Pi 1.5.0 adoption

Scope

Upgrade Agent Runtime's pi-agent-core, pi-ai, and pi-coding-agent, plus Desktop's pi-ai and pi-mcp, to exactly 1.5.0. Keep the Pi release group aligned through the lockfile and exact release-age exclusions, which now include the new pi-schema package. No Host protocol, database schema, persisted provider endpoint, permission policy, extension VERSION marker, or storage ownership changes.

Official release notes: Pi 1.5.0.

Compatibility work

• Rebase the three pnpm patches onto the published 1.5.0 package layouts. pi-agent-core applied unchanged. pi-ai needed manual ports where 1.5.0 replaced re-parsing with incremental tool-call parsing and moved compat schemas to pi-schema:
  • Anthropic hosted-search capture now registers server_tool_use and web_search_tool_result blocks in the adapter's openBlocks map and streams the query through ToolCallArgumentsStream, instead of an index / inputJson scratch on the block.
  • requiresNonEmptyReasoningReplay is declared with the pi-schema builders.
  • Local request-error mapping is re-inserted where 1.5.0 removed the per-block scratch cleanup from the Anthropic and OpenAI Responses catch paths.
  • pi-coding-agent only needed its Api type import re-applied.
• pi-ai no longer re-exports TypeBox. Agent Runtime imports Type from its existing direct typebox dependency. Tool schemas stay plain JSON Schema, so validation on pi-schema accepts them unchanged.
• Bump Agent Runtime's direct undici from 8.10.2 to 8.12.3 to match pi-coding-agent 1.5.0. Two 8.x copies split pi-ai into two peer instances and pulled undici into the sidecar entry (891 KiB, over the 500 KiB gate); one copy restores 315 KiB and the 1.4 instance topology.

Inherited without Desktop code changes

• Incremental tool-call argument parsing in providers and agent proxy streams (lower CPU and memory for large streamed tool calls).
• pi-schema coercion for plain JSON Schema tools, including MCP tools (numeric strings, single value to array). Validation messages shown to the model change wording.
• JPEG-first image resizing in the bundled coding-agent resize worker.
• Provider fixes: Gemini const to enum, Bedrock strict-tool retry, OpenRouter reported costs, LiteLLM overflow detection, Claude system-after-empty-assistant ordering and capped effort, Kimi K3.1 settings, GPT-5.4/5.5 tool namespace after model switch, OpenCode endpoints.
• GPT-6.1 Terra and the models.dev v2 catalog in pi-ai's built-in models. Desktop's bundled models.dev catalog and overlays remain the source of what the UI offers.

Adopted 1.5 features

• Billed vs. estimated cost. usageFromPi records a total marked usage.cost.totalFromProvider as costStatus: "reported", even when the model has no catalog rates, and keeps unmarked totals estimated. Replay back into Pi restores the marker. Persisted cost keeps its five numeric fields.
• Served model. Agent Runtime stores Pi's responseModel as the additive responseModelId message field when it differs from the requested model and is not a dated snapshot of it (response-model.ts). Host-core round-trips it through record meta with no SQL migration; older builds ignore it. The model chip shows the served model and labels the requested one (chat.servedByModel in all shipped locales), and error details show both.
• Extension pi.exec(). extensions/managed-exec.ts now decodes output across chunk boundaries, caps stdout and stderr combined in bytes (default 10 MiB), terminates on overflow with truncated: true and a nonzero code, and resolves a spawn failure with code 1 and the spawn error in stderr. ExtensionExecResult gains truncated.
• Zhipu/Z.ai 1308. Desktop's classifier treats the JSON error code 1308 (plan usage limit, sent as 429) as terminal PROVIDER_ERROR, so neither retry budget replays it.
• GPT-6.1 Terra. The bundled models.dev catalog is refreshed to the current snapshot, which lists gpt-6.1-terra for OpenAI and Azure (1.05M context, reasoning). The refresh also drops openai/gpt-image-1 from chat records; image settings metadata lives in settings-operation-metadata.json and is unaffected. ChatGPT (Codex) accounts list models from the live /codex/models endpoint; CODEX_MODELS_CLIENT_VERSION is unchanged because verifying a newer client version needs a real account.
• Not applicable to Desktop: TUI rendering, pi update / self-update command, extension package changelogs, terminal keybinding hints.

Explicitly unchanged

• Existing user data needs no migration. No persisted format, RPC, Plugin SDK, or permission contract changes.
• pi-coding-agent remains transitional; it is not the normal Desktop agent loop, compaction engine, or durable session store.

Validation

Task candidate: chore/pi-1-5-upgrade rebased onto 28d5d07de18f (origin/main, 2026-11-06), run from the dedicated request worktree with the primary checkout's Electron and a worktree-built debug Host.

• pnpm check:pr-base, pnpm build:js, Desktop typecheck, lint, and pnpm docs:check passed. Pi dependency and patch gates and patch-hash fixture tests passed.
• Agent Runtime: 122 test files, 1,769 tests passed. Shared: 1,402 passed. i18n: 31 passed.
• Desktop: 4,864 tests, 4,861 passed, 3 skipped, 0 failed. On the earlier base three timing-sensitive tests (macos-signing-watchdog, npm-executable, remote-host-ssh-transport) timed out once under the full parallel run and passed 41/41 alone.
• Host-core: cargo fmt --check passed; 1,210 of 1,211 tests passed. The one failure, review_rpc_records_workspace_scratch_and_approved_external_write_edit, fails identically on the unmodified base commit. Clippy reports two existing warnings in review.rs and transcripts.rs.
• Trusted-extension E2E passed 57/57.
• Hosted-search E2E passed 10/10 only with a local, uncommitted fake-host addition for tools.fileBatchCapabilities. Without it every scenario stops with Unsupported fake-host RPC: bdc35fa12 added that runtime RPC but not the handler in scripts/e2e/hosted-search-sidecar.mjs, which this change does not touch. That fixture gap is a separate fix.
• No real provider or live service was used.