oh-my-openagent v5.0.0-beta.13

OMO 5.0.0-beta.13 — Recoverable Graphs, Side Conversations, and a Faster Senpi 2026.8.20-2

The center of gravity: a failed DAG run is no longer a total loss, /btw gives you a second conversation without burning your main context, and the bundled Senpi 2026.8.20-2 engine ships a large performance + Cursor-correctness wave.

🕸️ mass-ulw: retry, steer, and amend a live graph

A graph that lost one node to a flaky tool call used to mean re-running the whole thing and paying for every child again. Not anymore:

• dag retry returns failed, cancelled, and skip-cascaded nodes to a fresh attempt while every completed node keeps its cached result. Prompt-corrected retries (retry + prompt) are supported for exactly one node.
• dag send routes a message to one node's child with real delivery semantics — steer into a running child, revive a finished-but-resident child with its context intact, queued with a durable queue position when the child cannot take it yet. Non-continuable children get typed refusals (node_not_continuable, node_has_no_task; an unknown node gets node_not_found) that name the remedy.
• dag amend accepts an edited definition against the same run, SHA-256-fingerprint-diffs every node, and re-runs only changed + added nodes plus their transitive dependents. load_skills is deliberately outside the fingerprint, so a skills-only edit re-runs nothing.
• Per-node result artifacts are written atomically (temp file + fsync + rename), so a crash mid-retry can no longer leave a truncated result a reader would trust.
• Three new journaled event types — dag.node.retried, dag.node.steered, dag.definition.amended — record every deliberate re-run and amendment explicitly in the journal. All strictly additive: schemaVersion stays 1, documented in docs/reference/mass-ulw-protocol.md with an explicit no-break guarantee.
• The status widget shows an x<attempt> badge on re-run nodes; /dag shows amended x<n> and prefixes node errors with the error code. A lost amend race is now a clean run_still_active refusal instead of a run wedged on journal replay.
• The JS SDK gains retry, send, and amend wrappers, and the mass-ulw skill documents the full refusal matrix.

💬 /btw — ask a side question without polluting the main thread

/btw <question> (alias /side) opens a temporary session on the same model and agent. Ctrl+/ toggles between main and side; Ctrl+C on an empty side composer closes it. The side thread sees your recent main context as read-only background (capped at 64 messages / 64 KiB behind a boundary that tells the side model to keep its hands off files and subagents unless you explicitly ask), and its Q&A never enters the main transcript. Parent deletion, crash re-adoption, BTW-in-BTW nesting, and cancellation ordering are all explicitly handled, so side sessions don't leak. Guide: docs/guide/btw.md.

🔁 Model fallback stops going quiet (#6579)

Community fix by @MoerAI (#6611):

The repeated-error stall is fixed at its root: the retry-dedupe key didn't include the failing model, so the same error on a different model looked like a duplicate and was dropped — the session just went silent. Now:

• Dedupe keys are scoped per model; a skipped retry returns its key instead of burning it forever.
• Internal fallback retries carry an explicit marker so they stop being discarded as synthetic traffic.
• Model identity travels as provider/model plus variant through the whole fallback path — your reasoning-effort selection survives a fallback and a recovery.
• Watchdog timers are generation-scoped, so a timer from an abandoned attempt can never hijack a later turn.
• A manual model change resets the fallback watchdog, in-flight retry flag, and the retry-key set.

🖼️ One notice language across the transcript

Fallback-architect notices, tips, task-completion cards, team-member liveness warnings, and every memory notice (reflection, health, soul-update, write receipts, nudges) now render through one shared padded notice box. Completion cards lead with a tone glyph (●/✗/⚠/◐) and the task name, keep id · target/model · duration · tokens · tools · tps on a dim second row, and hold continuation hints behind expansion.

🩺 /doctor reflection health is deterministic

Reflection-health severity was a function of the wall clock — [warn] decayed to [ok] by pure passage of time. The check now reads an injectable clock, so the verdict is a pure function of your reflection data.

🧪 CI and tests stop lying

The Senpi adapter is now a first-class mandatory-QA surface with a machine-enforced evidence-path contract. Two DAG wait-hang classes (resident-cap saturation, terminalization during subscription setup) are pinned by bounded regression tests. Memory-test teardown retries narrowly on EFAULT only, and three Windows-slow suites got honest ceilings with waits left event-driven.

🧰 Bun 1.4.0 stable, everywhere

Bun 1.4 went stable today, and this release train rides it end to end:

• omo: every CI, publish, and platform-build workflow moved from Bun 1.3.14 to 1.4.0 (cache keys included), and the committed omo-senpi extension bundles were regenerated under stable 1.4.0 so the CI freshness gate and your installed extension agree byte-for-byte.
• senpi 2026.8.20-2 (same-day re-release on top of 2026.8.20): the compiled senpi binary is now built with Bun 1.4.0 stable (was 1.3.14), and npm publishing graduated from the Bun canary channel to stable 1.4.
• Safe dependency refresh riding along — omo: posthog-node 5.49, puppeteer-core 25.8, @clack/prompts 1.7, MCP SDK 1.30, js-yaml 4.3.1, picomatch 4.0.5, bun-types 1.4.0; senpi: Biome 2.5.9, @types/node 26.2.0, vitest + coverage 4.1.11, AWS Bedrock/Smithy patches, protobuf 2.14.0. Deliberately excluded: anything major, the Anthropic SDKs (OAuth-lane sensitivity), @google/genai, @mistralai.

---

⚙️ Bundled Senpi 2026.8.20-2 — the engine wave

Faster everywhere

• /resume no longer stalls on big sessions: exact byte-bounded streaming summaries are reused for unchanged files, and the visible transcript tail paints first while older messages warm progressively.
• Startup sheds 680 modules: the 1.2 MB Claude Agent SDK bundle and the jsdom/Readability/turndown HTML stack now load on first use, not at boot (14,203 → 13,523 modules).
• One process lighter per launch: cli.ts runs the agent in-process unless an Inspector option or custom exec args genuinely require a child (measured -75 ms, -6.2% on --help).
• On-disk compile cache: Node's enableCompileCache() is on for both CLI entries and published to children, so second launches skip recompiling the engine graph.
• Hot reload off the critical path: recursive config watchers run in a worker thread on macOS too, watcher teardown leaves the reload path (1.5–62 s of shutdown stall eliminated), and MCP reconnect no longer blocks a reload.
• Watch scope fixed: recursive watch targets now open one non-recursive subscription per in-scope directory — an extensions dir containing node_modules no longer drives fseventsd to 123% CPU and multi-GB RSS.
• Retry spinners throttle on sessions with 1,000+ persisted entries; the countdown stays.

The TUI stops freezing

• Goal footer tickers retire themselves on a stale extension context instead of ticking dead forever — the "Pursuing goal" label and continuation countdown no longer freeze the session after a reload.
• TUI mode switches detach live components instead of disposing them; spinners and widget intervals survive the switch.
• A vetoed or failed /reload no longer destroys live extension footers and task widgets.
• Assistant text that arrives after the last tool call renders below the tool cards, so approval questions stay visible.

claude-sdk-oauth: the $1,084 retry-storm fix (#723)

Stream-start-timeout retries now fork the SDK conversation at the last assistant boundary instead of re-sending the whole conversation — each retry re-bills only its own turn on a prefix-cache read. Before: cache writes grew ~8K/attempt, $25/6min, $1,084/3days on worker dispatch.

Cursor, seriously corrected (community wave — thanks @leeseunguk and @HeiTuz)

• 0-token resource_exhausted is treated as overflow: surface on first failure → compact before any rotation; rotation persists under the agent dir and survives TUI restarts; a rotation skip at the 3-rotation cap remints a fresh wire id instead of poisoning the session.
• Same-model retry after remint/compact instead of falling back to another provider; too-small overflow compacts drop to the last user turn; implausible billed cacheRead (3×/8× the live window) can no longer force useless compacts.
• Mid-turn compaction refuses while a native Run is live, so conversationId can't get poisoned mid-flight.
• Turns that end stop with pending toolCalls continue as toolUse; turns idle after completed tools end normally instead of hanging to the idle timeout.
• ANTML invoke recovery is skipped on cursor-agent, so Claude-named Cursor models stop rejecting parallel tool starts.
• Native todo calls persist even when the server resolves them without a local op; late tool_execution_end events draw their card instead of orphaning results; streamed task args survive an empty {} completion.
• Explicit thinking levels resolve to catalog suffix variant ids (claude-fable-5-thinking-low) on both the native and CLI lanes — no more Connect error not_found.
• Exec-bridge ownership is airtight: handlers bind to the owning run's abort signal, fail closed without a captured run, re-check ownership after approval prompts, and emit tool_execution_end errors so nothing dangles.
• Session titles use the session model's own summarization auth — no more unauthenticated title calls under an explicit compaction model.

Long-running sessions stop burning money

• Goal continuations are append-only: per-request history rewriting invalidated the provider's cache prefix, so long team-mode sessions paid a full uncached re-read every turn and drove themselves into 429 storms.
• Same-model 429 retries floor every wait with the exponential schedule — a provider answering with the same tiny retry-after hint can't pin the cadence at milliseconds anymore.

Config reload behaves in shared-agent-dir setups (thanks @Indosaram, #1006)

Routine-preference saves (like defaultModel) from another session no longer cascade a reload; the handoff snapshot is cleared unconditionally after requestReload() settles, so a stale plaintext settings snapshot can't survive.

For extension authors

The notice-box primitives are public API now: buildNoticeBox, noticeMessageRenderer, noticeEntryRenderer, and the NoticeSpec/NoticeLine/NoticeTone types — render your transcript notices in the shared visual family instead of re-implementing it.

Full engine changelog: senpi v2026.8.20 (+ v2026.8.20-2, the same-day Bun 1.4 re-release)

---

• 5d2742bf7 Merge pull request #7075 from code-yeongyu/fix/omo-ci-bun-1314-until-windows-tail
• b4d52a803 fix(test): build the absent-store fixture absent instead of deleting it back out
• 3d91936a9 fix(test): give the Windows setup fixtures a real absent store and a real home
• 38092867c fix(test): let win32 teardown leak an OS-owned temp root instead of failing
• 8cff72082 fix(test): widen the Windows EBUSY teardown budget with escalating backoff
• f92deb89b fix(test): make the Windows suites correct under bun 1.4
• 3a6fa0f65 test(omo-opencode): pin ResolveMessage identity instead of its prototype for bun 1.4
• 8d6a66fa0 Merge pull request #7074 from code-yeongyu/fix/btw-parent-remote-revalidation
• 8aee39439 test(qa): check in BTW revalidation evidence
• fc2c5aa09 fix(opencode): require remote BTW parent confirmation
• aff84c753 fix(qa): keep resolver imports side-effect free
• d92429e21 Merge pull request #7071 from code-yeongyu/feat/dag-node-resume-steer
• 37c5d6e5f Merge pull request #7068 from code-yeongyu/test/dag-happy-timeout-contract
• 5a6f21d9b docs(omo): state the additive compatibility stance for dag node-control events
• 1bbc2aa7d docs(omo): mass-ulw protocol additions for per-node retry/steer/amend
• 26b2506d1 Merge pull request #7067 from code-yeongyu/fix/tui-notice-block-style
• 56df6d7fc Merge pull request #7066 from code-yeongyu/fix/dag-wait-residency-regression
• 81344928f Merge pull request #6611 from MoerAI/fix/6579-runtime-fallback-dedup
• 3932c9c66 fix(runtime-fallback): mark reused internal retry parts
• b401eb4fc fix(runtime-fallback): clear first skipped retry key
• 51573a146 fix(runtime-fallback): scope retry marker to fallback generation
• 23057207d fix(runtime-fallback): recover skipped retry status keys
• d5c5918d5 test(qa): verify PR 6611 after dev rebase
• 2ad216f3a test(qa): record pre-push fork head
• 486844975 test(qa): record PR 6611 fallback evidence
• fc1c6312d fix(opencode): route only marked fallback retries
• b7ee41a64 fix(runtime-fallback): preserve explicit retry effort
• 62491bd6c fix(runtime-fallback): restore effective primary variants
• 3aba9d1d6 fix(runtime-fallback): index registered category models
• 891937ea1 fix(runtime-fallback): bind timeouts to state generations
• f603b5b37 fix(runtime-fallback): own queued retry generations
• 0b097b73b fix(runtime-fallback): reject stale timeout generations
• f4517c42b fix(runtime-fallback): acknowledge synthetic retries
• 0616cd628 fix(runtime-fallback): align created fallback identity
• 8bcfcc6c7 fix(runtime-fallback): retain queued fallback state
• 88d7cb4b4 test(lsp): isolate Windows compatibility checks
• 671f69b0e fix(runtime-fallback): honor explicit retry variants
• 895b9d469 fix(runtime-fallback): align retry reasoning state
• 06eab0f71 fix(runtime-fallback): isolate retry generations
• 3e24b3942 fix(runtime-fallback): preserve effective retry state
• c09d25876 test(runtime-fallback): wire generation acknowledgment
• c42be3ebd fix(runtime-fallback): align fallback generations
• 9af2747b4 fix(runtime-fallback): scope model-less retry dedupe
• 8c465e42b fix(runtime-fallback): preserve variant lifecycle state
• a5f5fee77 fix(runtime-fallback): persist effective fallback variant
• 8310d911d fix(runtime-fallback): preserve fallback variants
• f01536361 fix(runtime-fallback): read variant from chat output
• 5f7b005ab test(runtime-fallback): prove source-current OpenCode QA
• bee902117 test(runtime-fallback): record variant reset QA
• e6fa807d4 fix(runtime-fallback): reset retry keys on variant change
• b586cef81 fix(runtime-fallback): reset retry keys on manual model change
• 26a2a8fbe fix(runtime-fallback): retain active retry keys
• 5e3a12e74 fix(runtime-fallback): distinguish retry variants
• 11798f876 fix(runtime-fallback): dedupe model-less retry statuses
• 617400982 fix(runtime-fallback): prevent repeated-error stalls (fixes #6579)
• e676fef9d Merge pull request #7062 from code-yeongyu/refactor/doctor-health-clock-thread
• b8b7b0954 Merge pull request #7060 from code-yeongyu/test/memory-teardown-efault-retry
• 9dc8116e2 Merge pull request #7058 from code-yeongyu/feat/btw-codex-side-ui
• 6eb7e8cae fix(opencode): revalidate BTW parents
• 198a1963c docs(opencode): update BTW hook inventory
• 12807167b fix(opencode): preserve BTW cancellation semantics
• 28284f29c fix(opencode): reject deleted pending BTW
• e4ce3f658 docs(opencode): index BTW feature module
• bc2f49662 fix(opencode): preserve BTW route ownership
• 1be110282 fix(opencode): own every BTW cleanup
• 786d1955b fix(opencode): await pending BTW creation
• b174d9872 fix(opencode): await BTW cleanup on dispose
• cdc106126 fix(opencode): close final BTW boundary gaps
• ed8682862 fix(opencode): propagate BTW classification failures
• 369be3608 fix(opencode): retry BTW session classification
• f3a3c0b5b fix(opencode): restrict unverified BTW sessions
• 399ddb8fd fix(opencode): retry persisted BTW adoption
• e015f8e7a fix(opencode): validate persisted BTW parents
• 8b423d959 fix(opencode): enforce BTW execution bounds
• 113b49354 fix(opencode): bound BTW prompt inheritance
• 680973669 fix(opencode): block orphaned BTW nesting
• baf7d9255 fix(opencode): reject deleted BTW parents
• b3f4f0be1 fix(opencode): re-adopt persisted BTW sessions
• 3cbeeebc1 fix(opencode): cancel BTW on parent deletion
• a3981b98d fix(opencode): discard stale BTW adoption
• 4586d135b fix(opencode): cancel pending BTW navigation
• 960947c42 fix(opencode): harden BTW deletion lifecycles
• 559b3e22e fix(opencode): address BTW review feedback
• b5eb04896 feat(opencode): align BTW side conversation with Codex
• ec3d5afaa Merge pull request #7057 from code-yeongyu/fix/beta12-lockfile-sync
• 23db3d9cb fix(deps): sync bun.lock with the v5.0.0-beta.12 version bump
• 65f8540b3 Merge pull request #7056 from code-yeongyu/fix/windows-team-idle-wake-budget

Thank you to 1 community contributor:

• @MoerAI:
  • feat(opencode): align BTW side conversation with Codex
  • fix(opencode): address BTW review feedback
  • fix(opencode): harden BTW deletion lifecycles
  • fix(opencode): cancel pending BTW navigation
  • fix(opencode): discard stale BTW adoption
  • fix(opencode): cancel BTW on parent deletion
  • fix(opencode): re-adopt persisted BTW sessions
  • fix(opencode): reject deleted BTW parents
  • fix(opencode): block orphaned BTW nesting
  • fix(opencode): bound BTW prompt inheritance
  • fix(opencode): enforce BTW execution bounds
  • fix(opencode): validate persisted BTW parents
  • fix(opencode): retry persisted BTW adoption
  • fix(opencode): restrict unverified BTW sessions
  • fix(opencode): retry BTW session classification
  • fix(opencode): propagate BTW classification failures
  • fix(opencode): close final BTW boundary gaps
  • fix(opencode): await BTW cleanup on dispose
  • fix(opencode): await pending BTW creation
  • fix(opencode): own every BTW cleanup
  • fix(opencode): preserve BTW route ownership
  • docs(opencode): index BTW feature module
  • fix(opencode): reject deleted pending BTW
  • fix(opencode): preserve BTW cancellation semantics
  • docs(opencode): update BTW hook inventory
  • fix(opencode): revalidate BTW parents
  • Merge pull request #7058 from code-yeongyu/feat/btw-codex-side-ui
  • fix(runtime-fallback): prevent repeated-error stalls (fixes #6579)
  • fix(runtime-fallback): dedupe model-less retry statuses
  • fix(runtime-fallback): distinguish retry variants
  • fix(runtime-fallback): retain active retry keys
  • fix(runtime-fallback): reset retry keys on manual model change
  • fix(runtime-fallback): reset retry keys on variant change
  • test(runtime-fallback): record variant reset QA
  • test(runtime-fallback): prove source-current OpenCode QA
  • fix(runtime-fallback): read variant from chat output
  • fix(runtime-fallback): preserve fallback variants
  • fix(runtime-fallback): persist effective fallback variant
  • fix(runtime-fallback): preserve variant lifecycle state
  • fix(runtime-fallback): scope model-less retry dedupe
  • fix(runtime-fallback): align fallback generations
  • test(runtime-fallback): wire generation acknowledgment
  • fix(runtime-fallback): preserve effective retry state
  • fix(runtime-fallback): isolate retry generations
  • fix(runtime-fallback): align retry reasoning state
  • fix(runtime-fallback): honor explicit retry variants
  • test(lsp): isolate Windows compatibility checks
  • fix(runtime-fallback): retain queued fallback state
  • fix(runtime-fallback): align created fallback identity
  • fix(runtime-fallback): acknowledge synthetic retries
  • fix(runtime-fallback): reject stale timeout generations
  • fix(runtime-fallback): own queued retry generations
  • fix(runtime-fallback): bind timeouts to state generations
  • fix(runtime-fallback): index registered category models
  • fix(runtime-fallback): restore effective primary variants
  • fix(runtime-fallback): preserve explicit retry effort
  • fix(opencode): route only marked fallback retries
  • test(qa): record PR 6611 fallback evidence
  • test(qa): record pre-push fork head
  • test(qa): verify PR 6611 after dev rebase
  • fix(runtime-fallback): recover skipped retry status keys
  • fix(runtime-fallback): scope retry marker to fallback generation
  • fix(runtime-fallback): clear first skipped retry key
  • fix(runtime-fallback): mark reused internal retry parts
  • Merge pull request #6611 from MoerAI/fix/6579-runtime-fallback-dedup
  • fix(opencode): require remote BTW parent confirmation
  • test(qa): check in BTW revalidation evidence
  • Merge pull request #7074 from code-yeongyu/fix/btw-parent-remote-revalidation