@code-yeongyu/senpi 2026.10.16-2 → 2026.10.16-4

package/CHANGELOG.md

@@ -12,6 +12,54 @@
 
 ### Removed
 
+## [2026.10.16-4] - 2026-10-16
+
+### Breaking Changes
+
+### Added
+
+- A permission prompt for a tool without its own parser offers a wider "Allow always": for a URL, that domain in this project ("Allow <tool> on <host> always in this project"); for other tools, the tool in this project. Commands, paths and built-in tools never get it.
+
+- `RpcClient.abort(options)` sends a typed abort (`effectId`, `awaitIdle`, `clearQueue`, `graceMs`) and returns the host's answer (`stopped`, `idle`, `cleared`, receipt); `RpcClient.getEffectStatus(effectId, { sessionPath? })` reads an effect receipt. `abort()` without options is unchanged ([#3453](https://github.com/code-yeongyu/senpi/issues/3453)).
+
+### Changed
+
+### Fixed
+
+- Reopening a session right after parking it no longer fails with `session_path_in_use`. A host no longer reopens the sessions it parked itself (those are for its next generation), and an open that arrives while another open of the same path is still claiming it waits and attaches instead of being refused; if that claim stalls past the close grace window, the refusal says `claim_in_progress` with a retry delay, and no racing open gets a refusal without a reason ([#3485](https://github.com/code-yeongyu/senpi/issues/3485), [#3490](https://github.com/code-yeongyu/senpi/pull/3490)).
+
+- A slow model no longer sends every session to its fallback for up to 30 minutes. Failures that arrive while the fallback circuit is already open no longer double its cooldown: the circuit takes one backoff step per open episode. A session that never failed on the model keeps it; only a billing, quota or account-limit failure moves other sessions off it, while a Retry-After stays with the session that received it ([#3467](https://github.com/code-yeongyu/senpi/issues/3467)).
+
+- Bounded what context files and the skills catalog add to every request: AGENTS.md/CLAUDE.md files are cut at 64 KiB each and 128 KiB in total, nearest file first, including truncation notes and closing code fences in the budgets, with newline-safe cuts (`contextFiles.maxBytesPerFile` / `contextFiles.maxTotalBytes` raise the limits). The skills section is capped at 48 KiB: above 100 visible skills or 64 KiB of descriptions it uses one-line descriptions, then drops descriptions and lists only the first skills that fit if needed, naming the omitted count. Size warnings appear in TUI resource notices and once on stderr at startup in non-interactive modes ([#3460](https://github.com/code-yeongyu/senpi/issues/3460), [#3463](https://github.com/code-yeongyu/senpi/pull/3463)).
+
+- `anthropic-subscription`: a transient failure (overload, server error, network) no longer rotates the account or blocks it in the shared credential store. The account is still retried twice in place; after that the turn fails on the same account, and the session's own retry and model fallback take over. Moving accounts wrote the whole conversation to a cold prompt cache, and the persisted block moved every other session off a healthy account. A plain rate limit (429), a one-model limit or a refill breaker still rotates the turn that hit it, but no longer moves other sessions off the account; only a confirmed account usage limit, billing or an auth failure does ([#3466](https://github.com/code-yeongyu/senpi/issues/3466)).
+
+- A session with the Full access preset no longer asks "Trust project folder?": it trusts its project for that session without saving the decision (a saved "do not trust" still wins). Home and the folders above it no longer count as project folders, so opening home as a project doesn't ask to trust your own global config. ([#3420](https://github.com/code-yeongyu/senpi/pull/3420))
+
+- `anthropic-subscription`: a spent account is now blocked until the reset its usage-limit message names ("resets in 1 hour", "resets 5am"), or, when the message names none, until the reset its fully used limit window last reported. Before, every account-level usage limit was blocked for 60 s. A pinned account therefore went back to the same spent limit every minute, and each round trip re-forked the session on the healthy account and re-wrote its prompt cache. The long block applies only to Claude Code's own usage-limit error and never lasts past that limit's window (5 hours for a session or 5-hour limit). A working account that serves a turn has its block cleared, and `/claude-account unblock <id>` lifts one by hand. Any other limit text and a plain rate limit keep the 60 s block ([#3462](https://github.com/code-yeongyu/senpi/issues/3462)).
+
+### Removed
+
+## [2026.10.16-3] - 2026-10-16
+
+### Breaking Changes
+
+- RPC: the outcome for input dropped from the queue is now `cleared`, written by `clear_queue`, `unload_session { clearQueue: true }` and `abort { clearQueue: true }` alike; 2026.10.16-2 wrote it as `cleared_by_abort`, and only for `abort`. Receipts written under the old name are still read and reported as they were stored, so an input cleared on 2026.10.16-2 is never restored on reopen. Clients matching `outcome.kind` should accept both values ([#3412](https://github.com/code-yeongyu/senpi/issues/3412)).
+
+### Added
+
+- Added the `vertex-garden` provider for DeepSeek V4.1 Flash and GLM 5.2 through Vertex Model Garden's global OpenAI-compatible endpoint, using shared Google Application Default Credentials token resolution without a token script or proxy. Availability follows the ADC chain: the lane is selectable while ADC is present but never counts as configured, so it cannot take the automatic default over a provider you configured. The request project comes from `GOOGLE_CLOUD_PROJECT`/`GCLOUD_PROJECT`, the ADC quota project, or `gcloud`'s configured project, and both models carry their published MaaS list prices. Marketplace partner models such as Claude and Grok are excluded because promotional credits do not apply ([#3296](https://github.com/code-yeongyu/senpi/pull/3296) by [@sleegme](https://github.com/sleegme)).
+
+- A turn that stopped making progress no longer runs forever: a session-level inactivity lease (`turn_inactivity_ms`, default 60 min, renewed only by tool execution events and admitted user input) and a hard maximum runtime (`turn_max_runtime_ms`, default 6 h) end a stuck turn as a system abort that carries a `deadline` detail on `agent_end`. Partial text is kept, the reason names the setting and the action to resume, and a goal is never auto-continued past the stop until the user sends a message ([#3227](https://github.com/code-yeongyu/senpi/issues/3227)).
+
+- Argon sessions get their own system prompt preset (`gemini-4-argon`, auto-resolved for every catalog id shape and forceable from settings). It is the default prompt with four spots tuned to the provider's published prompting guidance: the identity names the model, the terminal routing line is declared plain prose with the tool calls following in the same response, the Working the Task opening makes one `eval` cell with real control flow and `parallel(thunks)` the way a step is done when `eval` is available, the error rule retries transient failures and changes strategy on others, plus a plain workstation wording and a search-grounding paragraph when `web_search` is active. Every other model's prompt is byte-identical to before ([#3385](https://github.com/code-yeongyu/senpi/issues/3385)).
+
+### Changed
+
+### Fixed
+
+### Removed
+
 ## [2026.10.16-2] - 2026-10-16
 
 ### Breaking Changes