The `task` spec (argv templates, access modes, output dialects) and the
capability flags headlessTask, structuredProgress, resumable and
liveWaggleDance described how a Waggle Room run drove a coding agent
without its UI. Nothing runs tools headless any more, so built-in
manifests, the third-party adapter loader and detection results no
longer carry them. An adapter file that still declares `task` is
rejected like any other unknown key.
The executor router could propose Claude Code, Codex or Hermes and
dispatch the task as a Waggle Room run through POST /api/tools/run, which
is gone with the collaboration runtime. Confirming such a proposal could
only fail.
- ExecutorRegistry offers the five v1 personas; it no longer detects
external tools.
- Route proposals dispatch persona chat turns only. The memory brief and
its egress disclosure existed only for external executors, so
executor-brief.ts, the egress fields and removeFrameIds go too.
- RouteProposalCard drops the memory disclosure, "Run without memory" and
the link to an external run in the Room.
Waggle Room is not part of the community edition (decision Q2, ADR-0002).
Its headless runs of Claude Code, Codex, Hermes and OpenClaw needed the
packaged collaboration CLI from the removed @waggle/hive-mind-cli, so
POST /api/tools/run answered 503 collaboration_runtime_missing.
- Remove POST /api/tools/run (routes/external-tool-runs.ts) with its
Room fan-out, WaggleDance publishing, OpenClaw provisioning and the
memory write-back of run results.
- Remove runExternalTool (agent/external-tool-runner.ts) and
resolveWaggleRuntime; the taskkill path helper moves into the process
tracker, its only remaining user.
- Launcher: drop the hidden "Run task" composer, the Room hand-off and
adapter.runExternalToolTask. Detect, launch, watch and stop stay.
- Drop the `room` flag from the memory feature map: it only gated the
Launcher task button.
- The viewer policy no longer reads participants[].workspaceIds, a body
shape only /api/tools/run accepted.
.env.example described Stripe, Clerk, MinIO and the team server, none of
which the community edition runs, and sent users to a Settings tab that
does not exist. The web UI no longer names a person as author, and the
installer package gets the README its files list promised.
- Key messages named a Settings > API Keys tab that does not exist; they
now point at Settings > Models (UF-061).
- The connector suggestion looked for vault.db, which never exists; it
now checks the vault for a connector entry (UF-062).
- /config and /preferences were registered with a leading slash and
never matched (UF-063).
- Pins and documents were written under ~/.waggle whatever the data
directory; they now use the server's data directory (UF-064).
- The Windows uninstaller named a Data & Privacy page that does not
exist (UF-065).
release.yml now has one build job with a matrix over the four release
targets (windows-x64, macos-arm64, macos-x64, linux-x64). Each leg stages
the sidecar, checks it, builds the web UI, packs the server archive for
the installer (ADR-0009), runs the installer smoke test, builds the
desktop installers (NSIS and MSI, app and DMG, deb and AppImage), runs
the packaged app until its sidecar answers /health, and uploads the
installers and archives. A pushed v* tag also gets a draft GitHub
release with a SHA256SUMS file, created with GITHUB_TOKEN only. Pull
requests on the desktop paths and manual runs build and test without
publishing, so tauri-build-pr.yml is removed.
Nothing is signed or notarized. The macOS app gets an ad-hoc signature
("-"), which needs no certificate; without it macOS reports a downloaded
Apple Silicon app as damaged.
Removed with the old pipeline, as nothing else used them: the Windows
signing handoff and signer (sign-windows-artifact.ps1 and its tests,
new-windows-signing-handoff.ps1), the pilot signing scripts
(sign-windows-pilot.ps1, apply-signing-config.mjs, signing-config.ts and
its test, sign-macos-adhoc.sh, tauri.build-override.conf.json and the
tauri:sign:* and tauri:build:win:pilot-signed scripts), the installer
certifier with its bootstrap-token helper, and the Windows release
publisher. check-sidecar-resources.mjs stays: tauri.conf.json runs it
before every desktop build and the workflow runs it before packing the
archive.
app/package.json gains tauri:build:linux. The tests that pinned the old
workflows, the certifier, the signer and the publisher are removed;
new tests pin the targets, the step order, the absence of signing and of
secrets other than GITHUB_TOKEN, and the draft-only, tag-only publish.
The Windows installer certifier opened a WebView2 CDP port through
WAGGLE_CERTIFIER_WEBVIEW_DEBUG_PORT. The certifier leaves with the
commercial release pipeline, so the main window is built straight from
its configuration and the shipped app never opens a debugging port.
The launcher is now a self-contained installer with install, run, update,
doctor and uninstall. Install downloads waggle-server-<version>-<target>
.tar.gz, checks its SHA-256, unpacks it into a per-user program directory
and writes a waggle-server command that runs the tool on the archive's own
Node. Nothing is built on the user's machine and the data directory stays
WAGGLE_DATA_DIR or ~/.waggle.
scripts/build-server-archive.mjs packs the staged desktop sidecar payload
with the built web UI and a manifest. The download location waits for the
repository decision; until then install takes --from or
WAGGLE_RELEASE_BASE_URL (ADR-0009).
Removes install.sh, scripts/waggle-server.sh and their CI job, which the
installer replaces, and the bin name clash with @waggle/cli.
Sample data in tests, prompts and UI placeholders uses Alex Morgan
instead of the maintainer's name; decision attributions say "owner".
Personal details in two test fixtures are replaced. Local paths
(D:/Projects/...) and paths into the maintainer's planning repository
are reworded. Clone and install URLs in CONTRIBUTING.md and the getting
started guide use placeholders until the public repository location is
decided (Q9). The Benchmarks page no longer names the maintainer's
hive-mind repository, and the upstream fixes log names the main Waggle
repository instead of an internal git host.
When no embedding provider answers, the banner told users to set
HIVE_MIND_EMBEDDING_PROVIDER, which only the hive-mind CLI read; Waggle
reads EMBEDDING_PROVIDER. A test checks the banner text.
Comments that pointed at hive-mind-core files, the hive-mind OSS
mirror and its drift triage now describe the code as it is. The
in-process reranker caches its model in ~/.waggle/models, like the
embedder, when the caller passes no directory; the sidecar always
passes one.
The bench compared mem0 with MindDB, FrameStore and HybridSearch from
the hive-mind substrate, which @waggle/core no longer exports and which
left the repository. It now runs mem0, mem0-rerank and cosine; the
recorded comparison stays in docs/oss/research/memory-search-quality.md.
The hooks for Claude Code, Codex and the other agents ran the
hive-mind-hooks-* packages, which left the repository; ADR-0002 hides
the feature in v1. Remove the hook installer surface from the tool
launcher, the POST /api/tools/hooks route, the hook status probe in
tool detection, and the hook fields of the tool manifest and detection
envelope (hookCapable, hookRoot, hookPointer, hooksInstalled,
hookPointerPath). Third-party adapter manifests no longer accept the
hook fields.
In the web UI, remove the hidden hook actions and the "How memory is
shared" explainer from the Launcher, adapter.manageHooks, the
hook-wired count in Mission Control and the agentHooks and memoryMcp
memory feature flags. Launch, watch and stop are unchanged; the
roadmap tool note no longer mentions hooks.
Add "license": "MIT" to the workspace manifests that lacked it (app,
apps/web, server, shared and the postponed team packages) and to the
Tauri crate. Remove the personal author and the maintainer repository,
homepage and bugs URLs; the public repository location is not decided
yet (Q9). package-lock.json is left as is and is regenerated later.
Delete apps/www, packages/admin-web, every packages/hive-mind-* package,
packages/memory-mcp and packages/wiki-compiler (ADR-0002), with what
existed only for them: their tests kept outside the packages, the hook
runtime build, the hive-mind OSS subtree-split and drift scripts, the
www deploy and hive-mind-cli workflows, the staged hook lifecycle step
in the Windows desktop workflows, and the root dev:www, build:www and
build:hook-runtime scripts.
Drop the matching excludes from the Vitest, ESLint and workspace
dependency checks, the next override, and the extraneous workspace
entries from package-lock.json. The smart-router qualification seed
now writes its spend to waggle.db, where the sidecar has read cost
traces since P1.5.
build-sidecar.mjs required a git checkout: it read HEAD, the tracked
file list and the marketplace database blob from git. Outside a checkout
it now records no source revision and ships the database from disk; a
checkout keeps the exact-revision checks.
Linux packages keep resources in /usr/lib/<app> (deb) or under $APPDIR
(AppImage), not next to the executable, so the shell never found
service.js there. Setup now records Tauri's resource_dir and the lookup
prefers it, keeping the executable-relative paths as a fallback.
Drop tauri-plugin-updater: it was declared but never registered, and the
updater stays disabled until release signing exists.
The first start copied Waggle's tables and the memories of the commercial
edition's .mind files into waggle.db and mem0. The product is still in
development and nobody has data to bring over, so both copies, their tests
and the legacy creation-time field are gone. ADR-0007 records the change.
Anti-slop audit-002, item 13 (R-02). Every em dash in text the user sees
(web UI labels, toasts, tooltips, empty and error states, chat steps,
API messages, notifications and channel replies) became a period, comma,
colon or parentheses. Model prompts, logs, comments, regexes and the lone
dash that marks an empty value are unchanged. Text the client matches
against server copy (the default workspace summary, the quiet-night dream
summary) accepts the old wording too, so cached and saved data still match.
Bees, sparkle glyphs, fire, check and warning marks decorated buttons,
toasts, activity steps and channel replies (anti-slop audit-002, item 18).
Marks that carry state (key set, importance, the hive glyph) stay.
Home and the sidebar read the user's name from the identity row, which the
sidecar fills with the assistant's identity named Waggle. Both now use the
profile name, with the identity name as a fallback that skips the default.
Only a successful recall reported its outcome, so an empty or failed
recall left "Recalling relevant memories..." as the last step after the
answer. Both cases now send a closing step and the turn goes on.
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.
POST /api/personas/generate posted to /api/anthropic/v1/messages, a route
that does not exist, with no session token and a port built from
address().toString(), so the URL did not even parse. It also pinned
claude-sonnet-4-6, which needs an Anthropic key.
The route now calls the built-in proxy through self-request.ts, reads the
OpenAI-format reply, and picks the model from the user's default with the
same fallback chat uses, so a local Ollama model works too. An end-to-end
test with localhost trust off checks the generated persona comes back.
Since the server stopped trusting loopback callers, the built-in model proxy
needs the session token like every other route, but the sidecar's own calls
to it sent none. The Models key check read the proxy's 401 as a rejected key,
so a working Anthropic key showed "Key found but not responding". Scheduled AI
tasks and loops, the dream diary narrative and the profile calls failed the
same way; the profile calls also built the port from address().toString(),
and the scheduled tasks used the configured port, which the desktop app's
port fallback can leave pointing at another Waggle server.
self-request.ts now builds the address from the bound port and adds the
session token; all these calls and command.ts use it, and the token is never
sent to Ollama. An end-to-end test runs the key check through the real
middleware and proxy with localhost trust off; it fails without the fix.
UF-045 records the persona generator, which posts to a route that does not
exist; it is left open.
The OAuth routes started a module-level cleanup interval that was never
cleared or unref'd. The service's SIGTERM/SIGINT handler closes the server
and relies on the event loop going empty, so the process kept running: a
first Ctrl+C in `npx waggle` or `npm run dev:server` left it alive, and
SIGTERM never ended it. The interval is now unref'd.
The launcher tests left two `waggle` processes behind on every run: they
stopped `npx` but not the waggle process it started. They now start it in
its own process group and stop the whole group, with a SIGKILL fallback,
and a new test checks that the built CLI exits by itself after SIGTERM.
Core keeps only shared infrastructure, so InstallAuditStore leaves it and
lives in packages/server/src/local/install-audit.ts. The trail itself stays:
the Audit and Activity tabs, the extend and export routes, and the threat
model entry keep working on the same waggle.db table.
- @waggle/shared gets RecordAuditInput and the narrow InstallAuditRecorder
(record only), which the agent's skill tools now depend on.
- The server creates the audit table before the one-time legacy copy, which
no longer constructs it in core.
- UF-040: GET /api/debug/logs called prepare() on the store's private db
wrapper, which has none, so the support bundle never carried recent
installs. It now reads them through getRecentSummary().
POST /api/memory/frames, /api/quick-capture and the Memory Center's
POST /api/memory saved a memory meant for an unknown workspace into
personal memory, which every workspace reads. They now answer 404
workspace_not_found and write nothing; a request without a workspace
still goes to personal memory.
- The workspace isolation scenarios store memories through the memory
store and check that one workspace's search never returns another's.
- The Ollama manifest test covers the Linux x64 and arm64 artifacts added
for P1.3b; an unsupported platform still resolves to nothing.
- The root run leaves out the tests that only exercise excluded packages
(the hook packages and the memory MCP server); they go with those
packages in P2.2.
The checkout-queue test searches memory for 'SSE resilience test frame'. The
text used to reach mem0 only through the background copy of the seeded
legacy personal.mind; write it with server.memory instead.
The live external-agent test (skipped unless WAGGLE_LIVE_EXTERNAL_AGENTS=1)
looked up run memory receipts as frames in server.multiMind and
server.mindCache, which the server no longer has. Look them up with
server.memory.get and check that personal and workspace receipts sit in
their own scopes. Also type the prompt assembler's bulk fixture as
RecallFrame.
- cli: delete real-session-simulation.ts, a manual script that populated a
hard-coded Windows .mind file through the old Orchestrator API.
- cli runtime: @waggle/core no longer depends on hive-mind-core, so the
packed closure does not include it.
- marketplace: rename the local marketplace schema constant so it is not
mistaken for the memory engine's SCHEMA_SQL.
- prompt assembler harnesses (v4 and v5): prime an in-memory contract store,
snapshot its memories and replay them into a fresh store per condition
instead of copying a MindDB file; the identity card is created per
orchestrator. They had also drifted from the current Orchestrator API
(config takes memory, buildSystemPrompt is async).
- prompt assembler unit test: build RecallFrame fixtures, the type the
assembler consumes, instead of legacy MemoryFrame rows.
- M3c integration: drop the knowledge-graph Ontology case; the knowledge
graph is not part of v1 and core no longer exports it.
- harvest-provenance: tested the removed harvest pipeline (RawArchive and
frame metadata), which no longer exists.
- benchmarks: drop the FTS5 frame search and frame/session write
benchmarks; they measured the legacy engine only.
- trust wiring: read install_audit from waggle.db through the server; the
old checks read a legacy personal.mind that nothing writes any more.
- data dir: a fresh install is marked by waggle.db, not personal.mind.
- backup: expect waggle.db and memory/ in the archive.
- cross-platform: replace the MindDB block with waggle.db checks (file, WAL,
foreign keys) and memory round trips (plain, Unicode, 100 KB) through
server.memory. FTS5 and session CRUD were legacy-engine features.
- store fixtures (install audit, cron) open WaggleDB instead of MindDB.
The server no longer opens personal.mind, so tests build it on an empty
data directory. The /memory command test seeds its two memories through
server.memory and now checks that the seeded text is found.
The persona memory-denial and cron Loop tests counted rows in the legacy
personal and workspace minds, which nothing writes any more. Count
server.memory per scope (and origin 'loop' for the truncated Loop report)
instead, and stop seeding a personal.mind the server no longer opens.
- install audit: build fixtures on WaggleDB, the store's real home; drop the
substrate SCHEMA_SQL parity half and the two MindDB-only recovery cases
for rows stranded in install_audit__mig_old.
- workspace manager: create() and ensure() no longer leave a workspace.mind;
getMindPath tests create the legacy file themselves. Drop the
MultiMindCache cases (reopen of a missing mind, post-resolution junction
swap): they tested the legacy cache, not the manager.
- transformers model load: use core's own in-process reranker.
createTeamTools and createLocalTeamTools called /api/teams/* and
/api/team/* on a team server. Nothing registered them: not the
sidecar, the CLI, the team server or the worker.
packages/server/src/benchmarks held a LoCoMo failure-mode aggregator and
an LLM judge ensemble for the old memory engine benchmarks. Nothing in
the sidecar imports it and no npm script runs it; its inputs lived in
the root benchmarks/ directory that P2.2 already removed. Judge and eval
loops are also out of the community edition (AGENTS.md, rule 4).
The command bar could still offer "WaggleDance" and "Team" (governance)
as navigation targets, but those screens and app ids are gone from the
web app, so the frontend could not route them. Also drop a comment that
pointed at the removed WaggleDance routes.
- The WaggleDance routes left the sidecar with team mode, so the
x-waggle-run-token header no longer opens /api/waggle-dance/signal or
/api/waggle-dance/signals. A run token still authenticates only as a
Bearer on /v1/chat/completions (OpenClaw).
- Remove the session inactivity timeout. It switched on only in team
mode, when CLERK_SECRET_KEY was set.