All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

• Structured plan-then-execute reasoning for jam run: the agent now generates a typed ExecutionPlan (ordered steps with success criteria) before acting, replacing the free-text ReAct loop
• Read-before-write gate: write tools are automatically blocked until the target file has been read, preventing silent overwrites of unread files
• Post-write shrinkage guard: if a write_file call produces a file suspiciously smaller than the original, the file is auto-restored from git and the model is redirected
• --yes flag on jam run for non-interactive auto-approval of all write operations
• StepVerifier to validate each plan step before execution
• Working memory + tool-result caching for the agent loop
• Critic evaluation and correction pass after the tool loop completes
• Past-session search and symbol index builder for richer context injection

• --provider CLI flag no longer inherits baseUrl from the active profile when switching providers (e.g. --provider openai no longer accidentally hits localhost:11434)
• Removed unnecessary type assertions in run.ts
• Removed unnecessary escape characters in agent.ts

• Embedded provider (--provider embedded): run SmolLM2-1.7B fully in-process via node-llama-cpp — no external server needed
• Default embedded model upgraded to smollm2-1.7b-instruct-q4_k_m (1.7B, q4_k_m) with 8192-token context window
• One-time model download from GitHub releases with progress reporting
• jam commit --provider embedded — commit message generation works offline with diff-stat fallback for large diffs
• supportsTools / contextWindow fields on ProviderInfo for capability-aware routing
• Lean system prompt path for small models that cannot follow tool-call JSON schemas

• Lint errors in embedded provider download stream handler (Unsafe array destructuring / Unsafe member access)

• Initial release of Jam CLI
• jam ask — one-shot AI questions with streaming output
• jam chat — interactive multi-turn chat REPL (Ink/React TUI)
• jam explain — AI-powered code explanation
• jam search — codebase search with ripgrep (JS fallback)
• jam diff — git diff review with AI analysis
• jam patch — AI-generated unified diffs with validation and apply
• jam run — agentic task workflow with tool-calling loop
• jam auth — provider authentication management
• jam config — configuration management (init, show)
• jam models list — list available models from provider
• jam history — chat session history (list, show)
• jam completion install — shell completion for bash/zsh
• jam doctor — system diagnostics and health checks
• Ollama provider with NDJSON streaming
• Pluggable provider architecture (adapter pattern)
• Layered configuration (global → repo → CLI flags)
• Named profiles for multiple provider/model configs
• Secure secrets via OS keychain (keytar) with env var fallback
• Model-callable tools: read_file, list_dir, search_text, git_status, git_diff, write_file, apply_patch
• Tool permission enforcement (ask_every_time, allowlist, never)
• Chat session persistence (JSON files)
• Log redaction for sensitive patterns
• Markdown rendering in terminal (marked + marked-terminal)

• Initial project setup
• Core CLI framework with Commander.js
• Ollama integration
• Basic tool system
• Configuration with Zod schema validation