code-yeongyu/oh-my-openagent v4.5.0 on GitHub

v4.5.0 — Publish Safety & Internal Refactor

This release hardens runtime safety (fallback loops, parent-wake stalls, publish payload) and lands the long-running prompts-core refactor that splits every agent/mode prompt into versioned markdown.

🛡️ Publish Safety Hardening

Three guard rails that close production-impact pre-publish gaps:

runtime-fallback no longer trusts isRetryable: true on 4xx errors. If a provider mis-tagged a 401/403/404 as retryable, every configured fallback model would burn in an infinite loop. The signal is now gated to absent status codes, 5xx, the configured retry_on_errors list, and {408, 425, 429}.
parent-wake escapes a second stuck state. When the assistant's text stream died mid-message with no pending tool call, parent-wake deferred forever and background-agent completions never reached the parent. A bounded escape path now dispatches after toolCallDeferMaxMs.
npm payload blocks internal-only assets. Nested .npmignore per published dot-directory + a new script/package-layout-exclusion.test.ts permanent guard. Drop a __draft-foo/SKILL.md into .opencode/skills/ and CI fails before it ships to npm.

🔁 Ralph-Loop Liveness

Three independent stuck-state escapes for ralph-loop:

Oracle dispatch wait is now time-bounded (30-minute hard ceiling); previously a crashed/OOM-killed oracle session would freeze the loop forever.
Loop auto-stops when the latest assistant turn made zero model progress (both token counts at zero AND no assistant content).
The double-fire race fix no longer ignores in-flight oracle dispatches that never complete.

📞 Background-Agent Delivery

Four fixes converge on one outcome — completion alerts get through:

Debounced active parent wake flushes (no more thundering herd on burst completion).
Completion wakes deliver during active parent turns instead of being silently dropped.
Dispatched parent wake failures are isolated so one failure does not cascade.
Atlas subagents retry on provider usage-limit errors (no more dead-end runs).

👥 Team-Mode UX

Leaders now wake on unread mailbox messages (#4464). Idle leads no longer sleep through worker pings.
Model overrides preserved across team-mode boundaries.
Lead closes teams on its own initiative — no more lingering teams burning session / mailbox / member-turn budget.

🧠 Prompts-Core Migration (50+ commits, zero user-facing API change)

Every agent and mode prompt now lives in versioned markdown under packages/prompts-core/:

Atlas: 5 variants (default / GPT / Gemini / Kimi / Opus-4.7)
Prometheus: full prompt + variant table
Ultrawork: 4 variants (default / GPT / Gemini / planner)
Mode prompts: search / analyze / team / hyperplan

Companion dist-bundle test scans the built bundle for unique prompt signatures; a future Bun upgrade that silently regresses markdown inlining will fail CI loudly instead of breaking installs at first agent load.

🔧 Continuation + Validators

fix(continuation): resolves registered agent names before dispatching prompts (#4417) — fixes the empty-string bug when display names like "Atlas - Plan Executor" were used.
fix(prompt-async-gate): retry object-form path on runtime type errors (#4417).
fix(thinking-block-validator): avoid duplicate thinking prepends (#4374) — kills the dual-thinking-streams race.

🩺 Doctor + Help JSON Schemas

oh-my-opencode doctor, sandbox, status, and acp now expose JSON schemas for their help output — enables shell completion, structured CLI introspection, and --help consistency.

🏗️ Internal Package Layering

Workspace split into 9 published packages: utils / model-core / prompts-core / rules-engine / agents-md-core / ast-grep-core / comment-checker-core / hashline-core / boulder-state. Internal cleanup — no breaking changes to the plugin API.

🌐 Web

Marketing site canonical domain migrated to omo.dev.
README translations refreshed for ja / ko / ru / zh-cn.

🙏 Reviewer Note

This release was gated by a publish-debate-vortex hyperultradebate — 6 hostile category-member agents (quick / unspecified-low / unspecified-high / deep / ultrabrain / artistry) cross-attacked the 103-commit changeset across 3 rounds. The 3 must-fix blockers it identified — runtime-fallback boolean guard, parent-wake stuck escape, npm payload exclusion — all landed in this release.

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-system default behavior alignment so omitted configuration behaves consistently across runtime paths.
Install and publish workflow hardening, including safer release sequencing and package/install fixes.

Commit Summary

52ba125 Merge branch 'fix/pre-publish-blockers-v4.5.0'
8e28e29 fix(package): block internal-only assets from publish payload
5c0cc0d @Hungdoan565 has signed the CLA in #4492
ec552bc fix(ralph-loop): stop loop when latest assistant turn made no model progress
9810bad docs(publish-skill): pin Jobdori bot ID in Discord release commands
91aa915 @fcmfcm01 has signed the CLA in #4482
be2ed2b Merge pull request #4479 from code-yeongyu/fix-4470-momus-cancels-main
3b421c6 fix(background-agent): isolate dispatched parent wake failures
b03e08b Merge pull request #4478 from code-yeongyu/fix-4164-bg-reminder-timing
e1fde71 Merge pull request #4477 from code-yeongyu/fix-4374-dual-thinking-streams
ec7ce51 fix(background-agent): debounce active parent wake flushes
6a56144 fix(background-agent): deliver completion wakes during active parent turns
e443e86 fix(thinking-block-validator): avoid duplicate thinking prepends
037d017 Merge pull request #4476 from code-yeongyu/fix-4464-team-leader-wake
3e09910 @2wndrhs has signed the CLA in #4475
58025f9 fix(team-mode): wake leaders on unread mailbox messages (#4464)
759b566 Merge pull request #4474 from code-yeongyu/fix-4441-omo-evidence-ignore
3e11594 chore(repo): untrack .omo/evidence and .omo/notepads per .gitignore
dc05074 @hanakokoizumi has signed the CLA in #4467
939f2b8 @SoShymKing has signed the CLA in #4469
03eb9ff Merge pull request #4465 from code-yeongyu/fix-4417-atlas-path-and-start-work
28efc4e fix(continuation): resolve registered agent name before dispatching prompt (#4417)
a5cee76 fix(prompt-async-gate): retry object-form path on runtime type error (#4417)
c25c338 Merge pull request #4454 from code-yeongyu/fix-4448-npm-pack-dot-dirs
94d9d93 Merge pull request #4453 from code-yeongyu/fix-4419-atlas-subagent-fallback
0d0be43 Merge pull request #4452 from code-yeongyu/fix-4449-tool-execute-after-stall
9c56bf2 test(script): typecheck package layout regression
72119b3 fix(package): ship dot-directory command assets
f57532c fix(tool-execute-after): narrow metadata warning tools
dc845a8 chore(web): migrate canonical domain to omo.dev
6551f38 fix(background-agent): retry Atlas subagents on usage limits
767f5a6 fix(model-core): retry OpenAI usage_limit_reached fallbacks
6343feb Merge pull request #4450 from code-yeongyu/fix-4447-subagent-task-deny
4a68db0 fix(tool-execute-after): gate metadata recovery warnings
75d1ff4 fix(plugin-handlers): deny task for read-only subagents
9be97de Merge pull request #4435 from code-yeongyu/fix/issue-4429-team-mode-model-override
9fd52cb feat(help): #689 acp help JSON schema
eed88e4 feat(help): #689 acp help JSON schema
bab9caa feat(help): #688 sandbox help JSON schema
1fdb88a feat(help): #688 sandbox help JSON schema
2c6c42f feat(help): #687 status help JSON schema
36e394a feat(help): #687 status help JSON schema
a73e4db feat(help): #686 doctor help JSON schema
21713dc fix(agents): preserve model overrides with team mode
776da68 feat(help): #686 doctor help JSON schema
ce4b11b fix(cli): add explicit helpOption configuration for consistent help-flag ordering
3e0a975 test(dist-bundle): assert inlined prompt content survives bundling
14b3523 fix(ralph-loop): time-bound oracle dispatch wait to prevent stall
69c955f fix(parent-wake): bound assistant-text defer to escape stuck sessions
f05e0cb fix(runtime-fallback): gate retryable signal on status-code allowlist
222adeb fix(runtime-fallback): honor retryable signal
de7f1d8 Merge pull request #4400 from code-yeongyu/refactor/prompts-core-oracle-cleanup
8f09c02 docs: scrub remaining references to migrated files
d18cadb test(ultrawork): add byte-exact characterization test
e0b37b4 test(prometheus): add byte-exact characterization test
9d20425 docs(keyword-detector): rewrite AGENTS.md for prompts-core migration
87e28d2 docs(prometheus): rewrite AGENTS.md for thin-loader structure
640d7f7 Merge pull request #4396 from code-yeongyu/fix/session-scoped-internal-wake
3d0b4c4 fix(parent-wake): block unfinished assistant wakes
2ea7c49 Merge pull request #4390 from code-yeongyu/refactor/prometheus-to-prompts-core
1f32b88 refactor(prometheus): remove migrated section TypeScript files
8fc429a refactor(prometheus): load prompt variants from prompts-core
3806d93 feat(prompts-core): wire Prometheus variant table
fa0500e feat(prompts-core): add Prometheus prompt markdown variants
4749eab Merge pull request #4389 from code-yeongyu/refactor/mode-prompts-to-prompts-core
803f030 Merge remote-tracking branch 'origin/dev' into refactor/mode-prompts-to-prompts-core
528f083 Merge pull request #4388 from code-yeongyu/refactor/atlas-to-prompts-core
63ef72d docs(agents): document prompts-core mode prompts
80361c0 chore(prompts-core): wire markdown prompt packaging
2c8e2da refactor(mode-prompts): migrate hyperplan prompt
e79baf5 refactor(mode-prompts): migrate team prompt
3b7f51d refactor(mode-prompts): migrate analyze prompt
5cefbdb refactor(mode-prompts): migrate search prompt
c187dba test(keyword-detector): capture mode prompt baselines
8451ecd docs(atlas): document prompts-core prompt ownership
582c768 refactor(atlas): remove migrated TypeScript prompts
4975c35 refactor(atlas): load prompt variants from prompts-core
96a9365 build(prompts-core): bundle markdown prompt variants
91351d7 feat(prompts-core): support bundled prompt sources
b77fed2 feat(prompts-core): add Atlas Opus 4.7 prompt markdown
2adcec8 feat(prompts-core): add Atlas Kimi prompt markdown
d5995b3 feat(prompts-core): add Atlas Gemini prompt markdown
1923761 feat(prompts-core): add Atlas GPT prompt markdown
565d099 feat(prompts-core): add Atlas default prompt markdown
a92e25a test(atlas): add prompt byte preservation baselines
6644524 Merge pull request #4386 from code-yeongyu/refactor/ultrawork-to-prompts-core
79beeee docs: document ultrawork prompt markdown location
d217efc build(prompts): inline markdown prompt imports
e772c46 refactor(ultrawork): move planner prompt to prompts-core
e51acd0 refactor(ultrawork): move gemini prompt to prompts-core
a818bbd refactor(ultrawork): move gpt prompt to prompts-core
9c366cc refactor(ultrawork): move default prompt to prompts-core
39b3cbc chore(ultrawork): capture prompt baseline hashes
46ad6bf Merge pull request #4385 from code-yeongyu/refactor/prompts-core-foundation
436c618 fix(prompts-core): block prompt path traversal
74d5f70 chore(prompts-core): wire package typecheck
da0fa83 test(prompts-core): audit opencode coupling
c44a6bd feat(prompts-core): add prompt loader
6301bcf feat(prompts-core): add variant resolver
4338ddb feat(prompts-core): add package skeleton
87de056 refactor(model-core): move model family detectors
e1e9f34 Merge pull request #4382 from code-yeongyu/feat/ultrawork-prompt-tdd-evidence-tightening
d073bf1 feat(ultrawork): enforce TDD, scenario contract, durable notepad, reviewer gate
39a549a Merge pull request #4381 from code-yeongyu/fix/team-mode-closure-prompts
5f1fb0c fix(team-mode): make lead close teams on its own initiative
a4dcbb6 docs(publish): use Jobdori bot for releases
1473f93 Merge pull request #4380 from code-yeongyu/fix/ralph-loop-oracle-double-fire-race
9a1dd75 fix(ralph-loop): skip handleFailedVerification when oracle dispatch is in flight (#4256)
3523dab @niStee has signed the CLA in #4378