$ git clone https://github.com/can1357/oh-my-pi.git
@oh-my-pi/pi-agent-core
Fixed
- Fixed agent startup and context compaction failures for models with unrecognized tokenizer encodings.
@oh-my-pi/pi-ai
Fixed
- Fixed automatic session retries for Anthropic-compatible streams that end prematurely without a completion signal.
- Fixed Gemini 3.x tool-call continuations through OpenAI-compatible endpoints.
- Fixed credential fallback for HTTP 402 payment-required and deactivated-workspace responses, preventing them from being misclassified as quota exhaustion.
- Fixed Perplexity email sign-in for accounts protected by authenticator-based two-factor authentication.
- Fixed Qianfan API-key login validation for keys that cannot access the validation model.
- Fixed Z.AI browser sign-in to report an occupied callback port before opening the browser.
@oh-my-pi/pi-catalog
Fixed
- Fixed Baseten GLM models, including GLM-5.3-Flash, being incorrectly classified as non-reasoning models.
- Fixed Cloudflare AI Gateway catalog refreshes so active Workers AI chat models, including newly released models, are discovered correctly.
- Fixed Cursor Kimi K3, Grok 4, and Composer 2.5 models being incorrectly treated as text-only when they support image attachments.
- Fixed OpenRouter routed model variants, including GLM-5.3 and dated DeepSeek V4 Pro, so their reasoning effort options are preserved correctly.
- Fixed MiniMax-M3 max output regressing to the upstream 512K pricing-tier boundary; all MiniMax providers keep the documented 128K output cap.
- Fixed GLM-5.3-Flash bundling the 50%-off launch promotion price; the catalog stays on the documented list price.
- Fixed Synthetic model discovery re-adding image input from the bundled reference when the route advertises text-only modalities.
@oh-my-pi/pi-coding-agent
Added
- Added gallery previews for composer and status-line components, with CLI filters for browsing by surface, composer, or segment.
Changed
- The status line now displays the thinking level as a compact icon alongside the model name by default; set
statusLine.compactThinkingLeveltofalseto restore the previous display.
Fixed
- Fixed MCP OAuth discovery for shared API gateways and authorization servers with nested paths, including Keycloak realms, so authentication targets the correct resource issuer and supports endpoint and dynamic client-registration discovery.
- Fixed credential rotation for HTTP 402 payment-required responses so sibling credentials are tried before model fallback without misclassifying informative non-quota errors.
- Transport errors after a complete, non-executed tool call can now retry through configured retry budgets and fallback chains when it is safe to do so, instead of ending the turn prematurely.
- Improved handling of truncated or otherwise undecodable images so they produce an actionable error and no longer permanently block subsequent requests or resumed sessions.
- Fixed Sharpshooter consolidation preserving memory files and queued changes when an empty replacement is returned.
- Fixed
omp plugin featuresso it discovers marketplace-installed plugins. - Fixed Escape handling when closing the
/sessioninformation panel; the panel now retains focus until dismissed. - Fixed the thinking-block visibility toggle so streamed reasoning is correctly hidden when thinking blocks are set to hidden.
- Reduced high idle CPU usage while the agent is working.
- Fixed resumed advisor subscription usage being displayed as a dollar amount instead of as a subscription.
- Fixed relative API addresses whose names end in image extensions being pasted as text instead of incorrectly treated as missing local image files.
- Fixed chat Markdown links and bare URLs so they become clickable OSC 8 hyperlinks when
tui.hyperlinks=alwaysis enabled. - Fixed unreadable composer text on light terminal backgrounds when using transparent composer styles.
- Fixed
retry.fallbackChainswarnings for valid selectors from providers whose model discovery is still pending; validation now updates after discovery completes. - Fixed visible browser windows launched by OMP so page content resizes with the operating-system window.
- Fixed Python evaluation hanging on Windows when importing native-extension modules such as NumPy.
- Fixed subagent extension context helpers so
ctx.getContextUsage()andctx.compact()operate on the child session. - Fixed
lsp diagnosticsincorrectly reporting success for project-aware pull-diagnostic servers when diagnostics time out or fail. - Corrected labels under
Settings > Context > Compaction Token Limit. - Fixed orphaned pages, iframes, and workers accumulating in the shared headless browser after abnormal OMP session termination.
@oh-my-pi/pi-mnemopi
Fixed
- Fixed working-memory recall returning too few or no results after matching memories were invalidated.
@oh-my-pi/pi-natives
Fixed
- Fixed staging and committing files through a reused Git repository handle, ensuring newly staged changes are correctly included in commits even on filesystems with coarse timestamp resolution.
@oh-my-pi/pi-tui
Added
- Added
setTerminalHyperlinks()to let hosts control OSC 8 hyperlink behavior in rendered Markdown links.
Fixed
- Fixed inline color swatches appearing for words with hex-like prefixes, such as
#each; swatches now appear only when the entire word is a valid color.
@oh-my-pi/pi-utils
Fixed
- Fixed runtime installation getting stuck for up to 60 seconds after an installer crash or forced termination, allowing subsequent installation attempts to proceed normally.
What's Changed
- fix(tui): hide thinking blocks emitted as append-only stable rows by
- fix(coding-agent): resolve marketplace plugin features by
- test(coding-agent): make issue-966 repro deterministic against stale gix index by
- fix(tui): honor tui.hyperlinks setting for chat markdown links by
- fix(sharpshooter): refuse all-empty replacement to prevent memory wipe by
- fix(browser): reap orphaned shared-browser targets from dead sessions by
- fix(catalog,coding-agent): strip OpenRouter route suffixes in bareModelId by
- fix(lsp): surface slow or failed pull diagnostics instead of reporting OK by
- fix(tui): keep transparent composer text readable by
- fix(mcp): support sub-path OAuth issuer discovery by
- fix(session): defer retry.fallbackChains warnings for pending discovery by
- fix(coding-agent): stop an undecodable image wedging a session by
- fix(ai): complete Perplexity TOTP login by
- fix(coding-agent): correct compact limit label by
- fix(ai): tolerate qianfan invalid_model during api key validation by
- Enable image input for supported Cursor models by
- fix(catalog): mirror Workers AI models into Cloudflare gateway by
- fix(browser): don't pin viewport on visible headless launches by
- fix(mnemopi): also exclude expired rows from FTS candidate slots by
- fix(extensions): restore subagent context callbacks by
- fix(catalog): Fix Baseten GLM models being marked as non-reasoning (including GLM 5.3 Flash) by
- fix(tui): preserve relative image URL pastes by
- fix(coding-agent): close /session panel with Esc without aborting the session by
- fix(ai): classify zero-content message_stop stream truncation as transient by
- fix(ai): retry empty Anthropic stream truncations by
- fix(utils): make runtime install lock crash-safe by
- fix(coding-agent): stop per-frame model-catalog scan in cost status segment by
- fix(pi-vcs): read the git index from disk to avoid stale snapshot by
- fix(auth): prevent unregistered Z.AI callback ports by
- fix(coding-agent): prefer RFC 9728 path-inserted protected-resource metadata by
- fix(coding-agent): auto-retry transport errors after proven-unexecuted tool calls by
- fix(ai): preserve Gemini thought signatures by
- fix(agent): fall back when native tokenizer encoding is unknown by
- docs: sync documentation with code and changelogs as of 18.0.10 by
- fix(ai): classify HTTP 402 as usage limit to rotate sibling OAuth credentials before model fallback by
New Contributors
Full Changelog: v18.0.10...v18.0.11
Vollständiger Original-Bericht
Ausführliche Details, Code-Beispiele & Hersteller-Stellungnahme auf github.com.
Wie bewertest du diesen Beitrag?
1 Klick Feedback 53 Fachleser & IT-Security Experten haben diesen Report heute geteilt
Teilen mit Netzwerk & Team:
Hat Ihnen dieser Tipp / Anleitung geholfen?
Community-Analysen & Experten-Meinungen 0
Verfasse deine eigene Analyse, teile Workarounds oder diskutiere diesen Vorfall im Blog.
Noch keine Community-Analyse verfasst. Markiere einen Textabschnitt oder klicke oben auf „ Eigene Analyse verfassen“!
Community Pulse: Relevanz-Einschätzung
1 Klick Experten-Votum 🔴 Akute Relevanz 57%
🟡 In Evaluierung 20%
🟢 Keine Auswirkung 17%
Spannende Innovation 6%
Verwandte Story-Cluster & Quellen (Vektor-KI)
6 Quellen
Black Hat Asia 2026 | Graph-Aware LLM for Windows Logon with a Closed-Loop Guarded Detection Agent
1 Quelle
Warmwind AI Employees Automate Typing, Clicking & Computer Use Like 1,000 Humans
1 Quelle
Who’s afraid of an open-weight model? GLM, context bombing and post-Black Hat attacks
Tipp: Mit Pfeiltasten [ ← ] und [ → ] blättern
SOCIAL SHARE CARD GENERATOR