Skip to main content
CHANGELOG

What shipped

Released versions of Harmony only, read from the product's Keep a Changelog file. Unreleased work is not listed here: if it is not in a version heading, it has not shipped.

Latest released v3.25.8 · · 65 releases

Download the current build from /download. Channel and exact version always come from the live release manifest.

Releases 11–15 of 65

Page 3 of 13

v3.20.0

Added

  • Experimental: hidden panes can skip output parsing (Settings → "Skip hidden pane rendering"). Even with the shared output scheduler, hidden agents' output was still parsed eventually — and measurement showed that parsing total is what drags the visible pane once several background agents stream at once (4 hidden flooders pulled the visible pane down to ~10–20fps). With this toggle on (daemon sessions only, default off), hidden panes' output is queued but never parsed: the renderer does no parsing work for panes you aren't looking at. A pane whose backlog outgrows its cap is marked stale and transparently re-synchronized from the daemon's session buffer when revealed — the daemon replays the authoritative bytes onto a reset terminal, so what you see on reveal is the pane's true current state, never a duplicate or a half-parsed frame. Agent-facing buffer reads (wmux_search_panes, terminal_read) hydrate a stale pane before reading so orchestrating agents never see old output. If a re-sync can't complete (dead session, legacy daemon), the pane degrades to its last-known screen instead of sticking or losing its identity.

  • Diff comments now wake the task agent (J4). Commenting on a hunk in a fan-out task's diff surface no longer just records a note — it @-mentions the task's agents on the mission-channel post, so the existing mention→wake loop nudges them to read and act on the feedback. Every non-human member of the mission channel (excluding you, the commenter) is mentioned at the workspace level, so multiple agent panes sharing one workspace all get woken; if every agent has left the channel the comment still posts, just without a mention. The post's body also carries a [diff: <file> @ <hunk>] <comment> prefix so an agent reading the channel over the CLI or MCP (which don't render the structured anchor) still sees which file and hunk the comment is about. The success message reports how many agents were pinged.

  • Fleet cards surface an agent's completion evidence. A fleet card now shows a small ✓ evidence n/m badge when the pane's most recently completed A2A task carries structured completion evidence — n is how many of the m evidence items are actually verified (a passed command, or a verified inspection/artifact). It's the "trust it ran unattended" proof made legible on the card: the check reads green once at least one item is verified and stays muted when nothing is (verified is a grade, not a claim), and the task title plus the evidence summary live in the badge's tooltip so the on-card text stays a single compact token. The badge reads existing task state only (no new store or round-trip), is addressed per-pane (a pane-pinned task shows on exactly that pane; a workspace-level task shows on the workspace's active pane), and simply isn't drawn when there's no such task.

Fixed

  • Multiple workspaces full of busy agents no longer stutter the visible terminal. Every pane used to push its PTY output straight into its own terminal the moment it arrived over IPC — including panes in hidden workspaces — so a fleet of background agents ran that many independent parse/render pipelines on the one renderer thread, and the pane you were actually typing into starved between them. Terminal output now flows through a single shared scheduler: the visible pane keeps the exact direct-write path it always had for ordinary output (zero added latency), while hidden panes' output is batched and drained cooperatively under a hard per-tick time budget, so no amount of background agent chatter can pin the UI. Even the visible pane's own output floods are chunked through that budget rather than parsed in one blocking pass, so watching a chatty agent stays responsive too. Nothing is dropped — a hidden pane's backlog is handed over in full when it becomes visible (before its reveal repaint), when a reconnect replay needs it, or if it ever exceeds the scheduler's memory cap (which simply restores the old behavior for that pane).

  • Diff-panel comments now actually post to the mission channel. The diff comment post omitted the sender identity the daemon requires, so every comment was rejected with a "코멘트 발사 실패" authorization error instead of being recorded. The comment now posts as the diff's owner workspace (its own mission-channel member row), which is also what lets the new @-mention wake the agent.

Security

  • events.poll no longer lets an agent eavesdrop on another workspace's channels (audit B3). The event-poll RPC previously scoped its results by a caller-supplied workspaceId, so a same-user pipe/MCP client could live-subscribe to any workspace's private channel messages, channel lifecycle, and A2A task pointers just by naming that workspace's id — no pane identity required. Those confidentiality-sensitive event types are now scoped to a server-resolved workspace derived from the caller's verified senderPtyId (the same identity anchor the a2a.channel.* mutations already use), and the caller-supplied workspaceId is ignored for them; an unresolvable caller receives none of these events (fail-closed). The bundled MCP wmux_events_poll tool forwards its own PID-walked senderPtyId, so a legitimately-placed agent still sees its own channels and tasks unchanged. The first-party operator surface (the app's own renderer/plugin host) keeps scoping across the local workspaces it names. Ordinary lifecycle events (pane/process/agent/workspace metadata) are unaffected — their all-workspace firehose was already reachable by any events.subscribe subscriber, so their workspace scope was never a confidentiality boundary and external lifecycle subscribers keep working.

v3.19.0

Added

  • Task lifecycle: close, one-click PR, and a cleanup list (J3). A fan-out task's diff surface now carries 닫기 (Close) and PR buttons, so you can finish a harvested task without touching the terminal. Close runs in a deliberate order — it removes the task's git worktree first and only commits the close (and archives the mission channel) once the worktree is gone, so you can never end up with a "closed" task whose output still litters disk. If the worktree is dirty, close is held: the task stays open, the output is preserved, and a toast tells you to review the diff and commit/PR or discard it. If there are committed-but-unpushed commits, close warns instead of silently dropping them. PR is one click (with a single confirm that names the branch and warns a pre-push hook may run): it gates on gh being installed and authenticated, refuses if the worktree is dirty (uncommitted work wouldn't be in the PR), pushes the branch, and opens a PR against the repo's default branch — and it's idempotent, so a second click after a half-finished attempt recovers the existing PR URL instead of erroring. The PR URL is recorded on the task and the PR-status cache is refreshed immediately. A new "태스크 정리 목록" (Task Cleanup List) command in the palette scans the dedicated worktree root against live tasks and surfaces four kinds of leftovers — unmaterialized-open, disk-missing, dirty-preserved, and orphaned directories (reverse-mapped by an on-disk task.json stamp so they're identifiable even after a closed task ages out of memory) — with an inline Close for the ones that are still open tasks. If a fan-out agent pane comes up but its prompt never fired, you now get a "프롬프트 미발사" toast with a 재발사 (re-fire) action that re-sends the task's original startup command (agent launch + prompt together, same sanitization as the normal path) after checking the prompt file still exists — it never pastes the raw prompt into a bare shell. Finally, a task workspace whose pane wanders outside its worktree boundary gets a small ⚠ 이탈 badge in the sidebar (best-effort, warning only — nothing is blocked).

  • Operators can now join private agent-made channels. The channels panel grows a collapsed discovery section listing every channel on the daemon — including private rooms agents created without inviting the human, and archived rooms for audit visibility — with a one-click join. Joining seats the operator as a regular member with full history, and appends a server-published, viewpoint-neutral system marker ("Operator joined this channel") to the channel as an audit row; the marker consumes a sequence number but owes no member an unread, so agents are not nudged by it. The join surface is strictly human-side: the RPC methods are unreachable from agent transports (pipe router unregistered, first-party MCP exclusion), pinned by boundary tests.

  • Fan-out missions are now visible in the sidebar and fleet panel. Workspaces created by a J1 fan-out now show up under a "Missions" group at the top of the sidebar (title, open/closed status, and a link into the mission's channel) — the group only appears when a workspace has fanned out, so ordinary workspaces are unaffected. The fleet panel's cards also grow a mission line when they belong to a fan-out task. The existing worktree badge (⊕) is untouched — it marks the low-level "this is a git worktree" fact, while the new Missions section marks the higher-level "this is a fan-out task" fact, and a workspace can carry both. Mission data is read-only and pulled (mount + workspace-set changes + a 15s background poll for status drift + an immediate refetch right after a fan-out completes), since the daemon doesn't push mission updates.

Changed

  • Fleet view is now always-on chrome instead of a full-screen modal. Ctrl+Shift+A still toggles it, but it now mounts as a fixed-width panel alongside the workspace sidebar and channel dock (mirroring the channel dock's existing flex-sibling layout) rather than a fixed overlay with a backdrop — other panes stay visible and interactive while it's open, and closing it no longer drops keyboard focus into <body>: the element that had focus when it opened is restored. The fleet/approvals/remote tabs, keyboard row-navigation, and approve/deny shortcuts are unchanged; the card grid narrows to fit the panel's width instead of a full-screen layout. Two focus bugs found in review were fixed before this landed: opening the panel now lands real DOM focus on the active card/row (not just the panel container, which used to leave keyboard users unable to reach any card when only one was present), and row shortcuts (Enter=approve, Backspace/Delete=deny) now only fire when the option row itself is focused — previously an auto-approve checkbox could steal focus and cause those keys to mis-fire as an approval/denial.

  • Type scale: apply the wave-1 semantic tokens to the always-visible chrome. The sidebar (WorkspaceItem, MiniSidebar), channel dock (ChannelsPanel, ChannelView, ChannelMembers), and fleet panel (FleetCard) now use .text-caption/.text-body instead of hardcoded text-[11px]/text-[13px] — swapped only where the token's actual size (caption=11px, body=13px) matches the literal exactly, so there is no size change. Elements that already carried an explicit font-*/leading-* utility are unaffected (utilities win over the token's own weight/line-height); a handful of small mono labels that had no explicit weight now pick up the caption token's weight 500 instead of the browser default 400 — a deliberate, disclosed exception, not a bug. 8px/9px/10px/12px literals in these six files are left untouched (no matching token without a size change) for a later pass.

  • Design tokens: promote hardcoded modal shadows, z-index literals, link accent, and typography to named tokens (visual-invariant). Internal design-system cleanup with no visual change: the six-way-duplicated 0 25px 60px rgba(0,0,0,0.75) modal shadow and the rgba(0,0,0,0.6) backdrop are now --shadow-modal/--backdrop-modal; eight ad-hoc z-[…] literals map to a named --z-* stacking scale (values and relative order unchanged); the link accent gains an accentSecondary token wired to the existing accent value across all eight built-in themes (a hook for future differentiation, currently identical); and a four-tier typography scale (--text-display/-title/-body/-caption) is defined with three representative applications. All values are byte-identical to the originals — verified against the pre-change literals by a three-model review — so themes render exactly as before. The sidebar's two bespoke "Copied!" DOM toasts (workspace-info copy and cwd copy), which each hand-built a bottom-center element and bypassed the canonical toast surface, now route through the shared toastSlice/ToastContainer so copy feedback is styled by one token-driven container instead of duplicated inline CSS (they adopt the app-wide bottom-right/5s presentation as a result). Four dark-only hardcoded hex values that broke the light themes are tokenized: the browser title bar and URL-bar resting state (#11111b → var(--bg-mantle)) and the browser-close / palette-item hovers (#3b1e1e/#2a2a3d → var(--bg-overlay)) now read correctly under hinomaru/taegeuk — these four spots intentionally normalize to the sibling components' tokens, so dark themes see a subtle shade shift there (e.g. #11111b → #181825, and the two outlier hover tints join the twenty sibling hovers already on --bg-overlay) rather than staying byte-identical. The custom-theme-editor, contrast-warning, and color-inspect chrome keep their fixed high-contrast hex by design (they must stay legible while the live theme is being edited/broken), and the webview inspector overlay keeps self-contained hex because it is injected into arbitrary guest pages that have no wmux theme variables.

Fixed

  • UI responsiveness: clicks no longer contend with a background re-render storm. Interaction latency ("every button feels sluggish") had two dominant causes, both fixed. (1) Renderer re-render fan-out: seventeen always-mounted components (sidebar, status bar, channels panel, composer, palette, fleet view, …) subscribed to the entire workspaces tree, which is replaced on every agent-output metadata tick — and the renderer had zero React.memo barriers, so agent activity re-rendered large components continuously and clicks landed on an already-busy render thread. Subscriptions are now minimal derived selectors backed by a reference cache (unchanged projections return the same array/element references, so components only commit when a field they actually display changes), workspace list items self-subscribe by id behind React.memo, title/cwd/git-branch metadata writes are coalesced to one store write per frame, and the 1-second status-bar clock is isolated into its own tiny component. A new re-render regression suite (React Profiler commit counting + selector reference-contract tests) pins the fix: unrelated workspace churn now produces zero commits in unrelated components. (2) Main-process stall: the 5-second periodic session autosave performed a synchronous atomic write on the main event loop, delaying whatever IPC a click had just issued. The periodic path is now an async atomic write with a write-epoch guard and post-write recovery — if an in-flight async write races a newer event-driven synchronous save (the reboot-survival path), the newer snapshot is re-committed immediately, so the final on-disk state matches the latest save under any interleaving (crash-loss window unchanged at ≤5s; exit paths still flush synchronously).

Added

  • Diff review & hunk adoption: harvest a fan-out task's output (J2). Fan-out tasks now have a fourth surface type — a diff surface — that reads a task worktree's uncommitted changes against its merge-base and lets you review, comment, and cherry-pick them into the target repo. Fan-out's result toast gains a "diff 열기" action that opens the diff for that task's workspace. The panel shows a file tree (numstat), a unified diff (+/- coloring only — no full IDE editor, by design), per-hunk checkboxes, and an adopt button. Adoption is all-or-nothing: the selected hunks are reassembled into a single patch (file headers and hunk bodies preserved byte-for-byte, only hunk line-counts recomputed) and applied with one git apply — the target is either fully changed or fully untouched, never half-applied. Adoption is gated hard: a target snapshot (HEAD/branch/dirty set) is captured at read time and re-verified at apply time (rejects if the target moved), any selected file that is dirty in the target is refused (conflict avoidance), a combined pre-apply --check is the gate (so hunks that only apply together aren't wrongly blocked), and hunks already applied to the target are surfaced as an explicit failure so you can deselect them. Untracked files are synthesized into proper new-file patches (regular files only — symlinks/FIFOs are labeled unsupported so a symlink can't leak a file from outside the repo); rename/copy/mode/binary changes and files over the 512KB/2MB caps are display-only (adoption refused, double-checked). File names with spaces, non-ASCII, or quotes are handled correctly (-z porcelain, quotepath off). Comments post to the task's mission channel with a diff-comment anchor (file + hunk header) and render inline under the matching hunk on reload; comments whose hunk header no longer matches the current diff drop into a "위치 이동됨" group (v1 anchor precision is hunk-header granularity — line-level anchors are deferred). The whole path is backed by a validation rig that proves adoption atomicity under a mid-apply kill and catches a re-serialization corruption (dropped no-newline marker) as a shipping blocker.

  • Perf harness: N-pane instrumentation + boolean consistency gates (W2, dev/CI-facing). Extends the existing A1 app benchmark (scripts/perf-bench.mjs + scripts/perf-compare.mjs, driven by .github/workflows/perf.yml) rather than adding a new harness, turning the B2 engine-resume decision from an undefined "feels blocked" call into recorded numeric + pass/fail gates. Four scenarios now run by default on a dedicated bench instance (isolated from the coldStart/input/RAM numbers): (1) N-pane concurrent-streaming frame budget — the 8-pane split loop is generalized to spawnPanes(client, page, n), and at N=4/8/16 every pane's PTY is flooded with continuous output while the renderer's rAF cadence is sampled; each N is gated independently (scenarios.frameBudget.N{n}.frameDeltaMs.p95, ratio 2.0 = the strategy doc's "budget 2×"). (2) Korean IME composition — since CDP/playwright-core cannot drive a real IME, the scenario synthesizes the DOM composition contract xterm's CompositionHelper consumes (compositionstart/compositionupdate/compositionend + input + textarea.value diff) on the focused pane's hidden helper-textarea and verifies the PTY echoes the composed string (안녕하세요) back byte-for-byte; self-validating (a non-equivalent synthesis would echo nothing and fail). (3) Long scrollback — reuses the existing --scrollback-lines flag as a run combination (no new logic). (4) WebGL context-loss/restore — forces WEBGL_lose_context.loseContext()/restoreContext() on the focused pane's canvas and measures recovery via the webglcontextrestored event + !isContextLost() (plus a live-canvas re-count), recording recoveryMs. perf-compare gains a BOOL_GATES array (baseline-independent: scenarios.ime.pass / scenarios.webglContextLoss.pass FAIL immediately when present-but-not-true) alongside the three new numeric frame-budget gates; both stay record-only until an owner blesses a CI baseline (existing bench/baseline-ci.json convention). New CLI flags: --frame-budget-panes 4,8,16, --skip-frame-budget, --skip-ime, --skip-webgl-recovery. Pure logic (frame-stat summary, IME echo comparison, gate judgment) is factored into scripts/perf-scenarios.mjs and unit-tested; the CDP-driven scenario bodies are validated on the Windows CI target only (this being a macOS worktree, they cannot run locally — an honest, documented limitation). No product-code (src/) changes.

  • Fan-out: one prompt → N isolated agent tasks (J1). The AgentToolbar gains a fan-out entry that spawns up to 8 WorkTask missions from a single prompt, each with worktree isolation by default: a dedicated git worktree under {wmux home}/worktrees/{repoHash}/{taskSlug} on a fresh wtask/{slug} branch, a dedicated task workspace (agent pane + shell pane, startupCwd pinned to the worktree), an auto-opened private mission channel (task workspace invited as a member), and the prompt delivered via a file-backed initialCommand (prompt body lives outside the worktree so task diffs stay clean; the path is shell-quoted for POSIX and PowerShell). The whole call is idempotency-keyed end to end — double-clicks and IPC retries can never mint duplicate worktrees — and a global preflight validates the repo and every task's slug/branch before any task or channel is created (unfit input rejects the batch with zero side effects). Per-task failures compensate individually (mission closed, channel archived, any created worktree preserved — never deleted) and surface in a per-task result report (materialization / channel-link state). Worktree operations are serialized per repo (no index.lock races), dirty worktrees refuse removal (preserve-and-list; no force-delete API exists), and bare/submodule/LFS repos fail closed. The daemon activates the reserved task.update materialization path (branch/worktreePath/paneGroupId, write-once monotonic, owner-or-CEO gated) and enforces the canonical-worktree-path exclusivity invariant. A separate broadcast-only action (send text to every terminal pane in the current workspace) is deliberately kept apart from fan-out — non-isolated "fan-out" does not exist. Includes a reboot-survival demo script (single task round-trip: daemon restart → projection restored, worktree intact on disk).

  • WorkTask mission channels: durable task canon + minimal mission-channel lifecycle (J0, dev-facing). Introduces WorkTask — the worktree-mission unit (domain:'task' in the append-only event log) that J1 fan-out and J2 diff will build on — as a projection-first daemon service (daemon/worktask/WorkTaskService), kept deliberately distinct from the A2A Task (different lifecycle + transition graph). Two new pipe RPCs plus their thin MCP tools (channel_mission_start / channel_mission_close) create a WorkTask AND a bound private mission channel in one call, and close flips the task to closed while archiving the channel. Ownership is server-constructed and born-owned (owner = createdBy, never caller-supplied); close authz is a task-level gate (owner OR CEO), the first line of defense over the channel gate. Identity rides the same senderPtyId → verifiedWorkspaceId server stamp as a2a.channel.* mutations (fail-closed on unresolvable identity). Crash-safety is enforced end-to-end: mission channels carry a wmux:mission:{taskId} topic anchor, boot runs a fixed replay → bidirectional reconcile → closed-GC order (an orphan channel from a crash between channel-create and task-append is archived; a closed task whose channel is still active is re-archived — both idempotent no-ops when already settled), and an append-failure on start triggers an immediate compensating archive (the empty-channel reaper cannot reap it — the creator remains a member). Start/close are idempotency-keyed so a lost-response retry never creates a duplicate mission + channel, and re-closing an already-closed mission is a no-op success. Closed tasks are GC'd from the projection after 7 days (log untouched — a view bound only), with archive-unconfirmed tasks exempt. J1+ materialization fields (branch/worktreePath/paneGroupId/prUrl) and the §6.M lease / born-pending contract are schema-reserved but not yet active; task.mission.list is pipe-only in J0 (MCP exposure deferred to J1). Renderer unchanged.

  • E0 conformance harness: recorder + corpus + differential runner (§6.A M1/M2, dev-facing). Introduces the terminal-emulator conformance harness under top-level core/harness/, the measurement scaffolding for the future clean-room VT core. M1 (recorder + corpus): a script-driven recorder (recorder.ts) spawns a real PTY via node-pty to exercise initial geometry + resize, then emits a deterministic recording.bin (raw bytes), events.jsonl (init/resize/reflow_mode trail with monotonic byte offsets), and meta.json (seed + workload-script sha256). PTY spawn, resize, and abnormal-exit failures are escalated (thrown) rather than swallowed, so a broken geometry-exercise path fails the gate instead of silently no-op'ing. The committed corpus (corpus/) is six deterministic synthetic workloads only — scroll flood, resize roundtrip (80→79→80, an explicit non-reflow control at 40 chars where no wrap occurs), resize reflow (120 chars that wrap into two rows, so the 80→79→80 roundtrip actually exercises the rewrap path — its golden pins xterm.js's observed deterministic post-roundtrip state, not an idealized restoration), alt-screen enter/exit, CJK/emoji/VS16/ZWJ width cases, and the SGR spectrum (16/256/truecolor + attribute flags) — each carrying ≥3 golden assertions next to its definition. A companion miner (miner.ts) scrubs {stateDir}/buffers/*.buf dumps (multi-layer: api-key/token/secret key=value, AWS uppercase-snake credential envs, URL userinfo, JSON "key": "…" credentials, PEM private-key blocks, known token prefixes sk-/ghp_/gho_/xox…, Bearer headers, OSC 52 payloads, and a base64 high-entropy heuristic) to a local-only, git-ignored output whose write root is pinned to core/harness/corpus-local/ (an isolation guard rejects any in-repo non-ignored path) — .buf preserves only the ring tail (no geometry), so mined output is for mid-stream robustness and fuzzer seeds, never the deterministic corpus. M2 (differential runner): differ.ts feeds a recording into @xterm/headless@6 (with @xterm/addon-unicode11 pinned to Unicode 11 as the baseline width model) behind a Subject interface (our E1 core and a third reference plug in later), extracts a full-cell grid snapshot (char, width, fg/bg + portable color booleans, 9 style flags, cursor, active buffer), and diffs two snapshots cell-by-cell into a report whose classification schema encodes the four-way ledger (our-bug / xterm-bug / spec-ambiguous / intended) — where intended is admitted only via an explicit approval list (intended-diffs.json, loaded onto the diff path via loadIntendedDiffs), never implicitly. The diff compares the active buffer (normal vs alternate) before cell comparison and excludes xterm.js's non-portable raw color-mode integers from cross-subject comparison; before replay, the event stream is validated (first event is init, byte offsets are monotonic non-decreasing in original order and within range) and violations throw rather than being hidden by sorting; reflow_mode events encountered during replay are honestly recorded on the result. The four-part baseline gate ships as tests: determinism (two xterm.js runs identical) — including a chunk-boundary robustness check that feeds each recording one byte at a time and requires an identical layout to whole-buffer feed (a narrow, documented ZWJ-joiner-at-write-boundary char difference is the only tolerated exception; widths/cursor/colors/flags must match) — no-crash full-corpus completion, golden-assertion pass, and record→replay round-trip stability that reads the committed corpus into memory first and regenerates into a separate temp dir (the gate never writes the repo corpus, so the drift check is no longer a self-comparison). Throughput is recorded as the xterm.js baseline (steady-state feed MB/s + full-cell extraction time). Wired as a fourth vitest lane (vitest.harness.config.ts, tsconfig.harness.json, npm run test:harness). Zero product-code changes; existing test lanes and typecheck unaffected.

Show the other 10 changes in v3.19.0

Added

  • Append-only event log: crash-safe primitives (envelope PR1). Introduces the segmented NDJSON append-only log (daemon/eventlog/AppendOnlyLog) and the shared event-envelope schema (shared/eventlog) — the foundation for rewiring the channels and A2A canonical state to a crash-safe commit log (§6.L). Key properties: fsync coalescing (group-commit batches), single-ftruncate per-batch rollback, boot-time forward-scan recovery (trim at the first corrupt byte, no partial promotion), Lamport/seq high-watermark resume (reuse forbidden, gaps permitted), and fail-stop on truncation failure rather than silently diverging coordinates. Includes machine-id minting and recovery, and a durable option for atomicWrite (fsync sequence). No service is wired to this log yet — that lands in subsequent PRs.

  • Event log migration engine (envelope PR2). Adds the zero-downtime boot gate (daemon/eventlog/migrateToEventLog) that promotes legacy channels.json to log mode, plus the durable-only EventLogManifest (atomic migration-complete marker) and SnapshotStore (latest → .bak → reseed → genesis fallback chain). Detection uses three branches: inexplicable state is quarantined under quarantine/ and retried rather than silently accepted. Conversion failures leave the legacy file intact and are idempotent on retry. Downgrade detection uses a Lamport + state-hash watermark — a record of an older daemon's writes triggers a reseed snapshot. Compaction safety: no truncation before durable confirmation; genesis and reseed snapshots are never truncated. Not wired into daemon boot yet.

  • A2A tasks are now durable in the daemon event log (envelope PR4). Canonical A2A task state moves from the renderer's in-memory store (30-min GC, lost on restart) into A2aTaskService in the daemon, persisted as domain:'a2a' envelopes in the append-only log. Create, transition, and cancel all reach the log under fsync commits; tasks survive restarts via projection replay. VALID_TRANSITIONS is enforced daemon-side — out-of-graph transitions are rejected at the canonical source. Background ClaudeWorker transitions (working / completed / failed) now route through the daemon rather than writing directly to the renderer, carrying completion evidence along. The renderer a2aSlice is demoted to a read cache that applies daemon commits verbatim without re-validation; when the daemon is unavailable the existing renderer validation path is the automatic fallback (no degraded behavior). Workspace close force-fails in-flight tasks in the log so they do not resurrect on restart; completed tasks are periodically pruned. Daemon canonical state wins over a stale cache on reconnect, including immediately after restart.

  • A2A event authContext is now server-stamped; daemon.ping exposes the active log format generation (envelope PR5). The authContext.principalId in every A2A task event (create, transition, cancel) is now derived by the daemon from stored task coordinates rather than accepted from the caller's claim — actor pane for transitions (to.paneId), caller-side pane for cancel/create, workspace fallback for headless workers or unpinned tasks. principalId and trustTier are display/routing/audit fields only; the authorization anchor remains the server-pinned verifiedWorkspaceId invariant. trustTier is always 'semi-trusted', resolved unilaterally by the server (the temporary caller-override field from PR4 is removed — callers cannot claim a trust tier). daemon.ping responses now carry eventLogFormatVersion additively: present when log mode is active (value = the active format version integer), absent in the legacy fallback. Absence signals a pre-envelope daemon to the auto-replacement logic, which treats unknown format generations fail-closed.

  • A2A completion evidence: schema and pure validator (§6.M P1). Introduces the CompletionEvidence schema and a pure, side-effect-free validator (shared/completionEvidence.ts). Gate = structure: non-empty summary, well-formed items, sanitized paths, DoS caps on body lengths and item counts. verifiedItemCount is derived honestly — an all-unverified completion is accepted at grade 0 rather than rejected (grade is observability, not a gate requirement). Path sanitization rejects colons, leading separators, .., and C0 control characters (undecoded literals enforced). Untrusted-wire normalization: plain-object check, hasOwn gating, fresh-object copy to prevent prototype pollution. Not wired to any transition at this point — gate activation is the next PR, after envelope PR4.

  • A2A completion evidence: production and transport wiring (§6.M P1). ClaudeWorker now produces structured completion evidence from its Claude run results. Both success and failure paths emit inspection + unverified self-report — run-success is never promoted to verified (no laundering). MCP a2a_task_update transports evidence via a dedicated evidence parameter; the contract is fixed in the tool description and coexists with the existing artifact channel. The renderer bridge normalizes untrusted wire shapes before they reach the store: a poisoned shape is stored as completion_evidence_malformed (additive-inert — no task state change at this stage), and server-only stamps like recordedBy are stripped on ingestion. No rejection gate yet — that is the next PR.

  • A2A completion-evidence gate activated (§6.M P1). completed/failed A2A task transitions now require structured completion evidence: completed needs a non-empty summary plus at least one well-formed item (command/inspection/artifact), and failed needs a summary (the failure reason). The daemon A2aTaskService.transition is the single enforcement point; the renderer fallback writer applies the same gate for pane-pinned tasks driven by a pane-identity caller or when the daemon is unavailable. Rejections return actionable reason codes (completion_evidence_missing, completion_evidence_no_items, completion_evidence_empty_summary, completion_evidence_invalid_item, failure_reason_missing) and leave task state unchanged with no log append. verifiedItemCount remains an honest grade rather than a gate requirement — an all-unverified completion is still accepted (grade 0). Workspace-teardown force-fail and verbatim application of daemon commits intentionally bypass the gate to prevent split-brain.

  • Completion evidence grade is now observable in A2A task events (§6.M P1). a2a.task events received via wmux_events_poll now carry verifiedItemCount (count of independently-verified evidence items; 0 = unverified completion) on completed and failed transitions. Event pollers can now distinguish an unverified completion (grade 0) from a graded one without querying the task separately. The count is derived from task.status.evidence at terminal transitions only — non-terminal transitions such as working carry no count. The renderer's primary publisher emits it; workspace-teardown force-fails emit a separate grade-0 event. The trust boundary admits only non-negative integers (forged or out-of-range values are dropped silently). created and cancelled pointers carry no grade field.

  • Validation rig: harness core + SIM smoke (§6.G, dev-facing). Introduces the self-verifying harness under top-level rig/. Components: run isolation (isolation.ts — fresh temp home per run, 4-env wipe of HOME/USERPROFILE/APPDATA/LOCALAPPDATA, WMUX_DATA_SUFFIX='-rig-{runId}'), headless daemon wrapper (daemon.ts — dist/daemon-bundle spawn with a detached process group, daemon.ping ready-poll, group tree-kill, respawn, explicit error on missing bundle), daemon pipe client (pipe.ts — persistent-socket JSON-RPC, dual-ok-layer unwrap, G6 honest-main discipline: one workspaceId binding per persona, throws on cross-workspace impersonation or reserved identity claims), state assertion helpers (assert.ts — seq integrity, full-body cross-check, unread counts, canonical coordinate comments), and deterministic seed (seed.ts). SIM scenario S1 (flood ×8 concurrent senders → getMessages full cross-check: all-delivered, seq-continuous, no-duplicate) lands as a third vitest lane (vitest.rig.config.ts, npm run test:rig:sim, requires npm run build:daemon first). Zero product-code changes; existing two test lanes unaffected.

  • Validation rig: simulator scenarios S2–S8 + SIM regression-detection evidence (§6.G, dev-facing). Completes the synthetic multi-agent simulator on top of the R1 harness. The persona framework (rig/harness/persona.ts) handles identity assignment, channel preamble, seed wiring, and member lifetime; behavioral scripts are owned by each scenario. Deterministic scenarios S2–S8 each run against an isolated daemon: S2 channel integrity under ping-pong load; S3 dead-member expiry — unread, membership, and message-ledger remnants asserted against the client-side cursor only (avoids cursor-circular derivation from lastReadSeq); S4 hung-member: post commits immediately with no infinite hold, unread stays accurate; S5 deliveryStatus receipt contract pinned at current behavior (ack-only pending→delivered); S6 cap-boundary ±1 at the wire level (body 8192 B, mention cap 64, evidence item count 64 / item string 4096 B — string overflow is too_large at the gate, item-count overflow is malformed at wire normalization); S7 SIGKILL mid-flood → respawn → one-way subset assertion {ok-commits} ⊆ replay (at-least-once tail promotion: "no uncommitted resurrection" is intentionally NOT asserted); S8 full A2A lifecycle (send→working→completed, gate-rejection→retry, idempotent resend) plus detection of the #354 idempotency-authz ordering bug (non-participant key-replay is blocked after authz, not before). EPERM chaos: chmod 000 on the Unix socket → client isolation, daemon survival, and recovery confirmed; skipped under root (DAC bypass). CL7 early gate opened via stage-1 detection evidence (rig/EVIDENCE.md): #354 fix reverted on a scratch branch → S8 red confirmed → main green restored. Dogfood script catalog (rig/CATALOG.md): 29 scripts triaged — absorb 4, keep 24, retire 1 (zero physical deletions). Zero product-code changes.

v3.17.0

Added

  • wmux now updates its own background daemon — no manual restart. When an upgraded app reconnects to a daemon left running by an older version, it replaces it automatically: the old daemon suspends every session durably (scrollback, running commands, agent conversations), a current-version daemon starts, and your panes restore themselves — scrollback replayed, supervised commands relaunched, agents resumed. Same session preservation as a full quit-and-restart, without the quit. A brief "Updating the background daemon" toast explains the pause. The 3.16.0 stale-daemon banner remains as the fallback for the cases the replacement deliberately refuses (a NEWER daemon is never downgraded; a daemon that won't shut down cleanly is left running rather than force-killed pre-save).
  • Every agent in a channel now has one honest name — owned by the server, not typed by the agent. Channel display names are derived by the daemon from its pane registry (the same auto-names you see on panes, like w26-1(claude)), so an agent can no longer post under an arbitrary label and two Claude panes can never collapse into one indistinguishable "Claude Code". Names even follow agent swaps: replace claude with codex in a pane and its next message posts under the new name automatically.
  • Recovered agents show up as invite and @-mention candidates right after launch. Previously a workspace you hadn't visited yet contributed nothing to the "Add an agent pane" picker until you clicked into it once; the app now asks the daemon which panes are running agents at startup.

Changed

  • Quitting the app during a daemon replacement now does the right thing for both quit flavors: a normal Quit leaves the fresh daemon running with your restored sessions (tmux-style persistence), while "Shut down wmux completely" guarantees no daemon survives — including one spawned mid-replacement.
  • While the daemon is shutting down for a replacement (or full shutdown), new pane creation is rejected with a clear error instead of silently creating a pane that would be lost in the handover.

Fixed

  • Agents no longer get re-nudged about their own messages. A CLI/MCP agent posting under a stale member id matched no roster seat, so its own post counted as its own unread and the wake worker kept poking it. Posts are now mapped onto the workspace's actual seat (when unambiguous) — and when a workspace has several seats and none match, the sender gets an explicit warning instead of a silent identity fork, including on idempotent retries.
  • The same pane can no longer hold two channel seats. Joining once via the GUI and once via the CLI (or joining before and after agent detection) used to create duplicate roster rows — double nudges, double delivery entries. Joins now converge onto the pane's canonical seat and name the existing seat when they collide.
  • CLI agents stopped colliding on the shared "agent" identity. Panes are spawned with a unique $WMUX_MEMBER_ID, wmux channel join requires an identity instead of silently defaulting, and the join reply reports the seat you actually got.
  • Channel mention nudges are no longer typed into a plain shell terminal. When a member's agent pane was busy (its real Claude pane owned by the on-screen window), the wake worker could auto-submit its wmux channel read … hint into an agent-less shell, where it ran as a stray command; it now stays silent there and leaves delivery to polling.

v3.16.0

Added

  • You are ONE person in channels now — everywhere. Your channel identity is a single app-wide seat instead of one seat per workspace: the roster shows just "Me" (no more "Me · Workspace 2"), your channel list / memberships / unread badges are identical no matter which workspace is open, and joining or creating a channel no longer stamps whichever workspace happened to be active. The daemon merges your previously scattered per-workspace rows into the one seat at boot (deterministic, crash-safe, keeps your earliest join date and furthest read position).
  • Upgrades can't silently wipe your channels anymore. wmux keeps the background daemon alive across app restarts by design, so an upgraded app could attach to an old daemon and channels would look missing (posts failed with no explanation). The channels panel now detects the stale daemon and shows a "quit wmux fully and start it again" banner; it clears itself after the restart.

Changed

  • The unread badge is honest now. Agent posts from the workspace you're looking at used to be silently muted (workspace-level self-mute); with the unified seat, only YOUR OWN posts stay quiet — an agent posting from any workspace counts as unread, because it's news to you.
  • Adding a whole workspace as a channel member is retired — you are already in your channels as one seat, and agents join as individual panes.

Fixed

  • Private agent-only channels no longer leak into your dock. A private channel between agents whose workspace happened to be active could bump your unread badge for a channel you can't even open (phantom badge). Display is now scoped to channels you are actually in.
  • The channel wake worker no longer sweeps the virtual human seat every tick (it owns no terminal, so the sweep was pure CPU drift that grew with history).

Security

  • The reserved human seat cannot be invited, claimed, or targeted from the agent pipe — an agent could previously seed a phantom "human" member row that force-injected its channel into your always-on view. Rejected at both the pipe router and the daemon, so a direct-socket caller cannot bypass it either.

v3.15.0

Added

  • You can now tell agents apart in a channel. Every message shows the sender's pane identity chip (Claude Code · w26-1(claude)) plus a per-workspace color badge (round = a human seat, square = an agent pane); human posts read "Me · ", and the roster labels only YOUR row "Me" (another workspace's human seat reads as its workspace name). Previously every Claude pane rendered as an identical "Claude Code" and every workspace's human row read "Me".
  • Hand-typed @mentions now deliver. Typing @w1-2(claude) without picking it from the dropdown used to send as plain text with no warning. Typed tokens that match a live agent pane are promoted to real mentions — including when typed flush against Korean text or punctuation (확인요@…, cc:@…) — and tokens that match nobody get an inline "didn't match anyone" warning instead of a silent drop. An empty @-dropdown now says "No agents to mention" (dismissible with Escape) instead of rendering nothing.
  • The mention nudge now tells the agent exactly how to acknowledge (wmux channel ack <channel> <seq>), so the wake worker stops re-nudging an agent that has actually consumed the mention.

Fixed

  • Mentioning an agent no longer delivers twice. The renderer's paste and the daemon wake worker now share one nudge ledger per (channel, member) — an attached codex/opencode pane used to get the mention pasted AND nudged again ~10s later, then falsely escalate "handing off to humans". One paste covering several queued mentions debits the ledger once.
  • Agent greeting loops are cut at the source. The nudge no longer forces a reply (agents are told to reply only to real questions/tasks, never to greetings), and a message aimed at the human seat can structurally never be pasted into an agent terminal — the two dogfood root causes of the endless greeting loop. Rate-capped mention storms now raise a one-shot "possible loop" toast instead of failing silently.
  • A mention no longer vanishes when its target agent restarts. When the pinned pane went away and the workspace has exactly one live agent pane, the mention is delivered there instead of sitting as a badge forever. Genuinely workspace-level mentions stay badge-only.
  • A mention held while you reload the app is no longer lost. Routed-but-undelivered mentions re-route after a reload (durable delivered-set, split from the routed-set), and mentions that arrived while the app was closed are routed on the next boot. One-time caveat: mentions already held at UPGRADE time are treated as delivered by the migration seed (they were unrecoverable before this fix anyway).
  • A hung agent can no longer hold a mention hostage forever. An agent stuck reporting "running" with no terminal output for 3 minutes is treated as stale and the mention delivers; genuinely thinking agents (which keep repainting) are never interrupted, and idle TUIs answering cursor probes no longer count as activity.