# oh-my-openagent changelog

> oh-my-openagent release notes.

- Vendor: oh-my-openagent
- Category: AI
- Official site: https://omo.dev/
- Tracked by: What's New (https://whatsnew.fyi/product/oh-my-openagent)
- Harvested from: GitHub (code-yeongyu/oh-my-openagent)
- Entries below: 10 (newest first)

What's New is an index, not a publisher: every entry below links to the vendor's own release notes, which are the authoritative source. Entries are labelled where they are hand-curated sample data, pre-releases, or drawn from a secondary source such as a developer blog.

## Releases

### v4.19.3

- Date: 2026-07-28
- Version: v4.19.3
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.19.3
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.19.3

- **removed** — Remove legacy ~/.omo/config.jsonc loader and consolidate omo.jsonc as the only config surface
- **changed** — Harden config migration to automatically reclaim stale migration locks instead of blocking startup
- **changed** — Preserve migration transform diagnostics so changes are visible during migration
- **changed** — Create backup directories only when migration actually runs instead of always
- **fixed** — Fix legacy config root discovery on Windows and in setups where project lives outside home directory
- **changed** — Route config migration through unified path for Codex-side integrations
- **changed** — Re-tier visual-engineering and artistry category chains for improved model routing to frontend, design, and creative problem-solving tasks
- **added** — Ground interaction and UI-behavior work with curated beui.dev-based reference patterns in shared skill library
- **fixed** — Rebuild four flaky CI test suites around explicit synchronization to reduce timing-dependent failures
- **changed** — Introduce single binary approval gate and hard completion contract for release workflow to prevent stalled half-announced releases
- **fixed** — Resolve reviewer-tier contradiction in ulw-loop planning

##### oh-my-openagent v4.19.3 — Config migration you can trust

A stability-focused patch: the unified `omo.jsonc` config surface is now the single source of truth, migration is far more resilient across platforms, model routing got a tune-up, and CI got noticeably more deterministic.

###### One config, safer migration

The legacy `~/.omo/config.jsonc` loader is gone — `omo.jsonc` is now the only config surface, and the docs walk through exactly how migration works. Migration itself got hardened where it used to bite: stale migration locks are reclaimed automatically instead of wedging startup, transform diagnostics are preserved so you can see exactly what changed, backup directories are only created when a migration actually runs, and legacy config roots are discovered correctly on Windows and in setups where your project lives outside your home directory. Codex-side config migration now flows through the same unified path.

###### Smarter model routing for design and creative work

The `visual-engineering` and `artistry` category chains were re-tiered, so frontend/design and creative problem-solving tasks route to the models that currently do them best.

###### Interaction design, grounded

Interaction and UI-behavior work is now anchored to a curated beui.dev-based reference in the shared skill library, giving agents concrete interaction patterns to draw from instead of improvising.

###### A calmer, more deterministic pipeline

Four flaky CI test suites were rebuilt around explicit synchronization instead of timing luck, and the release workflow itself now carries a single, binary approval gate plus a hard completion contract — releases can no longer stall half-announced.

Also included: ulw-loop planning fixes (reviewer-tier contradiction resolved, clearer team-mode decision rule).

---

### v4.19.2

- Date: 2026-07-26
- Version: v4.19.2
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.19.2
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.19.2

- **changed** — Background task output is now non-blocking and completion notifications are authoritative instead of blocking polls
- **added** — Team messages now use injection-driven delivery so leads and members can continue working instead of waiting for empty turns
- **changed** — CodeGraph now runs as a managed shared daemon by default with pinned-runtime validation and process-tree cleanup
- **added** — Task surfaces now expose clearer role, model, turn, token, and throughput details
- **changed** — Model routing and documentation have moved to Claude Opus 5 as the recommended model
- **added** — Sisyphus now has a dedicated Claude Opus 5 prompt tuned for evidence-driven execution and strict stop behavior
- **changed** — start-work --make-pr and --ship flags now carry delivery intent through continuations
- **fixed** — cmux-backed subagent panes attach directly, avoid duplicate respawns, protect authenticated contexts, and clean up native sessions correctly
- **added** — Session discovery for oh-my-pi and gajae-code session stores
- **fixed** — CodeGraph now skips sweep for excluded projects on Windows
- **fixed** — CodeGraph enforces managed process lifecycle to prevent stale helpers and mismatched binaries
- **fixed** — Goal command provenance is now preserved in native command handoff
- **fixed** — ast-grep version probe output no longer leaks to parent process stderr
- **added** — Ultrawork hidden directive injection for Senpi
- **fixed** — Ultrawork requires obsessive live todos in GPT and GLM variants
- **fixed** — Ultrawork mandates goal registration and history-mimicking atomic commits
- **changed** — TypeScript dependency upgraded to 7.0.2

##### oh-my-openagent v4.19.2 — Smarter handoffs, cleaner runtimes

This release tightens the machinery around long-running agent work. Completion and team messages now arrive through authoritative notifications instead of blocking polls, CodeGraph runs as a managed shared daemon by default, and terminal-backed subagents behave predictably in cmux environments.

###### Notification-driven agent coordination

Background task output is now an immediate status or transcript peek, while completion notifications carry the final result automatically. Team messages use the same injection-driven model, so leads and members can keep working instead of spending turns on empty waits. The task surfaces also expose clearer role, model, turn, token, and throughput details.

###### Managed CodeGraph by default

OpenCode and LazyCodex now prefer the shared CodeGraph daemon automatically. Pinned-runtime validation, process-tree cleanup, and tighter lifecycle ownership prevent stale helpers and mismatched binaries while preserving a configuration opt-out for isolated per-session mode.

###### Claude Opus 5-ready orchestration

Sisyphus now has a dedicated Claude Opus 5 prompt, and the recommended model routing and documentation have moved to Opus 5. The new prompt is tuned for the same evidence-driven execution and strict stop behavior used by the rest of the current model families.

###### Cleaner delivery workflows and terminals

`start-work --make-pr` and `--ship` carry delivery intent through continuations, while cmux-backed subagent panes attach directly, avoid duplicate respawns, protect authenticated contexts, and clean up native sessions correctly.

Also included: hidden ultrawork directive injection for Senpi, oh-my-pi and gajae-code session discovery, quieter ast-grep version probing, native goal-command provenance fixes, and more deterministic CI and live QA fixtures.

- 3f917a94c Merge pull request #6360 from code-yeongyu/release/v4.19.2-source-state
- 6fbdf4599 Merge pull request #6359 from code-yeongyu/fix/codegraph-windows-exclusion
- b3e263b61 fix(codex): skip codegraph sweep for excluded projects
- 7fcfeb57f Merge pull request #6358 from code-yeongyu/feat/team-steer-only
- 29c53c9dd docs(qa): note CI synchronize trigger in steer-only evidence
- b75d9ebed Merge pull request #6357 from code-yeongyu/feature/codegraph-daemon-default-on
- 5a41124d7 Merge remote-tracking branch 'origin/dev' into feature/codegraph-daemon-default-on
- c27bf70b8 feat(codegraph): enable daemon by default
- 53b775745 feat(omo-config): add codegraph daemon setting
- aa7d3c523 Merge pull request #6353 from code-yeongyu/docs/prompt-sweep-injection
- 6090d31ed Merge origin/dev into docs/prompt-sweep-injection
- be979a5a5 Merge pull request #6350 from code-yeongyu/feat/team-injection
- 0cd7ab9e1 Merge origin/dev into feat/team-injection
- 6f1025edb Merge origin/dev into feat/team-injection
- 1639e4d8f docs(prompts)!: teach injection-driven waiting across shipped prompt surfaces
- 27b3bbaee Merge pull request #6346 from code-yeongyu/feat/task-output-peek
- 2da97a0f8 feat(team): injection-driven member inbox and crash liveness events
- f69e07350 fix(team): deliver idle injections as follow-ups
- 0f7b86827 feat(team)!: replace team_wait with injected lead inbox delivery
- ec823b854 chore(qa): lane QA fixtures and scanner updates
- b17479d0e Merge pull request #6347 from code-yeongyu/feature/agent-sessions-oh-my-pi-gajae-code
- 57d6e01c4 test(omo-codex): classify the generated pi_family scanner copy
- d59aa6dd4 docs(shared-skills): document oh-my-pi and gajae-code session stores
- dfc912c7f feat(shared-skills): scan oh-my-pi and gajae-code session stores
- 430b7e68f feat(task): make task_output non-blocking and completion notifications authoritative
- 3a56234e3 Merge pull request #6344 from code-yeongyu/feat/ultrawork-hidden-injection
- 2e2b9b748 fix(shared-skills): realign the debugging description with its TS source
- cd6f6cbce feat(programming): dependency-upgr

_[Truncated at 4000 characters — full notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.19.2]_

### v4.19.1

- Date: 2026-07-22
- Version: v4.19.1
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.19.1
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.19.1

- **changed** — GPT defaults including the Oracle reviewer now route through GPT-5.6 Sol with aligned prompt identity and guardrails
- **removed** — Retired gpt-5.5 fallback rungs are purged to prevent silent fallback to older models
- **changed** — Builtin Kimi fallbacks are promoted to K3
- **added** — Family-based process sweep at startup cleans up stale codegraph and LSP helpers
- **added** — Stdio MCP server processes now carry a parent-liveness watchdog that exits when parent dies
- **added** — LSP daemon reaps older-version daemons on startup
- **changed** — Vendored CodeGraph binary is pinned to 1.4.1 with MCP init guidance aligned to new output shapes
- **added** — Opt-in shared codegraph daemon mode available via OMO_CODEGRAPH_DAEMON across OpenCode and Codex
- **fixed** — Silent language servers that never push diagnostics no longer stall behind freshness timeout with clean diagnostics resolution for no-pull servers
- **fixed** — LSP daemon reports cancelled and timed-out requests accurately instead of mislabeling them
- **added** — ulw-loop gained atomic batch steering with steer --proposals-json and validation-batch review boundaries
- **added** — Next-goal auto-advance is fused into the checkpoint so multi-goal loops advance with fewer round trips
- **changed** — ulw-plan now enforces its mandatory opening announcement
- **changed** — Stop-resume writes are contained to the session state directory
- **fixed** — bun test now exits cleanly on win32 with held-open streams destroyed and isProcessAlive never throws
- **fixed** — codex.cmd shim resolves for doctor spawns
- **changed** — Cold-start SessionStart hook timeouts are raised to 15 seconds
- **fixed** — Duplicate role registration is avoided
- **fixed** — npm-install lock drift after local dependency rewrites is fixed
- **fixed** — Lazycodex marketplace payload bundles component CLIs correctly

##### 🩹 oh-my-openagent v4.19.1 — Quieter, tougher, and routed through Sol

A stability-first patch. Nothing changes about how you drive omo, but a lot changes about how well it behaves underneath: GPT defaults now route through GPT-5.6 Sol, stray background processes get swept instead of piling up, and a wave of Windows and Codex install fixes lands so the rough edges from v4.19.0 are gone.

###### 🧠 GPT defaults now run on GPT-5.6 Sol

Active GPT defaults — including the **Oracle** reviewer — now route through **GPT-5.6 Sol**, with the prompt identity and guardrails aligned to match. The retired **gpt-5.5** fallback rungs are purged so nothing silently falls back to an older model, and the builtin **Kimi** fallbacks are promoted to **K3**. You get the newer, stronger routing automatically with no config change.

###### 🧹 No more orphaned background processes

omo now runs a **family-based process sweep** at startup that cleans up stale codegraph and LSP helpers, and every stdio MCP server (LSP, git-bash, codegraph bridge) now carries a **parent-liveness watchdog** that exits the moment its parent dies. The LSP daemon also **reaps older-version daemons** on startup. If you ever noticed leftover helper processes lingering after a session, that class of leak is closed.

###### 🔎 CodeGraph 1.4.1 + opt-in shared daemon

The vendored **CodeGraph** binary is pinned to **1.4.1**, with MCP init guidance aligned to its new output shapes. A new **opt-in shared codegraph daemon** mode (`OMO_CODEGRAPH_DAEMON`) is available across OpenCode and Codex for teams that want a single long-lived index instead of per-session spawns.

###### 🩺 Sharper LSP diagnostics

Silent language servers that never push diagnostics no longer stall behind a freshness timeout — omo now resolves **clean diagnostics** for no-pull servers and guards the fallback with a version-current pull cache. The LSP daemon also reports **cancelled and timed-out requests accurately** instead of mislabeling them.

###### ⚙️ ulw-loop batch efficiency

The ulw-loop gained **atomic batch steering** (`steer --proposals-json`), **validation-batch review boundaries**, and **next-goal auto-advance fused into the checkpoint**, so multi-goal loops advance with far fewer round trips. `ulw-plan` now enforces its mandatory opening announcement, and stop-resume writes are contained to the session state dir.

###### 🪟 Windows & Codex install hardening

A broad reliability pass for Windows and Codex users: `bun test` now exits cleanly on win32 (held-open streams destroyed, `isProcessAlive` never throws), the `codex.cmd` shim resolves for doctor spawns, cold-start SessionStart hook timeouts are raised to 15s, duplicate role registration is avoided, npm-install lock drift after local dep rewrites is fixed, and the lazycodex marketplace payload bundles its component CLIs correctly.

###### 🎯 Model routing fixes

`antigravity-gemini-3-flash` ids are no longer rewritten to an invalid `-preview` id, and GPT-5.6 (`sol`/`terra`/`luna`) now carry proper **372k PostCompact context budgets** so compaction stops applying the 200k fallback.

---

- 9c81de52a Merge pull request #6300 from code-yeongyu/release/v4.19.1-source-state
- 499b52358 Merge pull request #6293 from code-yeongyu/fix/config-watch-rejection-reregister
- 166d2e1ab Merge pull request #6294 from code-yeongyu/fix/windows-flaky-coalesced-flush-wait
- ad0085b2b test(background-agent): await guaranteed flush-settle signal, not a poll-cap race
- 97b05026f Merge pull request #6290 from code-yeongyu/fix/139-marketplace-root-manifest
- c3d0e683d fix(script): also write marketplace.json to the lazycodex repo root
- 4ae3be456 Merge pull request #6289 from code-yeongyu/test/136-codex-only-install-exit-status
- 85a1bec60 Merge pull request #6285 from code-yeongyu/fix/108-sync-component-clis
- 8cfd05ce8 Merge pull request #6287 from code-yeongyu/fix/137-npm-ci-lock-drift
- 66a174fe3 test(install): regression-lock codex-only install failure exit s

_[Truncated at 4000 characters — full notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.19.1]_

### v4.19.0

- Date: 2026-07-17
- Version: v4.19.0
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.19.0
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.19.0

- **added** — Add first-class support for Kimi K3 with dedicated prompt variants for Sisyphus, Sisyphus-Junior, and Atlas, plus model-family detection and fallback routing in model-core
- **added** — Add Goal feature with state, persistence, parser, config migration, and /goal command to replace Ralph Loop
- **added** — Ship hyperplan adversarial-planning skill in omo-senpi with Boulder state platform support and codegraph MCP registration
- **added** — Ship default shared and local skills out of the box in Senpi session platform
- **added** — Wire start-work continuation for Codex
- **changed** — Replace legacy Ralph Loop wiring with Goal feature, with automatic config migration for existing Ralph Loop configs
- **changed** — Update transition detection, plugin/config compatibility, and install surfaces across package detection
- **changed** — Update task and tool behavior including delegate-task contract and runtime registration behavior
- **fixed** — Fix Senpi residency claims to release correctly when tasks are marked lost
- **fixed** — Wire TTL cleanup into session start with proper live-handle guards in Senpi task system
- **fixed** — Improve record store reliability by caching and reusing file descriptors in Senpi
- **fixed** — Optimize lead poller to scan incrementally in Senpi task system
- **fixed** — Fix Codex V2 thread-cap preservation
- **fixed** — Add stricter ulw-plan parsing and validation in Codex
- **fixed** — Add Playwright MCP custom-arg support in Codex
- **fixed** — Improve codegraph git-marker handling safety in Codex
- **changed** — Harden install and publish workflow with safer release sequencing and package/install fixes

##### 🚀 oh-my-openagent v4.19.0 — Open models just took the crown

This release ships with first-class support for **Kimi K3**, the 2.8-trillion-parameter open-weight frontier model from Moonshot AI that is now beating Claude Fable 5 on the Arena.ai Frontend Code leaderboard. OSS must win — and with K3, it just did.

###### 🤖 Kimi K3 is a first-class citizen

Sisyphus, Sisyphus-Junior, and Atlas now carry dedicated Kimi K3 prompt variants calibrated for K3's reasoning style, plus model-family detection and fallback routing in `model-core`. Whether you route through OpenRouter or Moonshot's API directly, omo recognizes K3 and picks the right prompts automatically.

###### 🎯 Goals replace Ralph Loop

The legacy Ralph Loop wiring is gone. In its place: a clean **Goal** feature with state, persistence, parser, config migration, and a `/goal` command. Existing Ralph Loop configs migrate automatically, so nothing breaks.

###### 🛠️ Senpi task engine gets serious

The Senpi task system received a deep reliability pass: residency claims release correctly when tasks are marked lost, TTL cleanup is wired into session start with proper live-handle guards, the record store caches and reuses file descriptors, and the lead poller scans incrementally. These changes make long-running multi-agent workflows noticeably more stable.

###### 🔌 Senpi platform expansion

omo-senpi now ships with the **hyperplan** adversarial-planning skill, a Senpi session platform for Boulder state, codegraph MCP registration, and default shared + local skills out of the box. Start-work continuation is also wired for Codex.

###### 🐛 Lazycodex / Codex reliability fixes

Codex users get fixes for V2 thread-cap preservation, stricter `ulw-plan` parsing and validation, Playwright MCP custom-arg support, and safer codegraph git-marker handling.

###### 🌍 Why this release matters

Kimi K3 is the largest open-weight model ever released, and independent benchmarks place it ahead of Claude Opus 4.8 and GPT-5.5 while trading blows with Claude Fable 5 and GPT-5.6 Sol. The frontier is no longer a closed garden. This release makes sure omo users can ride that wave the moment it hits.

###### 📝 A personal note from the maintainer

Today I did not use Claude Opus 4.8 once. Kimi K3 and the open-weights wave handled the work.

---

##### Minor Compatibility and Stability Release

This release carries compatibility-facing behavior changes and operational hardening. Read the summary below before upgrading or publishing.

- Rename transition updates across package detection, plugin/config compatibility, and install surfaces.
- Task and tool behavior updates, including delegate-task contract and runtime registration behavior.
- Install and publish workflow hardening, including safer release sequencing and package/install fixes.

##### Commit Summary

- 14083b89f Merge pull request #6193 from code-yeongyu/release/v4.19.0-source-state
- b56483fc7 Merge pull request #6192 from code-yeongyu/feat/kimi-k3-junior-atlas-prompts
- 22b91b662 feat(agents): add Kimi K3 prompt variants for Sisyphus-Junior and Atlas
- 9aff020ab Merge pull request #6188 from code-yeongyu/docs/goal-replaces-ralph-docs
- 1b6c47740 docs: replace ralph-loop with goal feature in user docs
- 731bdaef6 Merge pull request #6187 from code-yeongyu/docs/init-deep-refresh-20260717
- cd407ed11 docs: refresh hierarchical AGENTS.md knowledge base to 7d664b96b
- fe1c3dfb1 Merge pull request #5948 from Hungdoan565/codex/docs-known-issue-5850-ulw-native-plan-20260707
- 582e57d58 Merge latest dev into PR 5948 repair
- 7d664b96b Merge pull request #6186 from code-yeongyu/fix/windows-team-wait-commit-order
- 215f92df2 Merge latest dev into PR 5948 repair
- 5459b09d0 Merge pull request #5949 from Hungdoan565/codex/docs-known-issue-5839-oracle-token-20260707
- 58a8c1c60 docs: cover both native plan storage paths
- 22b231e05 Merge origin/dev into codex/docs-known-issue-5839-oracle-token-20260707
- 7676ff474 Merge dev and clarify na

_[Truncated at 4000 characters — full notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.19.0]_

### v4.18.2

- Date: 2026-07-16
- Version: v4.18.2
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.18.2
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.18.2

- **fixed** — GPT-5.6 sessions no longer spend model rounds on empty polls; waiting discipline now replays accumulated context in one sized call, re-validates only when inputs change, and requires explicit STOP conditions for subagent spawns
- **fixed** — Interrupting an agent mid-task no longer silently re-arms the todo-continuation enforcer; continuation boundaries and interruption provenance are now tracked
- **changed** — Plan agent now inherits fallback_models from Prometheus for consistent degradation across all agents
- **fixed** — Prose tokens like @ts-ignore, @latest, or (@path) in skill and command bodies are no longer mangled; resolver leaves non-file tokens untouched while real file references still inline
- **fixed** — Category-skill reminder is now delivered as a synthetic message part instead of appended to tool.execute.after output, preventing reminder text from leaking into model-visible results
- **fixed** — LSP proxy, per-user daemon, and codegraph stdio bridge now bound their startup, cancel pending acquisitions on parent close, and propagate closed-stdio errors instead of hanging
- **changed** — Plugin discovery now honors XDG_CONFIG_HOME and treats OPENCODE_CONFIG_DIR additively across Linux, WSL, and Windows
- **fixed** — Windows port-probe test flake stabilized
- **changed** — Test-suite hook dependencies are now injected instead of relying on process-global module mocks, removing cross-file bun test pollution
- **changed** — Custom provider fallback priority is now preserved in plan agent configuration

##### oh-my-opencode v4.18.2 — Stability & Waiting-Discipline Patch

A stability-focused patch across both editions (OpenCode Ultimate + Codex Light / lazycodex). No config changes, no new flags — upgrade in place.

###### ⏱️ GPT-5.6 stops burning time on empty polls

Long GPT-5.6 sessions were spending real model rounds on nothing — one recent run logged 91 empty stdin polls at ~134k tokens each, plus full re-validation of every criterion on every increment. Ultrawork, ulw-loop, and Hephaestus now carry a **waiting discipline**: a status poll replays the accumulated context so a long command finishes in one sized call, validation re-runs only when its inputs actually changed (with one final full pass), and subagent spawns require an explicit STOP condition. Applied to both editions.

###### 🔁 Continuation no longer re-arms after you interrupt

Interrupting an agent mid-task could silently re-arm the todo-continuation enforcer and resume against your intent. Continuation boundaries and interruption provenance are now tracked, so an intentional interruption stays interrupted (#6109).

###### 🧩 Plan agents inherit their fallback models

The `plan` agent now inherits `fallback_models` from Prometheus, so a planning session degrades to the same backup chain as the rest of your agents instead of dropping fallback entirely. Custom provider fallback priority is also preserved.

###### 📝 Prose `@tokens` stay intact in skills and commands

Literal tokens like `@ts-ignore`, `@latest`, or `(@path)` in skill/command bodies were being mangled into `[file not found: ...]` fragments. The resolver now leaves any `@token` that isn't a real file untouched, while real `@file` references still inline (#5978).

###### 🧹 Category-skill reminders no longer corrupt tool output

The category+skill reminder was appended to `tool.execute.after` output, leaking reminder text into Read/Bash/edit results the model saw. It's now delivered as a single synthetic message part (#6001).

###### 🔌 LSP / MCP startup is bounded and cancellable

The LSP proxy, per-user daemon, and codegraph stdio bridge now bound their startup, cancel pending acquisitions on parent close, and propagate closed-stdio errors instead of hanging or swallowing them. Codex's final Boulder gate is enforced and review lanes are bound to the full commit SHA.

###### 🪟 Plugin discovery respects XDG_CONFIG_HOME

`loadOpencodePlugins` now reuses the shared config-dir resolver, so plugin discovery honors `XDG_CONFIG_HOME` and treats `OPENCODE_CONFIG_DIR` additively (custom wins over default) across Linux, WSL, and Windows.

###### 🛠️ Also in this release

- Hephaestus delegation-table contract aligned; GPT-5.6 Copilot capability + OpenAI fast-tier alias handling.
- Windows port-probe test flake stabilized.
- Test-suite hardening: hook dependencies are now injected instead of relying on process-global module mocks, removing cross-file `bun test` pollution that could produce order-dependent failures.

---

##### Minor Compatibility and Stability Release

This release carries compatibility-facing behavior changes and operational hardening. Read the summary below before upgrading or publishing.

- Rename transition updates across package detection, plugin/config compatibility, and install surfaces.

##### Commit Summary

- 23cb030bb Merge pull request #6156 from code-yeongyu/release/v4.18.2-source-state
- d87596e58 Merge pull request #6114 from code-yeongyu/code-yeongyu/fix-windows-port-utils-flake
- 64a069bcd test(utils): stabilize port probe sequence
- b9d448894 Merge pull request #6155 from code-yeongyu/fix/test-mock-pollution-publish-gate
- 4a33c1e8f Merge pull request #6082 from H-TTTTT/feat/6032-copilot-gpt-5-6
- fb2d5d9dc Merge pull request #6152 from code-yeongyu/code-yeongyu/fix-6087-hephaestus-delegation-contract
- 3962c9152 test(models): refresh GPT-5.6 fallback QA
- c21e8ed0f test(hooks): make mock.module victims order-independent under single-process bun test
- 82bf7be14 fix(hepha

_[Truncated at 4000 characters — full notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.18.2]_

### v4.18.1

- Date: 2026-07-14
- Version: v4.18.1
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.18.1
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.18.1

- **changed** — LSP tooling across OpenCode, Codex, and Senpi now converges on the same hardened daemon runtime with authenticated local IPC, private runtime directories, endpoint and ownership checks, version validation, and conservative stale-process cleanup
- **changed** — Diagnostics now wait for fresh results instead of returning stale state
- **changed** — Directory checks now preserve complete counts
- **changed** — Rename operations now preflight workspace edits, detect conflicts, and synchronize document state after a successful change
- **changed** — Codex upgrades will only terminate a legacy daemon when ownership is proven
- **fixed** — Continuing a background task that is still running now returns an actionable error instead of reporting a false success
- **changed** — Context overflow is now classified separately so OpenCode can complete native compaction without the runtime fallback interrupting it
- **changed** — Ultrawork and delegated tasks now carry explicit observable stop conditions and evidence requirements
- **changed** — Review and QA approvals are now bound to the exact content or review lane, reducing duplicate work without allowing stale approval to cover changed code
- **removed** — Senpi project-local .pi/lsp-client.json files can no longer provide executable command or env values
- **changed** — Senpi now uses the shared LSP daemon instead of its adapter-local implementation

##### Safer, Shared LSP Runtime

LSP tooling across OpenCode, Codex, and Senpi now converges on the same hardened daemon runtime. Diagnostics wait for fresh results instead of returning stale state, directory checks preserve complete counts, and rename operations preflight workspace edits, detect conflicts, and synchronize document state after a successful change.

The daemon now uses authenticated local IPC, private runtime directories, endpoint and ownership checks, version validation, and conservative stale-process cleanup. Codex upgrades will only terminate a legacy daemon when ownership is proven.

##### More Predictable Agent Continuations and Compaction

Continuing a background task that is still running now returns an actionable error instead of reporting a false success. Context overflow is also classified separately so OpenCode can complete native compaction without the runtime fallback interrupting it.

Ultrawork and delegated tasks now carry explicit observable stop conditions and evidence requirements. Review and QA approvals are bound to the exact content or review lane, reducing duplicate work without allowing stale approval to cover changed code.

##### Senpi LSP Migration Note

Senpi now ships and uses the shared LSP daemon instead of its adapter-local implementation. For safety, project-local `.pi/lsp-client.json` files can no longer provide executable `command` or `env` values. Move custom server commands and environment variables to the user-level `~/.pi/lsp-client.json`; project files may continue to define non-executable settings such as extensions, language IDs, and priorities.

##### Release Layers

- **Pure components:** shared daemon security, fresh diagnostics, safer workspace edits, and reusable post-edit diagnostics.
- **OpenCode:** shared runtime integration, correct running-task continuation errors, native compaction preservation, and tighter agent completion contracts.
- **Codex / LazyCodex:** packaged LSP runtime, ownership-safe legacy cleanup, and commit/tree-bound review and QA gates.

---

##### Minor Compatibility and Stability Release

This release carries compatibility-facing behavior changes and operational hardening. Read the summary below before upgrading or publishing.

- Task and tool behavior updates, including delegate-task contract and runtime registration behavior.
- Install and publish workflow hardening, including safer release sequencing and package/install fixes.

##### Commit Summary

- 80c3c81eb Merge pull request #6116 from code-yeongyu/release/v4.18.1-source-state
- e3754dbf6 Merge pull request #6096 from impartshadow/fix/5966-running-continuation-false-success
- dd6db6aa8 Merge pull request #6115 from code-yeongyu/docs/skill-verify-gates-20260714
- 56c57a4f2 test(background-agent): cover running resume rejection
- ab0411bb7 docs(prompts): bind ulw-loop evidence to tree hash instead of commit
- 2e5e9eb90 docs(skills): drop review-work gate from work-with-pr verify loop
- b19f65257 Merge pull request #6107 from naiba-forks/fix/context-overflow-native-compaction
- 18ad5cb32 Merge pull request #6112 from code-yeongyu/fix/dedupe-commit-bound-reviews
- 0657c2f29 docs(skills): fail Gate B from any lane, scoped or full-run
- 9d71ac54c fix(prompts): bind review reuse to lane/SHA pairs and lane-scoped Gate B dispatch
- 034fb56e3 fix(omo-codex): make the continuation final gate commit-bound
- 158402e48 fix(prompts): scope review reuse to lanes so security and context lanes still run
- 64eef1f4d Merge pull request #6111 from code-yeongyu/fix/ulw-gpt56-stop-line-binding
- e3dfc6746 Merge pull request #6108 from code-yeongyu/code-yeongyu/fix-windows-team-tasklist-flake
- e8bb713c1 fix(omo-codex): make start-work global review gate pass commit-bound
- d55b5761a docs(ulw-loop): final quality gate approval satisfies same-SHA PR review gates
- 12a47f12d fix(codex-rules): make hephaestus review-work approvals commit-bound
- 628dae95c docs(skills): bind work-with-pr Gate B review approval t

_[Truncated at 4000 characters — full notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.18.1]_

### v4.18.0

- Date: 2026-07-14
- Version: v4.18.0
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.18.0
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.18.0

- **fixed** — Fixed a Codex run that kept going for eight hours by coupling the intent gate and stop gate across the main agent, ultrawork, and subagents
- **changed** — Every GPT-5.6 Hephaestus intent line must now declare the exact observable condition that ends the turn
- **changed** — Ultrawork must now register the goal through create_goal, include a binding WHEN TO STOP, and rank user problem resolution above ledger ceremony
- **changed** — Spawned subagents must now receive GOAL, STOP WHEN, and EVIDENCE so the parent can control completion by an observable contract
- **added** — Added layout-mechanics layer in OMO's frontend skill with shared spatial vocabulary including stack, cluster, content limiter, sidebar, switcher, cover, frame, reel, imposter, overlay stack, fixed sidenav shell, list-detail, and sticky aside
- **changed** — Frontend layouts now declare what scrolls, how regions shrink, and how the screen survives empty, long, unbroken, narrow, and RTL content
- **added** — Added bun run install:codex-dev command to install local source build with an unmistakable dev version stamp
- **changed** — Development builds now display unmistakable dev version stamp in cache path, plugin metadata, per-turn (OmO dev) prefix, and omo get-local-version output, while skipping update checks for non-semver stamps
- **fixed** — The Codex installer now strips inherited npm 12 allow-scripts configuration only from dependency-install subprocesses, preventing an outer npx policy from breaking the plugin install

##### GPT-5.6 That Actually Knows When to Stop

We just fixed a real Codex run that kept going for eight hours. This was not one bad loop; it was the full runaway stack: a main-agent model that does not treat stopping as the default, an ultrawork prompt designed to keep summoning help, subagents that also keep going, and a harness whose subagent lifecycle is difficult to control once all of those forces combine.

We are genuinely proud of this one. OMO now couples the **intent gate and stop gate** instead of treating completion as an afterthought:

- Every GPT-5.6 Hephaestus intent line must declare the exact observable condition that ends the turn.
- Ultrawork must register the goal through `create_goal`, include a binding `WHEN TO STOP`, and rank “is the user’s problem actually solved?” above ledger ceremony.
- The moment the stop goal holds, the agent must answer and stop immediately: no extra verification loop, polish pass, review cycle, or bonus refactor.
- Every spawned subagent must receive `GOAL`, `STOP WHEN`, and `EVIDENCE`, so the parent can control completion by an observable contract instead of trusting a vague “done.”

This is specifically tuned for `gpt-5.6-sol` and the persistent behavior that makes it such a strong autonomous worker. OMO keeps that persistence where it is useful, then gives it a hard, coordinated stopping boundary across the main agent, ultrawork, and subagents. For GPT-5.6 users, this update is strongly recommended; this control layer is the reason the OMO Codex experience can be substantially more disciplined than a stock harness run.

##### Frontend Layout Mechanics, Inspired by StyleGallery

A big shoutout to [changeroa/StyleGallery](https://github.com/changeroa/StyleGallery). Its layout work directly inspired the new layout-mechanics layer in OMO’s frontend skill, distilled part by part rather than copied as a generic style preset:

- [Layout Brief](https://github.com/changeroa/StyleGallery/blob/main/guides/layout-brief.md) inspired declaring scroll ownership, fixed/sticky/static regions, constraints, viewport-vs-container responsiveness, and content stress before CSS is written.
- [scroll-body-shell](https://github.com/changeroa/StyleGallery/blob/main/patterns/viewport-shell/scroll-body-shell.md) inspired the bounded app-shell contract: `auto minmax(0, 1fr) auto`, a single named scroll body, and the load-bearing `min-block-size: 0` fix.
- [ram-grid](https://github.com/changeroa/StyleGallery/blob/main/patterns/grid-repetition/ram-grid.md) inspired the overflow-safe intrinsic grid using `repeat(auto-fit, minmax(min(16rem, 100%), 1fr))`.
- StyleGallery’s pattern catalog inspired OMO’s shared spatial vocabulary: stack, cluster, content limiter, sidebar, switcher, cover, frame, reel, imposter, overlay stack, fixed sidenav shell, list-detail, and sticky aside.
- [Layout Gate](https://github.com/changeroa/StyleGallery/blob/main/quality/gates/layout.md) and the [Dashboard recipe](https://github.com/changeroa/StyleGallery/blob/main/recipes/dashboard.md) inspired the strict boundary between spatial mechanics and visual taste, plus source-order, scroll, reflow, and decision-priority checks.

OMO keeps its own visual direction, motion rules, image workflow, and QA system. The adopted lesson is narrower and extremely useful: app layouts must declare what scrolls, how regions shrink, and how the screen survives empty, long, unbroken, narrow, and RTL content.

##### Safer Codex Development and Installation

Codex contributors can now run `bun run install:codex-dev` to install the local source build with an unmistakable `dev` version stamp. The cache path, plugin metadata, per-turn `(OmO dev)` prefix, and `omo get-local-version` output all reveal that the development build is active, while update checks are skipped for non-semver stamps.

The Codex installer also strips inherited npm 12 `allow-scripts` configuration only from dependency-install subprocesses, preventing an outer `npx` policy from breaking

_[Truncated at 4000 characters — full notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.18.0]_

### v4.17.1

- Date: 2026-07-13
- Version: v4.17.1
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.17.1
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.17.1

- **fixed** — Interrupted process-backed tasks now resume only from trusted current runtime state
- **fixed** — Team members retain their scoped messaging runtime after recovery
- **fixed** — Corrupt team state fails closed
- **fixed** — Independent team waits cannot consume one another's messages
- **changed** — Batch task launches are capped before work is queued
- **fixed** — Failed child start is reported per item instead of hiding already-started work
- **changed** — Senpi QA gate now requires real child completion and output
- **fixed** — Publication is now pinned to the exact prepared source commit across npm, platform binaries, and LazyCodex
- **fixed** — Explicit GPT-5.6 Sol overrides retain their intended reasoning level across gateway model IDs
- **changed** — Rename transition updates across package detection, plugin/config compatibility, and install surfaces
- **changed** — Install and publish workflow hardening, including safer release sequencing and package/install fixes

##### Safer Durable Task Recovery

Interrupted process-backed tasks now resume only from trusted current runtime state. Team members retain their scoped messaging runtime after recovery, corrupt team state fails closed, and independent team waits cannot consume one another’s messages.

##### Bounded, Observable Task Fan-Out

Batch task launches are capped before work is queued. A failed child start is reported per item instead of hiding already-started work, and the Senpi QA gate now requires real child completion and output.

##### Release and Routing Integrity

Publication is now pinned to the exact prepared source commit across npm, platform binaries, and LazyCodex, so provenance and payloads agree. Explicit GPT-5.6 Sol overrides also retain their intended reasoning level across gateway model IDs.

---

##### Minor Compatibility and Stability Release

This release carries compatibility-facing behavior changes and operational hardening. Read the summary below before upgrading or publishing.

- Rename transition updates across package detection, plugin/config compatibility, and install surfaces.
- Install and publish workflow hardening, including safer release sequencing and package/install fixes.

##### Commit Summary

- ed0241d1a Merge pull request #6080 from code-yeongyu/release/v4.17.1-source-state
- 616d30bdd Merge pull request #6079 from code-yeongyu/fix/publish-skill-fastpath
- 930664124 Merge remote-tracking branch 'origin/dev' into fix/publish-skill-fastpath
- e52cd8f47 test(ci): cover provenance dispatch summary
- 298b743f3 fix(publish): make /publish ship-only, stop auto-triggering pre-publish review
- 017d70c46 fix(release): pin publication provenance to source
- c1a741dd9 fix(codex): align model and team spawn contracts
- 23312138c Merge pull request #6068 from code-yeongyu/code-yeongyu/cleanup-codex-toml-parser
- 848ffce93 Merge origin/dev into code-yeongyu/cleanup-codex-toml-parser
- 82864f8e0 Merge pull request #6070 from code-yeongyu/fix/visual-qa-tag-emission
- 12e2637e8 Merge origin/dev into code-yeongyu/cleanup-codex-toml-parser
- e47aa9b87 fix(skills): emit [product]/[evidence] tag reviewers and unify backoff dashes
- c660246be Merge pull request #6069 from code-yeongyu/fix/qa-loop-contradictions
- af27b971a fix(skills): resolve QA-loop contradictions and evidence-vs-product churn
- 5cfa6066c Merge remote-tracking branch 'origin/dev' into code-yeongyu/cleanup-codex-toml-parser
- 734b1b420 refactor(omo-codex): reuse TOML string parser
- 6e4c82627 Merge pull request #6065 from code-yeongyu/code-yeongyu/cleanup-shared-skill-helpers
- 955214b12 Merge remote-tracking branch 'origin/dev' into code-yeongyu/cleanup-shared-skill-helpers
- e1e09aef6 feat(team-core): lead unread recipient + leadSessionId + consumed-ledger + consumer lease
- 58000e267 Merge pull request #6064 from code-yeongyu/fix/ulw-loop-test-timeouts
- 4ccba96df fix: increase bun:test timeout for load-sensitive async tests in ulw-loop
- 5071a7c82 Merge pull request #6063 from code-yeongyu/fix/release-4-17-1-blockers
- 20887c5d4 Merge pull request #6057 from code-yeongyu/code-yeongyu/cleanup-root-contract-tests
- fbb56162b docs(skills): align review and Hephaestus guidance
- d66fe09e2 fix(release): build from prepared installer version
- 7c55a25f4 fix(skills): preserve full evidenced planning scope
- 8bbabd65f fix(ulw-loop): route planning by uncertainty
- e38bc3fa1 fix(teammode): gate fallback on visible capabilities
- b5040f937 Merge remote-tracking branch 'origin/dev' into code-yeongyu/cleanup-shared-skill-helpers
- d8a71f0a0 refactor(skills): reuse path resolver
- 99f124606 refactor(skills-loader): reuse initiator marker
- e564bd4ec refactor(utils): own skill path resolution
- efc99280a Merge origin/dev into code-yeongyu/cleanup-root-contract-tests
- f7eb7dd2d Merge pull request #6060 from code-yeongyu/fix/plan-prompts-scope-narrowing
- 19aabe5e1 fix(skills): make full-request scope the ulw-plan default
- d0e2e0863 fix(agents): remove scope-nar

_[Truncated at 4000 characters — full notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.17.1]_

### v4.17.0

- Date: 2026-07-11
- Version: v4.17.0
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.17.0
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.17.0

- **changed** — LazyCodex now selects low-, medium-, or high-difficulty workers based on change size and risk instead of routing every implementation through one general executor
- **changed** — New installs default to GPT-5.6 Sol with a 372K context window
- **added** — Exploration and deep-analysis categories use new Luna and Terra routes for stronger reasoning
- **added** — Interrupted ULW runs can now resume automatically through a bounded Stop hook
- **changed** — Evidence is isolated by attempt and tied to the commit that produced it, checked before the final reviewer starts
- **added** — Spawn guards and reviewer preflight checks prevent runaway fan-out and stop stale evidence from being relabeled as proof for a newer HEAD
- **changed** — Senpi task TUI now shows resolved model details and contextual status, control, and output excerpts
- **fixed** — Rows are width-bounded and terminal control sequences are sanitized to prevent narrow terminals and untrusted task output from distorting the interface
- **fixed** — Codex bundles no longer ship nested component MCP manifests with invalid relative daemon paths
- **changed** — Prompt and skill changes are now tested at behavioral seams—parsed values, runtime sentinels, and shipped-artifact equality—rather than fragile sentence matching or word-count tests

##### Codex Work That Scales to the Task

LazyCodex now selects low-, medium-, or high-difficulty workers from the actual change size and risk instead of routing every implementation through one general executor. New installs default to GPT-5.6 Sol with a 372K context window, while exploration and deep-analysis categories use the new Luna and Terra routes. The result is cheaper handling for narrow work and stronger reasoning where a change genuinely needs it.

##### ULW Loops Recover and Prove the Right Revision

Interrupted ULW runs can now resume automatically through a bounded Stop hook. Evidence is isolated by attempt, tied to the commit that produced it, and checked before the final reviewer starts. Spawn guards and reviewer preflight checks prevent runaway fan-out and stop stale evidence from being relabeled as proof for a newer HEAD.

##### More Useful, Safer Task Output

The Senpi task TUI now shows resolved model details and contextual status, control, and output excerpts. Rows are width-bounded and terminal control sequences are sanitized, so narrow terminals and untrusted task output no longer distort the interface.

##### Cleaner Plugin Installs and Honest Prompt QA

Codex bundles no longer ship nested component MCP manifests with invalid relative daemon paths. Prompt and skill changes are now tested at behavioral seams—parsed values, runtime sentinels, and shipped-artifact equality—rather than fragile sentence matching or word-count tests that block legitimate prose improvements.

---

##### Minor Compatibility and Stability Release

This release carries compatibility-facing behavior changes and operational hardening. Read the summary below before upgrading or publishing.

- Rename transition updates across package detection, plugin/config compatibility, and install surfaces.
- Task and tool behavior updates, including delegate-task contract and runtime registration behavior.
- Install and publish workflow hardening, including safer release sequencing and package/install fixes.

##### Commit Summary

- b4673bf07 Merge pull request #6050 from code-yeongyu/fix/ulw-size-tier-guidance
- 9741e995a feat(omo-codex): size-drive the ulw worker-tier selection
- d6a4631f6 Merge pull request #6049 from code-yeongyu/fix/ulw-agent-tier-config
- cdc710aae feat(omo-codex): emphasize code-mode parallel fan-out in hephaestus gpt-5.6 discovery
- 339e85b49 refactor(omo-codex): remove redundant lazycodex-executor agent; gate the workers
- e0734a8a6 fix(ultrawork): add answer-when-answerable stop rule to codex directive
- 7e26cae44 Merge pull request #6047 from code-yeongyu/fix/ulw-loop-evidence-commit-identity
- be7ea8f95 fix(ulw-loop): make the tool surface V2-native for GPT-5.6 (sol/terra)
- 940849332 Merge pull request #6048 from code-yeongyu/fix/codex-ulw-explore-first-plan-gate
- 1cc0be6c4 fix(ulw-loop): bind QA evidence to its capture commit to stop stale-HEAD relabeling
- 1a079089e test(ultrawork): pin explore-first scope-gated planning invariants
- 647cf1f29 fix(ultrawork): gather context before scope-gated planning in codex directive
- 09514c223 Merge pull request #6046 from code-yeongyu/fix/omo-codex-drop-nested-mcp-manifest
- 367d3b007 chore(package): ship git-bash-mcp package.json for vendored-MCP parity
- 661f9a3fe fix(lazycodex-sync): drop nested component .mcp.json from the marketplace bundle
- 740f034cf fix(codex-install): drop nested component .mcp.json from the plugin cache
- 28024d963 Merge pull request #6045 from code-yeongyu/fix/team-mode-deflake-windows-ci
- 7baf2bdb1 test(team-mode): give concurrency cases an explicit 30s per-test timeout
- 860e01cdc test(team-mode): make launch-concurrency probe waits event-driven
- 1a2c5dcdd Merge pull request #6044 from code-yeongyu/feat/lazycodex-gate-reviewer-high
- e278bc735 Merge pull request #6041 from code-yeongyu/feat/lazycodex-default-model-gpt-5.6-sol
- b70edfc76 feat(lazycodex): drop gate-reviewer reasoning effort to high
- 13addbf1d Merge pull request #6040 from co

_[Truncated at 4000 characters — full notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.17.0]_

### v4.16.3

- Date: 2026-07-10
- Version: v4.16.3
- Original notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.16.3
- Permalink: https://whatsnew.fyi/product/oh-my-openagent/releases/v4.16.3

- **changed** — Hephaestus, Momus, deep, and ultrabrain now prefer GPT-5.6 Sol with role-appropriate reasoning levels, with GPT-5.5 as compatibility fallback
- **changed** — OpenCode and LazyCodex select GPT-5.6-aware prompts and rule variants tuned to the model's tool surface
- **changed** — GLM fallback chains move to GLM 5.2
- **added** — LazyCodex codex team mode can now run on native MultiAgentV2 agents with Codex App threads as fallback transport
- **changed** — Team state records one immutable transport and validates agent and thread identities accordingly while preserving durable worktree, messaging, integration, and archive workflow
- **changed** — LazyCodex now resolves multi-agent behavior from the same custom model_catalog_json that Codex uses, including relative catalog paths beside config.toml
- **fixed** — GPT-5.6/V2 installs remove stale V2 disables and incompatible agents.max_threads settings
- **fixed** — Unknown root models no longer receive a risky thread-limit key
- **changed** — Managed Explorer, Librarian, and Momus defaults advance to their new model-appropriate efforts without overwriting custom model or effort choices
- **fixed** — Symlinked plugin-cache entrypoints execute correctly
- **fixed** — Optional lifecycle tools are no longer assumed to exist
- **fixed** — Dynamic bundled rules are reinjected when a session changes model
- **fixed** — OpenCode background tasks now finalize and notify their parent when provider, model, or credential errors are terminal even if the failed session shell still exists
- **fixed** — Live-listener routing verifies that the listener actually owns the target session before sending a parent wake, preventing completed work from disappearing into the wrong server instance

#### GPT-5.6 Orchestration, Hardened End to End

This patch makes GPT-5.6 the first-class path for OMO's deepest work while tightening the handoffs that keep long-running agent workflows alive.

##### GPT-5.6 for Deep Work and Review

Hephaestus, Momus, `deep`, and `ultrabrain` now prefer GPT-5.6 Sol with role-appropriate reasoning levels, while GPT-5.5 remains the compatibility fallback. OpenCode and LazyCodex also select GPT-5.6-aware prompts and rule variants, so the model gets guidance tuned to its actual tool surface instead of inheriting older assumptions. GLM fallback chains move to GLM 5.2.

##### Native MultiAgentV2 Teams in LazyCodex

Codex team mode can now run on native MultiAgentV2 agents when that complete tool surface is available, with Codex App threads retained as the fallback transport. Team state records one immutable transport, validates agent and thread identities accordingly, and preserves the existing durable worktree, messaging, integration, and archive workflow across both paths.

##### Safer Codex Installation and Rule Switching

LazyCodex now resolves multi-agent behavior from the same custom `model_catalog_json` that Codex uses, including relative catalog paths beside `config.toml`. GPT-5.6/V2 installs remove stale V2 disables and incompatible `agents.max_threads` settings, while unknown root models no longer receive a risky thread-limit key. Managed Explorer, Librarian, and Momus defaults advance to their new model-appropriate efforts without overwriting custom model or effort choices. Symlinked plugin-cache entrypoints execute correctly, optional lifecycle tools are no longer assumed to exist, and dynamic bundled rules are reinjected when a session changes model.

##### Background Agents Stop Waiting Forever

OpenCode background tasks now finalize and notify their parent when provider, model, or credential errors are terminal even if the failed session shell still exists. Live-listener routing also verifies that the listener actually owns the target session before sending a parent wake, preventing completed work from disappearing into the wrong server instance.

No manual configuration migration is required. Explicit GPT-5.5 selections, custom LazyCodex agent efforts, and Codex App thread teams remain supported.

---

##### Minor Compatibility and Stability Release

This release carries compatibility-facing behavior changes and operational hardening. Read the summary below before upgrading or publishing.

- Rename transition updates across package detection, plugin/config compatibility, and install surfaces.
- Task and tool behavior updates, including delegate-task contract and runtime registration behavior.
- Install and publish workflow hardening, including safer release sequencing and package/install fixes.

##### Commit Summary

- 46a55adb1 Merge pull request #6022 from code-yeongyu/fix/codex-v2-catalog-guards-qa
- 7e45232fa chore(codex): regenerate installer bundle
- cb11103e5 test(codex): pin bundled agent model matrix
- 60a8c5584 fix(codex): advance managed agent efforts
- 6d69dc03d fix(codex): clean stale V2 installer guards
- 5eb76107b fix(codex): resolve V2 runtime catalog paths
- 916916105 Merge pull request #6021 from code-yeongyu/fix/preserve-explicit-gpt55-pins
- bdf7bb49f fix(config): preserve explicit GPT-5.5 pins
- 5c795d27b chore(skills): pin pre-publish review to Sol
- f33957d8d Merge pull request #6018 from code-yeongyu/fix/release-model-default-docs
- 3af258783 Merge pull request #6020 from code-yeongyu/fix/release-bg-terminal-errors
- 0eb507b3b Merge pull request #6019 from code-yeongyu/fix/release-rules-dynamic-model
- 7beada44e Merge pull request #6017 from code-yeongyu/fix/release-rules-hephaestus-truncation
- cf9cc748a docs: align root model defaults
- 8054baec4 docs: fix ultrabrain release model claim
- 39692c963 test(background): cover alive terminal cleanup
- effa6990c fix(background): classify terminal provider errors
- f12a6b6b6 Merge pull request #6016 from code-yeo

_[Truncated at 4000 characters — full notes: https://github.com/code-yeongyu/oh-my-openagent/releases/tag/v4.16.3]_