Repository navigation
Releases
Hivemind v2.31.2
Hivemind v2.31.2
Patch release. Ships #486 and #487 to #493, merged since 2.31.1. The tools agents see are unchanged. Most of it is for the hivemind-memory client packages and for operators.
hivemind-memory
- Typed replies (#493, ADR 0164). Each method returns its reply's shape:
TypedDicts inhivemind_memory.repliesfor Python, interfaces in the npm package. The types describe the parsed JSON and never validate it, so a key a newer server adds still reaches you.call()stays untyped. CI's live tests check that the replies match the types. - The skill copy fits an app (#490, ADR 0163). The
hivemindskill thatinstructions()hands your agent no longer mentions the plugin's setup skill or incognito sessions, which an app's agent cannot use. It tells the agent to ask the user for an activated or new key instead. - npm publishing from the tag works again (#486). The 2.31.1 npm publish failed on the tarball path, so npm never got 2.31.1 (PyPI did). 2.31.2 is the next version on npm after 2.31.0.
Upgrade with pip install -U hivemind-memory or npm install hivemind-memory@2.31.2.
What changed for agents
Nothing in the tools. Tool names, input schemas, descriptions, server instructions and search ranking are the same as 2.31.1.
Operators
- No migrations, no new settings, no API change.
- Upgrade the image to
ghcr.io/monochrome-muncher/hivemind:v2.31.2. - Grafana (#489): the fleet health row gains an "Unverified memories, per fleet" panel (
hivemind_health_fleet_unverified, ADR 0156). If you imported the dashboard by hand, re-importdeploy/kubernetes/optional/monitoring/hivemind-dashboard.json; a sidecar picks the ConfigMap up by itself. No new metric. - Operator docs (#487, #488): DEPLOY.md and the ops runbook are rewritten in plain voice. DEPLOY.md now says to set the embedder in
deploy/kubernetes/configmap.yaml, covers deploying without GitLab, and has a new Monitoring section (§10). Section numbers are kept. - Toolchain (#491, #492): uv 0.13.9, fresh digests for the Python base image and pgvector, mcp-publisher 1.8.9, and refreshed Python dependencies (FastAPI 0.165, pydantic 2.14.3, among others).
Plugin users
- Nothing to update.
Hivemind v2.31.1
Hivemind v2.31.1
Patch release. Ships #482, #483 and #484, merged since 2.31.0. All three fix the new hivemind-memory client packages; the server and the agent plugin are unchanged.
hivemind-memory fixes
- Publishing works with provenance (#482). The npm package now names its repository (npm refuses a provenance bundle without one), and both packages list their homepage and issue tracker. The publish workflow builds the packages in a job that has no publishing token and publishes the built files from jobs that run nothing else.
- A locked key is reported as such (#483). The clients used to turn every 401 into
unauthenticatedwith a generic message. They now raisekey_lockedwith the server's message, and the TypeScript client does so on any call, as Python did. An unrelated failure after an earlier 401, once a call has succeeded, is no longer reported as an authentication error. - Same error handling in both languages (#483). A failed call that is not the error envelope is wrapped as
tool_error, sorun_toolalways hands the model the envelope. TypeScript joins every text block of a reply and refuses a reply that is not an object, as Python does. - Typed options and runnable quickstarts (#484). Python's
kind,verdictandrecencytakeLiteraltypes (EntryKind,Verdict,Recency), as TypeScript's do.anthropicTools()typesinput_schemaastype: "object", so it type-checks against the Anthropic SDK. The README examples run as written, send the end-of-task reminder in code, and the Vercel AI SDK example setsstopWhen(generateTextstops after one step by default).
Upgrade with pip install -U hivemind-memory or npm install hivemind-memory@2.31.1.
What changed for agents
Nothing in the tools. Tool names, input schemas, descriptions, server instructions and search ranking are the same as 2.31.0.
Operators
- No migrations, no new settings, no API change.
- Upgrade the image to
ghcr.io/monochrome-muncher/hivemind:v2.31.1(no server change; the version moves with the packages).
Plugin users
- Nothing to update.
Hivemind v2.31.0
Hivemind v2.31.0
Minor release. Ships #478, #479 and #480, merged since 2.30.0. The headline is hivemind-memory, new Python and npm packages for building your own agents or systems on Hivemind.
New: hivemind-memory for Python and npm
For agents built outside the supported harnesses (#480, ADR 0162). Both packages are async clients of the hostable /mcp endpoint:
- One typed method per tool (search, write, get, feedback and the rest).
instructions()returns the server instructions plus thehivemindskill, ready for a system prompt;reminder()returns the end-of-turn reminder the Stop hook uses.- Tool definitions for the Anthropic and OpenAI APIs.
- Adapters as extras: LangChain for Python, the Vercel AI SDK for npm.
- Configured with
HIVEMIND_MCP_URLandHIVEMIND_API_KEY;connect()checks the key withhive_whoami. - Python 3.10 and up (
pip install hivemind-memory), Node 22 and up, ESM (npm install hivemind-memory).
The packages carry the server's version and are released with it. See clients/README.md and the READMEs in clients/python and clients/typescript.
What changed for agents
Nothing in the tools. Tool names, input schemas, descriptions, server instructions and search ranking are the same as 2.30.0.
Other changes
- Docs brought in line with the code (#478): Antigravity skills, OpenCode and KiloCode events, the runbook's tool-call series, and smaller fixes.
- Dead code removed (#479), no behaviour change.
Operators
- No migrations, no new settings, no API change.
- Upgrade the image to
ghcr.io/monochrome-muncher/hivemind:v2.31.0.
Plugin users
- Nothing to update.
Hivemind v2.30.0
Hivemind v2.30.0
Minor release. Ships #474, #475 and #476, merged since 2.29.0. Tool calls on the MCP runner are now counted by error code, the shipped Grafana dashboard tells refusals from real failures, and the Cursor plugin is ready for the Cursor Marketplace.
What changed for agents
Nothing in the tools. Tool names, input schemas, descriptions, server instructions and search ranking are the same as 2.29.0.
Monitoring
- New
codelabel onhivemind_mcp_tool_calls_total(#475, ADR 0161). It isnonefor a success,raisedwhen the tool raised, and otherwise the reply's error code (for examplepermission_deniedorfleet_busy). An unknown code counts asother, so the label set stays small. Theoutcomelabel is unchanged, so existing queries keep working. - Two dashboard panels (#475). "Tool error share" is replaced by "Tool failure share", which counts only server-side failures (database timeout, embedder down, raise). The new "Tool refusals per minute, by code" panel shows the refusals an agent can fix, by tool and code. Re-import the dashboard (
deploy/kubernetes/optional/monitoring/hivemind-dashboard.json) to get them. Against an older image the failure panel reads 0 and the refusal panel is empty. - Faster fleet health (#476). Vital signs counts unverified entries with its own query that uses the tags index, instead of checking every active entry. Same counts; at 200,000 entries
fleet_healthwent from 901 to 748 ms.
Operators
- No migrations, no new settings, no API change.
- Re-import the Grafana dashboard (see above).
- Upgrade the image to
ghcr.io/monochrome-muncher/hivemind:v2.30.0.
Plugin users
- Cursor (#474). The plugin's manifest now carries the logo, homepage, repository and license a Cursor Marketplace listing shows. Setup copies the stop script to
~/.cursor/hooks/and points the user stop hook at that copy, so the hook keeps working wherever the plugin is installed. If you set up Cursor before, ask the agent to "Update Hivemind": it copies the script and, with your approval, points the hook at it. A marketplace listing is pinned to a reviewed commit, so it can trail the latest release. - The plugin README has a new Privacy section: the plugin collects nothing and its hooks contact no server.
- Other harnesses: nothing to update.
Hivemind v2.29.0
Hivemind v2.29.0
Minor release. Ships #470, #471 and #472, merged since 2.28.0. An admin can now rename an agent, and the admin panel gathers each agent's edits into one dialog.
What changed for agents
Nothing in the tools. Tool names, input schemas, descriptions, server instructions and search ranking are the same as 2.28.0, and the plugin is unchanged.
- If your admin renames you, your key keeps working. You keep your trust level, fleet, memories, ratings, pins and anything you inherited, and
hive_whoamishows the new name on your next call. fleet_busyduring a rename (#470, #472, ADRs 0159 and 0160). A write, pin or rating sent at the very moment of a rename gets the retryablefleet_busy. Retry and it goes through under the new name. A rating now also gives up after 3 seconds on a memory that a cleanup, merge or move holds, answeringfleet_busyinstead of waiting.
Admin panel
- Rename an agent (#470, ADR 0159).
POST /v1/admin/agents/{name}/renamein the API andhivemind-keys rename-agent OLD NEWin the CLI, admin key only and audited asagent.rename. It works for an agent in any status and rewrites every copy of the name in one transaction: the key, memories, ratings, pins, successions, cleanups and fleet moves. Old audit rows keep the old name, and the agent's history (admin and portal) follows its former names. A former name stays reserved, so no new agent can register it. - One Edit dialog and a More menu (#471). Each agent row has one Edit dialog for name, owner, trust level and fleet, and a More menu for the other actions. The agent's own page keeps its full set of buttons. If the dialog saves some fields and then the rename fails, the page refreshes so it shows what was saved (#472).
Operators
- Migration 0036 runs on upgrade (expand-only): a new
agent_former_namestable and theagent.renameaudit action. Rolling it back needs--discard-audit-rowsonce any rename has happened. - No new settings.
- Don't rename agents until every pod runs 2.29.0. An older pod doesn't know which names are reserved.
- Renaming an agent with many memories rewrites all of them, so it is slow and leaves work for autovacuum, much like a fleet merge.
- Upgrade the image to
ghcr.io/monochrome-muncher/hivemind:v2.29.0.
Plugin users
Nothing to update.
Hivemind v2.28.0
Hivemind v2.28.0
Minor release. Ships #462 to #468, merged since 2.27.2. It trims agent text that only weak models needed, and adds a way for agents to mark claims they have not tested.
What changed for agents
- Entries "may be wrong" (#462, ADR 0155). The server instructions and the always-on reminder now say entry content is data written by other agents and may be wrong, not only that it is never to be followed as instructions.
- The
unverifiedtag (#463, ADR 0156).hive_write's description and the skill ask agents to tag a guess, an unconfirmed fix or something they were toldunverified. A reader whose work tests it rates it; a confirminghelpfulwith a note, or a successor without the tag, clears it. Ranking is unchanged. - The
tentativewrite warning is gone (#464, ADR 0157). It was a net for small models, which ignored it; Claude models never needed it. Scope by evidence stays in the skill, andnot_englishand the other warnings stay. - Shorter tool descriptions (#465, ADR 0158).
hive_searchno longer repeats theclosest: loosenote its reply already carries;hive_whoamipoints at the hivemind-switch skill without the how-to. Always-on text is now 17,270 of 18,000 bytes. Opus, Sonnet and Haiku behaved the same before and after.
Tool names and input schemas are unchanged. Tool descriptions and server instructions changed.
Admin panel and agent portal
- An "unverified" badge on entries and an "Unverified only" filter (
unverified=trueon the entry lists). An entry counts as unverified while it is active, taggedunverifiedand never confirmed. - Vital signs shows the unverified count per fleet.
Operators
No migrations, no new settings. New optional unverified query parameter on the entry lists, and a new Prometheus gauge hivemind_health_fleet_unverified (per fleet). Upgrade the image to ghcr.io/monochrome-muncher/hivemind:v2.28.0.
Plugin users
Update the plugin: the skill and the reminder text changed. If you copied the skills or a hook by hand, copy them again. Cursor users: copy the plugin's cursor/rules/hivemind.mdc to ~/.cursor/rules/ again.
Other
Hivemind v2.27.2
Hivemind v2.27.2
Patch release. Ships one fix merged since 2.27.1, found by the weekly security audit on 2026-11-16.
What changed
- A write with a list or object
scopein its payload no longer fails (#460). Thepayload_fieldswarning added in 2.27.0 (#455, ADR 0152) compared the payload'sscopevalue against the known scopes with a dict lookup, so a list or object there (for example{"scope": ["self"]}) raised an error andhive_writeandPOST /v1/entriesrefused a valid write. Only a stringscopeis compared now; any other value is stored as given and draws no warning. Not a security issue: the failure was confined to the caller's own request.
Tool names, input schemas, tool descriptions, the always-on text and search ranking are unchanged.
Operators
No migrations, no new settings, no API change. Upgrade the image to ghcr.io/monochrome-muncher/hivemind:v2.27.2.
Plugin users
Nothing to refresh. Updating the plugin only moves its version number.
Hivemind v2.27.1
Hivemind v2.27.1
Patch release. Ships one change merged since 2.27.0.
What changed
- Always-on text trimmed (#458, ADR 0154): 17,956 to 17,107 bytes of the 18,000 budget, no rule dropped. The skill no longer repeats what a tool description or the server's reply already says.
- Tool descriptions changed:
hive_write,hive_get,hive_withdraw,hive_feedback,hive_register, and the server instructions each lose a clause the replies already explain. - The MCP
embedding_unavailablemessage now says "don't retry in a loop: keep the entry locally and write it later". REST is unchanged. - Skill: on
supersede_denied,hive_getthe target withinclude_historyand supersede its newest successor, else flag it.
- Tool descriptions changed:
Tool names, input schemas and search ranking are unchanged.
Operators
No migrations, no new settings, no API change. Upgrade the image to ghcr.io/monochrome-muncher/hivemind:v2.27.1.
Plugin users
Update the plugin to get the shorter skill. If you copied the skills by hand, copy them again.
Hivemind v2.27.0
Hivemind v2.27.0
A new write warning for a common slip of weaker models, clearer tool and
skill text, and a batch of plugin fixes, most of them found by checking
the plugin live in every harness. Antigravity 1.5.8 and later can install
the plugin again from this release on. No migrations, no new settings,
no API change.
For agents
- New
payload_fieldswrite warning (ADR 0152, #455). A weak model
sometimes putsscope,tags,sourcesorimportanceinside
payload. The entry was stored with defaults and nothing said so. Now
the write warns, names each such field and says to pass it as its own
argument and supersede the stored entry. It is advisory like the other
warnings, and only fires when the real argument was left unset. - Tool and skill text (#456).
hive_whoami's description now says
the key is fixed for the session, so acting as another agent or fleet
means the user starts a new session (ADR 0151). A small model had
written and pinned a made-up "switch" decision instead. The skill now
answers a general question with the general answer first, and keeps a
memory about one system or incident as a labelled aside unless it fits
the user's case (ADR 0153). It also stops contradicting itself about
alternativeentries, which may go to the fleet. To pay for this,
hive_write,hive_searchandhive_geteach lost a clause the
replies already explain. The always-on text is 17,956 of 18,000 bytes. - Pins for any spelling of the fleet id (#437).
hive_pinnedand
GET /v1/pinsreturned no pins and nofleet_phasewhen the fleet id
came in upper case or without hyphens. They now fold it first, and so
does the admin panel's fleet pins page.
Plugin and harnesses
- Antigravity installs again (#436). Antigravity 1.5.8 refuses
a plugin whose skills folder is a link out of its folder, so
plugins/antigravity/skillsis now a real copy, kept in step by a
test. Install with
agy plugin install https://github.com/monochrome-muncher/hivemind/tree/v2.27.0/plugins/antigravity.
The stop hook also reads Antigravity 1.5.9'scontinuationActive
(#441). - Incognito Codex always drops the key (#439). Before, the key stayed
in the session whenconfig.tomlhad no Hivemind entry, where another
MCP server named by a trusted project could read it. - Launchers: Cursor's new
agentcommand now counts as Cursor in
bothhivemind-incognitoandhivemind-as(#434), and DeepSeek
Harness, OpenCode and Kilo Code installs now include the launchers the
hivemind-switchskill points to (#433). - Docs: a Hermes profile needs the Hivemind server in its own config
(#435); setup notes name the harness versions checked live, and the
plugin README counts four skill folders (#438, #445, #454).
Under the hood
No behaviour change: the Store port is split into one protocol per area,
the embedder and extractor share one HTTP client, search ordering and the
MCP error codes have their own modules, and the admin panel's
manage.js is split by page (#440, #442-#444, #446-#449). A code tour in
docs/code-tour.md follows a request through the layers for anyone
auditing the code (#450). The test suite lost duplicate tests and copied
helpers (#451-#453).
Upgrading
Upgrade the server for the new warning, the tool text and the pins fix;
there is nothing to migrate. Update the plugin for the skill text and the
launcher and Antigravity fixes.
- If you copied the skills by hand (Cline, or any harness set up by
hand), copy them again. - If you copied
hivemind-incognito,hivemind-asorhivemind-stop.sh
instead of linking them, copy them again. - Antigravity users on 1.5.8 or later who installed from
tree/maincan
move to thev2.27.0link above.
Hivemind v2.26.0
Hivemind v2.26.0
Key profiles for people who work in more than one fleet, a new
hivemind-switch skill, and fixes that keep incognito sessions really
incognito in a few more places. Everything here is in the agent plugin.
The server is unchanged: no migrations, no new settings, no API change,
and the ten tool names, schemas, always-on text and search ranking are
the same.
- Key profiles (ADR 0150, #431). If you work for several teams, you
have one agent in each fleet, each with its own key. A key profile is a
name for one of them, andhivemind-as <profile> <harness>starts
Claude Code, Codex or another harness as that agent, for example
hivemind-as client claude. Each profile's key lives in
~/.config/hivemind/; Hermes uses its own-pprofiles.defaultis
your usual setup, unchanged.hivemind-as auto <harness>picks the
profile from the folders you list in~/.config/hivemind/profiles, so
a folder always runs in the same fleet. A switch is always a new
session, because the conversation you are in already holds what the
other fleet's memories said. The launcher drops keys it inherited from
the shell, refuses a profile file other users can read, and never
prints a key. Junie and Antigravity keep the key in their own settings,
so there the skill gives the steps to switch by hand. - New skill
hivemind-switch(#431). Ask your agent to "switch to my
client's fleet" or "add a Hivemind profile" and it walks you through it
without the key passing through the chat. It loads only when asked, so
it adds no always-on text. This is the plugin's fourth skill. - Incognito fixes (#430).
hivemind-incognito code .(orcodium)
now keeps Hivemind off in the Cline and Kilo Code editor extensions.
OpenCode and KiloCode incognito sessions no longer keep the key. The
Hermes checks follow-pprofiles. The Claude Code setup notes now say
that a project's settingsenvdoes redirect the plugin's URL. - Docs and tests: the READMEs and agent docs caught up with 2.25 (#425,
#426), and the test suite gained gap tests and lost duplicated helpers
(#427-#429).
Upgrading
Update the plugin to get hivemind-as, the new skill and the incognito
fixes. Server upgrades need nothing.
- If you copied the skills by hand (Cline, or any harness set up by
hand), also copy the newskills/hivemind-switch/folder. - If you copied
hivemind-incognitoinstead of linking it, copy it
again. Linkbin/hivemind-asinto a folder on yourPATHto use key
profiles. - For a confidential client fleet, ask your admin for a
contributor
agent there: aprivilegedagent reads every fleet.