First release as an oh-my-openagent

This is a first release as an oh-my-openagent. We have changed our name to make it less confused; OmO and Sisyphus is about the whole architecture- not just a plugin. Just to help opencode community spending whole day answering OmO questions, we are changing our project name. We know that the name is ultralong, so just call our plugin as 'oh-mo' or Sisyphus.

The GPT Era Begins — Sisyphus, Multimodal-Looker, and the Full Agent Catalog Now GPT-Native

Dear Ultraworkers,
Finally we have GPTPhus - he got sprits of Sisyphus, but with the powers of Hephaestus.

This release marks a fundamental shift in our model strategy. After extensive real-world testing—including burning through two ChatGPT Pro accounts worth of tokens—we've systematically elevated GPT-5.4 to first-class citizenship across the entire agent ecosystem. This isn't a simple model swap; it's a recognition that GPT-5.4 has earned its place as a serious alternative to Claude for complex orchestration work.

Sisyphus Goes GPT-Native (Finally)

For the first time ever, Sisyphus is properly tuned for GPT-series models. We've crafted dedicated GPT-5.4 prompts with an 8-block architecture that rivals our Opus-native prompts in sophistication. The results are remarkable: GPTPhus (GPT-powered Sisyphus) delivers orchestration quality that genuinely impressed us during testing—so much so that we burned through two ChatGPT Pro accounts validating it.

The Verdict: Opus Sisyphus remains the gold standard for general-purpose orchestration. Claude Opus 4.6's unique architectural strengths are unmatched for certain reasoning patterns. But GPTPhus is more than "good enough"—it's genuinely excellent, and for users with OpenAI-only setups, it's now a first-class option rather than a fallback.

New Fallback Chain: claude-opus-4-6 max → k2p5 → kimi-k2.5 → gpt-5.4 medium → glm-5 → big-pickle

Complete Agent Catalog Migration: GPT 5.2 → GPT 5.4

We've completed the migration of our entire agent catalog from GPT-5.2 dependencies to GPT-5.4. Every agent that previously relied on GPT capabilities now routes to the 5.4 family:

Important Distinction: GPT-5.3-codex and GPT-5.4 are fundamentally different models with different strengths. While we've migrated most agents to 5.4, Hephaestus continues to recommend GPT-5.3-codex for deep autonomous work—its specific training for extended coding sessions remains unmatched for that agent's particular workflow.

ULW-Loop: Oracle Verification Becomes Mandatory

The ultrawork loop (/ulw-loop) now requires explicit Oracle verification before completing any task. This is a significant behavioral change:

• Before: ULW loop would attempt self-verification and potentially declare completion
• After: Oracle agent must explicitly verify completion criteria are met
• Impact: Higher confidence in ULW-loop results, slightly longer tail-end execution

We've also added explicit Oracle session tracking to prevent the "interrupted verification" edge cases some users reported. The system now maintains full context of any ongoing Oracle verification and handles parent session retry logic gracefully when verification fails.

Category System: GPT-5.4 Defaults

Our category-to-model mapping now defaults to GPT-5.4 for high-effort tasks:

Note that ultrabrain and deep categories retain GPT-5.3-codex as their primary—their specific training for extended reasoning sessions remains optimal for those task types.

Model Cleanup: kimi-k2.5-free Removal

We've removed kimi-k2.5-free from all fallback chains. This model variant was deprecated by the provider and was causing unnecessary fallback delays.

Multimodal-Looker: Now GPT-5.4 Native

The multimodal-looker agent—responsible for image analysis and visual understanding—now uses GPT-5.4 medium as its primary model (was GPT-5.3-codex). This aligns visual analysis capabilities with our latest GPT-native prompts.

Full Changelog

feat

• sisyphus: Add GPT-5.4 variant with 8-block prompt architecture; extend fallback chain with k2p5, kimi-k2.5, gpt-5.4 medium
• multimodal-looker: Upgrade primary model from gpt-5.3-codex to gpt-5.4 medium
• oracle: Add GPT-5.4 variant prompt with model-based routing
• momus: Add GPT-5.4 variant prompt with model-based routing and QA scenario executability checks
• metis: Add QA scenario executability checks alongside analysis capabilities
• sisyphus-junior: Add GPT-5.4 and GPT-5.3-codex model-specific prompt routing
• hephaestus: Add generic GPT prompt fallback with model-specific routing
• ulw-loop: Require Oracle verification before completion; add explicit Oracle session tracking
• ultrawork: Enforce manual QA execution and acceptance criteria workflow
• delegate-task: Add model field to TUI metadata; make sync subagent timeout configurable; use explicit high variant for unspecified-high category
• atlas: Add Final Verification Wave to orchestration workflow; enforce checkbox completion before next task
• gpt-5.4: Amplify parallel tool-calling with XML behavioral contracts
• start-work: Add auto_commit config option
• look-at: Add automatic image format conversion for HEIC/RAW/PSD files
• dispatch: Wire marketplace plugin commands into slash command dispatch

fix

• atlas: Skip compaction in last-agent recovery; add full eligibility checks to delayed retry callback
• ulw-loop: Retry parent session after failed verification
• plugin: Ignore compaction session agent updates
• preemptive-compaction: Allow re-compaction after context grows and use model-specific limits
• librarian: Correct tool name from websearch_exa_web_search_exa to websearch_web_search_exa
• tmux: Correct health check endpoint to /global/health
• session-notification: Ensure notification fallback on terminal-notifier failure
• doctor: Quote paths and respect version channels in fix messages
• config: Initialize config context in plugin runtime to prevent warnings
• chat-headers: Skip x-initiator override for @ai-sdk/github-copilot models
• tool-config: Stop overriding external_directory permission (respect user setting)
• background-task: Clarify timeout unit is milliseconds in tool description
• model-fallback: Apply transformModelForProvider in getNextFallback
• task: Disambiguate background task_id metadata; align with opencode contract
• install: Build dist for git-based plugin installs
• look-at: Add catch block to prevent TUI crash on unexpected errors; temp dir cleanup, Windows compat
• delegate-task: Resolve timeout detection and config drift in sync poller
• model-resolution: Normalize model format and remove dead config flag

refactor

• kimi-k2.5-free: Remove from all fallback chains and reorder multimodal-looker accordingly
• model-resolution: Deduplicate model resolution utility functions
• installer: Remove dead Antigravity auth plugin code
• ultrawork: Rename gpt5.2.ts to gpt.ts and align with 5.4 style
• sisyphus: Extract prompt builders into subdirectory with GPT-5.4 variant
• prompts: Align all GPT prompts with 5.4 system prompt style across agents

docs

• agent-model-matching: Refresh Sisyphus chain description and model tables
• installation: Refresh provider model mappings
• orchestration: Align model tables with current runtime
• AGENTS.md: Regenerate hierarchy across all modules
• README: Remove security warning banners and fix table formatting

Migration Notes

For OpenAI-Only Users

Sisyphus now works out of the box with OpenAI-only configurations. Previously you would fall back to Hephaestus; now you get full Sisyphus orchestration with GPT-5.4 medium.

For Users with Kimi k2.5 Free

The kimi-k2.5-free model has been removed from all fallback chains. If you were relying on this (free) model, the system will now fall through to the next available model in your configured providers.

For ULW-Loop Users

Tasks will now take slightly longer to complete as Oracle verification is mandatory. The tradeoff is significantly higher confidence in completion correctness.

Thank You to Our Contributors

17 community PRs merged in this release, including fixes for flaky tests, doctor paths, model format normalization, notification click targets, and more. Special thanks to all first-time contributors!

Contributors: @rluisr, @MoerAI, @devxoul, @wousp112, @JimMoen, @acamq, @mertyldrm, @YLRong, @1noilimrev, @ualtinok, @Wangmerlyn, @Vacbo, @Romanok2805, @RaviTharuma, @brandonwebb-vista, @guazi04, @SeeYouCowboi, @chan1103, @SwiggitySwerve, @markarranz, @yhc509, @janghoon-ju, @wangjingu, @ilovingjny, @nous-labs