@code-yeongyu/senpi 2026.7.31 → 2026.8.3

package/docs/mcp.md

@@ -153,10 +153,8 @@ local `tool_search` for the session.
 | search (Tier-B) | `exposure:"search"` or `auto` above the threshold | Full catalog registered, ~135 tokens resident (`tool_search` only); matches promote next turn; promotions survive resume/compaction |
 | proxy (Tier-C) | `exposure:"proxy"` only — never `auto` | One `mcp_<server>` gateway tool (`search`/`describe`/`call` with JSON-string args); cheapest, but no provider-side argument validation |
 
-Skills can carry MCP servers too (an `mcp.json` sidecar next to SKILL.md, or a
-`mcp:` frontmatter block): those servers register with zero active tools and
-reveal their `includeTools` matches when the skill loads. A name collision
-with a configured server resolves in favor of your config.
+Skills can carry MCP servers too — see
+[Skill-carried MCP servers](#skill-carried-mcp-servers).
 
 ### Server sources
 
@@ -174,6 +172,67 @@ A name collision resolves as follows: trusted user config (including an
 declaration replaces an untrusted placeholder and records a diagnostic.
 Extension-declared servers use bare names (no prefix).
 
+## Skill-carried MCP servers
+
+A skill can bundle the MCP servers it needs, so the workflow and its tools
+travel as one package. Skill-declared servers register lazily with **zero
+active tools** — they cost no prompt tokens until the skill loads. Loading the
+skill (a `/skill:<name>` command, or the model reading its SKILL.md) reveals
+the tools matching its `includeTools` globs for the rest of the session.
+
+Two declaration forms exist; the sidecar wins when both are present.
+
+**`mcp.json` sidecar** (Amp-compatible) next to SKILL.md. Servers accept the
+same fields as [`mcpServers.<name>`](#server-fields-mcpserversname), plus
+`includeTools`:
+
+```json
+{
+  "mcpServers": {
+    "exa": {
+      "command": "npx",
+      "args": ["-y", "exa-mcp-server"],
+      "env": { "EXA_API_KEY": "${EXA_API_KEY}" },
+      "includeTools": ["web_search*"]
+    }
+  }
+}
+```
+
+The `mcpServers` wrapper is optional — a bare server-name map works too.
+
+**Frontmatter `mcp:` block** in SKILL.md:
+
+````markdown
+---
+name: exa-search
+description: Web search via Exa. Use for finding current documentation.
+mcp:
+  exa:
+    command: npx
+    args: ["-y", "exa-mcp-server"]
+    env:
+      EXA_API_KEY: ${EXA_API_KEY}
+    includeTools: ["web_search*"]
+---
+````
+
+Semantics:
+
+- `includeTools` is a glob allowlist (`*` wildcards) over server-side tool
+  names; default `["*"]` (every tool). Skill servers are forced into search
+  mode with no `directTools`, so nothing is active before the skill loads
+  regardless of any declared `exposure`.
+- Several skills declaring the same server name: the first configuration
+  wins; `includeTools` union-merges across the declaring skills, and loading
+  any one of them reveals the merged matches.
+- A name collision with a configured server resolves in favor of your config,
+  so a skill never overrides a server you already trust.
+- An unreadable sidecar or frontmatter `mcp:` block is skipped with a
+  warning; the skill itself still loads.
+- There is no unload signal: tools revealed by a skill stay active until the
+  session ends.
+
 ## Resources and prompts
 
 - `mcp_list_resources` / `mcp_read_resource` register automatically when a

package/docs/models.md

@@ -544,6 +544,7 @@ For providers with partial OpenAI compatibility, use the `compat` field.
 | `supportsDeveloperRole` | Use `developer` vs `system` role |
 | `supportsReasoningEffort` | Support for `reasoning_effort` parameter |
 | `supportsUsageInStreaming` | Supports `stream_options: { include_usage: true }` (default: `true`) |
+| `supportsFinishReason` | Whether streamed responses include `finish_reason`. When `false`, pi infers `stop` or `toolUse` when the stream ends. Default: `true`. |
 | `maxTokensField` | Use `max_completion_tokens` or `max_tokens` |
 | `requiresToolResultName` | Include `name` on tool result messages |
 | `requiresAssistantAfterToolResult` | Insert an assistant message before a user message after tool results |

package/docs/providers.md

@@ -34,17 +34,52 @@ Use `/logout` to clear credentials. Tokens are stored in `~/.pi/agent/auth.json`
 
 Anthropic subscription auth is active for Claude Pro/Max accounts. Third-party harness usage draws from [extra usage](https://claude.ai/settings/usage) and is billed per token, not against Claude plan limits.
 
-### Claude Agent SDK
+### Claude SDK OAuth
 
-The `claude-agent-sdk` provider routes LLM calls through the official [Claude Agent SDK](https://www.npmjs.com/package/@anthropic-ai/claude-agent-sdk) - it spawns the real Claude Code engine - while senpi executes every tool itself. Subscription usage flows through Anthropic's official Claude Code surface.
+The `claude-sdk-oauth` provider routes LLM calls through the official [Claude Agent SDK](https://www.npmjs.com/package/@anthropic-ai/claude-agent-sdk) - it spawns the real Claude Code engine - while senpi executes every tool itself. Subscription usage flows through Anthropic's official Claude Code surface.
 
-- Run `/login claude-agent-sdk` to sign in with your Claude Pro/Max subscription (PKCE, same OAuth client as the Claude Code CLI). An existing Anthropic OAuth credential is offered as an import.
-- Multiple accounts: each `/login claude-agent-sdk` adds another named account. `CLAUDE_CODE_OAUTH_TOKEN` (and `_2`..`_N`) are honored as read-only env accounts. `/claude-account` lists, adds, removes, and pins accounts; `--claude-account <name>` pins one for the session; `claudeAgentSdkProvider.pinnedAccount` pins one in settings.
+- Run `/login claude-sdk-oauth` to sign in with your Claude Pro/Max subscription (PKCE, same OAuth client as the Claude Code CLI). An existing Anthropic OAuth credential is offered as an import.
+- Multiple accounts: each `/login claude-sdk-oauth` adds another named account. `CLAUDE_CODE_OAUTH_TOKEN` (and `_2`..`_N`) are honored as read-only env accounts. `/claude-account` lists, adds, removes, and pins accounts; `--claude-account <name>` pins one for the session; `claudeSdkOauthProvider.pinnedAccount` pins one in settings.
 - Session affinity: one senpi session sticks to one account (rendezvous hashing), which keeps Anthropic's prompt cache warm - accounts never rotate mid-session except on automatic failover. Rate limits and auth errors block the account (with cooldown) and retry on the next account, before any visible output; once output has started, the error surfaces instead of replaying.
 - Default lane: **ambient** - with no `tokenInjection` setting the provider inherits the environment like the upstream extension (Claude Code CLI login or `ANTHROPIC_API_KEY`). Managed lanes (`oauth-slots`, `config-dir`) are opt-in via one settings line until the live subscription spike proves a managed default.
-- Settings (`claudeAgentSdkProvider`): `appendSystemPrompt`, `settingSources` (filesystem settings load only in the ambient lane, so they cannot override your selected account), `strictMcpConfig`, `pinnedAccount`, `tokenInjection` (`oauth-slots` | `config-dir` | `ambient`).
+- Settings (`claudeSdkOauthProvider`):
+  - `systemPromptMode` — controls how the system prompt is delivered. **`full`** (default) sends senpi's composed system prompt verbatim; the lane no longer rebuilds from the SDK `claude_code` preset, so all prompt regions (project rules, response-language instructions, etc.) reach the model. **`preset-append`** is the previous behaviour (deprecated, kept for one release; emits a one-time warning). **`override`** loads the system prompt from a file (`systemPromptFile`). The legacy `appendSystemPrompt` key still works: `false` → `preset-append`, `true`/unset → `full`; setting both keys makes `systemPromptMode` win and warns.
+  - In `full` and `override` modes, `settingSources` defaults to `[]` on every lane because senpi's prompt already carries project context — loading the SDK's own CLAUDE.md would double-inject it. The CLI always prepends its own `"You are a Claude agent, built on Anthropic's Claude Agent SDK."` block, which senpi cannot suppress; `full` means the prompt is delivered intact, not that it is the only system-prompt text.
+  - `settingSources` (filesystem settings load only in the ambient lane, so they cannot override your selected account), `strictMcpConfig`, `pinnedAccount`, `tokenInjection` (`oauth-slots` | `config-dir` | `ambient`), `resumeMode` (`auto` default | `off` restores per-turn sessions), `systemPromptFile`.
+- **Environment overrides** (precedence: `env > project settings > global settings > default`; no new CLI flags):
+
+  | Variable | Purpose |
+  |----------|---------|
+  | `SENPI_CLAUDE_SDK_OAUTH_SYSTEM_PROMPT_MODE` | `full` \| `preset-append` \| `override` |
+  | `SENPI_CLAUDE_SDK_OAUTH_SYSTEM_PROMPT_FILE` | Path to the system prompt file (used with `override` mode) |
+  | `SENPI_CLAUDE_SDK_OAUTH_RESUME` | `auto` (default) \| `off` — disables session reuse, restoring per-turn SDK queries |
+  | `SENPI_CLAUDE_SDK_OAUTH_TOKEN_INJECTION` | `oauth-slots` \| `config-dir` \| `ambient` |
+  | `SENPI_CLAUDE_SDK_OAUTH_SETTING_SOURCES` | Overrides `settingSources` |
+  | `SENPI_CLAUDE_SDK_OAUTH_PINNED_ACCOUNT` | Overrides `pinnedAccount` |
+
+  Every `SENPI_*` variable is stripped from the Claude Code subprocess environment on all three lanes; other inherited variables are preserved.
+- **Session reuse.** By default one long-lived SDK query spans the entire senpi session instead of a fresh one per turn, so conversations continue with only the new delta sent. Reuse fails closed to a fresh session on compaction, branch/fork navigation, account failover, an aborted turn, or any configuration change. Idle sessions retire after 30 minutes (max 32 resident); a session with an in-flight turn is never evicted. After a senpi process restart the lane always starts fresh. Set `resumeMode: "off"` (or `SENPI_CLAUDE_SDK_OAUTH_RESUME=off`) to restore the old per-turn behaviour. Accepted values are `"auto"` (default) and `"off"`; any other value is silently ignored.
 - Account state is exposed to desktop/automation clients: RPC `get_provider_accounts`, `account_pin`, `account_remove` and the `auth_accounts_changed` / `account_failover` events, mirrored through the app-server protocol. Token material is never included.
 
+#### Session continuity self-check
+
+Every main turn records one continuity decision. Healthy turns are silent in the transcript; degraded turns print a muted one-line notice. To watch the decisions directly, tail the session log and filter on the event prefix:
+
+```bash
+# Global agent dir (default):
+tail -f "${SENPI_CODING_AGENT_DIR:-$HOME/.senpi/agent}/logs/session.log" | rg claude_sdk_oauth_session_
+# Project-local agent dir (when the session uses one — getAgentDir prefers it):
+# tail -f .senpi/agent/logs/session.log | rg claude_sdk_oauth_session_
+```
+
+Each line is JSON with `kind` (`bootstrap`, `delta`, `reattach`, `fork`, `flatten`, `disabled`), a sanitized `reason`, and `count` (messages submitted this turn). A healthy conversation shows one `bootstrap` followed by `delta` lines; repeated `flatten` lines mean the lane is resending the whole conversation and losing prompt-cache hits. `SENPI_SESSION_DEBUG=1` mirrors the same lines to stderr.
+
+To confirm from the SDK side, list the transcript files Claude Code keeps per project - one long-lived session should keep appending to a single file rather than creating one per turn:
+
+```bash
+ls -lt ~/.claude/projects/*/ | head
+```
+
 ### GitHub Copilot
 
 - Press Enter for github.com, or enter your GitHub Enterprise Server domain
@@ -59,6 +94,7 @@ The `claude-agent-sdk` provider routes LLM calls through the official [Claude Ag
 
 - Run `/login openrouter`, then select **Sign in with OpenRouter** to open the OpenRouter PKCE authorization flow
 - The authorization creates a user-controlled OpenRouter API key billed from your OpenRouter credits
+- On remote/headless machines (e.g. over SSH) the browser cannot reach the loopback callback; paste the final redirect URL (or the authorization code) into the login prompt instead
 - `OPENROUTER_API_KEY` remains available through **Use an API key**
 
 ### Radius

package/docs/release-guide.md

@@ -72,6 +72,8 @@ The source machine additionally needs:
 - the repository dependencies
 - `tmux` for the POSIX TUI smoke channel
 
+Binary builds use the Bun 1.4+ canary channel. Local builders should run `bun upgrade --canary` first to ensure the toolchain matches CI.
+
 Verify the local toolchain:
 
 ```bash
@@ -79,6 +81,7 @@ cd /Users/yeongyu/local-workspaces/senpi
 uname -m
 node --version
 npm --version
+bun upgrade --canary
 bun --version
 git status --short --branch
 ```
@@ -99,7 +102,7 @@ This release uses CalVer:
 2026.7.30-2
 ```
 
-The eight lockstep package manifests are:
+The seven lockstep package manifests are:
 
 - `packages/ai/package.json`
 - `packages/agent/package.json`
@@ -108,7 +111,6 @@ The eight lockstep package manifests are:
 - `packages/pty/package.json`
 - `packages/senpi-codemode/package.json`
 - `packages/tui/package.json`
-- `packages/web-ui/package.json`
 
 `@earendil-works/pi-storage-sqlite-node` remains on its independent semver line and is not
 part of the CalVer bump.
@@ -166,7 +168,7 @@ The authoritative `builtinExtensions` registration order contains 31 extensions:
 26. `goal`
 27. `ttsr`
 28. `btw`
-29. `claude-agent-sdk`
+29. `claude-sdk-oauth`
 30. `config-reload`
 31. `mcp`
 
@@ -255,7 +257,7 @@ npm --prefix packages/coding-agent exec vitest -- \
   test/suite/vendored-builtins.test.ts \
   test/extensions/loader-concurrency.test.ts \
   test/mcp/ \
-  test/suite/claude-agent-sdk-extension.test.ts \
+  test/suite/claude-sdk-oauth-extension.test.ts \
   test/suite/terminal-extension.test.ts \
   test/compaction/ \
   test/ttsr/
@@ -624,7 +626,7 @@ For the next release:
 4. Duplicate user-facing AI, agent, and TUI changes into the coding-agent changelog.
 5. Query registry-verifiable plugin dependencies for newer exact versions.
 6. Preserve vendored snapshots when no authoritative upstream source is available.
-7. Update the eight lockstep package versions.
+7. Update the seven lockstep package versions.
 8. Regenerate package-lock, publish-deps lock, and install lock.
 9. Run focused plugin tests.
 10. Run `check`, `build`, and the full test suite.

package/docs/session-format.md

@@ -121,6 +121,8 @@ interface Usage {
 }
 ```
 
+The exported pi-ai `StopReason` type also includes `"pending"`, but that value is reserved for partial messages in streaming events. Terminal `done`/`error` messages replace it with a completion reason before Senpi persists the assistant message, so `"pending"` should never appear in session JSONL.
+
 ### Extended Message Types (from @code-yeongyu/senpi)
 
 ```typescript

package/docs/settings.md

@@ -84,7 +84,7 @@ Permission rules are a confirmation policy, not a sandbox. Senpi, extensions, pa
 | `defaultProvider` | string | - | Default provider (e.g., `"anthropic"`, `"openai"`) |
 | `defaultModel` | string | - | Default model ID |
 | `defaultThinkingLevel` | string | - | `"off"`, `"minimal"`, `"low"`, `"medium"`, `"high"`, `"xhigh"`, `"max"` |
-| `promptPreset` | string | `"auto"` | Force a system prompt preset: `"auto"`, `"kimi-k2-6"`, `"kimi-k2-7"`, `"kimi-k3"`, `"glm-5.2"`, `"claude-fable-5"`, `"claude-opus-5"`, `"claude-opus-4-5"`, `"claude-opus-4-6"`, `"claude-opus-4-7"`, `"claude-opus-4-8"`, `"gpt-5"`, `"gpt-5.2"`, `"gpt-5.3-codex"`, `"gpt-5.4"`, `"gpt-5.5"`, or `"gpt-5.6"` |
+| `promptPreset` | string | `"auto"` | Force a system prompt preset: `"auto"`, `"kimi-k2-6"`, `"kimi-k2-7"`, `"kimi-k3"`, `"glm-5.2"`, `"claude-fable-5"`, `"claude-opus-5"`, `"claude-opus-4-5"`, `"claude-opus-4-6"`, `"claude-opus-4-7"`, `"claude-opus-4-8"`, `"deepseek-v4-flash"`, `"deepseek-v4-flash-0731"`, `"deepseek-v4-pro"`, `"gpt-5"`, `"gpt-5.2"`, `"gpt-5.3-codex"`, `"gpt-5.4"`, `"gpt-5.5"`, or `"gpt-5.6"` |
 | `hideThinkingBlock` | boolean | `false` | Hide thinking blocks in output |
 | `showCacheMissNotices` | boolean | `false` | Show transcript notices for significant prompt-cache misses |
 | `thinkingBudgets` | object | - | Custom token budgets per thinking level |
@@ -135,6 +135,8 @@ When this value is anything other than `"auto"`, it overrides any model-level `p
 | `outputPad` | number | `1` | Horizontal padding for user messages, assistant messages, and thinking (0 or 1) |
 | `autocompleteMaxVisible` | number | `5` | Max visible items in autocomplete dropdown (3-20) |
 | `showHardwareCursor` | boolean | `false` | Show the terminal cursor while TUI positions it for IME support |
+| `uiMode` | string | `"regular"` | Interactive UI mode: `"regular"` or experimental `"fullscreen"`. Changes from `/settings` apply after restart; `--ui-mode` overrides this setting for one run |
+| `fullscreenScrollbar` | string | `"auto"` | Fullscreen transcript scrollbar: `"auto"` shows it temporarily while scrolling, `"always"` reserves the rightmost column and keeps it visible, and `"hidden"` hides it. Has no effect in regular UI mode |
 
 For VS Code, include `--wait` so pi resumes after the editor exits:
 

package/docs/skills.md

@@ -13,6 +13,7 @@ Senpi implements the [Agent Skills standard](https://agentskills.io/specificatio
 - [Skill Commands](#skill-commands)
 - [Skill Structure](#skill-structure)
 - [Frontmatter](#frontmatter)
+- [Carrying MCP Servers](#carrying-mcp-servers)
 - [Validation](#validation)
 - [Example](#example)
 - [Skill Repositories](#skill-repositories)
@@ -96,6 +97,7 @@ A skill is a directory with a `SKILL.md` file. Everything else is freeform.
 ```
 my-skill/
 ├── SKILL.md              # Required: frontmatter + instructions
+├── mcp.json              # Optional: MCP servers bundled with the skill
 ├── scripts/              # Helper scripts
 │   └── process.sh
 ├── references/           # Detailed docs loaded on-demand
@@ -173,6 +175,33 @@ Poor:
 description: Helps with PDFs.
 ```
 
+## Carrying MCP Servers
+
+A skill can bundle the MCP servers it uses, shipping the workflow and its
+tools as one package. Servers declared by a skill register lazily with zero
+active tools — they cost no prompt tokens until the skill loads, and the
+tools matching the skill's `includeTools` globs activate for the session when
+it does.
+
+Declare servers in an `mcp.json` sidecar next to SKILL.md (or in a `mcp:`
+frontmatter block):
+
+```json
+{
+  "mcpServers": {
+    "exa": {
+      "command": "npx",
+      "args": ["-y", "exa-mcp-server"],
+      "env": { "EXA_API_KEY": "${EXA_API_KEY}" },
+      "includeTools": ["web_search*"]
+    }
+  }
+}
+```
+
+See [Skill-carried MCP servers](mcp.md#skill-carried-mcp-servers) for the
+full declaration forms, collision rules, and reveal semantics.
+
 ## Validation
 
 Senpi validates skills against the Agent Skills standard. Most issues produce warnings but still load the skill:

package/docs/themes.md

@@ -71,6 +71,7 @@ vim ~/.senpi/agent/themes/my-theme.json
     "text": "",
     "thinkingText": "secondary",
     "selectedBg": "#2d2d30",
+    "scrollbarThumb": "#555566",
     "userMessageBg": "#2d2d30",
     "userMessageText": "",
     "customMessageBg": "#2d2d30",
@@ -140,13 +141,13 @@ vim ~/.senpi/agent/themes/my-theme.json
 
 - `name` is required, must be unique, and must not contain `/`.
 - `vars` is optional. Define reusable colors here, then reference them in `colors`.
-- `colors` must define all 51 required tokens. `thinkingMax` is optional and falls back to `thinkingXhigh`.
+- `colors` must define all 51 required tokens. `thinkingMax` is optional and falls back to `thinkingXhigh`; `scrollbarThumb` is optional and falls back to `selectedBg`.
 
 The `$schema` field enables editor auto-completion and validation.
 
 ## Color Tokens
 
-Every theme must define all 51 required color tokens. `thinkingMax` is optional for compatibility with existing themes; when omitted, it uses `thinkingXhigh`.
+Every theme must define all 51 required color tokens. `thinkingMax` and `scrollbarThumb` are optional for compatibility with existing themes; when omitted, they use `thinkingXhigh` and `selectedBg`, respectively.
 
 ### Core UI (11 colors)
 
@@ -164,11 +165,12 @@ Every theme must define all 51 required color tokens. `thinkingMax` is optional
 | `text` | Default text (usually `""`) |
 | `thinkingText` | Thinking block text |
 
-### Backgrounds & Content (11 colors)
+### Backgrounds & Content (11 required, 1 optional)
 
 | Token | Purpose |
 |-------|---------|
 | `selectedBg` | Selected line background |
+| `scrollbarThumb` | Fullscreen scrollbar thumb background; optional, falls back to `selectedBg` |
 | `userMessageBg` | User message background |
 | `userMessageText` | User message text |
 | `customMessageBg` | Extension message background |

package/docs/usage.md

@@ -243,12 +243,17 @@ pi --no-extensions -e ./my-extension.ts
 |--------|-------------|
 | `--system-prompt <text>` | Replace default prompt; context files and skills are still appended |
 | `--append-system-prompt <text>` | Append to system prompt |
+| `--ui-mode <mode>` | UI mode: `regular` (default) or experimental `fullscreen` |
 | `--verbose` | Force verbose startup |
 | `-a`, `--approve` | Trust project-local files for this run |
 | `-na`, `--no-approve` | Ignore project-local files for this run |
 | `-h`, `--help` | Show help |
 | `-v`, `--version` | Show version |
 
+In `fullscreen` mode, the transcript scrolls inside the terminal viewport while queued messages, working status, extension widgets, editor, and footer remain fixed at the bottom. Mouse/trackpad input scrolls the region under the pointer; keyboard viewport actions always remain available. Inline images work in terminals that support the Kitty graphics protocol, including Kitty and Ghostty. In iTerm2 they render as text placeholders because its inline-image protocol cannot delete or crop placements during application-owned scrolling. In `regular` mode, pi uses the main screen and terminal-owned scrollback, and iTerm2 inline images continue to render normally.
+
+To use fullscreen mode by default, set **UI mode** to `fullscreen` in `/settings`. The change applies after restarting pi.
+
 ### File Arguments
 
 Prefix files with `@` to include them in the message:

package/examples/extensions/custom-compaction.ts

@@ -14,7 +14,6 @@
  */
 
 import { uuidv7 } from "@earendil-works/pi-ai";
-import { complete } from "@earendil-works/pi-ai/compat";
 import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
 import { convertToLlm, serializeConversation } from "@earendil-works/pi-coding-agent";
 
@@ -88,7 +87,7 @@ ${conversationText}
 
 		try {
 			// Pass signal to honor abort requests (e.g., user cancels compaction)
-			const response = await complete(
+			const response = await ctx.modelRegistry.complete(
 				model,
 				{ messages: summaryMessages },
 				{

package/examples/extensions/custom-provider-anthropic/index.ts

@@ -353,7 +353,7 @@ function streamCustomAnthropic(
 				totalTokens: 0,
 				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
 			},
-			stopReason: "stop",
+			stopReason: "pending",
 			timestamp: Date.now(),
 		};
 
@@ -546,8 +546,14 @@ function streamCustomAnthropic(
 			if (options?.signal?.aborted) {
 				throw new Error("Request was aborted");
 			}
+			if (output.stopReason === "pending") {
+				throw new Error("Anthropic stream ended without a stop reason");
+			}
+			if (output.stopReason === "error" || output.stopReason === "aborted") {
+				throw new Error(output.errorMessage || "An unknown error occurred");
+			}
 
-			stream.push({ type: "done", reason: output.stopReason as "stop" | "length" | "toolUse", message: output });
+			stream.push({ type: "done", reason: output.stopReason, message: output });
 			stream.end();
 		} catch (error) {
 			for (const block of output.content) delete (block as any).index;

package/examples/extensions/doom-overlay/wad-finder.ts

@@ -1,11 +1,12 @@
 import { existsSync, writeFileSync } from "node:fs";
 import { dirname, join, resolve } from "node:path";
 import { fileURLToPath } from "node:url";
+import { gunzipSync } from "node:zlib";
 
 // Get the bundled WAD path (relative to this module)
 const __dirname = dirname(fileURLToPath(import.meta.url));
 const BUNDLED_WAD = join(__dirname, "doom1.wad");
-const WAD_URL = "https://distro.ibiblio.org/slitaz/sources/packages/d/doom1.wad";
+const WAD_URL = "https://www.gamers.org/pub/idgames/idstuff/doom/doom-1.8.wad.gz";
 
 const DEFAULT_WAD_PATHS = ["./doom1.wad", "./DOOM1.WAD", "~/doom1.wad", "~/.doom/doom1.wad"];
 
@@ -42,8 +43,11 @@ export async function ensureWadFile(): Promise<string | null> {
 		if (!response.ok) {
 			throw new Error(`HTTP ${response.status}`);
 		}
-		const buffer = await response.arrayBuffer();
-		writeFileSync(BUNDLED_WAD, Buffer.from(buffer));
+		const wad = gunzipSync(Buffer.from(await response.arrayBuffer()));
+		if (wad.subarray(0, 4).toString("ascii") !== "IWAD") {
+			throw new Error("Downloaded file is not a valid IWAD");
+		}
+		writeFileSync(BUNDLED_WAD, wad);
 		return BUNDLED_WAD;
 	} catch {
 		return null;

package/examples/extensions/handoff.ts

@@ -13,8 +13,7 @@
  */
 
 import type { AgentMessage } from "@earendil-works/pi-agent-core";
-import { uuidv7 } from "@earendil-works/pi-ai";
-import { complete, type Message } from "@earendil-works/pi-ai/compat";
+import { type Message, uuidv7 } from "@earendil-works/pi-ai";
 import type { ExtensionAPI, SessionEntry } from "@earendil-works/pi-coding-agent";
 import { BorderedLoader, convertToLlm, serializeConversation } from "@earendil-works/pi-coding-agent";
 
@@ -134,7 +133,7 @@ export default function (pi: ExtensionAPI) {
 						timestamp: Date.now(),
 					};
 
-					const response = await complete(
+					const response = await ctx.modelRegistry.complete(
 						ctx.model!,
 						{ systemPrompt: SYSTEM_PROMPT, messages: [userMessage] },
 						{

package/examples/extensions/overlay-test.ts

@@ -103,7 +103,7 @@ class OverlayTestComponent implements Focusable {
 
 		// Edge cases - full width lines to test compositing at boundaries
 		lines.push(row(` ${th.fg("dim", "─── Edge Cases (borders should align) ───")}`));
-		lines.push(row(` Wide: ${th.fg("warning", "中文日本語한글テスト漢字繁體简体ひらがなカタカナ가나다라마바")}`));
+		lines.push(row(` Wide: ${th.fg("warning", "中文日本語汉字テスト漢字繁體简体ひらがなカタカナ中文汉字")}`));
 		lines.push(
 			row(
 				` Styled: ${th.fg("error", "RED")} ${th.fg("success", "GREEN")} ${th.fg("warning", "YELLOW")} ${th.fg("accent", "ACCENT")} ${th.fg("dim", "DIM")} ${th.fg("error", "more")} ${th.fg("success", "colors")}`,

package/examples/extensions/qna.ts

@@ -7,7 +7,7 @@
  * 3. Loads the result into the editor for user to fill in answers
  */
 
-import { complete, type UserMessage } from "@earendil-works/pi-ai/compat";
+import type { UserMessage } from "@earendil-works/pi-ai";
 import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
 import { BorderedLoader } from "@earendil-works/pi-coding-agent";
 
@@ -87,7 +87,7 @@ export default function (pi: ExtensionAPI) {
 						timestamp: Date.now(),
 					};
 
-					const response = await complete(
+					const response = await ctx.modelRegistry.complete(
 						ctx.model!,
 						{ systemPrompt: SYSTEM_PROMPT, messages: [userMessage] },
 						{ apiKey: auth.apiKey, headers: auth.headers, env: auth.env, signal: loader.signal },

package/examples/rpc-extension-ui.ts

@@ -18,7 +18,16 @@ import { spawn } from "node:child_process";
 import { dirname, join } from "node:path";
 import * as readline from "node:readline";
 import { fileURLToPath } from "node:url";
-import { type Component, Container, Input, matchesKey, ProcessTerminal, SelectList, TUI } from "@earendil-works/pi-tui";
+import {
+	type Component,
+	Container,
+	Input,
+	matchesKey,
+	ProcessTerminal,
+	SelectList,
+	type TUI,
+	TuiMainScreen,
+} from "@earendil-works/pi-tui";
 
 const __dirname = dirname(fileURLToPath(import.meta.url));
 
@@ -267,7 +276,7 @@ async function main() {
 	// -- TUI setup --
 
 	const terminal = new ProcessTerminal();
-	const tui = new TUI(terminal);
+	const tui: TUI = new TuiMainScreen(terminal);
 
 	const outputLog = new OutputLog();
 	const loadingIndicator = new LoadingIndicator();

package/examples/sdk/02-custom-model.ts

@@ -4,13 +4,12 @@
  * Shows how to select a specific model and thinking level.
  */
 
-import { getModel } from "@earendil-works/pi-ai/compat";
 import { createAgentSession, ModelRuntime } from "@earendil-works/pi-coding-agent";
 
 const modelRuntime = await ModelRuntime.create();
 
 // Option 1: Find a specific built-in model by provider/id
-const opus = getModel("anthropic", "claude-opus-4-5");
+const opus = modelRuntime.getModel("anthropic", "claude-opus-4-5");
 if (opus) {
 	console.log(`Found model: ${opus.provider}/${opus.id}`);
 }

package/examples/sdk/12-full-control.ts

@@ -41,7 +41,9 @@ const resourceLoader: ResourceLoader = {
 	getAgentsFiles: () => ({ agentsFiles: [] }),
 	getSystemPrompt: () => `You are a minimal assistant.
 Available: read, bash. Be concise.`,
+	getSystemPromptSource: () => undefined,
 	getAppendSystemPrompt: () => [],
+	getAppendSystemPromptSources: () => [],
 	extendResources: () => {},
 	reload: async () => {},
 };

package/node_modules/@code-yeongyu/senpi-codemode/CHANGELOG.md

@@ -12,6 +12,51 @@
 
 ### Removed
 
+## [2026.8.3] - 2026-08-03
+
+### Breaking Changes
+
+### Added
+
+### Changed
+
+### Fixed
+
+### Removed
+
+## [2026.8.1] - 2026-08-01
+
+### Breaking Changes
+
+### Added
+
+### Changed
+
+### Fixed
+
+- Preserve rich live and terminal `eval` details when peeking detached cells,
+  including code, title, output, phase, status events, tool-call summaries,
+  duration, and structured displays; cancellation now remains authoritative
+  over late completion races
+  ([#603](https://github.com/code-yeongyu/senpi/pull/603)).
+
+### Removed
+
+## [2026.7.31-2] - 2026-07-31
+
+### Breaking Changes
+
+### Added
+
+### Changed
+
+- Include a live elapsed label in detached `eval` footer status. The ticker updates only when the rendered duration
+  changes and is disposed when the cell completes, fails, or is stopped.
+
+### Fixed
+
+### Removed
+
 ## [2026.7.31] - 2026-07-31
 
 ### Breaking Changes

package/node_modules/@code-yeongyu/senpi-codemode/README.md

@@ -11,9 +11,12 @@ task-tool names are known.
 - Persistent JavaScript, Python, Ruby, and Julia cells. State survives later
   cells in the same language until reset, restart, or session disposal.
 - Timeout detachment for interactive `eval`: long pure-compute cells return a
-  handle and continue in their existing kernel. Completion is injected with the
-  final value/error and buffered output; use `eval({ action: "peek"|"stop",
-  cell_id })` to inspect or terminate a detached cell.
+	handle and continue in their existing kernel. Completion is injected with the
+	final value/error and buffered output; use `eval({ action: "peek"|"stop",
+	cell_id })` to inspect or terminate a detached cell. A running peek preserves
+	the original code and title together with current output, phase, status
+	events, tool-call summaries, elapsed duration, and structured display state;
+	a terminal peek preserves the exact final result.
 - Loopback, bearer-authenticated kernel bridge with bounded JSONL frames.
 - Structured status events for file operations, environment access, phases,
   bridge activity, and delegated task progress.

package/node_modules/@code-yeongyu/senpi-codemode/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@code-yeongyu/senpi-codemode",
-	"version": "2026.7.31",
+	"version": "2026.8.3",
 	"description": "Source-only senpi extension package for codemode evaluation tools",
 	"type": "module",
 	"main": "./src/index.ts",
@@ -30,14 +30,14 @@
 	},
 	"dependencies": {
 		"@babel/parser": "8.0.4",
-		"@earendil-works/pi-ai": "^2026.7.31",
+		"@earendil-works/pi-ai": "^2026.8.3",
 		"typebox": "1.3.8"
 	},
 	"peerDependencies": {
 		"@code-yeongyu/senpi": "*"
 	},
 	"devDependencies": {
-		"@code-yeongyu/senpi": "2026.7.31"
+		"@code-yeongyu/senpi": "2026.8.3"
 	},
 	"keywords": [
 		"senpi",