Files
waggle-os/docs/oss/upstream-fixes.md
shumeiko.it 1e826aa220 fix(desktop): stop the sidecar with SIGTERM first, and free a stale WhatsApp profile (UF-046)
On macOS and Linux the shell stopped the sidecar with Child::kill(), which is
SIGKILL, so the sidecar never ran its shutdown: the processes it started (MCP
servers, a managed Ollama, the WhatsApp browser) kept running. After an app
restart the orphaned WhatsApp browser still held the session profile, and the
channel failed with "The browser is already running".

The shell now sends SIGTERM, waits up to 5 seconds for the sidecar to exit,
then sends SIGKILL; Windows keeps its kill-on-close job. For crashes, where no
shutdown runs, the WhatsApp channel stops a browser left over from an earlier
run before starting its own. It only stops a process whose command line uses
this profile and whose parent is gone; a browser another running Waggle server
owns is left alone, and the channel says so instead of retrying.
2026-10-26 12:03:31 +01:00

52 KiB

Upstream fixes for the main Waggle project

This log lists defects in the main Waggle project that we found while preparing the open-source demo edition. Each entry describes what we observed, why it happens, what we did, and what we recommend changing in the main project.

Changes made only to meet the OSS scope (removing billing, tiers, KVARK, team mode and similar) are not defects and are not listed here.

Baseline: commit b20b138 of git.os-dev.tech/oleg.m/waggle-os.

Status values

  • Fixed: fixed in the OSS edition. The same fix is recommended for the main project.
  • Workaround: worked around without changing the code. The main project needs a proper fix.
  • Open: found, not fixed yet.

Entry template

### UF-NNN. Short title
- Status: Fixed | Workaround | Open
- Found: YYYY-MM-DD
- Where: path/to/file.ts:line
- Symptom: what the user or developer sees
- Cause: why it happens
- What we did: change in the OSS edition, or workaround
- Recommendation: what to change in the main project

Summary

ID Title Status Area
UF-001 Vendored PptxGenJS runtime is missing from the repository Fixed Build, git
UF-002 Setup instructions skip the package build step Open Docs, build
UF-003 npm run dev:web cannot authenticate against the sidecar Workaround Web UI, sidecar
UF-004 Packaged macOS and Linux apps cannot find the sidecar Fixed on macOS, open on Linux Desktop shell
UF-005 Live Stripe webhook secret in git history Open Security
UF-006 Conflicting license files in weaver and optimizer Open Licensing
UF-007 Onboarding Industry dropdown opens invisibly under the scrim Fixed Web UI
UF-008 Local model downloads are cancelled after five minutes Fixed Sidecar, local models
UF-009 Model recommendations assume a smaller quantization than the one downloaded Open Local models
UF-010 Brand fonts are declared but never loaded Open Web UI
UF-011 Token-based shadows compile to nothing Open Web UI
UF-012 Border-radius scale is inverted Open Web UI
UF-013 Notification fetch failure shows "All caught up" Open Web UI
UF-014 Timeline "Edit" button has no action Open Web UI
UF-015 Ctrl+Shift+digit app shortcuts never fire Open Web UI
UF-016 Tailwind dark: variants never apply Open Web UI
UF-017 Startup text is invisible in the light theme Fixed Web UI
UF-018 Default model shown in the top bar disagrees with Model Pilot Open Web UI, models
UF-019 Memory count pluralization: "1 Memories" Open Web UI
UF-020 WhatsApp connector bundles GPL-3.0 code into the sidecar Fixed Licensing
UF-021 Settings visual-regression baseline is ignored by git Fixed Tests, git
UF-022 Unit tests write into the developer's real ~/.waggle Fixed Tests
UF-023 Four unit-test files fail on macOS Open Tests, macOS
UF-024 Workspace packages import packages they do not declare Fixed Build, packaging
UF-025 Workspace chat writes entities and error-path messages into personal memory Open Memory, privacy
UF-026 Three memory tools of the agent are never wired Open Agent, memory
UF-027 Self-learning loops run on data nothing collects Open Agent, self-learning
UF-028 The Settings prompt-shape choice has no effect on chat Open Web UI, chat
UF-029 The Settings "Import Data" button does nothing Open Web UI
UF-030 WhatsApp setup hint suggests pairing from the "Message yourself" chat, which is never read Fixed Web UI, channels
UF-031 Team sync shares the save_memory confirmation instead of the memory Open Team mode, memory
UF-032 Team capability policies are always read from the team "default" Open Team mode, governance
UF-033 Command palette memory search uses whichever workspace the Memory app opened last Fixed Sidecar, command palette
UF-034 Artifact Center "related sessions" never finds a chat session Fixed Sidecar, Artifact Center
UF-035 Timeline search and the chat recall toast send the workspace id as the search scope Fixed Web UI, memory, privacy
UF-036 The web client reads "important" and "temporary" memories as normal importance Fixed Web UI, memory
UF-037 Restore overwrites database files the running sidecar holds open Open Sidecar, backup
UF-038 Marketplace installs never notify the running sidecar Open Marketplace, sidecar
UF-039 A memory saved to an unknown workspace lands in personal memory Fixed Sidecar, memory, privacy
UF-040 The support bundle never includes recent installs Fixed Sidecar, support
UF-041 The CLI server does not exit on Ctrl+C or SIGTERM Fixed Sidecar, CLI
UF-042 Sidebar hints repeat the visible labels and are partly hidden behind the main area Fixed Web UI
UF-043 The chat composer draws two focus outlines Fixed Web UI
UF-044 A working API key shows as not responding, and other internal model calls fail Fixed Sidecar, models, auth
UF-045 Generating a persona from a description calls a route that does not exist Fixed Sidecar, agents
UF-046 On macOS and Linux the desktop app kills the sidecar without its shutdown Fixed Desktop shell, sidecar

Entries

UF-001. Vendored PptxGenJS runtime is missing from the repository

  • Status: Fixed
  • Found: 2026-09-27
  • Where: .gitignore:14 (rule dist); vendor/pptxgenjs/package.json (main points to dist/pptxgen.cjs.js)
  • Symptom: on a fresh clone, npm run dev:server crashes with ERR_MODULE_NOT_FOUND for packages/agent/node_modules/pptxgenjs/dist/pptxgen.cjs.js.
  • Cause: the generic dist ignore rule keeps the vendored runtime file out of git. The file exists only on machines where it was created by hand.
  • What we did: restored the file locally from npm pptxgenjs@4.0.1. The vendored types/index.d.ts and LICENSE are byte-identical to that release, so the runtime is assumed to be unmodified. SHA-256 of the restored file: 873d182a8e2e1c0b5e522ef146117936b96b9b2024667bd4c1de59e2b031d27a. On 2026-10-15 we added the exception !vendor/*/dist/ to .gitignore, marked vendor/** as -text in .gitattributes so git keeps the upstream bytes, and committed the file.
  • Recommendation: add !vendor/**/dist/ to .gitignore (or anchor the rule to /packages/*/dist) and commit the original vendored file. Confirm whether it differs from upstream 4.0.1.

UF-002. Setup instructions skip the package build step

  • Status: Open
  • Found: 2026-09-27
  • Where: README.md, section "Run from source (development)"
  • Symptom: following the README (npm install, then npm run dev:server) fails with ERR_MODULE_NOT_FOUND for node_modules/@waggle/core/dist/index.js.
  • Cause: workspace packages export their compiled dist/ output, which only npm run build:packages produces. The README does not mention this step.
  • What we did: ran npm run build:packages before starting the sidecar.
  • Recommendation: add npm run build:packages to the README, or run it automatically (for example a predev:server script), or export package sources for development.

UF-003. npm run dev:web cannot authenticate against the sidecar

  • Status: Workaround
  • Found: 2026-10-13
  • Where: apps/web/vite.config.ts (proxy option changeOrigin: true); packages/server/src/local/index.ts:459 (browserBootstrapAuthorityAllowed)
  • Symptom: the UI on http://localhost:8080 loads, but every API call fails. The console shows "Session bootstrap origin does not match the local service authority" (HTTP 403, SESSION_BOOTSTRAP_ORIGIN_MISMATCH), followed by HTTP 401 on the other endpoints.
  • Cause: the session-token bootstrap only accepts requests whose page origin equals the request Host. The Vite proxy runs with changeOrigin: true and rewrites Host to 127.0.0.1:3333, so the check always fails in the documented dev setup.
  • What we did: built the UI with npm run build and opened it through the sidecar on http://localhost:3333, the same way the Playwright suite and install.sh run it. No code changes.
  • Recommendation: set changeOrigin: false for the /api and /health proxies so Host stays localhost:8080, or allow the dev-server origin explicitly. Confirm the security intent of the bootstrap check first. Update the README accordingly.

UF-004. Packaged macOS and Linux apps cannot find the sidecar

  • Status: Fixed on macOS (confirmed on 2026-10-24 by launching a packaged build: the window stays on "Starting Waggle…"); open on Linux
  • Found: 2026-09-27
  • Where: app/src-tauri/src/service.rs, functions resolve_node_path and resolve_service_script
  • Symptom: expected error on start of a packaged macOS or Linux build: the shell cannot locate node or resources/service.js.
  • Cause: both functions look for resources next to the executable (std::env::current_exe()). That matches the Windows NSIS layout only. On macOS resources live in Contents/Resources, on Linux under /usr/lib/<app>. The code never uses Tauri's resource_dir(). CI builds the macOS app but never launches it.
  • What we did: bundled_resources_dir() in service.rs looks next to the executable first (Windows) and then in the app bundle's Contents/Resources/resources (macOS), for node, service.js, native/ and node_modules/. The Linux layout follows with the Linux packages in P1.3d.
  • Recommendation: resolve resources through app.path().resource_dir() and add a CI smoke test that launches packaged macOS and Linux builds.

UF-005. Live Stripe webhook secret in git history

  • Status: Open (action for Egzakta)
  • Found: 2026-09-27
  • Where: first commit 0c3e2ea, file docs/launch/drafts/2026-05-12-apps-www-deployment-readiness.md. The value is redacted in b20b138, but it remains in history.
  • Symptom: anyone with access to the repository history can read the live webhook signing secret. The same document says the live Stripe secret key was left in a chat log and not rotated.
  • Cause: a secret was committed in a planning document.
  • What we did: the OSS edition will be published from a new repository with a single clean commit; the current history will not be pushed publicly.
  • Recommendation: rotate the live webhook signing secret and the live secret key, and treat the existing history as compromised.

UF-006. Conflicting license files in weaver and optimizer

  • Status: Open
  • Found: 2026-09-27
  • Where: packages/weaver/LICENSE, packages/optimizer/LICENSE; packages/weaver/package.json, packages/optimizer/package.json
  • Symptom: the LICENSE files say "All rights reserved. This software is proprietary and confidential", while both package.json files declare "license": "MIT".
  • Cause: inconsistent licensing metadata.
  • What we did: nothing yet. The OSS edition needs a written decision (question Q6 for Egzakta).
  • Recommendation: decide the license for both packages and make the LICENSE files and package.json agree.

UF-007. Onboarding Industry dropdown opens invisibly under the scrim

  • Status: Fixed
  • Found: 2026-10-13
  • Where: apps/web/src/components/os/overlays/onboarding/WhoAreYouStep.tsx (Industry SelectContent); apps/web/src/components/os/overlays/OnboardingWizard.tsx:298 (.onboarding-scrim, z-[9999]); apps/web/src/components/ui/select.tsx (SelectContent, z-50)
  • Symptom: on the "Tell us who you are" step, clicking the Industry select shows no options. Clicking just below the select still picks a value.
  • Cause: SelectContent is portaled to <body> with the primitive's z-50. The onboarding scrim is a fixed full-screen layer with z-[9999], a 92% opaque background and pointer-events: none. The options render under the scrim, so they are invisible, while clicks pass through the scrim to them.
  • What we did: added className="z-[10000]" to the Industry SelectContent and a regression test in WhoAreYouStep.test.tsx that opens the dropdown and checks that it stacks above the scrim. Checked in the browser: the options are visible and selectable. No other portaled popup renders inside a layer above z-50 (the PersonaSwitcher tooltip is not portaled).
  • Recommendation: apply the same fix. Longer term, define a shared z-index scale for overlays and popovers instead of independent values, and add a browser-level (Playwright) check for the onboarding flow: the jsdom test can only check classes, not real stacking.

UF-008. Local model downloads are cancelled after five minutes

  • Status: Fixed. Verified three ways: (1) unit tests; (2) a real download of llama3.1:8b (4.7 GB) through the fixed route returned HTTP 200 with a verified digest and a passing generation probe; (3) a scaled-down experiment with Node's real fetch and a 2-second header timeout: stream: false failed with UND_ERR_HEADERS_TIMEOUT, stream: true completed a 5-second pull.
  • Found: 2026-10-13
  • Where: packages/server/src/local/routes/local-inference.ts, route POST /api/local-inference/pull
  • Symptom: downloading a large local model (observed with qwen3:30b-a3b, about 18.6 GB) never finishes. The request in the browser stays without a response, and the download stops after exactly five minutes. Ollama's partial files showed 12 of 20 parts last written at 18:05:47, five minutes after the start at 18:00:47; a retry stopped again about five minutes later. The download reached 68% over two attempts.
  • Cause: the route called Ollama's /api/pull with stream: false. In that mode Ollama sends no response headers until the whole model is downloaded. Node's built-in fetch (undici) aborts a request after 300 seconds without headers, and Ollama cancels the pull when its client disconnects. Every pull longer than five minutes was therefore cancelled. Ollama keeps the downloaded parts, so each retry resumed, but again for five minutes only.
  • What we did: the route now requests stream: true and reads Ollama's NDJSON progress to the end. Response headers arrive at once, and progress lines keep the connection alive. An {"error": ...} line or a stream that ends without success returns HTTP 502 with Ollama's message. Added two regression tests in packages/server/tests/local/local-inference-route.test.ts; all 23 route tests pass and tsc for packages/server is clean.
  • Recommendation: apply the same fix. Note that the route still has a 45-minute overall timeout and the web UI waits 50 minutes (apps/web/src/lib/adapter.ts, pullLocalModel). Very large models on slow connections can still hit those limits; a retry resumes, but relaying the streamed progress to the UI would remove the problem and show the user a progress bar.
  • Follow-up observed on 2026-10-13: after the repeated five-minute interruptions, resuming the same download with the fixed route failed inside Ollama with remove …/sha256-58574f2e…-partial-0: no such file or directory. Ollama's partial download state had become inconsistent and the previously downloaded data was lost. The fixed route reported this error to the client within two minutes instead of hanging. Consider detecting this error and offering to reset the partial download.

UF-009. Model recommendations assume a smaller quantization than the one downloaded

  • Status: Open
  • Found: 2026-10-13
  • Where: GET /api/local-inference/models (ranking from packages/agent/src/cookbook/*); apps/web/src/components/os/model-gate/ModelGate.tsx:393 (adapter.pullLocalModel(name))
  • Symptom: on a Mac with 16 GB of memory, the model picker ranks qwen3:30b-a3b first with fit "good", estimating 9.2 GB of memory for quantization Q2_K. Downloading it fetches the default tag, which is about 18.6 GB (Q4_K_M), so the recommended model does not fit the machine once downloaded.
  • Cause: the fit engine scores each model at its best-fitting quantization (bestQuant), but the UI pulls the bare catalog name, which resolves to Ollama's default quantization.
  • What we did: nothing yet.
  • Recommendation: either pull the tag that matches the scored quantization (where Ollama publishes one), or score models at the quantization of the default tag. Show the download size and memory requirement for exactly what will be downloaded.

Entries UF-010 to UF-019 come from the UI audit anti-slop/audit-001-2026-10-14.md (found 2026-10-14). "Live" means confirmed in the running app; "code" means found by reading the code and the built CSS.

UF-010. Brand fonts are declared but never loaded

  • Status: Open (live)
  • Where: apps/web/tailwind.config.ts, apps/web/src/index.css:213-215
  • Symptom: the UI asks for Hanken Grotesk and JetBrains Mono, but the page registers 0 @font-face rules, so most users see the system font.
  • Recommendation: bundle the fonts (self-hosted @font-face) or import them, as the design handoff does (docs/design_handoff_waggle_app/design-files/styles/waggle.css:7).

UF-011. Token-based shadows compile to nothing

  • Status: Open (code)
  • Where: 31 uses of shadow-[var(--shadow-*)], e.g. OvernightHero.tsx:32, HomeCockpit.tsx:271, 374
  • Symptom: Tailwind treats the arbitrary value as a shadow colour, so no box-shadow is drawn. Intended card elevation and honey focus glows do not render.
  • Recommendation: define the shadows in tailwind.config.ts (boxShadow) or use shadow-[...] with a full shadow value.

UF-012. Border-radius scale is inverted

  • Status: Open (code)
  • Where: apps/web/tailwind.config.ts (xl mapped to --r-xl = 26px)
  • Symptom: rounded-xl (26px) is larger than rounded-2xl (16px) and rounded-3xl (24px); small dropdowns get 26px corners.
  • Recommendation: restore a monotonic scale and replace the 154 arbitrary radius values with scale steps.

UF-013. Notification fetch failure shows "All caught up"

  • Status: Open (code)
  • Where: apps/web/src/hooks/useNotifications.ts:14, NotificationInbox.tsx
  • Symptom: on a failed request the list is set to empty, and the inbox tells the user everything is read. There is no loading or error state.
  • Recommendation: keep loading and error states separate from the empty state, with a retry.

UF-014. Timeline "Edit" button has no action

  • Status: Open (code)
  • Where: apps/web/src/components/os/apps/memory/TimelineTab.tsx:237
  • Symptom: an icon button labelled Edit has no onClick and no accessible name; it does nothing.
  • Recommendation: wire it to editing or remove it.

UF-015. Ctrl+Shift+digit app shortcuts never fire

  • Status: Open (code)
  • Where: apps/web/src/hooks/useKeyboardShortcuts.ts:92, help text in KeyboardShortcutsHelp.tsx:12-21
  • Symptom: the handler compares e.key with digits, but with Shift held e.key is "!", "@" and so on (US layout), so the documented shortcuts do not work.
  • Recommendation: match e.code (Digit0…Digit9).

UF-016. Tailwind dark: variants never apply

  • Status: Open (code)
  • Where: apps/web/tailwind.config.ts:4 (darkMode: ["class"]), theme switching via data-theme
  • Symptom: the .dark class is never set, so dark: variants (e.g. ui/alert.tsx:12, chart theme in ui/chart.tsx:7) never take effect.
  • Recommendation: use darkMode: ["selector", '[data-theme="dark"]'] or set .dark together with data-theme.

UF-017. Startup text is invisible in the light theme

  • Status: Fixed
  • Where: apps/web/src/main.tsx:12
  • Symptom: the "Starting Waggle…" text has a fixed colour #f6f1e4, while the light theme background is #f7f1e4.
  • What we did: the startup screen takes its styles from index.html, with a dark text colour for the light theme. It is also fixed to the window, so the unstyled body no longer shows scroll bars around it.
  • Recommendation: use a theme token for the startup text.

UF-018. Default model shown in the top bar disagrees with Model Pilot

  • Status: Open (live)
  • Where: top bar model pill; Settings → Models → Model Pilot
  • Symptom: the top bar shows "Default model: Llama3.1 (8b)" while Model Pilot shows Primary "Not set" next to "You have a working model".
  • Recommendation: show the same source of truth in both places, or explain that Model Pilot is unset and the default comes from elsewhere.

UF-019. Memory count pluralization: "1 Memories"

  • Status: Open (live)
  • Where: Memory → Trust tab header
  • Symptom: with one memory the header reads "1 Memories in this hive".
  • Recommendation: pluralize by count.

UF-020. WhatsApp connector bundles GPL-3.0 code into the sidecar

  • Status: Fixed
  • Found: 2026-10-15
  • Where: packages/server/package.json:30 and packages/launcher/package.json:42 (@whiskeysockets/baileys ^7.0.0-rc13); packages/server/src/local/channels/whatsapp-adapter.ts:118,260; scripts/build-sidecar.mjs (EXTERNAL list)
  • Symptom: Baileys itself is MIT, but it depends on libsignal 6.0.0 (WhiskeySockets/libsignal-node), which is GPL-3.0. Baileys is not in the esbuild EXTERNAL list, so the sidecar build bundles Baileys together with libsignal into service.js, which ships inside the desktop installer. This follows from the build configuration; we have not inspected a built installer.
  • Cause: the licenses of transitive dependencies are not checked.
  • What we did: the OSS edition replaces Baileys with whatsapp-web.js (Apache-2.0), per Egzakta decision Q5 (P1.7, ADR-0005). libsignal and Baileys are gone from the lockfile, and whatsapp-web.js stays outside the esbuild bundle. Its own dependency node-webpmux is LGPL-3.0-or-later and ships as a separate package; that is for the license review.
  • Recommendation: include this in the license review (Q6). Shipping GPL-3.0 code inside a binary distributed under other terms is a compliance risk that needs a legal opinion. Options: replace Baileys in the main product as well, or move the WhatsApp connector into a separately installed component. Add a license check for transitive dependencies to CI.

UF-021. Settings visual-regression baseline is ignored by git

  • Status: Fixed
  • Found: 2026-10-15
  • Where: .gitignore:230 (rule s*.png); tests/visual/views.spec.ts:19 (view settings)
  • Symptom: the visual suite screenshots seven views, but the repository holds baselines for six. The settings-dark.png and settings-light.png baselines match the s*.png ignore rule, so they never reach git and the Settings view has no committed baseline.
  • Cause: a broad ignore rule meant for stray screenshots also matches baseline file names.
  • What we did: added the exception !tests/visual/baselines/**/*.png to .gitignore. The Settings baselines still need to be generated with npm run test:visual -- --update-snapshots and committed.
  • Recommendation: apply the same exception in the main project and commit the Settings baselines.

UF-022. Unit tests write into the developer's real ~/.waggle

  • Status: Fixed
  • Found: 2026-10-15
  • Where: packages/marketplace/src/installer.ts:47-50 (~/.waggle/plugins/registry.json); packages/marketplace/src/security.ts:183 (~/.waggle/security-cache); vitest.setup.ts
  • Symptom: npm run test creates fixture files such as evil-skill-*.json and safe-skill-*.json in ~/.waggle/security-cache and creates or overwrites ~/.waggle/plugins/registry.json with an empty registry. The developer's own Waggle data directory is modified by the test suite.
  • Cause: these modules compute their paths from os.homedir() when they are imported, the tests do not override them, and the Vitest setup does not isolate the home directory.
  • What we did: vitest.setup.ts now points HOME and USERPROFILE at one temporary directory per test worker before any test module loads, and turns off npm's update notice, which a fresh home would otherwise print into CLI test output. Verified: after the full root suite nothing new appears under the real ~/.waggle or ~/.claude.
  • Recommendation: apply the same setup change in the main project, and resolve these paths from the configured data directory at call time instead of at import time.

UF-023. Four unit-test files fail on macOS

  • Status: Open
  • Found: 2026-10-15
  • Where: packages/server/tests/local/managed-ollama-runtime.test.ts; packages/hive-mind-hooks-openclaw/tests/handler.test.ts:330; packages/server/tests/local/files.test.ts:721; packages/server/tests/local/external-tool-runs.test.ts:1270
  • Symptom: on macOS arm64 the root suite has 41 failing tests in these four files. 38 managed Ollama runtime tests fail with "Official runtime archive did not contain ollama.exe". The OpenClaw handler test expects /var/folders/... and receives /private/var/folders/.... A storage move test is rejected with "escapes workspace root through symlink". An OpenClaw run test times out with "OpenClaw run did not settle". All four files fail the same way when run on their own.
  • Cause: the runtime fixtures appear to assume the Windows archive layout, and the path checks do not account for /var being a symlink to /private/var on macOS. The OpenClaw timeout is not investigated yet.
  • What we did: nothing yet. The failures are recorded as known in docs/oss/baseline.md.
  • Recommendation: make the runtime fixtures platform-aware, compare resolved paths (fs.realpathSync) in path assertions, and run the unit suite on macOS and Linux in CI.

UF-024. Workspace packages import packages they do not declare

  • Status: Fixed
  • Found: 2026-10-15
  • Where: package.json of packages/server, packages/agent, packages/launcher, packages/worker, packages/cli and apps/web
  • Symptom: inside the monorepo everything works because npm hoists all dependencies to the root node_modules, but a package installed on its own fails to resolve imports. server imports @waggle/marketplace, @waggle/sdk, @waggle/weaver, zod, @anthropic-ai/sdk and @ax-llm/ax without declaring them; agent imports pdf-parse and playwright-core; worker imports @waggle/waggle-dance; launcher imports @waggle/server; apps/web imports @waggle/shared. server declares @waggle/hive-mind-shim-core and cli declares @waggle/weaver without using them.
  • Cause: no check compares imports with declared dependencies.
  • What we did: declared the missing dependencies (playwright-core as an optional peer, because agent loads it only when available; @waggle/server as a dev dependency of launcher, because its build bundles @waggle/*), removed the unused ones, and added scripts/check-workspace-deps.mjs (npm run check:deps) to CI.
  • Recommendation: apply the same declarations in the main project and run the check in CI.

UF-025. Workspace chat writes entities and error-path messages into personal memory

  • Status: Open
  • Found: 2026-10-16
  • Where: packages/server/src/local/index.ts:1298 (createSessionOrchestrator builds every session orchestrator on multiMind.personal); packages/server/src/local/routes/chat.ts:3435 and :3786
  • Symptom: in a workspace chat, entities extracted from the agent's answer and, on the error path, the user's message are written to the personal mind, not to the workspace mind. The comment at chat.ts:3428 says they go to "the knowledge graph of the active workspace mind". Workspace content leaks into personal memory, which follows the user into every other workspace.
  • Cause: getKnowledge(), getFrames() and getSessions() on the orchestrator return the layers of the database it was built on, which is always the personal one; the workspace mind is only attached with setWorkspaceMind.
  • What we did: nothing in the old code. Found by reading the code, not reproduced at runtime. In the community edition these writes move to the memory contract in P1.5, where the scope is explicit on every call.
  • Recommendation: write through the workspace layers when a workspace mind is attached, and add a test that a workspace chat leaves the personal mind unchanged.

UF-026. Three memory tools of the agent are never wired

  • Status: Open
  • Found: 2026-10-16
  • Where: packages/agent/src/orchestrator.ts:187 (the only createMindTools call); packages/agent/src/tools.ts:96-109, :314, :474, :657
  • Symptom: search_all_workspaces searches only personal memory although its description promises all workspaces; correct_knowledge always answers "Feedback handler not available. Cannot correct knowledge."; the embedding-based duplicate check in save_memory never runs.
  • Cause: createMindTools is called without getAllWorkspaceSearches, feedback and embedder, and each tool silently falls back when its dependency is missing.
  • What we did: nothing in the old code. In the community edition these tools are rebuilt on the memory contract in P1.5.
  • Recommendation: pass the three dependencies, or remove the tools whose dependency cannot be provided so the agent is not offered a tool that cannot work.

UF-027. Self-learning loops run on data nothing collects

  • Status: Open
  • Found: 2026-10-20
  • Where (commit b734c1d; these files are removed from the community edition): packages/agent/src/optimization-capture.ts:33 (captureInteraction); packages/server/src/local/index.ts:753 (createSkillTools), :2044 (prompt_optimization cron), :2138, :2366 (monthly_assessment cron); packages/server/src/local/monthly-assessment.ts:60
  • Symptom: the nightly prompt-optimization job, the monthly agent assessment and the agent_insights tool report empty or zero activity however much the agent is used. Skill promotion, auto-extraction and retirement never leave the improvement signals their tool descriptions promise.
  • Cause: captureInteraction, the only function that writes interactions into optimization_log, has no caller; the only write to that table is the GEPA job storing its own prompt variants. The server builds the skill tools without improvementSignals, so their signal code never runs.
  • What we did: nothing in the old code. Self-learning is excluded from the community edition (ADR-0004), so these modules were removed.
  • Recommendation: when reworking self-learning, call captureInteraction after each chat turn and pass the signal store to createSkillTools, with a test that a chat turn produces an optimization_log row.

UF-028. The Settings prompt-shape choice has no effect on chat

  • Status: Open
  • Found: 2026-10-20
  • Where: apps/web/src/lib/adapter.ts:1060 (sendMessage sends shape); packages/server/src/local/routes/chat.ts (never reads it); packages/server/src/local/routes/agent-run.ts:253 (the only route that honours a shape)
  • Symptom: the Prompt Shape dropdown in Settings saves the choice and the client sends it with every chat message, but chat answers do not change. The client comment says the sidecar "ignores shape until A3.1 wires it".
  • Cause: /api/chat does not read the shape field; only /api/agent/run applies a prompt-shape override.
  • What we did: nothing yet. The setting stays in the community edition until its fate is decided.
  • Recommendation: pass shape into the chat agent loop as agent-run.ts does, or hide the dropdown until it works.

UF-029. The Settings "Import Data" button does nothing

  • Status: Open
  • Found: 2026-10-23
  • Where: apps/web/src/components/os/apps/SettingsApp.tsx:772 (commit ad2ca87), the Backup section
  • Symptom: Settings → Backup shows an "Import Data — Import from ChatGPT or Claude export" button; clicking it has no effect.
  • Cause: the <button> has no onClick handler and is not wired to the import flow.
  • What we did: import history is hidden in the community edition v1 (decision Q2), so the button is hidden behind the memory feature map rather than wired.
  • Recommendation: connect it to the existing import flow (the Memory → Imports screen) or remove it.

UF-030. WhatsApp setup hint suggests pairing from the "Message yourself" chat, which is never read

  • Status: Fixed
  • Found: 2026-10-24
  • Where: at commit b20b138, apps/web/src/components/os/settings/ChannelsSettings.tsx:48 (setupHint) and packages/server/src/local/channels/whatsapp-adapter.ts:417 (if (!jid || raw.key.fromMe) return null)
  • Symptom: Settings → Channels tells the user to "Message yourself or the linked number /pair ". A /pair sent in the "Message yourself" chat of the linked account gets no reply and pairs nothing.
  • Cause: messages the linked account sends, including those to itself, arrive with fromMe set, and the adapter drops every fromMe message (correctly, so the bot never reads its own replies).
  • What we did: the hint now says to send /pair <code> to the linked number from another WhatsApp account and that messages to yourself are not read. The whatsapp-web.js adapter drops own messages the same way.
  • Recommendation: change the hint in the main project the same way, or support the self-chat explicitly (accept fromMe messages in the account's own chat and skip the bot's replies there).

UF-031. Team sync shares the save_memory confirmation instead of the memory

  • Status: Open
  • Found: 2026-10-24
  • Where: at commit f3eb6b6, packages/server/src/local/routes/chat.ts:2997-3027 (the save_memory TeamSync push in onToolResult)
  • Symptom: in a team workspace, every save_memory call sends the team server a frame whose content is the tool's reply, for example "Memory saved to workspace mind (importance: normal, ...)", instead of the fact the agent saved. Replies to blocked, rate-limited and duplicate saves ("Memory save blocked: ...", "Memory already exists ... skipped duplicate save") are shared too, because only replies starting with "Error" are skipped.
  • Cause: the push builds the frame from result (the tool output string) with a synthetic id, not from the frame that save_memory created.
  • What we did: nothing in the old code. Team mode is postponed in the community edition (ADR-0003) and this push was removed with it in P1.1.
  • Recommendation: push the frame that save_memory created (as the orchestrator's pattern write-back does), and only when the save succeeded.

UF-032. Team capability policies are always read from the team "default"

  • Status: Open
  • Found: 2026-10-24
  • Where: at commit f3eb6b6, packages/server/src/local/routes/chat-governance.ts:46, called from packages/server/src/local/routes/chat.ts:2591; the same fallback in packages/server/src/local/routes/team.ts:416 (GET /api/team/governance/permissions)
  • Symptom: tools blocked for a role in a team's capability policies stay available in that team's workspaces; the chat route applies the policies of a team with the slug "default", if one exists.
  • Cause: the slug is read from teamServer.teamSlug in config.json, a field TeamServerConfig does not have and POST /api/team/connect never writes, so it always falls back to 'default'. The workspace's teamId is not used.
  • What we did: nothing in the old code. Team mode is postponed in the community edition (ADR-0003) and the lookup was removed with it in P1.1.
  • Recommendation: fetch /api/teams/<workspace teamId>/capability-policies for the workspace the turn runs in, and cache per team rather than per workspace.

UF-033. Command palette memory search uses whichever workspace the Memory app opened last

  • Status: Fixed
  • Found: 2026-10-24
  • Where: at commit b20b138, packages/server/src/local/routes/command.ts:111 (server.multiMind.search(q, searchScope, perFacet)); the workspace is set only by ensureWorkspaceMind() in packages/server/src/local/routes/memory.ts:184-192
  • Symptom: memory results in the Ctrl+K palette (scope 'all' or 'workspace') come from the workspace last viewed in the Memory app, not from the workspace the user is in; before the Memory app is opened, workspace memory is not searched at all. If the mind cache has evicted that workspace, the memory group is silently empty.
  • Cause: MultiMind holds one borrowed workspace handle that only the memory routes set; the palette reuses it without choosing a workspace.
  • What we did: the palette searches personal memory and the memory of the active workspace (agentState.activeWorkspaceId) through the memory contract.
  • Recommendation: pass the workspace the palette was opened in (or the active workspace) to the memory search instead of relying on shared MultiMind state.
  • Status: Fixed
  • Found: 2026-10-24
  • Where: at commit b20b138, packages/server/src/local/routes/artifacts.ts:204-222 (GET /api/artifacts/search-related, the sessions group)
  • Symptom: the "Related" section of an artifact lists no chat sessions, even when a chat of the workspace discusses the artifact's topic.
  • Cause: the route matches the query against the memory engine's session rows (SessionStore.getActive(), their summary and gop_id). Those rows are named write buckets such as 'harvest', 'agent-runs' or 'loop', not chat sessions; chat transcripts are JSONL files under <dataDir>/workspaces/<id>/sessions/.
  • What we did: the route runs searchSessions() over each workspace's chat session files, the same search as GET /api/workspaces/:id/sessions/search, and returns the session id, title, workspace and the first matching snippet.
  • Recommendation: do the same in the main project.

UF-035. Timeline search and the chat recall toast send the workspace id as the search scope

  • Status: Fixed
  • Found: 2026-10-24
  • Where: at commit b20b138, apps/web/src/hooks/useMemory.ts:25 (searchMemory(filters.searchQuery, workspaceId)), apps/web/src/components/os/apps/ChatApp.tsx:670 (scope = workspaceId ?? 'global'), packages/server/src/local/routes/memory.ts:259-278 (/api/memory/search)
  • Symptom: a search in Memory → Timeline, and the "I just remembered something relevant" toast in a workspace chat, do not search the open workspace. They return personal memories plus those of whichever workspace mind the sidecar attached last, which can belong to another workspace.
  • Cause: both callers pass the workspace id as scope and send no workspace parameter. The route treats an unknown scope as 'all' and, without a workspace parameter, searches multiMind with the workspace mind it still holds from an earlier request.
  • What we did: the client sends scope=all&workspace=<id>; the memory contract version of the route searches exactly the scopes it is asked for (personal memory plus the named workspace, personal only when none is named).
  • Recommendation: pass the workspace as its own parameter in both callers, and make the route resolve an unknown scope to personal memory only instead of reusing a previously attached workspace mind.

UF-036. The web client reads "important" and "temporary" memories as normal importance

  • Status: Fixed
  • Found: 2026-10-24
  • Where: at commit b20b138, apps/web/src/lib/adapter.ts:223 (IMPORTANCE_MAP) and :252 (normalizeFrame)
  • Symptom: the Timeline shows important memories with the same importance dots as normal ones, its minimum-importance filter cannot separate them, and the login briefing, which ranks highlights by importance, ranks important memories like normal ones.
  • Cause: the sidecar returns the frame-store importance values (critical, important, normal, temporary), but the client maps only low, normal, high and critical, so 'important' and 'temporary' fall back to 2 (normal).
  • What we did: the map now covers temporary (1) and important (3), and keeps low and high for the Tauri command.
  • Recommendation: map the importance values the sidecar actually returns.

UF-037. Restore overwrites database files the running sidecar holds open

  • Status: Open
  • Found: 2026-10-24
  • Where: at commit b20b138, packages/server/src/local/routes/backup.ts:380-542 (POST /api/restore)
  • Symptom: restoring a backup while Waggle runs writes personal.mind, the workspace minds and the other SQLite files over the files the sidecar has open. The running process keeps its old view of the data and can write over the restored files, or leave a WAL file that no longer matches them.
  • Cause: the restore writes every file of the archive in place (only marketplace.db is skipped) without closing the open databases first.
  • What we did: nothing yet. In the community edition the same applies to waggle.db and the mem0 files in <dataDir>/memory/.
  • Recommendation: stage the restored files and apply them at the next start, before any database opens, as the data-erase marker does; or close the databases, restore, and restart the sidecar.

UF-038. Marketplace installs never notify the running sidecar

  • Status: Open
  • Found: 2026-10-24
  • Where: at commit b20b138, packages/marketplace/src/installer.ts:94 (API_BASE = process.env.WAGGLE_API_URL || 'http://localhost:3000') and notifyServer() at :960, called after skill and plugin installs and uninstalls (:590, :667, :764, :773)
  • Symptom: after a skill or plugin is installed from the Marketplace, the hot-reload the installer asks for does not happen: the sidecar endpoints it calls (PUT /api/skills/:name reloads agentState.skills, POST /api/plugins/install registers and enables the plugin) never receive the request. Not verified end to end whether another path reloads them before a restart.
  • Cause: the installer posts to port 3000, while the sidecar listens on 3333 by default and nothing sets WAGGLE_API_URL; the request also carries no session token, which the sidecar requires. The error is swallowed (// Server not running).
  • What we did: nothing yet.
  • Recommendation: when the installer runs inside the sidecar (the Marketplace routes), call the reload logic directly instead of over HTTP; for the standalone CLI, take the port and token from the running sidecar.

UF-039. A memory saved to an unknown workspace lands in personal memory

  • Status: Fixed
  • Found: 2026-10-24
  • Where: at commit b20b138, packages/server/src/local/routes/memory.ts:417-427 (POST /api/memory/frames: getWorkspaceMindDb(workspace) returns null for an unknown id, and the frame goes to server.multiMind.personal); the same fallback in POST /api/quick-capture and POST /api/memory (Memory Center)
  • Symptom: a memory sent for a workspace that does not exist, for example one deleted in another window or a mistyped id, is saved without an error into personal memory, which every workspace reads. Searching from any other workspace then returns it. The e2e isolation tests (competitive-benchmarks B2.4, user-behavior U2.4) passed only because the legacy full-text search rejected their hyphenated queries with HTTP 400.
  • Cause: the routes treat an unknown workspace the same as no workspace.
  • What we did: the three write routes save to personal memory only when no workspace is named and answer 404 workspace_not_found for a workspace that does not exist (writeScopeForWorkspace in memory-records.ts).
  • Recommendation: refuse writes for unknown workspaces in the main project as well.

UF-040. The support bundle never includes recent installs

  • Status: Fixed
  • Found: 2026-10-25
  • Where: at commit b20b138, packages/server/src/local/index.ts:2752-2760 (GET /api/debug/logs)
  • Symptom: the support log download in Settings (waggle-logs-*.json) has no auditRecent field, although the route promises recent install events.
  • Cause: the route reads the private db field of the audit store and calls prepare on it. That field is the database wrapper (MindDB), which has no prepare method, so the call throws and the empty catch drops the error.
  • What we did: the server's InstallAuditStore got getRecentSummary(), which returns only the timestamp, action and capability type of the latest 50 events, and the route calls it. A test checks the field and its columns.
  • Recommendation: read the trail through the store's own API in the main project as well, and log errors in this route instead of dropping them.

UF-041. The CLI server does not exit on Ctrl+C or SIGTERM

  • Status: Fixed
  • Found: 2026-10-25
  • Where: at commit b20b138, packages/server/src/local/routes/oauth.ts:67-73 (a module-level setInterval), together with the signal handler in packages/server/src/local/service.ts:305-312
  • Symptom: after Ctrl+C in npx waggle or npm run dev:server the HTTP server closes, but the process keeps running until a second Ctrl+C. After SIGTERM, for example from kill or a process manager, it never exits. The launcher's own test left two waggle processes running after every test run. The desktop app is not affected: it stops the sidecar with SIGKILL.
  • Cause: the OAuth routes start a cleanup interval when the module loads and never clear or unref it. The service installs SIGTERM and SIGINT handlers, which replace Node's default exit, closes the server, and relies on the event loop going empty; the interval keeps it alive.
  • What we did: the interval is unref'd. A launcher test checks that the built CLI exits by itself after SIGTERM, and the launcher tests now stop npx and the waggle process it starts as one process group.
  • Recommendation: unref housekeeping timers or clear them when the server closes, and consider ending the signal handler with an explicit exit.

UF-042. Sidebar hints repeat the visible labels and are partly hidden behind the main area

  • Status: Fixed
  • Found: 2026-10-25
  • Where: at commit b20b138, apps/web/src/components/os/Sidebar.tsx:76-78, 129, 165, 181, 196 (a HintTooltip on every item, unportaled per apps/web/src/components/ui/tooltip.tsx), with the sidebar at Sidebar.tsx:126 and <main> at AppShell.tsx:497, both relative z-10
  • Symptom: with the full-width sidebar every item shows a hover hint that only repeats its visible label, for example "New Agent" next to the "New Agent" button. The hint is cut off where the main area has content to the right of the sidebar.
  • Cause: the sidebar shows the same hints at every width, although they are only needed on the icon rail below lg, as a code comment already noted. The hint renders inside the nav, and the nav and <main> are sibling z-10 layers, so <main>, later in the DOM, paints over the part of the hint that reaches into it.
  • What we did: sidebar hints show only on the icon rail (lg:hidden) and render in document.body through a new opt-in portal flag on HintTooltip, so <main> no longer covers them. Other hints are unchanged, because several of them sit inside overlays above z-50, where a portaled hint would end up under the overlay. A test checks both properties.
  • Recommendation: apply the same change in the main project, and review other hints that repeat a visible label.

UF-043. The chat composer draws two focus outlines

  • Status: Fixed
  • Found: 2026-10-26
  • Where: at commit b20b138, apps/web/src/components/os/apps/ChatApp.tsx:1760 (the composer textarea), inside the box at ChatApp.tsx:1742
  • Symptom: when the message field has focus, a square amber ring appears around the text area inside the rounded composer box, which turns amber at the same time.
  • Cause: the textarea has its own focus-visible:ring-2 while its box already shows focus with focus-within:border-[var(--focus-ring)] and a glow.
  • What we did: removed the ring from the textarea. The box's border stays the single focus indicator; it measures 8.14:1 against the composer in the dark theme and 4.92:1 in the light theme. The composer test now checks for one indicator.
  • Recommendation: apply the same change in the main project.

UF-044. A working API key shows as not responding, and other internal model calls fail

  • Status: Fixed
  • Found: 2026-10-26
  • Where: at commit b20b138, packages/server/src/local/routes/settings.ts:309-360 (POST /api/settings/probe-model, the request at :344 sends no session token); the same unauthenticated requests to the built-in proxy in packages/server/src/local/index.ts:1901, 2117, 2297 and 2428 (scheduled AI tasks and loops), routes/dreams.ts:31 (dream diary narrative) and routes/profile.ts:289, 346 and 423 (style analysis and related profile calls); the bearer check in security-middleware.ts, which trusts loopback callers only with WAGGLE_TRUST_LOCALHOST=1
  • Symptom: Settings → Models shows "Key found but not responding" for an Anthropic key that works, so the key looks as if it keeps dropping off. On a real machine the check answered rejected, while the same request sent with the session token got a normal model reply. Scheduled AI tasks and loops, the dream diary narrative and the profile calls fail with HTTP 401 the same way. The profile calls also built the proxy address from address().toString(), which is [object Object], and the scheduled tasks used the configured port, which in the desktop app (port fallback) can belong to another Waggle server.
  • Cause: since the server stopped trusting loopback callers (D1), the built-in proxy needs the session token like every other route, but these internal calls do not send it. The key check reads the proxy's own 401 as a rejected key. vitest.setup.ts turns localhost trust on for every test, so the suites never saw it; routes/command.ts had the same bug fixed locally.
  • What we did: packages/server/src/local/self-request.ts builds the address from the port the server is bound to and adds the session token. The key check, the scheduled tasks, the dream diary, the profile calls and command.ts use it; the token is never sent to Ollama. A route test checks the header, and an end-to-end test with localhost trust off checks that a working key verifies through the real middleware and proxy; it fails without the fix.
  • Recommendation: apply the same change in the main project, and run the auth-sensitive suites with localhost trust off.

UF-045. Generating a persona from a description calls a route that does not exist

  • Status: Fixed
  • Found: 2026-10-26
  • Where: at commit b20b138, packages/server/src/local/routes/personas.ts:160 (POST /api/personas/generate, used by the agent templates view)
  • Symptom: generating an agent persona from a text description cannot get an answer from the model.
  • Cause: the route posts to /api/anthropic/v1/messages, which no route serves. The request also has no session token, and its port comes from address().toString(), which is [object Object], so the URL does not even parse. The model was fixed to claude-sonnet-4-6, so without an Anthropic key the call could not work either.
  • What we did: the route calls the built-in proxy (/v1/chat/completions through self-request.ts, UF-044) and reads its OpenAI-format reply with parseOpenAiTextCompletion. The model starts from the user's default and falls back through resolveUsableModel, the way chat does, so a local Ollama model works too. An end-to-end test with localhost trust off checks that the persona the model wrote comes back; before the fix it got HTTP 500.
  • Recommendation: apply the same change in the main project.

UF-046. On macOS and Linux the desktop app kills the sidecar without its shutdown

  • Status: Fixed
  • Found: 2026-10-26
  • Where: at commit b20b138, app/src-tauri/src/service.rs:1569 (stop_service_sync) and :1632 (watchdog restart), both launch.child.kill()
  • Symptom: quitting or restarting the desktop app leaves the processes the sidecar started running, because their stop code runs only on a graceful shutdown: MCP servers (mcpRuntime.stopAll() in onClose, index.ts:3319) and the managed Ollama (stopTree on SIGTERM, managed-ollama-runtime.ts:138). In the community edition the WhatsApp browser is one of them: after a restart the channel fails with "The browser is already running for …/channels/whatsapp/session", because the orphaned browser still holds the session profile.
  • Cause: Child::kill() sends SIGKILL on Unix, so the sidecar gets no chance to shut down. Windows is not affected: there the sidecar runs in a job object that ends its whole process tree.
  • What we did: on Unix the shell now sends SIGTERM, waits up to 5 seconds for the sidecar to exit (it exits on SIGTERM since UF-041) and only then sends SIGKILL; Rust tests cover both paths. For the crash case, where no shutdown runs at all, the WhatsApp channel now stops a browser left over from an earlier run before it starts its own (whatsapp-profile-lock.ts): it reads the profile's SingletonLock, and stops the process only if its command line uses this profile and the process that started it is gone. A browser that another running Waggle server still owns is left alone, and the channel says so instead of retrying.
  • Recommendation: stop the sidecar with SIGTERM first in the main project as well.