snapshots visible pane text, then kills the session.

What's New in v0.5.10

  • RMUX tool — Standard tool backed by rmux-sdk for durable terminal sessions, pane input, wait-for-text synchronization, and pane snapshots.
  • Tool visibility — rmux ships as a Standard tool so coordinator/default tool definitions expose terminal sessions without extra activation.

What's New in v0.5.9

  • stream_options.include_usage — OpenAI-compatible streaming requests now request final-usage chunks, fixing the token/ctx widget for providers (vLLM, SGLang) that omit streamed usage by default
  • Module changes — tui/render.rs split into render/{mod, messages, overlays}.rs; new tui/effects.rs for motion + value animation

What's New in v0.5.8

  • CRAP score reduction — decomposed 10 high-complexity functions across agent, TUI, and CLI modules; extracted ~40 focused helpers
  • Render decomposition — split render.rs (4287 LOC) into render/{mod.rs, messages.rs, overlays.rs}
  • Test coverage expansion — 60 new tests for 0%-coverage functions; 1779 total workspace tests across 18 suites; 61.09% line coverage baseline

What's New in v0.5.0

  • Session store — SQLite in WAL mode with FTS5 and JSON migration; JSONL branching with parent_id (depth capped at 5); session labels and bookmarks
  • Agent pool — concurrent agents with semaphore bounding; agent definitions with YAML frontmatter; 6 agent types, 300s timeout
  • Parallel tool execution — bounded concurrency (max_parallel_tools); batch tool (25 concurrent calls)
  • Bash permission tiers — tree-sitter based, feature-gated with main/sub/lua audience bitflags
  • Doom-loop detection — configurable backoff multiplier; retry policy with exponential backoff + jitter
  • Auto-compaction — LLM summarization via /compact; strategies: default (10 msgs), aggressive (5 msgs), conservative (20 msgs)
  • Memory system — consolidation (Jaccard similarity), retrieval, prompt injection scanner (6 patterns), SessionScopedMemory fencing
  • Tool registry overhaul — Tool::execute now async; on_pre_compress hook for context pre-processing; sync_turn on return
  • Module splits — coordinator/types.rs extracted; tools/git.rs → tools/git/ (5 files); tools/native.rs split into native_search, mise, lsp_tool
  • Full core test suite passing before release

Features

  • Agent Engine — Multi-turn conversation with LLM backends (NVIDIA NIM, Ollama, OpenAI, MLX)
  • Tool System — 37 tools in 3 tiers (Core/Standard/Extended) with tiered visibility, including RMUX-backed terminal panes
  • Configuration — Pluggable config resolution via aegis or local files
  • Healing — Auto-recovery from tool failures, context overflow, model errors
  • Thinking Budget — Per-model thinking mode dispatch (Qwen, Gemma, Mistral, DeepSeek)
  • Session Management — Save, load, and resume conversation sessions
  • Model Selection — Dynamic model switching with search and filtering

Usage

use pawan::{PawanAgent, PawanConfig};
use std::path::PathBuf;

let config = PawanConfig::load(None)?;
let _agent = PawanAgent::new(config, PathBuf::from("."));

This crate is the foundation — use pawan (the CLI binary) for the full experience.

License

MIT