Release v1.0.73 build 162

✨ Features

• feat: add MiniMax provider support
• Claude auth: accept Claude Code OAuth tokens alongside API keys
• Claude OAuth: prepend Claude Code identity system block
• Claude models: OAuth-aware fetch, add Opus 4.7 to defaults

🐛 Bug Fixes

• Fix edit_file diff syntax highlighting in activity log
• Fence edit_file diff log with language for syntax coloring
• Skip Apple triage in TaskExecution when Accessibility is off
• Skip Apple triage when Accessibility permission is not granted
• Claude auth: strip whitespace/newlines from pasted credentials

🚀 Improvements

• Remove per-tool-result and per-message caps — full output to LLM
• read_file: remove 50K char cap — return full file to LLM
• Remove hardcoded default model lists — fetch is the only truth
• index_create/recreate/append/continue: return full JSONL body in same turn

📚 Documentation

• revert: restore full README (undo 6250603 short version)
• docs: clarify Launch Agent is not privileged
• Docs: add CLAUDE_CODE_OAUTH.md — how to mint + paste a claude setup-token

🔖 Release

• Bump version to 1.0.72 (161)

🔀 Merges

• Merge pull request #6 from octo-patch/feature/add-minimax-provider
• Merge branch 'main' of https://github.com/macOS26/Agent

🔧 Recent Fixes

• Add Codex LLM provider (CodexService.swift) + provider setup wiring
• New docs: docs/CODEX.md, docs/CODEX_OAUTH_RESEARCH.md
• TaskExecution refinements: Setup, ToolBatch, Guards, ErrorHandler, FileTools
• AgentViewModel: StuckGuard, LLMServices, ScriptTabs, ModelFetching, SubAgent, Logging improvements
• Apple Intelligence mediator + Claude service updates
• UI: FallbackChainView, SettingsView, ToolsView, NewMainTabSheet, ThinkingIndicatorView polish
• LaunchAgent/LaunchDaemon plist tweaks

🚀 Agent! v1.0.71 (build 160)

Release date: 2026-04-17
Tag: v1.0.71.160

✨ Highlights

• Full output to the LLM — removed per-tool-result, per-message, and read_file 50K char caps. Index and file reads return the complete body in a single turn.
• Claude Code OAuth support — paste a Claude Code setup token as an alternative to an API key. OAuth-aware model fetch and Claude Code identity system block are prepended automatically on OAuth sessions.
• MiniMax provider — new LLM provider alongside Claude / OpenAI / Ollama / Apple Intelligence.
• Project index returns inline — index_create, recreate, append, and continue all return the full JSONL body in the same turn they wrote it.

🤖 LLM / Providers

• Claude: OAuth token flow (CLAUDE_CODE_OAUTH.md), whitespace-tolerant credential paste, Opus 4.7 added to the default list.
• OpenAI: stopped filtering fetched models — every model the API returns is shown. Legacy GPT-3.5 / dated GPT-4 snapshots no longer special-cased.
• Removed hardcoded default model lists — the fetch endpoint is the only source of truth.
• MiniMax provider merged (PR #6).

🔧 Tool Layer

• read_file: 50K char cap removed.
• Per-tool-result and per-message caps removed — LLM always sees the full tool output.
• index tool: create / recreate / append / continue return the written JSONL in the same response (no second read round-trip).

🧠 Memory / Project Layout

• New memory tool — Claude-compatible /memories filesystem with global and project scopes.
• Consolidated per-project hidden dirs under .agent/{index,memory,worktrees,plans}.
• .agent-index/ removed from tracking — project index output is not source.

🎨 UI / UX

• Long Cmd+V pastes are captured as removable chips instead of inlining into the TextField.
• Pasted-text chips show first + last character preview, stacked over three lines with trailing ellipsis.
• Multi-line log messages drop onto their own line below the timestamp for readability.
• Suggestion matching is skipped for long task input to prevent main-thread stalls.

🍎 Apple Intelligence

• Accessibility triage is skipped when a project folder is active (unless a system app is named), cutting latency for coding sessions.

📚 Docs

• CLAUDE_CODE_OAUTH.md — how to mint and paste a Claude Code setup token.
• README: new hero intro and refreshed GitHub metadata.
• Clarified that the Launch Agent is not privileged (privileged path is AgentHelper).

🏗️ Build

• Marketing version: 1.0.71
• Build number: 160
• Deployment target: macOS 26.0, Swift 6.2
• Development Team: 469UCUB275

📝 Commits since v1.0.69 (build 158)

ebc0c080 chore: bump version to 1.0.71 (build 160)
e8761fe6 index_create/recreate/append/continue: return full JSONL body in same turn
4cbacac3 Remove per-tool-result and per-message caps — full output to LLM
e683f9ce read_file: remove 50K char cap — return full file to LLM
54f5f1c2 Remove hardcoded default model lists — fetch is the only truth
54ffcafc Claude models: OAuth-aware fetch, add Opus 4.7 to defaults
62506030 docs: clarify Launch Agent is not privileged
83862d6a Claude OAuth: prepend Claude Code identity system block
aeb34b8c Claude auth: strip whitespace/newlines from pasted credentials
5237f164 Docs: add CLAUDE_CODE_OAUTH.md
6da55309 Claude auth: accept Claude Code OAuth tokens alongside API keys
ed107cff feat: add MiniMax provider support
ab807399 Bump version to 1.0.70 (build 159)
a3cb7be8 OpenAI model fetch: stop filtering — return every model the API lists
1a8e9802 OpenAI model filter: drop GPT-3.5/legacy GPT-4 families + dated snapshots
b3404e29 OpenAI model fetch: future-proof filter + blocklist non-chat models
f1c09596 Apple AI: skip accessibility triage when project folder is active

What's Changed

OpenAI Model List Improvements

• Stop filtering OpenAI models — now returns every model the API lists, no more hidden models (a3cb7be)
• Dropped legacy GPT-3.5 and dated GPT-4 snapshot filtering (1a8e980)
• Future-proofed model filter with blocklist for non-chat models (b3404e2)

Apple Intelligence

• Skip accessibility triage when a project folder is active unless a system app is explicitly named — reduces unnecessary UI queries (f1c0959)

Full Changelog: https://github.com/toddbruss/Agent/compare/v1.0.69.158...1.0.70.159

What's New in v1.0.69.158

🧠 Project Index & Memory

• Project Index tool — portable JSONL index per file (lines, doc block, top-level symbols)
• Memory tool — persistent /memories filesystem with global and project scopes
• Consolidated per-project hidden dirs under .agent/{index,memory,worktrees,plans}
• Prompt version now derived from marketing + build number (no more hardcoded revision)

✏️ Editing & Diff

• diff_apply fix — splice source into file instead of replacing entire content
• Fix diff_apply truncation and diff-formatted input handling
• Edit reliability: drop truncation guard, add cache invalidation + no-op detect
• Stuck-edit nudge: lower threshold to 3, make actionable

🍎 Apple Intelligence

• Apple AI: persist accessibility agent session across turns
• Apple AI: use Apple AI itself to classify follow-up prompts
• Apple AI: add run_agent tool for launching agents via FoundationModels
• Apple AI: disable shell and applescript tools — accessibility only
• Fix Apple AI misrouting non-agent tool names to run_agent
• Fix Apple Intelligence isEnabled defaulting to false

🎨 UI Polish

• Long Cmd+V pastes captured as removable chips with first+last char preview
• LLM output: stop last-word flickering with cursor
• LLM output: don't snap to bottom on mouse-exit when not streaming
• Tab labels: main tab shows current model, LLM tabs preserve raw model case
• ToolStepsView: persist expanded state across Cmd+B and new steps
• Multi-line log messages drop onto their own line below timestamp

🔧 Fixes & Cleanup

• Fix LLM tab provider/model clobber + per-tab usage breakdown
• Fix tool output formatting: newline prefix and :0:0 warning cleanup
• Replace force-unwrap with if-let in diff_and_apply
• Pin AgentTools 2.50.10 — adds PROJECT INDEX prompt section
• Efficient-action system prompt rules to curb redundant file re-reads
• Anti-confabulation: dedup tool names, prune aliases, harden read guard
• Summarize comments across all major Swift files
• Wrap long lines with SwiftFormat across the project

📦 Dependencies

• AgentTools 2.50.10
• AgentAccess 2.9.5 (screenshots no longer block main thread)

DMG to follow.

✨ 🦾 Agent! What's New in 1.0.68?

🔧 diff_apply Reliability

• Fixed truncation bug: source is now spliced into the file instead of replacing entire content
• Fixed LLMs sending diff-formatted input to diff_apply
• Replaced force-unwrap with safe if-let in range note parsing
• Updated AgentTools to 2.50.7 with truncation fix
• Bumped prompt revision to 86 for improved diff_apply prompting

🧠 Apple Intelligence Fixes

• Fixed non-agent tool names being misrouted to run_agent
• Tightened run_agent trigger to exact phrase match only
• Fixed duplicate log entries for accessibility tasks

🖥️ UI Improvements

• Paste chips: Long Cmd+V pastes now appear as removable preview chips instead of flooding the text field
• Tab labels: Main tab shows current model name; LLM tabs preserve raw model casing
• Per-tab usage: Added individual usage breakdown to each LLM tab
• Log formatting: Multi-line messages now drop below the timestamp on their own line

🧹 Code Quality

• Consolidated vision model detection into shared isVisionModel() (removed two duplicate private methods)
• Summarized verbose comments across 10 source files
• Skipped suggestion matching for long inputs to prevent main-thread stalls
• Fixed tool output formatting (newline prefix and :0:0 warning cleanup)
• Added iteration cap enforcement

What's Changed

🤖 Apple Intelligence

• Use Apple AI itself to classify follow-up prompts
• Persist accessibility agent session across turns
• Disable shell and applescript tools for Apple AI — accessibility only
• Better app name recognition — fixes Photo Booth confusion
• Teach Apple AI accessibility agent to discover button labels before clicking
• Verify Apple AI tool outputs before claiming success — forward to LLM on failure
• Log Apple AI tool actions to activity log — expose what Apple AI does
• Harden Apple AI triage — never claim completion on failure or uncertainty
• Fix Apple Intelligence isEnabled defaulting to false
• Fix: inject previous conversation turn into Apple Intelligence rephrase prompt

🛠️ Agent Scripts

• Fix agent script run routing: script tabs run on same tab, main tab spawns separate tab
• Fix AgentScript.swift: restore missing function declaration, compileCmd guards, and closing braces
• Route Apple Intelligence tool logs to correct script tab
• Pass project folder to Apple AI shell tool — fixes "not a git repository"

🔒 Security & Guardrails

• Add action_not_performed guardrail at app and prompt layers
• README + SECURITY.md: document action_not_performed three-layer defense
• README: add Tool Execution Gating to Defense Layers

⚙️ Internal

• Bump version to 1.0.67 (build 155)
• Version bump to 1.0.65 (build 153)
• Version bump to 1.0.64 (build 152)
• Fix bump_version/get_version to use selected project instead of Xcode open documents
• README updates and improvements
  Full Changelog: 1.0.61.149...1.0.67.155

What's Changed

Version Management

• Bump version fixes — bump_version/get_version now correctly uses the selected project instead of Xcode open documents
• Version bumps: 1.0.64 (152) → 1.0.65 (153)

Security & Safety

• action_not_performed guardrail — three-layer defense against false-action claims (app layer, prompt layer, documentation)
• README + SECURITY.md updated to document the action_not_performed defense layers

Apple Intelligence Hardening

• Verify Apple AI tool outputs before claiming success — forwards to LLM on failure
• Log Apple AI tool actions to activity log for transparency
• Harden Apple AI triage — never claim completion on failure or uncertainty

Documentation

• README: Tool Execution Gating added to Defense Layers section
• README: Tip Jar section added

Full Changelog: https://github.com/toddbruss/Agent/compare/1.0.61.149...1.0.65.153

🦾 Agent! Agentic AI for your  Mac Desktop (or Laptop)

Agent! Requires macOS 26.4 or later. M series preferred for Apple Intelligence. We don't know if Apple AI works on Intel. Also Apple AI is not supported inside a Mac VM. It's optional, but does help move things along by doing several tasks locally, saving you some tokens.

🔧 Bug Fixes

• Fix double echo — skip appendLog for shell commands that already streamed output
• Remove tool result cache — eliminate read loop blocker that caused stale responses
• Remove file read dedup cache — LLMs work around stub responses anyway; cleaner tool flow

🧠 Apple AI Improvements

• Skip accessibility triage for file-path and coding prompts — faster task routing
• Add ! bypass prefix — force Apple AI to skip triage when you know what you want
• Add AppleScript + shell tools alongside accessibility for Apple AI agent

🖥️ UI & UX

• Propagate parent tab project folder to spawned script tabs — scripts inherit the right working directory
• Add provider tooltip to LLM Usage popup model rows — see which provider serves each model

📖 Documentation

• Add Defense Layers section to README covering security features
• Add build-without-developer-account instructions + build.sh script
• Add emojis to toolbar icon table, remove SF Symbol text names
• Fix GLM model sizes — 744B/754B MoE, not 32B
• Clarify GLM-5/5.1 are cloud-only (may be possible to run FP8 locally); GLM-4.7-Turbo (32B) runs locally

⬆️ Upgrade Notes

• If upgrading from v1.0.60, no migration needed — settings and tabs are preserved.

🦾 Agent! Agentic AI for macOS 26.4+  Mac

🖥️ Agentic Ai for your Apple Mac Desktop or Laptop

🍎 Works with macOS 26.4+

⚡ Improvements

🔄 Fallback chain: trigger after 2 failures instead of 3 for faster recovery
🚦 429 handling: record every failure with fallback chain, flat 10s retry (3 attempts × 10s)
📝 429 logging: log actual API response body + add clearRetryAfter helper
🤖 Z.ai: proper dual-API URLs — code and vision endpoints registered separately
🧠 BigModel: same dual-API URLs as Z.ai (code + vision endpoints)
🍎 Apple AI triage: don't intercept shell commands (open, ls, git, etc.)
🖥️ FallbackChainView: show model display name (with -Code suffix for Z.ai)
🔀 Per-tab provider: sync Settings picker to active tab's LLMConfig
🔧 Model refresh: consolidate 3 model-refresh functions into 1 fetchModelsIfNeeded(for:force:)

🐛 Bug Fixes

🍎 Apple AI refusal: fix not falling through to cloud LLM
📦 batch_commands: remove per-step delimiters, run as single script
🔄 batch: fix duplicate output + guard empty git branch name
📱 manage_app: fix action overwrite + read app from input[app] too
✏️ edit_file: show line numbers in activity log status line

🏗️ Refactoring

📂 Organize AgentViewModel + Views into subfolders, strip redundant prefixes
✂️ Shorten multi-line comments across 94 files (-460 lines)
📏 Wrap long lines across project with SwiftFormat
⬆️ Bump to macOS 26.4+ with 5 new Apple FoundationModels features
🔨 Upgrade to Xcode version bump docs