t4code-17.0.5-appserver-9

@oh-my-pi/pi-agent-core

Added

• Added a per-message token estimation cache to optimize performance by reusing token counts for settled message history, with automatic cache invalidation on message mutation.

Changed

• Improved tool execution control by making tool interruptibility resolvable per call, allowing side-effecting operations to complete while passive waits can yield to queued steering.

@oh-my-pi/pi-ai

Changed

• Changed Anthropic API-key requests to default to a 1-hour prompt-cache retention (using the extended-cache-ttl-2025-04-11 beta) to prevent cold-misses during idle sessions, with support for PI_CACHE_RETENTION values "short" and "none" to override this behavior.

Fixed

• Fixed transient OpenAI stream truncations by retrying once before output becomes replay-unsafe, preventing recoverable transport errors from failing the turn.
• Fixed native Kimi Code K3 thinking being disabled during named function selection by utilizing generic required tool choice.
• Fixed /login moonshot validating China-platform API keys against the international host instead of honoring MOONSHOT_BASE_URL.
• Fixed Anthropic session stickiness suppressing usage-based re-ranking indefinitely by gating stickiness on a 1-hour cache warmth window (configurable via ANTHROPIC_SESSION_STICKY_CACHE_WARM_MS) to restore proactive multi-account load balancing after long idle periods.
• Fixed credential ranking where clockless Anthropic usage windows incorrectly outranked clocked sibling credentials.
• Fixed tool request failures (HTTP 400) on local grammar-constrained OpenAI-compatible backends (such as llama.cpp, LM Studio, and vLLM) by widening bare boolean subschemas into a value-accepting primitive union.
• Fixed custom OAuth Anthropic-compatible endpoints receiving generated Claude Code fingerprint headers even when explicit header overrides were provided.
• Fixed active sessions for plan-gated OpenAI Codex models (Sol/Luna) silently re-routing to sibling OAuth accounts when usage headroom changed, ensuring session stickiness is preserved as long as the preferred credential remains usable and eligible.

@oh-my-pi/app-wire

Added

• Add the local-only project.reveal desktop command without exposing absolute folder paths to clients.
• Add a bounded, ordered browser preview contract with backend-neutral authority metadata, capture chunking, tab controls, input actions, policy checks, cooperative leases, and human handoff.
• Add the negotiated transcript.search feature with strict, bounded host search and session context commands.
• Add the strict, bounded SessionRef.attention summary for cross-session pending questions, approvals, plan reviews, and latest root outcomes.

Changed

• Replace the fork-owned wire implementation with a compatibility export of the checksum-pinned
  T4 host-wire artifact. The active contract now includes bounded backward transcript paging.

@oh-my-pi/pi-catalog

Added

• Added an Anthropic compatibility flag to allow non-official OAuth endpoints to opt into configured Claude Code fingerprint header overrides.

Fixed

• Fixed a security issue where sensitive provider-defined request headers (such as API keys or credentials) were serialized in plaintext within the model cache (models.db). The cache now omits these headers, securely invalidates older cached rows, and restores or refetches them dynamically.
• Fixed OpenAI Codex discovery to respect caller-supplied fetch configurations (such as proxies or custom CAs) and correctly replace stale bundled models with the authenticated account catalog.
• Fixed stream timeouts and retry loops during long prefills on local loopback or RFC1918 backends (such as litellm proxies fronting local servers) by applying the local stream-timeout floor to these backends.
• Fixed Kimi K3 models served through generic OpenAI-compatible routes exposing unsupported reasoning efforts instead of the mandatory low/high/max scale.

@oh-my-pi/pi-coding-agent

Added

• Add native Finder reveal support for local appserver projects without exposing their absolute paths.
• Added profile-local cross-session transcript search with redacted snippets, bounded historical context, and filters for project, role, archive state, and time.
• Added redacted OpenAI Codex transport state to RPC session diagnostics, including actual transport, fallback, context-shape, and request-size counters without credentials or provider payloads.
• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.
• Added support for Codex (ChatGPT subscription) in generate_image via the providers.image: "openai-codex" option, including automatic subscription detection and fallback logic.
• Added an optional provider parameter to generate_image to override the global image provider setting for a single request.
• Added OpenTelemetry log and metric export capabilities alongside existing trace exports, supporting standard OTLP environment variables.
• Added support for id-prefixed targets and keys in retry.fallbackChains wildcards (e.g., "openrouter/google/*").
• Added support for Shift+Enter in the session tree selector (/tree, /branch) to summarize and switch branches in a single step.
• Added the PI_CONFIG_FILES environment variable to load settings overlays before --config overlays.

Changed

• Replaced the fork's copied appserver and app-wire implementations with a thin OMP authority
  launcher over checksum-pinned T4 host artifacts, including bounded transcript tail pagination.
• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.
• Changed bundled TTSR rules to warn instead of interrupting generation.
• Renamed the system prompt's project-context section wrapper from <context> to <repo-rules> to prevent XML tag collisions with in-band tool dialects.
• Renamed the /extensions dashboard tab "Agents (standard)" to "Agent Dirs (.agent/.agents)" to clarify its purpose.
• Optimized performance by reducing concurrent subagent update CPU usage, skipping unnecessary title generation in non-interactive hosts, and memoizing convertToLlm conversions over settled history.
• Improved the display of read xd:// calls by rendering them in a compact grouped view instead of full tool-execution cards.
• Made the hashline seen-line guard opt-in and off by default via edit.enforceSeenLines.

Fixed

• Fixed appserver restart loops when a crashed owner's Unix socket was inactive but its PID still appeared alive because it had not been reaped or had been reused. Recovery now confirms the completed socket is unreachable twice and revalidates every ownership file before reclaiming it.
• Fixed large session histories blocking appserver socket startup and repeatedly parsing every transcript body. Startup now publishes the transport before discovery, sends a complete initial inventory after bounded previews finish, and loads a transcript body only when that session is opened. Oversized non-header records are skipped without hiding the whole session.
• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver ...

t4code-17.0.5-appserver-8

@oh-my-pi/pi-agent-core

Added

• Added a per-message token estimation cache to optimize performance by reusing token counts for settled message history, with automatic cache invalidation on message mutation.

Changed

• Improved tool execution control by making tool interruptibility resolvable per call, allowing side-effecting operations to complete while passive waits can yield to queued steering.

@oh-my-pi/pi-ai

Changed

• Changed Anthropic API-key requests to default to a 1-hour prompt-cache retention (using the extended-cache-ttl-2025-04-11 beta) to prevent cold-misses during idle sessions, with support for PI_CACHE_RETENTION values "short" and "none" to override this behavior.

Fixed

• Fixed transient OpenAI stream truncations by retrying once before output becomes replay-unsafe, preventing recoverable transport errors from failing the turn.
• Fixed native Kimi Code K3 thinking being disabled during named function selection by utilizing generic required tool choice.
• Fixed /login moonshot validating China-platform API keys against the international host instead of honoring MOONSHOT_BASE_URL.
• Fixed Anthropic session stickiness suppressing usage-based re-ranking indefinitely by gating stickiness on a 1-hour cache warmth window (configurable via ANTHROPIC_SESSION_STICKY_CACHE_WARM_MS) to restore proactive multi-account load balancing after long idle periods.
• Fixed credential ranking where clockless Anthropic usage windows incorrectly outranked clocked sibling credentials.
• Fixed tool request failures (HTTP 400) on local grammar-constrained OpenAI-compatible backends (such as llama.cpp, LM Studio, and vLLM) by widening bare boolean subschemas into a value-accepting primitive union.
• Fixed custom OAuth Anthropic-compatible endpoints receiving generated Claude Code fingerprint headers even when explicit header overrides were provided.
• Fixed active sessions for plan-gated OpenAI Codex models (Sol/Luna) silently re-routing to sibling OAuth accounts when usage headroom changed, ensuring session stickiness is preserved as long as the preferred credential remains usable and eligible.

@oh-my-pi/app-wire

Added

• Add the local-only project.reveal desktop command without exposing absolute folder paths to clients.
• Add a bounded, ordered browser preview contract with backend-neutral authority metadata, capture chunking, tab controls, input actions, policy checks, cooperative leases, and human handoff.
• Add the negotiated transcript.search feature with strict, bounded host search and session context commands.
• Add the strict, bounded SessionRef.attention summary for cross-session pending questions, approvals, plan reviews, and latest root outcomes.

@oh-my-pi/pi-catalog

Added

• Added an Anthropic compatibility flag to allow non-official OAuth endpoints to opt into configured Claude Code fingerprint header overrides.

Fixed

• Fixed a security issue where sensitive provider-defined request headers (such as API keys or credentials) were serialized in plaintext within the model cache (models.db). The cache now omits these headers, securely invalidates older cached rows, and restores or refetches them dynamically.
• Fixed OpenAI Codex discovery to respect caller-supplied fetch configurations (such as proxies or custom CAs) and correctly replace stale bundled models with the authenticated account catalog.
• Fixed stream timeouts and retry loops during long prefills on local loopback or RFC1918 backends (such as litellm proxies fronting local servers) by applying the local stream-timeout floor to these backends.
• Fixed Kimi K3 models served through generic OpenAI-compatible routes exposing unsupported reasoning efforts instead of the mandatory low/high/max scale.

@oh-my-pi/pi-coding-agent

Added

• Add native Finder reveal support for local appserver projects without exposing their absolute paths.
• Added profile-local cross-session transcript search with redacted snippets, bounded historical context, and filters for project, role, archive state, and time.
• Added redacted OpenAI Codex transport state to RPC session diagnostics, including actual transport, fallback, context-shape, and request-size counters without credentials or provider payloads.
• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.
• Added support for Codex (ChatGPT subscription) in generate_image via the providers.image: "openai-codex" option, including automatic subscription detection and fallback logic.
• Added an optional provider parameter to generate_image to override the global image provider setting for a single request.
• Added OpenTelemetry log and metric export capabilities alongside existing trace exports, supporting standard OTLP environment variables.
• Added support for id-prefixed targets and keys in retry.fallbackChains wildcards (e.g., "openrouter/google/*").
• Added support for Shift+Enter in the session tree selector (/tree, /branch) to summarize and switch branches in a single step.
• Added the PI_CONFIG_FILES environment variable to load settings overlays before --config overlays.

Changed

• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.
• Changed bundled TTSR rules to warn instead of interrupting generation.
• Renamed the system prompt's project-context section wrapper from <context> to <repo-rules> to prevent XML tag collisions with in-band tool dialects.
• Renamed the /extensions dashboard tab "Agents (standard)" to "Agent Dirs (.agent/.agents)" to clarify its purpose.
• Optimized performance by reducing concurrent subagent update CPU usage, skipping unnecessary title generation in non-interactive hosts, and memoizing convertToLlm conversions over settled history.
• Improved the display of read xd:// calls by rendering them in a compact grouped view instead of full tool-execution cards.
• Made the hashline seen-line guard opt-in and off by default via edit.enforceSeenLines.

Fixed

• Fixed appserver restart loops when a crashed owner's Unix socket was inactive but its PID still appeared alive because it had not been reaped or had been reused. Recovery now confirms the completed socket is unreachable twice and revalidates every ownership file before reclaiming it.
• Fixed large session histories blocking appserver socket startup and repeatedly parsing every transcript body. Startup now publishes the transport before discovery, sends a complete initial inventory after bounded previews finish, and loads a transcript body only when that session is opened. Oversized non-header records are skipped without hiding the whole session.
• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver sessions losing their RPC child when inline image payloads made a lifecycle or durable-entry notification exceed the one MiB line ceiling. Managed children now omit only redundant image bytes from their internal stdout notifications and mark the frame, while the full images remain unchanged in OMP's session and model context; even tiny canonical image payloads are externalized for transcript ...

t4code-17.0.5-appserver-10

@oh-my-pi/pi-agent-core

Added

• Added a per-message token estimation cache to optimize performance by reusing token counts for settled message history, with automatic cache invalidation on message mutation.

Changed

• Improved tool execution control by making tool interruptibility resolvable per call, allowing side-effecting operations to complete while passive waits can yield to queued steering.

@oh-my-pi/pi-ai

Changed

• Changed Anthropic API-key requests to default to a 1-hour prompt-cache retention (using the extended-cache-ttl-2025-04-11 beta) to prevent cold-misses during idle sessions, with support for PI_CACHE_RETENTION values "short" and "none" to override this behavior.

Fixed

• Fixed transient OpenAI stream truncations by retrying once before output becomes replay-unsafe, preventing recoverable transport errors from failing the turn.
• Fixed native Kimi Code K3 thinking being disabled during named function selection by utilizing generic required tool choice.
• Fixed /login moonshot validating China-platform API keys against the international host instead of honoring MOONSHOT_BASE_URL.
• Fixed Anthropic session stickiness suppressing usage-based re-ranking indefinitely by gating stickiness on a 1-hour cache warmth window (configurable via ANTHROPIC_SESSION_STICKY_CACHE_WARM_MS) to restore proactive multi-account load balancing after long idle periods.
• Fixed credential ranking where clockless Anthropic usage windows incorrectly outranked clocked sibling credentials.
• Fixed tool request failures (HTTP 400) on local grammar-constrained OpenAI-compatible backends (such as llama.cpp, LM Studio, and vLLM) by widening bare boolean subschemas into a value-accepting primitive union.
• Fixed custom OAuth Anthropic-compatible endpoints receiving generated Claude Code fingerprint headers even when explicit header overrides were provided.
• Fixed active sessions for plan-gated OpenAI Codex models (Sol/Luna) silently re-routing to sibling OAuth accounts when usage headroom changed, ensuring session stickiness is preserved as long as the preferred credential remains usable and eligible.

@oh-my-pi/app-wire

Added

• Add the local-only project.reveal desktop command without exposing absolute folder paths to clients.
• Add a bounded, ordered browser preview contract with backend-neutral authority metadata, capture chunking, tab controls, input actions, policy checks, cooperative leases, and human handoff.
• Add the negotiated transcript.search feature with strict, bounded host search and session context commands.
• Add the strict, bounded SessionRef.attention summary for cross-session pending questions, approvals, plan reviews, and latest root outcomes.

Changed

• Replace the fork-owned wire implementation with a compatibility export of the checksum-pinned
  T4 host-wire artifact. The active contract now includes bounded backward transcript paging.

@oh-my-pi/pi-catalog

Added

• Added an Anthropic compatibility flag to allow non-official OAuth endpoints to opt into configured Claude Code fingerprint header overrides.

Fixed

• Fixed a security issue where sensitive provider-defined request headers (such as API keys or credentials) were serialized in plaintext within the model cache (models.db). The cache now omits these headers, securely invalidates older cached rows, and restores or refetches them dynamically.
• Fixed OpenAI Codex discovery to respect caller-supplied fetch configurations (such as proxies or custom CAs) and correctly replace stale bundled models with the authenticated account catalog.
• Fixed stream timeouts and retry loops during long prefills on local loopback or RFC1918 backends (such as litellm proxies fronting local servers) by applying the local stream-timeout floor to these backends.
• Fixed Kimi K3 models served through generic OpenAI-compatible routes exposing unsupported reasoning efforts instead of the mandatory low/high/max scale.

@oh-my-pi/pi-coding-agent

Added

• Add native Finder reveal support for local appserver projects without exposing their absolute paths.
• Added profile-local cross-session transcript search with redacted snippets, bounded historical context, and filters for project, role, archive state, and time.
• Added redacted OpenAI Codex transport state to RPC session diagnostics, including actual transport, fallback, context-shape, and request-size counters without credentials or provider payloads.
• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.
• Added support for Codex (ChatGPT subscription) in generate_image via the providers.image: "openai-codex" option, including automatic subscription detection and fallback logic.
• Added an optional provider parameter to generate_image to override the global image provider setting for a single request.
• Added OpenTelemetry log and metric export capabilities alongside existing trace exports, supporting standard OTLP environment variables.
• Added support for id-prefixed targets and keys in retry.fallbackChains wildcards (e.g., "openrouter/google/*").
• Added support for Shift+Enter in the session tree selector (/tree, /branch) to summarize and switch branches in a single step.
• Added the PI_CONFIG_FILES environment variable to load settings overlays before --config overlays.

Changed

• Replaced the fork's copied appserver and app-wire implementations with a thin OMP authority
  launcher over checksum-pinned T4 host artifacts, including bounded transcript tail pagination.
• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.
• Changed bundled TTSR rules to warn instead of interrupting generation.
• Renamed the system prompt's project-context section wrapper from <context> to <repo-rules> to prevent XML tag collisions with in-band tool dialects.
• Renamed the /extensions dashboard tab "Agents (standard)" to "Agent Dirs (.agent/.agents)" to clarify its purpose.
• Optimized performance by reducing concurrent subagent update CPU usage, skipping unnecessary title generation in non-interactive hosts, and memoizing convertToLlm conversions over settled history.
• Improved the display of read xd:// calls by rendering them in a compact grouped view instead of full tool-execution cards.
• Made the hashline seen-line guard opt-in and off by default via edit.enforceSeenLines.

Fixed

• Fixed appserver restart loops when a crashed owner's Unix socket was inactive but its PID still appeared alive because it had not been reaped or had been reused. Recovery now confirms the completed socket is unreachable twice and revalidates every ownership file before reclaiming it.
• Fixed large session histories blocking appserver socket startup and repeatedly parsing every transcript body. Startup now publishes the transport before discovery, sends a complete initial inventory after bounded previews finish, and loads a transcript body only when that session is opened. Oversized non-header records are skipped without hiding the whole session.
• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver ...

t4code-17.0.5-appserver-7

@oh-my-pi/pi-agent-core

Added

• Added a per-message token estimation cache to optimize performance by reusing token counts for settled message history, with automatic cache invalidation on message mutation.

Changed

• Improved tool execution control by making tool interruptibility resolvable per call, allowing side-effecting operations to complete while passive waits can yield to queued steering.

@oh-my-pi/pi-ai

Changed

• Changed Anthropic API-key requests to default to a 1-hour prompt-cache retention (using the extended-cache-ttl-2025-04-11 beta) to prevent cold-misses during idle sessions, with support for PI_CACHE_RETENTION values "short" and "none" to override this behavior.

Fixed

• Fixed transient OpenAI stream truncations by retrying once before output becomes replay-unsafe, preventing recoverable transport errors from failing the turn.
• Fixed native Kimi Code K3 thinking being disabled during named function selection by utilizing generic required tool choice.
• Fixed /login moonshot validating China-platform API keys against the international host instead of honoring MOONSHOT_BASE_URL.
• Fixed Anthropic session stickiness suppressing usage-based re-ranking indefinitely by gating stickiness on a 1-hour cache warmth window (configurable via ANTHROPIC_SESSION_STICKY_CACHE_WARM_MS) to restore proactive multi-account load balancing after long idle periods.
• Fixed credential ranking where clockless Anthropic usage windows incorrectly outranked clocked sibling credentials.
• Fixed tool request failures (HTTP 400) on local grammar-constrained OpenAI-compatible backends (such as llama.cpp, LM Studio, and vLLM) by widening bare boolean subschemas into a value-accepting primitive union.
• Fixed custom OAuth Anthropic-compatible endpoints receiving generated Claude Code fingerprint headers even when explicit header overrides were provided.
• Fixed active sessions for plan-gated OpenAI Codex models (Sol/Luna) silently re-routing to sibling OAuth accounts when usage headroom changed, ensuring session stickiness is preserved as long as the preferred credential remains usable and eligible.

@oh-my-pi/app-wire

Added

• Add the local-only project.reveal desktop command without exposing absolute folder paths to clients.
• Add a bounded, ordered browser preview contract with backend-neutral authority metadata, capture chunking, tab controls, input actions, policy checks, cooperative leases, and human handoff.
• Add the negotiated transcript.search feature with strict, bounded host search and session context commands.
• Add the strict, bounded SessionRef.attention summary for cross-session pending questions, approvals, plan reviews, and latest root outcomes.

@oh-my-pi/pi-catalog

Added

• Added an Anthropic compatibility flag to allow non-official OAuth endpoints to opt into configured Claude Code fingerprint header overrides.

Fixed

• Fixed a security issue where sensitive provider-defined request headers (such as API keys or credentials) were serialized in plaintext within the model cache (models.db). The cache now omits these headers, securely invalidates older cached rows, and restores or refetches them dynamically.
• Fixed OpenAI Codex discovery to respect caller-supplied fetch configurations (such as proxies or custom CAs) and correctly replace stale bundled models with the authenticated account catalog.
• Fixed stream timeouts and retry loops during long prefills on local loopback or RFC1918 backends (such as litellm proxies fronting local servers) by applying the local stream-timeout floor to these backends.
• Fixed Kimi K3 models served through generic OpenAI-compatible routes exposing unsupported reasoning efforts instead of the mandatory low/high/max scale.

@oh-my-pi/pi-coding-agent

Added

• Add native Finder reveal support for local appserver projects without exposing their absolute paths.
• Added profile-local cross-session transcript search with redacted snippets, bounded historical context, and filters for project, role, archive state, and time.
• Added redacted OpenAI Codex transport state to RPC session diagnostics, including actual transport, fallback, context-shape, and request-size counters without credentials or provider payloads.
• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.
• Added support for Codex (ChatGPT subscription) in generate_image via the providers.image: "openai-codex" option, including automatic subscription detection and fallback logic.
• Added an optional provider parameter to generate_image to override the global image provider setting for a single request.
• Added OpenTelemetry log and metric export capabilities alongside existing trace exports, supporting standard OTLP environment variables.
• Added support for id-prefixed targets and keys in retry.fallbackChains wildcards (e.g., "openrouter/google/*").
• Added support for Shift+Enter in the session tree selector (/tree, /branch) to summarize and switch branches in a single step.
• Added the PI_CONFIG_FILES environment variable to load settings overlays before --config overlays.

Changed

• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.
• Changed bundled TTSR rules to warn instead of interrupting generation.
• Renamed the system prompt's project-context section wrapper from <context> to <repo-rules> to prevent XML tag collisions with in-band tool dialects.
• Renamed the /extensions dashboard tab "Agents (standard)" to "Agent Dirs (.agent/.agents)" to clarify its purpose.
• Optimized performance by reducing concurrent subagent update CPU usage, skipping unnecessary title generation in non-interactive hosts, and memoizing convertToLlm conversions over settled history.
• Improved the display of read xd:// calls by rendering them in a compact grouped view instead of full tool-execution cards.
• Made the hashline seen-line guard opt-in and off by default via edit.enforceSeenLines.

Fixed

• Fixed large session histories blocking appserver socket startup and repeatedly parsing every transcript body. Startup now publishes the transport before discovery, sends a complete initial inventory after bounded previews finish, and loads a transcript body only when that session is opened. Oversized non-header records are skipped without hiding the whole session.
• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver sessions losing their RPC child when inline image payloads made a lifecycle or durable-entry notification exceed the one MiB line ceiling. Managed children now omit only redundant image bytes from their internal stdout notifications and mark the frame, while the full images remain unchanged in OMP's session and model context; even tiny canonical image payloads are externalized for transcript reads.
• Fixed large appserver turns losing their terminal event when the redundant aggregate agent_end exceeded the RPC child's line or bounded-JSON structural limits. RPC now keeps a valid newest-message suffix with the original count and terminal status after durable entries; reade...

t4code-17.0.5-appserver-6

@oh-my-pi/pi-agent-core

Added

• Added a per-message token estimation cache to optimize performance by reusing token counts for settled message history, with automatic cache invalidation on message mutation.

Changed

• Improved tool execution control by making tool interruptibility resolvable per call, allowing side-effecting operations to complete while passive waits can yield to queued steering.

@oh-my-pi/pi-ai

Changed

• Changed Anthropic API-key requests to default to a 1-hour prompt-cache retention (using the extended-cache-ttl-2025-04-11 beta) to prevent cold-misses during idle sessions, with support for PI_CACHE_RETENTION values "short" and "none" to override this behavior.

Fixed

• Fixed transient OpenAI stream truncations by retrying once before output becomes replay-unsafe, preventing recoverable transport errors from failing the turn.
• Fixed native Kimi Code K3 thinking being disabled during named function selection by utilizing generic required tool choice.
• Fixed /login moonshot validating China-platform API keys against the international host instead of honoring MOONSHOT_BASE_URL.
• Fixed Anthropic session stickiness suppressing usage-based re-ranking indefinitely by gating stickiness on a 1-hour cache warmth window (configurable via ANTHROPIC_SESSION_STICKY_CACHE_WARM_MS) to restore proactive multi-account load balancing after long idle periods.
• Fixed credential ranking where clockless Anthropic usage windows incorrectly outranked clocked sibling credentials.
• Fixed tool request failures (HTTP 400) on local grammar-constrained OpenAI-compatible backends (such as llama.cpp, LM Studio, and vLLM) by widening bare boolean subschemas into a value-accepting primitive union.
• Fixed custom OAuth Anthropic-compatible endpoints receiving generated Claude Code fingerprint headers even when explicit header overrides were provided.
• Fixed active sessions for plan-gated OpenAI Codex models (Sol/Luna) silently re-routing to sibling OAuth accounts when usage headroom changed, ensuring session stickiness is preserved as long as the preferred credential remains usable and eligible.

@oh-my-pi/app-wire

Added

• Add a bounded, ordered browser preview contract with backend-neutral authority metadata, capture chunking, tab controls, input actions, policy checks, cooperative leases, and human handoff.
• Add the negotiated transcript.search feature with strict, bounded host search and session context commands.
• Add the strict, bounded SessionRef.attention summary for cross-session pending questions, approvals, plan reviews, and latest root outcomes.

@oh-my-pi/pi-catalog

Added

• Added an Anthropic compatibility flag to allow non-official OAuth endpoints to opt into configured Claude Code fingerprint header overrides.

Fixed

• Fixed a security issue where sensitive provider-defined request headers (such as API keys or credentials) were serialized in plaintext within the model cache (models.db). The cache now omits these headers, securely invalidates older cached rows, and restores or refetches them dynamically.
• Fixed OpenAI Codex discovery to respect caller-supplied fetch configurations (such as proxies or custom CAs) and correctly replace stale bundled models with the authenticated account catalog.
• Fixed stream timeouts and retry loops during long prefills on local loopback or RFC1918 backends (such as litellm proxies fronting local servers) by applying the local stream-timeout floor to these backends.
• Fixed Kimi K3 models served through generic OpenAI-compatible routes exposing unsupported reasoning efforts instead of the mandatory low/high/max scale.

@oh-my-pi/pi-coding-agent

Added

• Added profile-local cross-session transcript search with redacted snippets, bounded historical context, and filters for project, role, archive state, and time.
• Added redacted OpenAI Codex transport state to RPC session diagnostics, including actual transport, fallback, context-shape, and request-size counters without credentials or provider payloads.
• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.
• Added support for Codex (ChatGPT subscription) in generate_image via the providers.image: "openai-codex" option, including automatic subscription detection and fallback logic.
• Added an optional provider parameter to generate_image to override the global image provider setting for a single request.
• Added OpenTelemetry log and metric export capabilities alongside existing trace exports, supporting standard OTLP environment variables.
• Added support for id-prefixed targets and keys in retry.fallbackChains wildcards (e.g., "openrouter/google/*").
• Added support for Shift+Enter in the session tree selector (/tree, /branch) to summarize and switch branches in a single step.
• Added the PI_CONFIG_FILES environment variable to load settings overlays before --config overlays.

Changed

• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.
• Changed bundled TTSR rules to warn instead of interrupting generation.
• Renamed the system prompt's project-context section wrapper from <context> to <repo-rules> to prevent XML tag collisions with in-band tool dialects.
• Renamed the /extensions dashboard tab "Agents (standard)" to "Agent Dirs (.agent/.agents)" to clarify its purpose.
• Optimized performance by reducing concurrent subagent update CPU usage, skipping unnecessary title generation in non-interactive hosts, and memoizing convertToLlm conversions over settled history.
• Improved the display of read xd:// calls by rendering them in a compact grouped view instead of full tool-execution cards.
• Made the hashline seen-line guard opt-in and off by default via edit.enforceSeenLines.

Fixed

• Fixed large session histories blocking appserver socket startup and repeatedly parsing every transcript body. Startup now publishes the transport before discovery, sends a complete initial inventory after bounded previews finish, and loads a transcript body only when that session is opened. Oversized non-header records are skipped without hiding the whole session.
• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver sessions losing their RPC child when inline image payloads made a lifecycle or durable-entry notification exceed the one MiB line ceiling. Managed children now omit only redundant image bytes from their internal stdout notifications and mark the frame, while the full images remain unchanged in OMP's session and model context; even tiny canonical image payloads are externalized for transcript reads.
• Fixed large appserver turns losing their terminal event when the redundant aggregate agent_end exceeded the RPC child's line or bounded-JSON structural limits. RPC now keeps a valid newest-message suffix with the original count and terminal status after durable entries; reader failures reap their child; crashes remain closed until child exit, then become restartable without a state probe; confirmed close settles transient state, waits for child exit, and is idempotent; and the de...

t4code-17.0.5-appserver-5

@oh-my-pi/pi-agent-core

Added

• Added a per-message token estimation cache to optimize performance by reusing token counts for settled message history, with automatic cache invalidation on message mutation.

Changed

• Improved tool execution control by making tool interruptibility resolvable per call, allowing side-effecting operations to complete while passive waits can yield to queued steering.

@oh-my-pi/pi-ai

Changed

• Changed Anthropic API-key requests to default to a 1-hour prompt-cache retention (using the extended-cache-ttl-2025-04-11 beta) to prevent cold-misses during idle sessions, with support for PI_CACHE_RETENTION values "short" and "none" to override this behavior.

Fixed

• Fixed transient OpenAI stream truncations by retrying once before output becomes replay-unsafe, preventing recoverable transport errors from failing the turn.
• Fixed native Kimi Code K3 thinking being disabled during named function selection by utilizing generic required tool choice.
• Fixed /login moonshot validating China-platform API keys against the international host instead of honoring MOONSHOT_BASE_URL.
• Fixed Anthropic session stickiness suppressing usage-based re-ranking indefinitely by gating stickiness on a 1-hour cache warmth window (configurable via ANTHROPIC_SESSION_STICKY_CACHE_WARM_MS) to restore proactive multi-account load balancing after long idle periods.
• Fixed credential ranking where clockless Anthropic usage windows incorrectly outranked clocked sibling credentials.
• Fixed tool request failures (HTTP 400) on local grammar-constrained OpenAI-compatible backends (such as llama.cpp, LM Studio, and vLLM) by widening bare boolean subschemas into a value-accepting primitive union.
• Fixed custom OAuth Anthropic-compatible endpoints receiving generated Claude Code fingerprint headers even when explicit header overrides were provided.
• Fixed active sessions for plan-gated OpenAI Codex models (Sol/Luna) silently re-routing to sibling OAuth accounts when usage headroom changed, ensuring session stickiness is preserved as long as the preferred credential remains usable and eligible.

@oh-my-pi/app-wire

Added

• Add a bounded, ordered browser preview contract with backend-neutral authority metadata, capture chunking, tab controls, input actions, policy checks, cooperative leases, and human handoff.
• Add the negotiated transcript.search feature with strict, bounded host search and session context commands.
• Add the strict, bounded SessionRef.attention summary for cross-session pending questions, approvals, plan reviews, and latest root outcomes.

@oh-my-pi/pi-catalog

Added

• Added an Anthropic compatibility flag to allow non-official OAuth endpoints to opt into configured Claude Code fingerprint header overrides.

Fixed

• Fixed a security issue where sensitive provider-defined request headers (such as API keys or credentials) were serialized in plaintext within the model cache (models.db). The cache now omits these headers, securely invalidates older cached rows, and restores or refetches them dynamically.
• Fixed OpenAI Codex discovery to respect caller-supplied fetch configurations (such as proxies or custom CAs) and correctly replace stale bundled models with the authenticated account catalog.
• Fixed stream timeouts and retry loops during long prefills on local loopback or RFC1918 backends (such as litellm proxies fronting local servers) by applying the local stream-timeout floor to these backends.
• Fixed Kimi K3 models served through generic OpenAI-compatible routes exposing unsupported reasoning efforts instead of the mandatory low/high/max scale.

@oh-my-pi/pi-coding-agent

Added

• Added profile-local cross-session transcript search with redacted snippets, bounded historical context, and filters for project, role, archive state, and time.
• Added redacted OpenAI Codex transport state to RPC session diagnostics, including actual transport, fallback, context-shape, and request-size counters without credentials or provider payloads.
• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.
• Added support for Codex (ChatGPT subscription) in generate_image via the providers.image: "openai-codex" option, including automatic subscription detection and fallback logic.
• Added an optional provider parameter to generate_image to override the global image provider setting for a single request.
• Added OpenTelemetry log and metric export capabilities alongside existing trace exports, supporting standard OTLP environment variables.
• Added support for id-prefixed targets and keys in retry.fallbackChains wildcards (e.g., "openrouter/google/*").
• Added support for Shift+Enter in the session tree selector (/tree, /branch) to summarize and switch branches in a single step.
• Added the PI_CONFIG_FILES environment variable to load settings overlays before --config overlays.

Changed

• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.
• Changed bundled TTSR rules to warn instead of interrupting generation.
• Renamed the system prompt's project-context section wrapper from <context> to <repo-rules> to prevent XML tag collisions with in-band tool dialects.
• Renamed the /extensions dashboard tab "Agents (standard)" to "Agent Dirs (.agent/.agents)" to clarify its purpose.
• Optimized performance by reducing concurrent subagent update CPU usage, skipping unnecessary title generation in non-interactive hosts, and memoizing convertToLlm conversions over settled history.
• Improved the display of read xd:// calls by rendering them in a compact grouped view instead of full tool-execution cards.
• Made the hashline seen-line guard opt-in and off by default via edit.enforceSeenLines.

Fixed

• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver sessions losing their RPC child when inline image payloads made a lifecycle or durable-entry notification exceed the one MiB line ceiling. Managed children now omit only redundant image bytes from their internal stdout notifications and mark the frame, while the full images remain unchanged in OMP's session and model context; even tiny canonical image payloads are externalized for transcript reads.
• Fixed large appserver turns losing their terminal event when the redundant aggregate agent_end exceeded the RPC child's line or bounded-JSON structural limits. RPC now keeps a valid newest-message suffix with the original count and terminal status after durable entries; reader failures reap their child; crashes remain closed until child exit, then become restartable without a state probe; confirmed close settles transient state, waits for child exit, and is idempotent; and the desktop catalog exposes that close command.
• Fixed the desktop catalog omitting implemented session creation and lifecycle commands, which caused catalog-driven clients to hide working create, rename, archive, restore, and delete controls.
• Fixed RPC mode exiting on process termination, extension shutdown, or stdin EOF without disposing its session, which left a fres...

t4code-17.0.5-appserver-3

@oh-my-pi/pi-agent-core

Added

• Added a per-message token estimation cache to optimize performance by reusing token counts for settled message history, with automatic cache invalidation on message mutation.

Changed

• Improved tool execution control by making tool interruptibility resolvable per call, allowing side-effecting operations to complete while passive waits can yield to queued steering.

@oh-my-pi/pi-ai

Changed

• Changed Anthropic API-key requests to default to a 1-hour prompt-cache retention (using the extended-cache-ttl-2025-04-11 beta) to prevent cold-misses during idle sessions, with support for PI_CACHE_RETENTION values "short" and "none" to override this behavior.

Fixed

• Fixed transient OpenAI stream truncations by retrying once before output becomes replay-unsafe, preventing recoverable transport errors from failing the turn.
• Fixed native Kimi Code K3 thinking being disabled during named function selection by utilizing generic required tool choice.
• Fixed /login moonshot validating China-platform API keys against the international host instead of honoring MOONSHOT_BASE_URL.
• Fixed Anthropic session stickiness suppressing usage-based re-ranking indefinitely by gating stickiness on a 1-hour cache warmth window (configurable via ANTHROPIC_SESSION_STICKY_CACHE_WARM_MS) to restore proactive multi-account load balancing after long idle periods.
• Fixed credential ranking where clockless Anthropic usage windows incorrectly outranked clocked sibling credentials.
• Fixed tool request failures (HTTP 400) on local grammar-constrained OpenAI-compatible backends (such as llama.cpp, LM Studio, and vLLM) by widening bare boolean subschemas into a value-accepting primitive union.
• Fixed custom OAuth Anthropic-compatible endpoints receiving generated Claude Code fingerprint headers even when explicit header overrides were provided.
• Fixed active sessions for plan-gated OpenAI Codex models (Sol/Luna) silently re-routing to sibling OAuth accounts when usage headroom changed, ensuring session stickiness is preserved as long as the preferred credential remains usable and eligible.

@oh-my-pi/app-wire

Added

• Add the negotiated transcript.search feature with strict, bounded host search and session context commands.
• Add the strict, bounded SessionRef.attention summary for cross-session pending questions, approvals, plan reviews, and latest root outcomes.

@oh-my-pi/pi-catalog

Added

• Added an Anthropic compatibility flag to allow non-official OAuth endpoints to opt into configured Claude Code fingerprint header overrides.

Fixed

• Fixed a security issue where sensitive provider-defined request headers (such as API keys or credentials) were serialized in plaintext within the model cache (models.db). The cache now omits these headers, securely invalidates older cached rows, and restores or refetches them dynamically.
• Fixed OpenAI Codex discovery to respect caller-supplied fetch configurations (such as proxies or custom CAs) and correctly replace stale bundled models with the authenticated account catalog.
• Fixed stream timeouts and retry loops during long prefills on local loopback or RFC1918 backends (such as litellm proxies fronting local servers) by applying the local stream-timeout floor to these backends.
• Fixed Kimi K3 models served through generic OpenAI-compatible routes exposing unsupported reasoning efforts instead of the mandatory low/high/max scale.

@oh-my-pi/pi-coding-agent

Added

• Added profile-local cross-session transcript search with redacted snippets, bounded historical context, and filters for project, role, archive state, and time.
• Added redacted OpenAI Codex transport state to RPC session diagnostics, including actual transport, fallback, context-shape, and request-size counters without credentials or provider payloads.
• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.
• Added support for Codex (ChatGPT subscription) in generate_image via the providers.image: "openai-codex" option, including automatic subscription detection and fallback logic.
• Added an optional provider parameter to generate_image to override the global image provider setting for a single request.
• Added OpenTelemetry log and metric export capabilities alongside existing trace exports, supporting standard OTLP environment variables.
• Added support for id-prefixed targets and keys in retry.fallbackChains wildcards (e.g., "openrouter/google/*").
• Added support for Shift+Enter in the session tree selector (/tree, /branch) to summarize and switch branches in a single step.
• Added the PI_CONFIG_FILES environment variable to load settings overlays before --config overlays.

Changed

• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.
• Changed bundled TTSR rules to warn instead of interrupting generation.
• Renamed the system prompt's project-context section wrapper from <context> to <repo-rules> to prevent XML tag collisions with in-band tool dialects.
• Renamed the /extensions dashboard tab "Agents (standard)" to "Agent Dirs (.agent/.agents)" to clarify its purpose.
• Optimized performance by reducing concurrent subagent update CPU usage, skipping unnecessary title generation in non-interactive hosts, and memoizing convertToLlm conversions over settled history.
• Improved the display of read xd:// calls by rendering them in a compact grouped view instead of full tool-execution cards.
• Made the hashline seen-line guard opt-in and off by default via edit.enforceSeenLines.

Fixed

• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver sessions losing their RPC child when inline image payloads made a lifecycle or durable-entry notification exceed the one MiB line ceiling. Managed children now omit only redundant image bytes from their internal stdout notifications and mark the frame, while the full images remain unchanged in OMP's session and model context; even tiny canonical image payloads are externalized for transcript reads.
• Fixed large appserver turns losing their terminal event when the redundant aggregate agent_end exceeded the RPC child's line or bounded-JSON structural limits. RPC now keeps a valid newest-message suffix with the original count and terminal status after durable entries; reader failures reap their child; crashes remain closed until child exit, then become restartable without a state probe; confirmed close settles transient state, waits for child exit, and is idempotent; and the desktop catalog exposes that close command.
• Fixed the desktop catalog omitting implemented session creation and lifecycle commands, which caused catalog-driven clients to hide working create, rename, archive, restore, and delete controls.
• Fixed RPC mode exiting on process termination, extension shutdown, or stdin EOF without disposing its session, which left a fresh writer lock behind and caused appserver lifecycle operations to report session_locked after quiescing an idle child.
• Fixed appserver sessions briefly returning to idle at intermediate tu...

t4code-17.0.5-appserver-2

@oh-my-pi/pi-agent-core

Added

• Added a per-message token estimation cache to optimize performance by reusing token counts for settled message history, with automatic cache invalidation on message mutation.

Changed

• Improved tool execution control by making tool interruptibility resolvable per call, allowing side-effecting operations to complete while passive waits can yield to queued steering.

@oh-my-pi/pi-ai

Changed

• Changed Anthropic API-key requests to default to a 1-hour prompt-cache retention (using the extended-cache-ttl-2025-04-11 beta) to prevent cold-misses during idle sessions, with support for PI_CACHE_RETENTION values "short" and "none" to override this behavior.

Fixed

• Fixed transient OpenAI stream truncations by retrying once before output becomes replay-unsafe, preventing recoverable transport errors from failing the turn.
• Fixed native Kimi Code K3 thinking being disabled during named function selection by utilizing generic required tool choice.
• Fixed /login moonshot validating China-platform API keys against the international host instead of honoring MOONSHOT_BASE_URL.
• Fixed Anthropic session stickiness suppressing usage-based re-ranking indefinitely by gating stickiness on a 1-hour cache warmth window (configurable via ANTHROPIC_SESSION_STICKY_CACHE_WARM_MS) to restore proactive multi-account load balancing after long idle periods.
• Fixed credential ranking where clockless Anthropic usage windows incorrectly outranked clocked sibling credentials.
• Fixed tool request failures (HTTP 400) on local grammar-constrained OpenAI-compatible backends (such as llama.cpp, LM Studio, and vLLM) by widening bare boolean subschemas into a value-accepting primitive union.
• Fixed custom OAuth Anthropic-compatible endpoints receiving generated Claude Code fingerprint headers even when explicit header overrides were provided.
• Fixed active sessions for plan-gated OpenAI Codex models (Sol/Luna) silently re-routing to sibling OAuth accounts when usage headroom changed, ensuring session stickiness is preserved as long as the preferred credential remains usable and eligible.

@oh-my-pi/pi-catalog

Added

• Added an Anthropic compatibility flag to allow non-official OAuth endpoints to opt into configured Claude Code fingerprint header overrides.

Fixed

• Fixed a security issue where sensitive provider-defined request headers (such as API keys or credentials) were serialized in plaintext within the model cache (models.db). The cache now omits these headers, securely invalidates older cached rows, and restores or refetches them dynamically.
• Fixed OpenAI Codex discovery to respect caller-supplied fetch configurations (such as proxies or custom CAs) and correctly replace stale bundled models with the authenticated account catalog.
• Fixed stream timeouts and retry loops during long prefills on local loopback or RFC1918 backends (such as litellm proxies fronting local servers) by applying the local stream-timeout floor to these backends.
• Fixed Kimi K3 models served through generic OpenAI-compatible routes exposing unsupported reasoning efforts instead of the mandatory low/high/max scale.

@oh-my-pi/pi-coding-agent

Added

• Added redacted OpenAI Codex transport state to RPC session diagnostics, including actual transport, fallback, context-shape, and request-size counters without credentials or provider payloads.
• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.
• Added support for Codex (ChatGPT subscription) in generate_image via the providers.image: "openai-codex" option, including automatic subscription detection and fallback logic.
• Added an optional provider parameter to generate_image to override the global image provider setting for a single request.
• Added OpenTelemetry log and metric export capabilities alongside existing trace exports, supporting standard OTLP environment variables.
• Added support for id-prefixed targets and keys in retry.fallbackChains wildcards (e.g., "openrouter/google/*").
• Added support for Shift+Enter in the session tree selector (/tree, /branch) to summarize and switch branches in a single step.
• Added the PI_CONFIG_FILES environment variable to load settings overlays before --config overlays.

Changed

• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.
• Changed bundled TTSR rules to warn instead of interrupting generation.
• Renamed the system prompt's project-context section wrapper from <context> to <repo-rules> to prevent XML tag collisions with in-band tool dialects.
• Renamed the /extensions dashboard tab "Agents (standard)" to "Agent Dirs (.agent/.agents)" to clarify its purpose.
• Optimized performance by reducing concurrent subagent update CPU usage, skipping unnecessary title generation in non-interactive hosts, and memoizing convertToLlm conversions over settled history.
• Improved the display of read xd:// calls by rendering them in a compact grouped view instead of full tool-execution cards.
• Made the hashline seen-line guard opt-in and off by default via edit.enforceSeenLines.

Fixed

• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver sessions losing their RPC child when inline image payloads made a lifecycle or durable-entry notification exceed the one MiB line ceiling. Managed children now omit only redundant image bytes from their internal stdout notifications and mark the frame, while the full images remain unchanged in OMP's session and model context; even tiny canonical image payloads are externalized for transcript reads.
• Fixed large appserver turns losing their terminal event when the redundant aggregate agent_end exceeded the RPC child's line or bounded-JSON structural limits. RPC now keeps a valid newest-message suffix with the original count and terminal status after durable entries; reader failures reap their child; crashes remain closed until child exit, then become restartable without a state probe; confirmed close settles transient state, waits for child exit, and is idempotent; and the desktop catalog exposes that close command.
• Fixed the desktop catalog omitting implemented session creation and lifecycle commands, which caused catalog-driven clients to hide working create, rename, archive, restore, and delete controls.
• Fixed RPC mode exiting on process termination, extension shutdown, or stdin EOF without disposing its session, which left a fresh writer lock behind and caused appserver lifecycle operations to report session_locked after quiescing an idle child.
• Fixed appserver sessions briefly returning to idle at intermediate turn boundaries during tool-driven runs; the active prompt now remains owned until final agent.end, a correlated local-only result or failure, successful cancellation, closure, or child termination.
• Fixed RPC prompt failures emitting a second response with an already-settled request id, which caused strict child supervisors to terminate otherwise healthy sessions. Late failures now use an exact-ID asynchronous prompt_result frame, which appserver...

t4code-17.0.4-appserver-5

@oh-my-pi/pi-ai

Fixed

• Fixed Kimi Code usage reports dropping the 5h window reset time (omp usage showed no "resets in …" for the 5h limit): the API returns resetTime on the limit detail, not on window, so the parsed row-level reset is now carried onto the window when the window itself has none.
• Made Kimi device-id persistence best-effort: a missing or unwritable ~/.omp/agent directory no longer throws during Kimi header construction, which silently nulled every kimi-code usage probe on fresh installs.
• Coerced boolean tool-schema subschemas to MFJS object forms for native Moonshot/Kimi endpoints, preventing the task tool's outputSchema field from causing HTTP 400 responses (#5952).

@oh-my-pi/pi-catalog

Changed

• Kimi-family models now use MFJS tool schema on all hosts, including proxies like OpenRouter that forward schemas to Moonshot

@oh-my-pi/pi-coding-agent

Added

• Added redacted OpenAI Codex transport state to RPC session diagnostics, including actual transport, fallback, context-shape, and request-size counters without credentials or provider payloads.
• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.

Changed

• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.

Fixed

• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver sessions losing their RPC child when inline image payloads made a lifecycle or durable-entry notification exceed the one MiB line ceiling. Managed children now omit only redundant image bytes from their internal stdout notifications and mark the frame, while the full images remain unchanged in OMP's session and model context; even tiny canonical image payloads are externalized for transcript reads.
• Fixed large appserver turns losing their terminal event when the redundant aggregate agent_end exceeded the RPC child's line or bounded-JSON structural limits. RPC now keeps a valid newest-message suffix with the original count and terminal status after durable entries; reader failures reap their child; crashes remain closed until child exit, then become restartable without a state probe; confirmed close settles transient state, waits for child exit, and is idempotent; and the desktop catalog exposes that close command.
• Fixed the desktop catalog omitting implemented session creation and lifecycle commands, which caused catalog-driven clients to hide working create, rename, archive, restore, and delete controls.
• Fixed RPC mode exiting on process termination, extension shutdown, or stdin EOF without disposing its session, which left a fresh writer lock behind and caused appserver lifecycle operations to report session_locked after quiescing an idle child.
• Fixed appserver sessions briefly returning to idle at intermediate turn boundaries during tool-driven runs; the active prompt now remains owned until final agent.end, a correlated local-only result or failure, successful cancellation, closure, or child termination.
• Fixed RPC prompt failures emitting a second response with an already-settled request id, which caused strict child supervisors to terminate otherwise healthy sessions. Late failures now use an exact-ID asynchronous prompt_result frame, which appserver surfaces as a sanitized turn.error without allowing stale results to settle newer work.
• Fixed accepted prompts disappearing from desktop and mobile clients while compaction or reconnect recovery was in progress. Prompt, steer, and follow-up text now enters a bounded ordered projection immediately, survives the sending client disconnecting, and retires through an exact durable-entry settlement or explicit discard event.
• Fixed one slow appserver client delaying transcript and pending-prompt delivery to other attached clients.
• Fixed session.attach acknowledgement and replay races by preparing output before success, catching up from its baseline, rebuilding cached delivery after revalidating session existence, and bounding large snapshots and replays so desktop and remote clients remain connected.
• Fixed compiled appserver welcome frames reporting placeholder local identities instead of the owning OMP and appserver versions and build kinds.
• Fixed active or locked sessions disappearing from observational session lists while keeping writable resume selection lock-safe.
• Fixed one malformed or crash-truncated transcript entry hiding an otherwise valid session from appserver discovery.
• Fixed concurrent Time-Traveling Stream Rules attaching AST reminders to the wrong tool result when multiple tool calls completed together.
• Fixed explicit session rewrites and stale-lock takeover failures retaining lock ownership after the write or cleanup failed.
• Fixed detached and clone session managers releasing another manager's live writer lock after restoring shared session state; the source keeps ownership until normal disposal, and the clone acquires the lock only if it becomes the late result writer.
• Included dependency patch files in Bun CI cache keys so a changed patch cannot reuse stale unpatched packages.
• Fixed appserver discovery counting nested advisor and subagent transcripts as main sessions.
• Fixed remote fast-mode changes bypassing the controller-lease checks used by other session mutations.
• Fixed desktop session lifecycle races by fencing mutations before asynchronous work, refusing active or queued sessions, closing owned terminals and RPC children before removal, and keeping host-wide session indexes in sync with external transcript changes.
• Fixed remote frame transforms reordering lifecycle deltas and responses; each connection now preserves send order and fails closed when a transform stalls.
• Fixed bundled Linux ffmpeg recording by selecting its available ALSA input when PulseAudio support is absent, and surfaced recorder stderr when capture fails (#5907).
• Session load now skips the recursive async blob-ref resolver for entries with no blob:sha256: references. A cheap synchronous precheck gates the walk per entry (preserving the previous per-entry initiation order under synchronous store mutation), so text-heavy histories no longer pay the Promise.all tree descent for every non-session entry (#5922).
• Fixed task tool schemas emitting boolean subschemas that llama.cpp grammar generation cannot parse (#5957).
• Fixed the transcript keeping finalized assistant blocks in the live compose walk after their rows entered native terminal scrollback, making each stream tick's TranscriptContainer.render depth-linear in session length. Fully committed finalized blocks are now compacted out of the local frame regardless of post-finalize version tracking; a later mutation no longer recommits on ordinary frames (no duplication) and rehydrates on the next destructive full replay (no loss). Compose cost for a live tail tick is now flat as depth grows (bench/transcript-compose.bench.ts:...

t4code-17.0.4-appserver-4

@oh-my-pi/pi-ai

Fixed

• Fixed Kimi Code usage reports dropping the 5h window reset time (omp usage showed no "resets in …" for the 5h limit): the API returns resetTime on the limit detail, not on window, so the parsed row-level reset is now carried onto the window when the window itself has none.
• Made Kimi device-id persistence best-effort: a missing or unwritable ~/.omp/agent directory no longer throws during Kimi header construction, which silently nulled every kimi-code usage probe on fresh installs.
• Coerced boolean tool-schema subschemas to MFJS object forms for native Moonshot/Kimi endpoints, preventing the task tool's outputSchema field from causing HTTP 400 responses (#5952).

@oh-my-pi/app-wire

Added

• Add a frozen Agent View lifecycle corpus covering started, running, completed, parked, resumed, and cancelled worker states for cross-client compatibility verification.

@oh-my-pi/pi-catalog

Changed

• Kimi-family models now use MFJS tool schema on all hosts, including proxies like OpenRouter that forward schemas to Moonshot

@oh-my-pi/pi-coding-agent

Added

• Added an owner-socket appserver drain-if-idle maintenance fence that atomically refuses new client work only after the exact appserver identity is confirmed idle.
• Added desktop appserver controls for session model, thinking level, and fast mode, with live settings, model-role, and task-agent catalog metadata.
• Added desktop session archive, restore, and delete backed by private profile metadata and crash-recoverable same-filesystem tombstones.
• Added the /continue-in-t4 slash command, which persists the current TUI session and exits through the normal clean teardown so the session can be continued in T4; nothing is taken over while the TUI is still running.
• Added lock-aware appserver observation of sessions owned by another OMP process, such as an external TUI. The session is followed read-only from complete transcript records on disk; the appserver takes writable control only when the owner's lock is freshly missing — live, suspect, stale, or malformed locks are never seized — and the session stays read-only in a reconciling state until the promoted child reports it loaded the exact transcript the observer last saw.

Changed

• Bounded incremental subagent transcript RPC reads to complete 384 KiB JSONL chunks, with byte cursors and an option to omit the redundant message view.
• Appserver now admits one unresolved session.prompt per session and returns session_busy before a second prompt reaches the RPC child; active clients can use session.steer or session.followUp to add work to the running agent.
• Bounded completed appserver command outcomes to a five-minute replay window and a 1,024-entry least-recently-used cache. Replays do not extend expiry, and pending commands are never evicted.
• Reused content-addressed native addons from prior successful branch builds and split Rust checks from native packaging, shortening appserver-only release runs without weakening the release gates.
• Appserver remote listener mode, bind address, port, and browser origins now persist in host-local agent/local/config.yml instead of synchronized profile config.
• Session-scoped Bash and Python execution now read output limits from the active session settings instead of process-global defaults.

Fixed

• Fixed appserver Agent View cancellation aborting the parent RPC process instead of the selected live subagent. Cancellation now targets only validated worker subagents, survives the requesting client disconnecting after dispatch, publishes a non-resumable terminal cancelled lifecycle state before hard release, and cannot revive or park an agent after that release begins.
• Fixed macOS appserver startup rejecting Node's /var/folders temporary paths because /var is a system symlink, while retaining rejection of nested user-controlled symlinks.
• Fixed appserver durable tool results dropping structured content, renderer details, and error state. Bounded, redacted canonical fields now survive transcript projection without embedding image bytes, while legacy text output remains available.
• Fixed image-bearing appserver sessions losing their RPC child when inline image payloads made a lifecycle or durable-entry notification exceed the one MiB line ceiling. Managed children now omit only redundant image bytes from their internal stdout notifications and mark the frame, while the full images remain unchanged in OMP's session and model context; even tiny canonical image payloads are externalized for transcript reads.
• Fixed large appserver turns losing their terminal event when the redundant aggregate agent_end exceeded the RPC child's line or bounded-JSON structural limits. RPC now keeps a valid newest-message suffix with the original count and terminal status after durable entries; reader failures reap their child; crashes remain closed until child exit, then become restartable without a state probe; confirmed close settles transient state, waits for child exit, and is idempotent; and the desktop catalog exposes that close command.
• Fixed the desktop catalog omitting implemented session creation and lifecycle commands, which caused catalog-driven clients to hide working create, rename, archive, restore, and delete controls.
• Fixed RPC mode exiting on process termination, extension shutdown, or stdin EOF without disposing its session, which left a fresh writer lock behind and caused appserver lifecycle operations to report session_locked after quiescing an idle child.
• Fixed appserver sessions briefly returning to idle at intermediate turn boundaries during tool-driven runs; the active prompt now remains owned until final agent.end, a correlated local-only result or failure, successful cancellation, closure, or child termination.
• Fixed RPC prompt failures emitting a second response with an already-settled request id, which caused strict child supervisors to terminate otherwise healthy sessions. Late failures now use an exact-ID asynchronous prompt_result frame, which appserver surfaces as a sanitized turn.error without allowing stale results to settle newer work.
• Fixed accepted prompts disappearing from desktop and mobile clients while compaction or reconnect recovery was in progress. Prompt, steer, and follow-up text now enters a bounded ordered projection immediately, survives the sending client disconnecting, and retires through an exact durable-entry settlement or explicit discard event.
• Fixed one slow appserver client delaying transcript and pending-prompt delivery to other attached clients.
• Fixed session.attach acknowledgement and replay races by preparing output before success, catching up from its baseline, rebuilding cached delivery after revalidating session existence, and bounding large snapshots and replays so desktop and remote clients remain connected.
• Fixed compiled appserver welcome frames reporting placeholder local identities instead of the owning OMP and appserver versions and build kinds.
• Fixed active or locked sessions disappearing from observational session lists while keeping writable resume selection lock-safe.
• Fixed one malformed or crash-truncated transcript entry hiding an otherwise valid session from appserver discovery.
• Fixed concurrent Time-Traveling Stream Rules attaching AST reminders to the wrong tool result when multiple tool calls completed together.
• Fixed explicit session rewrites and stale-lock takeover failures retaining lock ownership after the write or cleanup failed.
• Fixed detached and clone session managers releasing another manager's live writer lock after restoring shared session state; the source keeps ownership until normal disposal, and the clone acquires the lock only if it becomes the late result writer.
• Included dependency patch files in Bun CI cache keys so a changed patch cannot reuse stale unpatched packages.
• Fixed appserver discovery counting nested advisor and subagent transcripts as main sessions.
• Fixed remote fast-mode changes bypassing the controller-lease checks used by other session mutations.
• Fixed desktop session lifecycle races by fencing mutations before asynchronous work, refusing active or queued sessions, closing owned terminals and RPC children before removal, and keeping host-wide session indexes in sync with external transcript changes.
• Fixed remote frame transforms reordering lifecycle deltas and responses; each connection now preserves send order and fails closed when a transform stalls.
• Fixed bundled Linux ffmpeg recording by selecting its available ALSA input when PulseAudio support is absent, and surfaced recorder stderr when capture fails (#5907).
• Session load now skips the recursive async blob-ref resolver for entries with no blob:sha256: references. A cheap synchronous precheck gates the walk per entry (preserving the previous per-entry initiation order under synchronous store mutation), so text-heavy histories no longer pay the Promise.all tree descent for every non-session entry (#5922).
• Fixed task tool schemas emitting boolean subschemas that llama.cpp grammar generation cannot parse (#5957).
• Fixed the transcript keeping finalized assistant blocks in the live compose walk after their rows entered native terminal scrollback, making each stream tick's TranscriptContainer.render depth-linear in session length. Fully committed finalized blocks are now compacted out of the local frame regardless of post-finalize version tracking; a later mutation no longer recommits on ordinary frames (no duplication) and rehydrates on the next destructive full replay (no loss). Compose cost for a live tail tick is now flat as depth grows (`bench/transcript-compose...