OpenAgent v4.1.4 — Bundle version metadata fix

OpenAgent v4.1.4

This patch release fixes stale CLI version metadata in the published bundle.

Fixed

• Rebuilt core generated version metadata before bundling.
• Rebuilt the npm bundle from a clean output directory.
• Verified that the bundled executable and telemetry metadata report 4.1.4.

Installation

npm install -g @haseeb_heaven/open-agent@4.1.4

OpenAgent v4.1.3 — Fast provider failover

OpenAgent v4.1.3

This release improves response latency for explicit model selections and makes
free-provider failover bounded and observable.

Performance

• Explicit Gemini model selections bypass the auto-model classifier on fresh
  turns, while model changes during an active sequence retain re-routing.
• OPENAGENT_CLI_FAST_MODE=1 caps compatible-provider completions at 1024
  tokens, skips the optional next-speaker follow-up, bounds provider requests to
  3 seconds, and limits fast failover to one alternate candidate.
• Model registry and derived free catalog setup are cached across turns.
• Live model matrices support bounded concurrency and report wall-clock, p50,
  p95, target misses, and per-model outcomes.

Reliability

• Fast-mode deadline failures enter the existing free-provider routing failure
  path instead of waiting through the normal retry budget.
• Normal mode retains existing retry, fallback, and model-quality behavior.

Validation

• Provider/client regression suite: 175 tests passed.
• Full core run: 8,108 tests passed; one existing workspace-policy test timed out
  in the local environment.
• A2A server suite: 140 tests passed.
• Real filesystem benchmarks were run across Google, Cerebras, Groq, OpenRouter,
  and Hugging Face providers.

Installation

npm install -g @haseeb_heaven/open-agent@4.1.3

v4.1.1 — Free-model picker, extensions market & post-release fixes

Ships the develop-branch feature set previously (mis-)tagged as an unreleased v4.1.0, renumbered to avoid colliding with main's already-released v4.1.0 (provider crash / test-isolation fixes — see that release).

Features (from the develop v4.1.0 feature set)

• feat(cli): /free-models interactive picker for free-tier models, plus /free shortcut
• feat(cli): /extensions market — browse and install extensions from the registry

Bugfixes (this release)

• fix(tui): /free-models key-entry dialog now rejects structurally-invalid input (leftover slash-command keystrokes, strings under 16 chars) before writing to .env, instead of silently persisting garbage as an API key
• fix(providers): resolveRegistryPath now walks upward from cwd for configs/models.toml before falling back to a bundle-relative search, so running from a repo subdirectory no longer risks pinning a stale bundle/configs/models.toml copy for the process lifetime
• fix(test): resolve a hanging/flaky kitty-protocol test suite in gemini.test.tsx
• fix(test): clear a stray force-exit timer on test teardown that was surfacing as unhandled process.exit exceptions in unrelated test runs
• fix(scripts): live-test harness (run-live-tests.mjs) retries once on timeout before grading a scenario FAIL, and flags scenarios whose prompt references a missing media-dir subfolder

Full details in CHANGELOG.md.

v4.1.0 — Provider crash & test-isolation fixes

Rolls up recent fixes: provider crash on zero-API-key first-run auth, and elimination of a homedir-mock test-isolation bypass. Full packages/cli suite re-verified green (468 files, 6952 tests) before release.

• fix(tui): rename the interactive model picker command to /models (with /model retained as a compatibility alias); unavailable paid-provider models now offer their provider-specific .env key setup directly in the dialog
• fix(providers): preserve unique registry-key routing and endpoint overrides when aliases share a LiteLLM model id; free sessions now rotate through the fallback catalog after rate-limit or free-router failures
• fix(providers): stop crashing on first-run auth when zero API keys are configured
• fix(tests): eliminate a homedir-mock bypass in the extension manager and related suites that let tests leak real developer-machine state (~/.gemini / ~/.openagent) instead of the mocked temp home
• test(providers): add fallback-chain coverage and validate OpenRouter's openai/gpt-oss-20b:free with both complete and streaming live requests
• docs: update the branch-specific clone command; document /models, --resume, and --yolo

Note: this release is from the main branch. A separate, unrelated v4.1.0 was shipped on develop (free-model picker + extensions marketplace) — see v4.1.1 for that lineage, renumbered to avoid a version collision.

Interpreter v3.6.0 — Stability Fixes & Live Testing Expansion

Highlights

Stability fixes

• Key-exhaustion no longer crashes the classic REPL; structured AllKeysExhaustedError with provider/retry_after_eta attributes
• Model router attempts one free-model fallback before surfacing key exhaustion
• Persistent banner/status line never wraps or truncates on narrow terminals
• Fixed Windows rich.Console OSError: Bad file descriptor crash (forced legacy_windows=False)
• Root-caused and fixed a test-suite stdout-fd corruption bug (tests/interactive/helpers.py's mock interpreter)
• litellm.completion() calls now bounded by a 90s timeout to stop indefinite hangs
• Fixed web-search DDGS import order, a chart prompt-injection contradiction, and live-flake soft-skip scoping

Live testing expansion

• 10 new create/analyze/summarize/convert/edit scenario cases covering zip, mp3, java, sqlite, docx, svg, webm (stdlib-only, no external codec/toolchain dependency in CI)
• Model-router multi-key rotation regression test
• Filled remaining all-modes e2e smoke gaps (yolo, gemini-style, and other previously-uncovered modes)
• Raised combined libs coverage to 81% (≥80% CI gate)

New provider

• Added Cerebras (cloud.cerebras.ai) as a fully wired free/rate-limited model provider

TUI improvements

• No-args wizard answers now persist to ~/.code-interpreter/config.json; --config forces a re-run
• "Configure advanced options?" → no now genuinely skips every subsequent advanced prompt
• Wizard cancellation (Ctrl+C/Esc) exits cleanly instead of an unhandled traceback

See CHANGELOG.md for the full list.

Interpreter v3.4.0 — MCP, Sessions, Data/Science, Sandbox & CI

Interpreter 3.4.0

Major feature wave on top of v3.3.0 agentic/free-LLM foundation.

Highlights

• MCP + autonomy: native FS/shell ToolRegistry, --yolo, --mcp-server (#215)
• Streaming + vision: --stream / --no-stream, --image / /image (#216)
• Web search: --search / /search (DuckDuckGo / Tavily / Serper) (#217)
• Codegen modes: --mode generate / --mode project without execution (#212)
• Structured output: --output-format json|markdown|plain (#219)
• Persistent sessions: --session / /session (#218)
• Identity & onboarding: free/local positioning + first-run UX (#220)
• Local-first: --attach / /file, --ollama, --local (#221)
• Data analysis: EDA, charts, SQL, exports (#222)
• Science: notebooks, themes, ML helpers, PDF reports (#223)
• CI/coverage: matrix CI, coverage gate, Codecov (#224)
• Sandbox hardening: subprocess/Docker backends, --timeout, --safety, audit, secret scan (#225)
• Interactive tests: slash/REPL/session/live-exec coverage (#226)

Quick start

python interpreter.py --list-free
python interpreter.py --local --attach data.csv "summarize this"
python interpreter.py --cli --yes --output-format json -m local-model -f task.txt
python interpreter.py --sandbox docker --timeout 60 "analyze sales.csv"

Full history: CHANGELOG.md

Interpreter v3.3.0 — Agentic Free LLMs & Resilience

Interpreter 3.3.0

Foundation release for agentic free-LLM UX and production resilience.

Highlights

• Gemini-CLI-style --gemini-style ReAct REPL with free/cheap catalog (--free, --list-free, /free)
• Multi-key rotation, rate limiter, circuit breaker
• Metrics CLI (/key-status, /reload-keys, /metrics)
• Non-interactive --yes / INTERPRETER_YES for CI
• Multi-agent --agent and ReAct --agentic loops

See v3.4.0 for MCP, streaming, sessions, data/science, sandbox levels, and CI coverage that landed after this tag.

Full history: CHANGELOG.md

v3.2.3

Interpreter 3.2.3 Latest

@haseeb-heaven haseeb-heaven released this Jul 11, 2026

3.2.3

---

🔥 Release highlights:

• Resolved Command Injection vulnerability in file opening by replacing subprocess.call with os.startfile on Windows.
• Mitigated Path Traversal vulnerability in UtilityManager.get_full_file_path with strict boundary checks.
• Enhanced stability with timeouts on all external HTTP requests to prevent application hanging.
• Boosted performance by pre-compiling all regex patterns in ExecutionSafetyManager.
• Improved Terminal UI accessibility, ensuring prompt choices fallbacks are explicitly visible in non-TTY environments.
• Fixed Ollama "NoneType" Error, allowing robust extraction of direct string responses and dictionary outputs for local models like Mistral.
• Fixed Ollama API Key Error, intentionally bypassing HUGGINGFACE_API_KEY requirements for local and Ollama models.
• Updated legacy model configurations to point to the modern 2026 stable aliases (e.g. gpt-4.1, claude-sonnet-4-6).
• Expanded unit test coverage to 263 tests, directly validating security fixes, UX enhancements, and API Key robustness logic.

---

📜 Changelog:

• v3.2.3 - Fixed Windows command injection, resolved path traversal in utility manager, implemented HTTP timeouts, optimized SafetyManager by pre-compiling regexes, improved non-TTY fallback prompts, fixed Ollama/local model API Key extraction and output parsing, updated legacy model configurations, and expanded test suite with full coverage.
• v3.2.2 - Added sandbox mode (default ON) with /sandbox and /unsafe toggles, replaced --unsafe with --sandbox / --no-sandbox, improved subprocess security delegation, increased SAFE timeout to 300s, fixed watchdog timer issues, strengthened safe-mode protection, added process-group kill on timeout, improved Python detection using AST parsing, fixed multiple security vulnerabilities (P0/P1/P2).

---

📦 Assets:

• interpreter.zip
• Source code (zip)
• Source code (tar.gz)

---

v3.2.2 Sandbox Security and Code Interpreter Architecture

Full Changelog: v3.2.2...v3.3.0

VERSION v3.3.0 (Summary of Key Changes)

This update mainly improves security, stability, and how code runs in a controlled environment.

MAIN FEATURE: SANDBOX MODE

• A new sandbox system is now the default way code runs
• You can control it using:
  • /sandbox → safe execution
  • /unsafe → less restricted execution
• The old --unsafe flag is replaced with --sandbox / --no-sandbox
• Sandbox now has stronger protection against harmful commands and file access

SECURITY IMPROVEMENTS

• Better blocking of dangerous system-level commands
• Fixed ways users could bypass safety (file writes, absolute paths, symlinks)
• Added stricter checks for file operations like .write()
• Improved detection of unsafe patterns across Python and other scripts
• Safer subprocess handling and execution delegation

EXECUTION & PERFORMANCE

• Increased timeout limit in safe mode to 300 seconds for long-running code
• Improved handling of stuck or long processes using process group kill
• Better detection of Python code using AST parsing
• Fixed watchdog timer issues in sandbox

STABILITY & BUG FIXES

• Fixed syntax errors and test failures
• Cleaned up execution logic formatting
• Fixed temporary file execution issues
• Resolved multiple high-priority (P0/P1/P2) security bugs

BUILD & TOOLING

• Improved build_release.sh script with:
  • better error handling
  • cleaner structure
  • more reliable release process

Overall, this release focuses on making the interpreter safer, more reliable, and better at handling real-world code execution, especially with the new sandbox-first approach.

v3.2.1

What's Changed

• Enhance model catalog, execution safety, and TUI features by @haseeb-heaven in #24
• Update model configs to newer LLM releases and fix routing issues by @haseeb-heaven in #23

Full Changelog: v3.1.0...v3.2.1