inject current file contents + write instruction when user asks to edit

add project file names to prompt; handle <css> and <javascript> tags

use XML tags (<read>/<write>) instead of markdown code fences for file ops

minimal prompt; trim history to 20 entries

clearer system prompt with exact format examples

strip system prompt to bare minimum for TPM limit

match read/write inside any code block language tag

inject project file tree into prompt; more lenient read/write regex

switch default to llama-3.1-8b-instant

switch to gemma2-9b-it; shorten system prompt to reduce token usage

increase iter limit; fix prompt to push model to act; don't log intermediate file reads to UI

replace function calling with text-parsed read/write code fence blocks

remove function calling (unsupported by API proxy), use plain chat only

switch default to llama-3.3-70b-versatile for better function calling support

default to llama-3.1-8b-instant; trim system prompt and tool descriptions to reduce token overhead

add built-in read_file/write_file/list_files tools; re-add tool_choice to API calls

use INFERENCE_API_URL as-is, no endpoint construction

fix endpoint construction when INFERENCE_API_URL already includes /gen/chat/completions

add 503 fallback: retry without tools then fall back to Gen endpoint

fix default model to 'lightning' for InferencePort API compatibility

strip Dockerfile to bare minimum for faster free-tier builds

update docs: USE_PERSISTENT_STORAGE env var, fix README

fix: guard /data access behind USE_PERSISTENT_STORAGE env var to prevent startup hang on slow mounts

add startup diagnostics, explicit 0.0.0.0 bind, uncaught error handlers

add model dropdown and fix stuck thinking indicator

Initial commit - Forge: self-improving AI assistant with filesystem access for self-editing