Runfusion/Fusion v0.47.0

on GitHub

New

• Structured changeset format with AI-distilled release notes for cleaner, user-facing changelogs.

Internal

• Saved agent tool-output details now default off to reduce persisted log payloads, while timeline rows remain logged and detailed tool arguments/results stay available via the global persistAgentToolOutput: true opt-in.
• Harden the workspace per-repo land loop against partial-failure races. A lost landedSha DB write after a sub-repo's integration ref already advanced no longer silently continues — it escalates to a retryable partial-land error, and the landed predicate now recognizes an already-landed repo via its Fusion-Task-Id trailer on retry, so a re-run never produces a second squash commit. The land lease is now taskId-aware across registry kinds: a merging task can no longer clobber an executing task's acquire lease on a shared sub-repo (any foreign-task holder is treated as contention), and the active-session registry rejects foreign-task overwrites instead of silently clobbering. The transient merging status is always reset before any throw escapes the land loop (no stuck-merging leak), and finalize re-reads the latest task and no longer swallows the merge-details persist failure (no finalizing on a stale row).
• Address Phase C workspace merge-loop review feedback. A sub-repo recognized as already-landed via the Fusion-Task-Id trailer fallback (when its landedSha persist was lost) now resolves and re-records a concrete landedSha, so finalize no longer drops it and mis-reports a fully-landed workspace task as a no-op (mergeConfirmed:false). A manual merge that hits sub-repo land-lease contention now surfaces the busy error to the user without consuming the persisted mergeRetries quota (matching the auto path's separate busy counter). The partial-land retry persists the incremented retry count before arming the backoff timer — a failed write now fails closed instead of looping without consuming budget — and clears the stale busy-contention counter when a real partial land supersedes transient busy failures. The CLI and dashboard merge doors use the shared isWorkspaceTask predicate instead of re-inlining the workspace check, and integration-branch shell interpolation in base-commit capture uses POSIX single-quote escaping.
• Fix narrow right-sidebar Dev Server preview overlap by replacing the inline preview with an accessible modal launcher when the dock is very narrow, while keeping inline preview for full-page, mobile viewport, and expanded pop-out hosts.
• Fix ntfy test notifications to honor unsaved Settings form config so users can enable ntfy, enter a valid topic/server/token, and send a test notification before saving.
• Close task detail dialogs and embedded task-detail hosts immediately after delete confirmations complete, while delete requests continue reporting success or error toasts asynchronously.
• Stack task-detail Chat agent headers above output blocks in the List View split-pane detail pane while preserving full-width desktop chat layout.
• Merger unification (master-plan U0): runAiMerge (the FN-5633 clean-room AI merge path) is now the sole merge path. The engine dispatch, the fn task merge CLI command, and the UI-only (--no-engine) dashboard merge all route through runAiMerge; the legacy aiMergeTask pipeline is soft-deprecated (body retained, @deprecated). The merger.mode setting is now inert and deprecated — the type and field are retained as published surface, but the "deterministic" value no longer selects a different pipeline; observing it logs a one-time deprecation warning and proceeds via the unified AI merge path. A new shared assertNotWorkspaceTaskMerge guard rejects workspace-mode tasks (populated workspaceWorktrees) at every merge entry point with a clear error until per-repo merge support (master-plan U6) lands.
• Fix multiworkspace tasks failing to complete. task.workspaceWorktrees is now durably persisted (it previously had no SQLite column, so fn_acquire_repo_worktree's write was dropped on every persist and fn_task_done always reported "acquired no sub-repo worktrees"). Concurrent workspace tasks no longer collide on the shared browse-root active-session path — each task gets a task-scoped session key, so a second workspace task no longer fails with "active-session path … is held by …".
• Breaking: the WorkflowOptionalStep type, previously exported from @runfusion/fusion, is removed — any consumer importing it must migrate to optional-group nodes / ResolvedWorkflowOptionalStep.
• Add X-Session-Id and X-Session-Affinity request headers to all LLM chat completion requests. These let LLM gateways sticky-route consecutive requests from the same conversation to the same backend, and let observability tools (Langfuse, Arize, etc.) group the otherwise-stateless API calls of a session into a single multi-turn trace. Both headers carry the same stable identifier — the task id when available (stable across pause/resume), otherwise the pi session id. (#1675)
• Workflow editor: add a Help section to the node detail pane. Every node now documents what it does, how to configure it, and its inputs/outputs/edges — including the engine-managed merge-lifecycle nodes (auto-merge gate, branch-group member integration, branch-group promotion, PR and recovery nodes), which are surfaced read-only with an "Engine-managed" badge.
• Workflow editor: optional steps are now graph-native. A new optional-group container node (foreach/loop-style) holds a subgraph the executor runs once when the group is enabled for a task (per-task enabledWorkflowSteps + workflow defaultOn) and bypasses when disabled. All seven built-in add-ons (documentation-review, qa-check, security-audit, performance-review, accessibility-check, browser-verification, frontend-ux-design) are insertable from the node-editor palette as a node or wrapped in an optional-group. The built-in coding and stepwise-coding workflows now express browser-verification as an optional-group. Optional-group enable resolution correctly handles id collisions with add-on template ids, so a group's enable state is not silently bypassed during task creation/update. (The legacy declaration-based optional-steps model is retired in a sibling changeset; only the workflow-step seam infrastructure removal remains a follow-up.)
• Workspace tasks no longer render blank in the dashboard. Task cards and the task
• Add workspace mode: open a folder of git repositories as a single Fusion
• Workspace mode (Phase A / U2): harden per-repo worktree acquisition. Each sub-repo worktree now gets the task identity guard installed (single-repo parity), a per-repo base commit SHA captured local-first against that sub-repo's resolved integration branch (shared integrationBranch override stripped so each repo falls through to its own origin/HEAD), and same-sub-repo acquisition exclusivity registered in the path-keyed active-session registry. Re-acquiring an already-acquired (taskId, repo) is idempotent, and acquisition failures surface an error plus an audit event instead of silently stalling.
• Workspace mode (Phase C U3): serialize concurrent same-sub-repo lands with a per-repo file-scope lease. When two workspace tasks try to land onto the SAME sub-repo's local integration ref at the same time, the merge phase now registers the sub-repo's absolute path in the path-keyed active-session registry under a distinct workspace-repo-land kind before each land and releases it in a finally (on land success or failure — no stuck lock). A second task contending for the same sub-repo fast-fails with a retryable WorkspaceRepoLandBusyError, which the existing partial-land auto-retry-then-park dispatch handles (consume a mergeRetry, re-enqueue with backoff, then operator-park). Disjoint sub-repos lease different paths and never serialize against each other. The lease prevents clean-room ai-merge worktree collisions; ref correctness is already guaranteed by advanceIntegrationBranchRef's CAS (concurrent-advance → rebuild).
• Workspace mode Phase A (U1): executor session scoping. In workspace mode the executor now skips the root worktree acquisition and every rootDir git preflight (base-commit capture, contamination, worktree-liveness), runs the agent session rooted at the browse-only workspace root, and tracks acquired sub-repo worktrees as a per-task set. Single-repo tasks are unchanged (one-element set, byte-for-byte preflight parity).
• Workspace mode (Phase B, U1): per-repo post-session change capture, contamination detection, and worktree-invariant verification. In workspace mode the executor now loops task.workspaceWorktrees, reusing captureModifiedFiles per sub-repo (diffing each against its own baseCommitSha, with a merge-base fallback when undefined) to aggregate repo-prefixed task.modifiedFiles and surface per-repo contamination, and un-stubs verifyWorktreeInvariants to assert each acquired worktree's git toplevel and fusion/<id> branch. Single-repo behavior is unchanged.
• Workspace mode (Phase B, U2): per-repo review at both review entry points plus per-repo fn_task_done completion + scope-leak verification. In workspace mode both review call sites (the in-session fn_review_step tool and the step-inversion review seam) now loop the single-cwd reviewStep once per acquired sub-repo (cwd = each repo's worktree) and aggregate the repo-tagged verdicts as a conjunction — the task is reviewed only when every sub-repo approves, and the first failing sub-repo's verdict (with repo-tagged findings) drives the existing verdict→edge mapping. fn_task_done now verifies worktree invariants per acquired repo and iterates the scope-leak guard per sub-repo (cwd = repo worktree, repo baseCommitSha), blocking completion on any sub-repo carrying off-scope changes and naming the repo. Adds a minimal shared repo-prefix-derivation helper (workspace-paths.ts). Single-repo behavior is unchanged.
• Workspace mode Phase C (U1): per-repo merge loop. Extract landOneRepo from the
• Workspace mode Phase C (U2): per-repo landed predicate, finalize-once, and idempotent
• Workspace mode Phase D (U1): workspace-aware self-healing. The existing merging-status reconcilers no longer mis-finalize a partial-landed workspace task (recoverInterruptedMergingTasks now clears the transient merging status and re-enqueues the idempotent per-repo land instead of running the single-commit finalize over the non-git workspace root), and recoverMergeableReviewTasks now admits workspace tasks (task.worktree is null). Adds three reconcilers: partial-land recovery (re-enqueue via enqueueMerge, FORK-A unrecoverable → park failed; guarded by autoMerge:false + user-pause + workspace-aware liveness), phantom workspace-repo-land lease reclaim (new entriesByKind registry seam), and per-repo worktree cleanup from stored paths (no temp walk). New run-audit events: task:reconcile-workspace-partial-land(-no-action), task:reclaim-phantom-workspace-land-lease, task:reconcile-orphaned-workspace-worktree.