You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Prefer exposed pi__mcp and pi__subagent bridge tools for MCP work and delegation, with Cursor-configured MCP and Cursor-native subagents as fallbacks when the matching pi tool is not exposed or is unavailable.
Update the Pi development and platform-smoke baseline to 0.80.5, including native openai-codex/gpt-5.6-luna, gpt-5.6-sol, and gpt-5.6-terra metadata.
Enable guarded branch-scoped local Cursor SDK resume by default for local runtime; users can opt out with --cursor-no-local-resume, PI_CURSOR_LOCAL_RESUME=0, or local.resume: false.
Added
Refresh the reviewed Cursor fallback catalog with Cursor's GPT-5.6 Luna/Sol/Terra models and Claude Sonnet 5, including the SDK-advertised aliases, context variants, reasoning/effort levels, and fast selections.
Add /cursor-refresh-config to call the current pooled Cursor SDK agent's agent.reload() for filesystem Cursor config refreshes without recreating the agent.
Add explicit local Cursor safety controls: --cursor-auto-review / PI_CURSOR_AUTO_REVIEW and --cursor-sandbox / PI_CURSOR_SANDBOX, plus local.autoReview and local.sandboxOptions.enabled config. Defaults stay off.
Add one-shot manual local stuck-run recovery with --cursor-local-force / PI_CURSOR_LOCAL_FORCE, consumed only at the next actual Agent.send(..., { local: { force: true } }) so pre-send failures/aborts preserve it and session reload cannot rearm a consumed CLI flag; no persistent config default, retry loop, or automatic force recovery is added.
Add minimal explicit Cursor cloud runtime execution after first-use acknowledgement and safety preflight. Defaults stay local + loopback MCP; cloud runs use fresh context by default, skip the pi bridge/local MCP/env forwarding, leave cloud agents alive for Cursor-managed cleanup, and keep project config limited to the runtime default for the initial cloud slice.
Show Cursor runtime directly in the interactive status footer (cursor:local · fast:on|off|n/a, cursor:cloud · fast:n/a) so cloud opt-in is visible.
Name explicit Cursor cloud agents from the current pi session title when available so Cursor Cloud agent lists are easier to match back to pi sessions.
Add explicit Cursor-managed cloud environment selection with --cursor-cloud-env-type, --cursor-cloud-env-name, PI_CURSOR_CLOUD_ENV_TYPE, PI_CURSOR_CLOUD_ENV_NAME, and cloud.environment.type/name user config. This selects Cursor Cloud cloud / pool / machine environments without forwarding local env values.
Stream bounded display-only Cursor Cloud completion telemetry when available: agent/run IDs, pushed branch, repository and PR URL, passive artifact paths, and raw cloud usage without persisting it into transcript content or feeding that usage into pi accounting.
Add /cursor-cloud list, /cursor-cloud archive <bc-agentId>, and /cursor-cloud delete <bc-agentId> --yes for session-branch recorded cloud agents; archive/delete reject empty, non-bc-, wildcard, unrecorded, bulk, and unconfirmed delete requests before SDK calls.
Add guarded branch-scoped local Cursor SDK resume; it stores SDK agent IDs only in pi session custom entries, re-supplies the current model and loopback MCP bridge, bootstraps the current pi transcript once after process reattachment, and falls back to create+bootstrap with a display-only continuity note when resume is unavailable.
Fixed
Prevent Bun from terminating pi when @cursor/sdk 1.0.23's controlled-exec error path writes to an already closed internal iterable during a long bridged tool call. Suppression is limited to the exact WriteIterableClosedError name/message with Cursor SDK stack provenance for the Pi session lifecycle because controlled-exec can reject after the originating provider turn; Connect/network/abort suppression remains active-turn scoped, and unrelated failures remain fatal.
Stop applying Cursor SDK RunResult.usage to pi assistant messages when no per-turn turn-ended usage is available; long local sessions can report full-agent-context usage there, which can poison compaction/session token totals. Pi now falls back to bounded local estimates unless real per-turn SDK usage is present and fits the selected model window.
Pass the current Cursor model selection on every SDK agent.send() so resumed and pooled agents keep pi's active model as the source of truth.
Fail closed on invalid explicit Cursor runtime or cloud-context CLI/environment values, and report the effective runtime source or safety cap from /cursor-runtime instead of claiming a shadowed session value took effect.
Apply local.resume enable/disable changes to the next pooled-agent turn without recreating the agent.
Force-create the replacement local SDK agent after incremental-threshold or context-divergence rebootstrap while keeping resume persistence enabled for the replacement, instead of resuming the agent the reset just disposed.
Stop parsing process-wide --api-key values during startup discovery; provider turns keep Pi's scoped options.apiKey, while model refresh and Cloud lifecycle commands ask Pi's ModelRegistry for provider cursor, preventing another provider's key from reaching Cursor SDK calls while preserving stored and environment Cursor auth fallback.
Terminate full smoke-runner process trees on timeout/cleanup, give bounded Windows CIM orphan discovery enough time under VM contention, reject unknown or conflicting paid smoke lane arguments before starting a live run, and prepare a root-started wrapper around the unprivileged Ubuntu Node image so Crabbox can install its SSH bootstrap dependencies.
Persist Cursor Cloud agent intent immediately after Agent.create() and before debug/abort handling to a branch-bound, fsynced, newline-framed sidecar keyed by stable pi session ID, opened through regular-file descriptor identity checks without following symlinks (POSIX mode 0600, Windows user-directory ACL), and persist returned run IDs before abort/wait handling so rejected, cancelled, or first-turn-crashed sends remain cleanup-eligible; individually truncated records no longer hide later valid IDs, successive run/report records merge without losing metadata, anchored journals fsync Pi session JSONL first, archive/delete require auth before durable intent/result records, tolerate optional Pi-mirror failure after a durable result, and leave durable intent pending when the sidecar result fails, Cloud requires a persisted pi session, lifecycle persistence failures fail closed with bounded cancellation and dashboard guidance, and optional debug-write failures cannot orphan returned or live runs.
Verify and fsync exact local-agent cleanup intent in the Pi session JSONL before SDK deletion, fsync the result afterward, use Windows-compatible read-write flush descriptors, and keep both durable intent and current-process state blocking retry when result durability fails.
Make Cloud smoke cleanup harvest exact agent IDs from canonical session/lifecycle artifacts as well as optional debug metadata, archive the union, and retain artifacts on run or cleanup failure.
Keep canonical platform artifact transport below Crabbox's 64 KiB ceiling with gzip JSON/base64 inline bundles or checksum-verified 32 KiB no-sync chunk retrieval before lease release; large bundles now use only the exact CWD-relative .platform-artifact-bundle.gz final component, created and written exclusively through a no-follow identity-checked descriptor, and chunk/marker validation rejects every parent-bearing path. Bounded scans accept caller roots only when their real path matches the same relative path beneath the canonical real CWD or temp base (preserving expected macOS /var → /private/var aliasing while rejecting user-created intermediate links), then hold each traversed directory and recheck its identity plus immutable ctime/mtime/size/link/mode snapshot around traversal and file reads, so nested rename/symlink ABA cannot expose outside bytes. Scans still cover every regular artifact before transport exclusions, reject non-regular entries, prune node_modules/.git, fail closed on oversized or unscannable files, and any traversal failure emits only bounded failure evidence rather than findings or previously observed bytes; bundles cap compressed/inflated/per-file/file-count/aggregate extraction before host writes, reject duplicate/file-prefix bundle paths and symlinked or pre-existing host destinations, and extract only through exclusive identity-checked descriptors after securing destination directories. Compact output still uses --capture-stdout and preserves real scenario exit codes. Windows controllers reject nonempty extraction before mutation because Node has no handle-relative Windows creation API; Windows remains a supported matrix target from the POSIX controller, with target-side scan/spill identity guards kept free of POSIX-only flags.
Reject local resume handles that cross user messages already persisted at process startup, preventing a hard-crash restart from resending a prompt that the prior Cursor agent may already have accepted.
Reject non-HTTPS and credential/query/fragment-bearing Cursor Cloud repository URLs before Agent.create(), and redact URL/SCP-style userinfo from provider and maintainer output.
Validation
Current offline evidence is limited to the checked-in focused tests, full unit suite, platform-smoke syntax/tooling checks, and TypeScript checks; final command results belong in the release artifacts/status.
Required and pending release gates are npm publish --dry-run, npm run smoke:platform:all, npm run smoke:cloud, and the applicable live release smokes. Their final results belong in the release artifacts/status, not this changelog.