SoulForge 1.1.1

[1.1.1] — 2026-03-29

Bug Fixes

• analytics: fix sales report bugs, add date filtering, add API routes
• run install steps synchronously in quiet mode (fixes Homebrew killing bg jobs)
• detect user shell for PATH config, support fish/ksh/bash/zsh
• add --quiet flag to install.sh for Homebrew, skip shell RC and animations
• show subcommands in /font help description
• wizard UI improvements and bundle script updates
• wire agent-managed tools, rename DEFERRED_TOOL_CATALOG, fix npm publish docs
• correct neovim linux arm64 asset name
• npm pack tar extraction on macOS, tolerate published versions
• use npm pack for cross-platform native deps, retry uploads
• patch dynamic platform import in bundled JS before compile
• cross-platform bundle — stub native libs for compiled binaries
• use available CI runners for all platforms
• npm registry, publish workflow, and build plugin fixes
• dead barrel detection for Python packages\n\n- isForbidden() returned truthy when uninitialized, causing collectFiles()\n to skip all files and break repo map indexing in tests\n- Dead barrel edge check now excludes sibling files within the same\n package directory (e.g. core.py → init.py is internal, not external)\n- Fallback ref check also excludes refs from files inside the barrel dir
• worker thread missing initForbidden — scan found 0 files
• wait-for-repomap UX + timeout increase
• sync symbol cache on IntelligenceClient for buildSymbolLookup
• shell guard no longer blocks code strings in node -e / python -c
• token-budget-only pruning for forge, z.preprocess coercion for numeric tool params
• enable pruning
• TabBar mode label reads stale registry on Ctrl+D cycle
• project format action uses dedicated formatter instead of lint+fix
• StatusDashboard bar alignment and popup width
• StatusDashboard data matches topbar, bar style improved
• remove misleading ^K hint from InputBox
• remove deprecated baseUrl, fix web-search agent callback types
• remove non-null assertions in intelligence router
• move Soul Map from message pair to system block, memoize LlmSelector rows
• add compaction to NESTED_KEYS, replace last scattered keepRecent default
• seed compaction defaults in DEFAULT_CONFIG
• centralize compaction defaults, LlmSelector fetch timeout
• await flushPromise on close, .aiignore hot reload, soul tool warnings, first-run hint, headless model fallback logging
• LSP race dedup, failedServer cooldown, shell env filtering, config hardening, UX guards
• reindexTimer cleanup, dynamic tool guidance, config dir hardening, agent-bus tests
• add action descriptions to soul_impact and subagent soul_analyze/soul_impact
• final claim sweep — correct all stale numbers across README and docs
• claim verification — correct all README/docs numbers, wire OpenRouter provider
• contested file edit stops agent, faster claim release, lock icon cleanup
• tool registration audit — remove invalid entries, add missing tools
• reduce dispatch token waste — scarier description, strip rejected attempts
• reject single-task dispatch, disable desloppify/verify by default, fix /changes per-tab
• secure auth system in little_backend
• close audit gaps — subagent shell claims, compound pre-checks, test coverage
• cross-tab coordination hardening — git blocking, cache staleness, agent sweep
• smoother text streaming, forge mode per-tab, edit tool cleanup
• tab bar lock icon formatting, add trailing newline
• deduplicate multi_edit result display, fix completed time not showing
• lint — replace non-null assertion with guard clause
• tab numbering in context popup, planning effort level, enriched API error messages
• prevent CI timeout on read-file outline test
• lint errors and add tabId to subagent explore tools type
• prevent steering message bar from wrapping to two lines
• scope tasks to owning tab, add attention indicator for pending input
• robust parent agent re-read blocking with full invalidation
• extract human-readable output from edit tool results in message history
• bump subagent step limits +3 to compensate for forced final step
• prevent NoObjectGeneratedError on subagent final step
• increase timeout for read-file outline tests to prevent CI flakiness
• add explicit return in useEffect to satisfy strict typecheck
• increase timeout for read-file outline tests to prevent CI flakiness
• show edit results inline in tool call displays
• async session saves, agent improvements, reasoning block & UI fixes
• report directory creation in edit_file output
• auto-create parent directories when edit_file creates new files
• relax circuit breaker threshold for dispatch agents
• wire checkAndClaim into buildTools edit_file/multi_edit (was missing)
• add leading slash to claims command registrations
• remove duplicate plan message injection in useChat
• downgrade zod for structured outputs consistency
• update paste handlers to use PasteEvent.bytes API
• deep re-export chains, FTS rebuild, pruning input mutation
• edit_file rich errors with lineStart, editor auto-open and file navigation
• two-pass ref resolution, Python normalization bug, buildEdges precision
• LLM semantic summaries UI consistency
• repo map live updates for neovim user saves and shell commands
• remove ghost recall tool refs, fix token ratio consistency, WSM cap
• salvage partial results from failed dispatch agents
• repo map import specifier extraction, refs priority, steering render bug
• improve dispatch agent collaboration — context-aware desloppify/verifier, file overlap warnings
• reduce flush interval 150ms→50ms, remove 100ms throttle — real-time UI updates
• remove startTransition from streaming flush — was deferring UI updates indefinitely causing frozen display
• tasks always resolve — complete on success, reset on error/abort, taskId for per-agent updates
• auto-complete in-progress tasks when agent finishes streaming
• Ctrl+X abort preserves partial chat content instead of clearing it\n\nSnapshot liveToolCallsBuffer and streamSegmentsBuffer before abort()\nclears them, so the catch block can reconstruct in-flight tool calls\nand partial assistant messages. Previously the buffers were empty by\nthe time the catch block ran, causing content to vanish on cancel.
• alphabetize /help and autocomplete commands, add 7 missing entries to help
• dispatch UI freeze — mark toolCallsDirty on agent stats + multi-agent events so streaming display updates during dispatch
• steering flush includes in-progress tool calls — shows progress before steering message
• StatusIcon shows warning for failed tool results
• dedupe installed skills by name, prefer project-scoped over global
• project tool reports lint warnings as failures — agent sees issues and can auto-fix
• SystemBanner useMemo exhaustive deps
• SystemBanner hooks after early return, rename icon to bannerIcon
• lint — move biome-ignore comments to correct lines
• wrap all raw numbers in String() for OpenTUI text compatibility
• wrap hiddenCount in String() — OpenTUI rejects raw numbers as text children
• nudge-aware tokenStop eliminates race condition where a single step could jump past both nudge threshold and stopWhen budget
• UI stability — stop scroll leaks, timer blink, reasoning duplication, picker cursor drift
• update step-utils tests for cache-aware pruning + done removal
• dispatch cache bugs + subagent fallback + forge pruning revert
• dispatch cache bugs + subagent fallback + forge pruning revert
• flush token display on finish-step, rename dispatch agents label
• async plugin bootstrap on first editor launch
• update Mason registry URL — moved from raw content to release artifacts
• prevent UI freeze during repo map indexing
• ContextBar content preserved across re-renders via ref
• git commit modal background, update /commit description
• complete modal stacking fix for toggleModal and openCommandPicker
• modal stacking, transient renders, keyboard early returns, scan throttle
• show dark red border on input during loading/compaction instead of invisible gray
• ContextBar token reset on modal open, repo-map picker live updates, scan progress labels
• health check readSymbol probe picks valid identifier name
• tree-sitter grammar for ALL typescript/tsx files — no tree-sitter-typescript.wasm exists
• tree-sitter tsx grammar lookup in findImports/findExports/getFileOutline + wider diagnose popup
• tree-sitter tsx grammar mismatch + smarter health check readSymbol probe
• improve steering message injection — drain all queued messages at once, stronger framing
• ASCII fallback icons for all providers, remove hardcoded Nerd Font glyph
• teach subagents to use startLine/endLine when task provides line ranges\n\nAdds WORKFLOW hint to explore and code agent prompts: when the dispatch\ntask includes line numbers, use read_file with startLine/endLine to\nbypass the 500-line truncation cap and get exact content.\n\nAlso removes fixresearch.md — all fixes verified as implemented.
• dispatch UI — late agent seeding, broken tree connectors, render storms
• strip contextManagement from subagent provider options
• steering race conditions — abort gate, ref sync, postAction queue drain
• add missing @openrouter/ai-sdk-provider dependency
• SQLite "database is locked" crash on concurrent repo map access
• SSRF protection hardening, rename-symbol comment awareness, shell path parsing
• shell timeout tests — fallback resolve after SIGKILL, bump test timeout
• CI test failures — git default branch, clone dirs, spawn timeout
• lint issues and add test step to CI

Documentation

• correct tool/theme/command counts, fix agent roles in diagrams
• add theme attribution credits
• audit and polish Mintlify documentation
• add Mintlify documentation site
• add Mintlify documentation site
• expand README comparison table with full intelligence stack
• full license compliance audit
• update third-party licenses for LazyVim migration
• add cross-tab coordination doc, fix commands-reference
• fix help popup — add missing commands, remove mislabeled entry
• fix wrong claims, update counts, add missing features
• update README with new headless CLI flags
• update README and repo-map docs for universal language support
• roadmap — SoulForge intelligence as library, MCP server, headless CLI
• command reference (60 commands), expanded security, docs index
• update contact email and website in license files
• fix Vercel Gateway and Proxy provider info + links
• fix ECC link, remove false Claude Code inspiration claim
• fix provider table — links, correct env vars, LLM Gateway description
• add inspirations section — Aider, Claude Code, ECC, AI SDK, Neovim
• honest comparison table against Claude Code, Copilot CLI, Aider
• sharpen differentiators, add hero screenshot, roadmap
• fix architecture diagrams, add hero screenshot, deep-dive links
• comprehensive README with logo, mermaid diagrams, feature deep-dives
• update readme and relevant docs
• slim README to highlights with deep dive links
• comprehensive documentation overhaul and architecture deep dives

Features

• little_backend: add search, enrich products, fix notifications
• add X-Source header to LLM Gateway requests
• floating terminals, 22 builtin themes, install script, edit-stack API fix
• terminals panel, worker bundling, dispatch role transitions, headless fixes
• add floating terminal, ghostty integration, command restructuring & theme fixes
• theme system, UI rebrand, docs refresh, and headless color updates
• add ASCII visualization guidance to plan & architect modes\n\n- Architect mode now instructs use of dependency graphs, comparison\n tables, box diagrams, and flow charts for design analysis\n- Plan mode (both full and light) presents visual file change summaries\n and dependency diagrams before calling the plan tool
• tab bar UX improvements and subagent fix\n\n- Hide model label in tab when it matches the default model\n- Restyle tab model labels with brackets and move after edit count\n- Fix explore subagent provider options not being stripped for mini-forge
• extended git tools, mini-forge dispatch display, tab UX improvements
• first-run wizard, UI polish, and toolchain improvements
• per-model cost breakdown with accurate multi-provider pricing
• /lsp now opens full management popup with disable/enable support
• max file cap for repo map — 10k file limit with git recency prioritization
• worker architecture — intelligence & IO workers, async FS migration, RPC framework
• cache-safe architecture, accurate cost tracking, tools management
• repo map toggle, shared tool schemas, zod 4 + biome bump
• distinct UI for code execution (node -e, bun -e, python -c, etc.)
• token optimization — two-layer pruning, subagent context management, slim dispatch output
• parallel tool display, multi-edit line tracking, cache breakpoints, scan animation + fixes
• editing model routing, tool-loop detection, scan animation, pruning toggle + bug fixes
• release infrastructure, changelog config, and skills lock update
• release infrastructure, install docs overhaul, OpenRouter support, and UX improvements
• redesign /changes sidebar with tree connectors and git status
• command palette, popup consolidation, and UX overhaul
• modular per-family prompt system, Soul Map as user message, streaming fixes
• dispatch overhaul, token optimization, LSP-first tiers, key priority
• agent editor access control, /export all diagnostic, text drip streaming, repo map always-on
• LazyVim editor, multi-lang LSP warmup, headless markdown rendering, UI unification
• integrate shiki + marked for syntax highlighting and markdown rendering
• semantic summaries, multi-lang LSP warmup, repo map UX overhaul
• model picker refresh, binary detection, tool & UI improvements
• cross-platform bundle, version sync, lint fixes
• LazyVim editor, async repo map, multi-provider proxy, tool grouping
• add project format action — explicit alternative to lint --fix
• cross-tab coordination hardening — shell guards, prompt guidance, dispatch gates, memory safety
• dispatch reliability — auto-split oversized tasks + complexity warnings
• mechanical re-read blocking for parent agent read_file
• export clipboard, per-tab expand state, reasoning context, input history fix
• subagent step limits, coordinator hardening, and test improvements
• post-edit formatter integration — authoritative indent fix
• WorkspaceCoordinator cross-tab file coordination (Tier 2 Soft Claims)
• add disablePruning option and "disabled" compaction strategy
• headless --chat multi-turn mode, session resume, SIGINT cleanup
• modular headless CLI, undo stack for all edit tools
• instruction files system, headless events/mode/timeout, InputBox history improvements
• custom providers, InputBox paste/history improvements, headless provider support
• read_file outline mode for large code files, InputBox history fixes
• merge read_code into read_file, soul_grep dep search, textarea input, headless CLI, mtime cache invalidation\n\n- Merge read_code tool into read_file via target/name params (delete read-code.ts)\n- Update all references across agents, prompts, intercepts, tool display (~15 files)\n- soul_grep: add dep param to search node_modules/vendor dirs with --no-ignore\n- InputBox: replace + InputEditor with native <textarea> from @opentui/core\n - Paste collapse (4+ lines → placeholder, ^E toggle)\n - History navigation with isNavigatingHistory guard\n - Proper visual line tracking for char-wrap height\n- Headless CLI mode: --headless, --list-providers, --list-models, --set-key\n- Read tracker: mtime-based cache invalidation (re-read if file changed on disk)\n- Docs: CLI flags in README/CLAUDE.md, headless.md documentation
• add call graph, fix ref resolution, filter non-code refs
• semantic summaries — merged AST+LLM mode, smart targeting, lazy regen, token tracking
• comprehensive unused_exports with dead files, barrels, clusters, test-only detection
• CommonJS exports, export * wildcards, Go modules, tsconfig paths
• source-resolved refs for precise unused export detection
• expose repo map data — top files, packages, symbol signatures, symbol-by-kind queries
• universal language support for repo map, dead code accuracy improvements
• ReadTracker, skill injection, dispatch returnFormat
• link tasks to dispatch agents via taskId — auto-updates on agent start/done/error
• add /keys command — manage LLM provider API keys from UI
• steering flush + LSP fixes + forge improvements
• project tool — raw mode skips preset flags, failure hints suggest raw: true for version issues
• legacy flag fallback for lint fix — auto-retries with older syntax on unknown-flag errors
• extend project lint fix support — oxlint, dart, swiftlint, hlint, gofmt, clippy allow-dirty
• fix subagent budget, skills loading, task UI, context drop, stop logging
• expand LSP server registry — 30+ servers, auto-discovered from PATH/Mason
• context-aware subagent limits — no step caps, proportional token thresholds
• remove done tool — output schema is the sole structured result mechanism
• Output schema for guaranteed structured subagent results
• cap 5 files per explore agent — auto-split large tasks for done reliability
• guaranteed done results — auto-synthesize DoneToolResult when agents exhaust steps
• question-driven tool routing + grep→navigate code hint
• prohibition-style prompts — FORBIDDEN enforcement + turnover discipline
• strip markdown formatting from system prompt — save tokens
• cache-aware pruning — skip tool result compaction when context is low
• token efficiency overhaul — dispatch contract, done-call fixes, system prompt split, outline filtering
• token optimization — forge-level pruning, escalating read nudges, richer summaries
• verification specialist, auto mode, dispatch quality, compaction UX
• dispatch validation gates + destructive action approval + compaction fix
• nerd font auto-detection, UI polish, chat export, bundle improvements
• project toolchain hardening, pre-commit checks, monorepo discovery, v2 compaction improvements
• co-author shell injection, LSP uninstall UI, biome lint fixes
• add uninstall for soulforge-installed LSP servers
• LSP backend — implement findImports, findExports, getFileOutline, readSymbol
• intelligence health check — /diagnose command probes all backends
• LSP installer with Mason registry, refactor name-based extraction, context bar improvements
• quickfix list, terminal output capture, editor event wiring
• neovim deep integration, pane splitting, git improvements, UI polish
• new editor tools, fix co-author email, update readiness doc
• new neovim editor tools + few subtle fixes
• extract editor layout module, improve editor panel UX and performance
• auto-install fd + lazygit, add licenses, extract UI components
• add concrete read_file hint in edit error output
• auto-enrich dispatch tasks with symbol line ranges from repo map
• user steering, abort cleanup, shell abort signals, task reset, breakage tests
• production hardening, outside-CWD security, open-source readiness
• tool consolidation, clone detection, system prompt scaling, UI cleanup
• UI refinements — blue user accent, message padding, queued message display
• title-only memory, boot spinner, splash polish, safety fixes, 1M context windows
• token optimization — slim subagent prompts, tighter pruning, progressive dispatch UI
• UI polish — bordered input, user msg backgrounds, collapsible errors/plans, history fixes
• ECC-enforced dispatch improvements, scoped model selection, UI rendering fixes
• llmgateway provider, site link extraction, shell read-redirect, plan mode polish
• unified web access approval gate for fetch_page + EventTarget listener fix
• rolling tool result pruning with repo-map symbol enrichment
• isolated tabs, smooth streaming, borderless input, icon centralization
• responsive UI, popup overlays, /lsp command, boot granularity, web search fixes, dispatch thresholds
• v2 compaction, git branch/stash ops, SSRF protection, agent bus hardening, and comprehensive test suite
• context compaction, plan view overhaul, persistent system messages, and broad refinements
• repo map intelligence, compound tools, web scraper, and Ink → OpenTUI migration
• multi-tab chat, parallel agent dispatch, and provider config system

Miscellaneous

• add SHA256SUMS.txt checksum generation to release workflow
• untrack .agents/skills, remove stale homebrew/ copy
• wrap tab bar indicators in brackets\n\n- Edited file count, unread dot, and error markers now use bracket styling\n- Consistent visual language across all tab bar indicators
• upgrade deps
• remove completed/obsolete improvement docs
• add HTML coverage report script
• add test coverage reporting + lcov artifact upload
• biome formatting — normalize imports, indentation, line wrapping
• fix lint formatting across refactored files
• fix lint errors from dead code removal (trailing blank lines, let→const)
• delete PlanView.tsx (dead file), remove dead exports from splash.ts and types/index.ts
• remove 43 dead exports, delete highlight.ts (fully dead file)
• remove 10 dead files (780 lines)
• fix biome formatting across 5 files
• remove JetBrains Mono from bundle, keep Symbols Only
• biome format fixes

Performance

• audit fixes — granular modal selectors, store selectors, concurrency guard, listener cleanup
• audit fixes — smoothStream factory, unmount cleanup, abort-aware retry, error handling, memo & memoization
• comprehensive React performance audit — 21 fixes across 25+ files

Refactor

• prompt updates, async DiffView, compaction return type, cost breakdown improvements
• UI polish, context bar simplification, repo map token budget, worker memory tracking, lint fixes & test updates
• collapse re-export symbols in repo map render
• remove shell search redirect gate\n\nRemove checkSearchAntiPattern and the blocking redirect layer that\nprevented shell from running grep/cat/find commands. This gate was\noverly aggressive and incorrectly blocked legitimate git and shell\noperations. The softer post-success hints in shell.ts are retained.
• clean up stream options, subagent tools, tab instance, and context manager\n\nCo-Authored-By: SoulForge soulforge@proxysoul.com
• fix dumb tests, reduce plan eagerness, clean up truncation messages
• production-grade codebase cleanup — remove noise, deduplicate, tighten exports
• production-grade codebase restructure
• production-grade codebase restructure
• remove ReadTracker (re-read prevention at tool execution time)
• ReadTracker replaces RecallStore, remove dead code, cleanup agent wiring
• centralize language detection — single EXT_TO_LANGUAGE map in types.ts

Testing

• comprehensive tests for format detection across 18 ecosystems
• comprehensive tests for WorkspaceCoordinator and tool-wrapper
• comprehensive edge cases for custom providers
• comprehensive edge case tests for unused export detection
• update .vue extension test — now correctly detected as 'vue' via centralized map

Grep

• add output truncation, max-columns, and source map exclusions

License

• fix AGPL references across codebase to BSL 1.1
• switch from AGPL-3.0 to Business Source License 1.1

Merge

• worker architecture (phases 0-3) into main

Reliability

• tab-scoped tasks, cache hardening, agent retry, prompt tightening

Tee

• cap individual file size at 512KB

Ui

• add Forge label to assistant message header

V1.0.0

• bundled distribution, investigate agents, dispatch UI overhaul

Wip

• pre-cost-breakdown checkpoint + misc improvements