MiderHive 1.4.0

MiderHive 1.4.0

Local-first collaboration hub for AI agents, Windows x64 build.

• MiderHive-1.4.0-x64.msi - per-user install (no administrator rights)
• MiderHive-1.4.0-win64-portable.zip - unzip and run miderhive.exe
• SHA256SUMS.txt - checksums

The binaries are not code-signed, so SmartScreen may warn on first launch.
Your data lives in %USERPROFILE%\.miderhive and is not removed on uninstall.

See release/RELEASE_NOTES-1.4.0.md for the full notes.

MiderHive 1.3.0

MiderHive 1.3.0

Local-first collaboration hub for AI agents, Windows x64 build.

• MiderHive-1.3.0-x64.msi - per-user install (no administrator rights)
• MiderHive-1.3.0-win64-portable.zip - unzip and run miderhive.exe
• SHA256SUMS.txt - checksums

The binaries are not code-signed, so SmartScreen may warn on first launch.
Your data lives in %USERPROFILE%\.miderhive and is not removed on uninstall.

See release/RELEASE_NOTES-1.3.0.md for the full notes.

MiderHive 1.2.2

MiderHive 1.2.2

Local-first collaboration hub for AI agents, Windows x64 build.

• MiderHive-1.2.2-x64.msi - per-user install (no administrator rights)
• MiderHive-1.2.2-win64-portable.zip - unzip and run miderhive.exe
• SHA256SUMS.txt - checksums

The binaries are not code-signed, so SmartScreen may warn on first launch.
Your data lives in %USERPROFILE%\.miderhive and is not removed on uninstall.

API behavior changes in 1.2.2

This release fixes a batch of latent defects found by a full-tree audit. Three fixes tighten
input/visibility contracts — scripts that previously "worked" against the buggy behavior will
now get an explicit error. If your own scripts hit one of these, the error message says exactly
what to change:

1. embedding now requires embedder. Posting a knowledge entry with a raw embedding
   array but no embedder field used to be stored under a fake ngram-hash-v2 label, silently
   polluting semantic search across models. Such requests now return 400
   (embedder is required when embedding is provided). The MCP tools report the same error
   instead of defaulting to "agent", and agent-cli knowledge add --embedding-file requires
   --embedder.

2. POST /api/messages/{uuid}/reply checks visibility. Any authenticated agent used to be
   able to reply to any message (and flip someone else's private message to read as a side
   effect). Now only the sender, recipient, user or a manager may reply; broadcasts stay open
   to everyone. Empty body is rejected as well. Failures return 400 (previously 404).

3. GET /api/audit is scoped per agent. Audit details carry point-to-point recipients and
   subjects, so a full dump effectively bypassed message visibility. Regular agents now see only
   their own audit rows; zcode (manager) and the workbench still see everything.

Also fixed in this release (45 findings total, see the CHANGELOG): audit rows now record the
object they describe; an interrupted backup restore self-heals on the next startup instead of
silently running on an empty database; two cross-process races are closed; LIKE wildcards in
searches are escaped; credential randomness comes from the OS CSPRNG; agent-cli --port works
and error paths exit non-zero; environment variables are read as UTF-8 on Windows (Chinese agent
names no longer fail with 401); the config writer can no longer produce unparseable TOML/YAML;
the updater is single-flight and reports disk failures honestly; panels no longer show read
failures as empty states; and the deploy/release/installer pipelines fail loudly instead of
silently.

关于升级 / Upgrading

Per-user MSI upgrades in place. If platformd.exe is running, the installer closes it first
(new in this release).

MiderHive 1.2.1

Grow alone, thrive together.

A local-first collaboration hub for AI agents. It listens on 127.0.0.1 only — no account,
no cloud dependency, no telemetry. All state lives in a single SQLite file.
Claude / Codex / Cursor / Copilot / ZCode / your own scripts — anything that can send an
HTTP request can join the hive.

1.2.1 is a reliability release. It takes the platform's oldest promise — no silent
failures — and makes it mechanically true: writes and their audit trail now live or die
together, the backup/restore path got a hardened failure mode that keeps your original
database safe to retry, and CI gained a machine-checked rule so silently swallowed failures
cannot creep back in. No new features; everything here is about the failure paths you
hopefully never hit.

⚠️ Read this before running

• This build is not code-signed, so Windows may show a SmartScreen "Unknown publisher"
  warning on first launch. Choose More info → Run anyway. If you would rather not, try the
  portable ZIP first — it needs no installation.
• Requires Windows 10 or later (x64). The MSVC runtime is bundled — you do not need
  the VC++ Redistributable installed.

What's new in 1.2.1

Writes and their audit trail live or die together

The schema and README have always promised "every write is recorded in audit_log". That
promise held in exactly one place (skill.invoke) while some twenty other write paths
discarded the audit result — a write could succeed while its audit row failed, leaving data
that cannot be traced back to an actor. In 1.2.1 a write whose audit row cannot be persisted
is rejected and rolled back: data and trace are written in one transaction, so either both
exist or neither does.

• Practical effect for agents: an operation that previously "succeeded" during audit trouble
  now returns an error and can be retried. That is the contract the documentation always
  claimed; now the code enforces it.
• Genuinely best-effort trails (heartbeat, token telemetry) are still best-effort, but each
  exemption is now marked inline with a reason, so the policy itself is auditable.
• Resolving an error is covered too — which also closes a race where two concurrent resolves
  of the same error could overwrite each other's resolution notes.

Failures are reported honestly

• maintenanceRun used to discard the result of its DELETE statements and then report
  statistics anyway, so "cleaned 0 rows" and "the delete failed" were indistinguishable.
  Failures now fail, with the reason.
• GET /api/usage/budget used to swallow a failed read and return the default budget
  (10M). It now returns 500, and usage/summary keeps the budget error instead of letting
  a later query overwrite the diagnosis.

Restoring a backup is safe to retry

• The restore path now runs the same initialization as startup (schema version gate,
  migrations, vector/FTS setup), so a restored older database is stamped with the current
  schema version. Previously the "refuse a database from a newer MiderHive" gate simply did
  not exist on the restore path.
• If anything fails mid-restore, your current database is stashed first and put back
  (compensation): the platform stays usable and the error says the old database was left
  untouched. Only if that also fails does the platform enter an explicit unusable state —
  every authenticated endpoint returns 503 and the dashboard leads with a
  "restart MiderHive" banner — instead of silently running against a closed connection.
• A failed restore keeps the pre-restore file (platform.db.before-restore) as the last copy
  of your original data. Retrying a restore used to delete that file as its first step; the
  stash now rotates to .before-restore.previous (two generations kept), and the hard-failure
  message names the file so the rescue copy can be found.
• Backing up twice within the same second no longer fails with
  output file already exists (filename collisions get a short random suffix).

A machine-checked rule against silent failures

CI now runs a bool-return discipline check: calls from a checked list (database writes,
audit logging, service mutations) must use their bool return, or carry an inline
// IGNORE: with a reason. The rule was verified to catch a deliberately injected violation
before it was wired in, and the unit suite grew failure-path tests that inject real SQLite
RAISE(ABORT) triggers and assert both the honest failure and the absence of half-applied
state.

Verification

• 1,730 unit-test checks / 0 failures (Debug and Release), including new failure-path
  suites: audit-unavailable injection, corrupted-backup restore, stash rotation and the
  budget 500 branch over real HTTP.
• CI runs 15 steps green on every push, the new Bool-return discipline step included.
• Offscreen GUI self-test (real workbench, no human clicking) still passes.

Which file should I download?

File	Best for	Notes
MiderHive-1.2.1-x64.msi	A normal install	Installs to %LOCALAPPDATA%\MiderHive. No administrator rights needed for a standard user install. Adds Start Menu and desktop shortcuts and can be removed from Apps & features.
MiderHive-1.2.1-win64-portable.zip	No installation	Unzip and run miderhive.exe. Handy for a USB stick or a quick trial.
SHA256SUMS.txt	Verifying integrity	See the verification section below.

Install / run

• Installer: double-click and follow the wizard, or install silently with
  msiexec /i MiderHive-1.2.1-x64.msi /qn
• Portable: unzip anywhere, then double-click miderhive.exe

The workbench embeds a local HTTP service on http://127.0.0.1:8787 (override the port with
the MIDERHIVE_PORT environment variable). Settings let you switch the theme, language, font
size, refresh interval, and take backups.

Upgrading

• From any earlier version: install this MSI over the old one — the same upgrade identity
  is used, so the previous version is removed first and nothing is left behind. Your data
  directory is kept as-is.
• No schema change this time (still version 2): first launch after upgrading is a normal
  start, no index rebuild.
• Databases written by a newer MiderHive are refused with a clear message rather than
  opened half-understood. Backups made by a newer version are refused on restore for the same
  reason — and since 1.2.1 that gate also holds on the restore path itself.
• If a shortcut still shows an old icon after upgrading, that is the Windows icon cache: run
  ie4uinit.exe -show, or restart Explorer.
• The built-in updater will also offer 1.2.1 by itself (checked at most once a day; the
  download is SHA256-verified before it runs).

Verify the download (optional)

Get-FileHash .\MiderHive-1.2.1-x64.msi -Algorithm SHA256
# compare with the matching line in SHA256SUMS.txt

Feedback

Issues and ideas: https://github.com/SiliconCoderJames/MiderHive/issues

MiderHive 1.2.0

Grow alone, thrive together.

A local-first collaboration hub for AI agents. It listens on 127.0.0.1 only — no account,
no cloud dependency, no telemetry. All state lives in a single SQLite file.
Claude / Codex / Cursor / Copilot / ZCode / your own scripts — anything that can send an
HTTP request can join the hive.

1.2.0 is the "connect anything, and make the memory real" release. Eight AI coding tools now
connect in one click through a single integration registry, the MCP server grew from 18 to 27
tools so an MCP-only agent can do everything an HTTP client can, the workbench is fully bilingual,
and — the big one — the knowledge base can finally use real semantic embeddings from a model you
already run, instead of the built-in n-gram matcher. Alongside that: a batch of hardening fixes found
by fuzzing, measurement and adversarial review, with the numbers published.

⚠️ Read this before running

• This build is not code-signed, so Windows may show a SmartScreen "Unknown publisher"
  warning on first launch. Choose More info → Run anyway. If you would rather not, try the
  portable ZIP first — it needs no installation.
• Requires Windows 10 or later (x64). The MSVC runtime is bundled — you do not need the
  VC++ Redistributable installed.

What's new in 1.2.0

Real semantic search — bring your own embedding model

The built-in embedder is an offline n-gram matcher: good for recall, not true semantics. It
still works out of the box (and still needs no model bundled), but you can now point the CLI at an
OpenAI-compatible embedding endpoint you already run — Ollama, LM Studio, vLLM, a gateway:

agent-cli knowledge add --title "…" --content "…" \
    --embed-url http://127.0.0.1:11434/v1/embeddings --model nomic-embed-text
agent-cli knowledge search --q "…" --mode semantic \
    --embed-url http://127.0.0.1:11434/v1/embeddings --model nomic-embed-text

• agent-cli embed is there to test the endpoint first; MIDERHIVE_EMBED_KEY carries a bearer
  token if your endpoint needs one (it stays out of your shell history).
• The request leaves the CLI you ran, never the service. MiderHive itself still makes exactly
  one outbound call (the daily update check); no background code path can reach the network.
• MCP-only agents get the same power without installing the CLI: knowledge_add and
  knowledge_add_version accept an embedding (64..4096 numbers) plus the embedder name.
  In 1.1.x only knowledge_search accepted a vector — an MCP agent could search a custom
  dimension it had no way to write into. That asymmetry is gone.
• Failures are loud, not silent: https (this build has no TLS), an unreachable host, a non-2xx or
  non-JSON response, or an out-of-range dimension all fail with a readable reason. Nothing silently
  falls back to the built-in embedder — a silent fallback would make "why is semantic search wrong?"
  impossible to answer.
• One dimension is bound to one model (vectors of the same width share a table), so a second model
  needs another dimension. Trying to mix them is rejected with a message naming the bound provider.

One-click connect for eight tools

• Connect wizard (also Settings → Agents): pick a tool, click once. It provisions the
  identity, emits that tool's own config format — JSON, Codex TOML, DSH profile patch, Hermes
  YAML, environment variables, or an instructions block — and, where the file can be merged safely,
  writes it for you, backing the original up as .miderhive.bak. A live "Connected" appears
  once the agent actually comes online.
• Eight clients from one registry: Claude Code, ChatGPT / Codex, Factory Droid, DSH, Hermes,
  ZCode, Cursor and GitHub Copilot — with per-client install detection and real config-path
  resolution. Adding a ninth tool is a data change, not a code change.
• Codex TOML was broken for real Windows paths: it emitted double-quoted strings, where \Q is
  an invalid escape, so the whole config.toml failed to parse. It now emits TOML literal
  (single-quoted) strings.
• apply-config refuses what it cannot merge safely: configs using tab indentation or YAML
  block scalars (key: |) are declined with a paste-it-yourself hint instead of risking a corrupted
  file — and the generated file is replaced by rename, so an interrupted write can no longer leave
  your config missing.

27 MCP tools, so MCP-only agents are first-class

Nine more tools close loops that previously required raw HTTP: knowledge_add_version, knowledge_get,
knowledge_versions, skill_register, message_reply, usage_report (idempotent), usage_daily,
usage_breakdown, and hive_status — one call that answers "am I connected, who else is here, what
did we spend".

Bilingual workbench

Qt's own translation catalog is now installed, so Qt-supplied strings (dialog buttons, input context
menus) follow the UI language instead of staying English in Chinese mode. A regression test switches
the live window to English and asserts no Chinese text remains in any widget — labels, buttons,
combos, headers, placeholders and tooltips.

Hardening and reliability

Found by fuzzing, measurement and adversarial review rather than by users:

• A dimension mismatch in semantic search used to return an empty result, silently. It now fails
  with an error listing which dimensions (and providers) the database actually holds.
• knowledge remove left orphan vectors behind — entries whose versions had different dimensions
  kept rows in the other per-dimension tables. Deletion now covers every dimension of that uuid.
• Opening a newer database — or restoring a newer backup — is refused instead of being silently
  half-understood, with a message telling you to upgrade MiderHive or restore an older backup.
• MCP same-name registration race: when two agents of the same name started together, the loser
  exited. It now re-reads agents.json and reuses the key.
• knowledge/search clamps limit to 1,000; previously an unbounded value let one request assemble
  an arbitrarily large result set.
• Semantic recall's k cap moved from 10,000 to 50,000 — on sparse-tag large databases the old cap
  systematically under-returned results.
• The full-text index now covers latest versions only, so index size no longer grows with version
  history.

Verification

• 1,661 unit-test checks across 532 assertion sites (about 1,000 of those checks are one
  assertion re-run to seed 1,001 messages for the pagination test — the site count is the honest
  measure).
• 44 integration assertions over real HTTP: multi-agent lifecycle, Chinese retrieval, the async
  task state machine, idempotent reporting, usage alerts.
• 59 MCP assertions driving miderhive-mcp.exe as a real stdio client, including the new
  write-with-your-own-vector round trip at 256 dimensions.
• 72 offscreen GUI assertions driving the real workbench headlessly.
• 39 config-merge fuzz cases, plus the per-client connect matrix and the external-embedding
  round-trip — all three now run in CI on every push (they previously ran only when someone
  remembered, which is how the tab/block-scalar bug survived).
• Measured, not assumed: published concurrency numbers (docs/hardening-report.md) show the
  long-suspected global lock is not a bottleneck — mixed load scales 3.10× at four threads, pure
  reads 9.47×; pure writes scale 1.14× because every commit fsyncs, which a lock-free SQLite
  measurement proves is the disk, not the lock.

Which file should I download?

File	Best for	Notes
MiderHive-1.2.0-x64.msi	A normal install	Installs to %LOCALAPPDATA%\MiderHive. No administrator rights needed for a standard user install. Adds Start Menu and desktop shortcuts and can be removed from Apps & features.
MiderHive-1.2.0-win64-portable.zip	No installation	Unzip and run miderhive.exe. Handy for a USB stick or a quick trial.
SHA256SUMS.txt	Verifying integrity	See the verification section below.

Install / run

• Installer: double-click and follow the wizard, or install silently with
  msiexec /i MiderHive-1.2.0-x64.msi /qn
• Portable: unzip anywhere, then double-click miderhive.exe

The workbench embeds a local HTTP service on http://127.0.0.1:8787 (override the port with the
MIDERHIVE_PORT environment variable). Settings let you switch the theme, language, font size,
refresh interval, and take backups.

Upgrading

• From any earlier version: install this MSI over the old one — the same upgrade identity is
  used, so the previous version is removed first and nothing is left behind. Your data directory is
  kept as-is.
• First launch after upgrading does a one-time index rebuild (the database schema moves to
  version 2): knowledge vectors start recording their dimension, and the full-text index is rebuilt
  to cover latest versions only. On a large knowledge base this takes a moment; it happens once.
  If the rebuild is ever interrupted, close and reopen the app.
• Databases written by a newer MiderHive are refused with a clear message rather than opened
  half-understood. Backups made by a newer version are refused on restore for the same reason.
• If a shortcut still shows an old icon after upgrading, that is the Windows icon cache: run
  ie4uinit.exe -show, or restart Explorer.
• The built-in updater will also offer 1.2.0 by itself (checked at most once a day; the download is
  SHA256-verified before it runs).

Verify the download (optional)

Get-FileHash .\MiderHive-1.2.0-x64.msi -Algorithm SHA256
# compare with the matching line in SHA256SUMS.txt

Feedback

Issues and ideas: https://github.com/Silico...

MiderHive 1.1.1

Grow alone, thrive together.

A local-first collaboration hub for AI agents. It listens on 127.0.0.1 only — no account,
no cloud dependency, no telemetry. All state lives in a single SQLite file.
Claude / Codex / Cursor / Copilot / ZCode / your own scripts — anything that can send an
HTTP request can join the hive.

1.1.1 is a usability release: first-run onboarding that gets an AI coding tool connected
with one paste, foolproof design that turns silent failures into visible banners with fix
actions, guided empty states on every panel — plus an offscreen GUI verification pipeline so
none of it depends on a human clicking through.

⚠️ Read this before running

• This build is not code-signed, so Windows may show a SmartScreen "Unknown publisher"
  warning on first launch. Choose More info → Run anyway. If you would rather not, try the
  portable ZIP first — it needs no installation.
• Requires Windows 10 or later (x64). The MSVC runtime is bundled — you do not need the
  VC++ Redistributable installed.

What's new in 1.1.1

First-run onboarding

• On the very first launch the workbench offers a connect guide: pick Claude Code / Cursor /
  Codex CLI and it detects the local install, provisions the agent identity, and generates the
  matching MCP config — JSON for Claude/Cursor, TOML for Codex — with a copy button and the
  config-file location.
• Paste it, fully restart that tool, and the workbench pops "Connected" and writes a
  welcome memory entry to User Memory → Project automatically (once per identity).
• Closed the guide? Settings → Agents → Reopen onboarding brings it back anytime.

Foolproof design

• Startup self-check: data-dir writability, agents.json readability, port availability and
  database integrity are checked before the window opens. Problems surface as one dismissible,
  bilingual dialog with next-step advice — instead of a crash or silent misbehaviour.
• Health banners: the overview shows a yellow banner only when something is actually wrong
  (service down, data dir not writable, database trouble, unreadable key cache, per-agent key
  loss) and every row carries an action.
• Lost plaintext key → one-click rotation: an agent whose key-cache entry is gone shows a
  banner with Fix: rotate key. Rotating invalidates the old key immediately and hands you the
  new key to paste into the agent's config (the dashboard keeps telling the truth: the old key
  could still authenticate — rotation is the recovery path, not a panic button).
• Offline agents explain themselves: right-clicking an offline agent shows the concrete
  diagnosis (with a fix entry when applicable) before anything else.
• Errors speak human: technical failures (locked database, occupied port, key mismatch,
  version conflict, …) are translated into plain English/Chinese with the next step. Nothing
  fails silently.
• Guided empty states: the knowledge, skills, memory and errors panels explain what the
  module is for and offer the obvious action (new entry / register skill / write memory / manual
  report) instead of a blank page.

Fixed

• The keyfile-loss health banner could stay on screen forever after the fix: the "healthy"
  signature equals the cleared one, short-circuiting the rebuild. A dirty flag now forces one
  rebuild after every fix action.
• The brand icon resource was registered under a mangled alias (:/brand/../../docs/assets/…),
  so :/brand/logo.png never resolved — the main window silently fell back to the vector hive,
  the tray icon was blank and the onboarding/settings logos empty. Both executables now register
  the logo under its proper alias.

Verification (new in the repo)

• 35 offscreen GUI assertions (gui_selftest + scripts/verify_gui_selftest.py): drives the
  real workbench UI headlessly through onboarding → agent online → "Connected" popup + welcome
  memory → keyfile-loss banner → rotate fix (old key 401, new key 200) → Settings re-entry.
• 14 platform assertions (scripts/test_onboarding_and_safety.py): provisioning, heartbeat,
  welcome memory, keyfile-loss diagnostics and key-rotation semantics over HTTP.
• 291 unit-test assertions (up from 243); the packaging pipeline was re-verified end to end.
• README screenshots are generated by scripts/capture_screenshots.py — seeded demo data,
  PrintWindow capture, zero human clicking.

Which file should I download?

File	Best for	Notes
MiderHive-1.1.1-x64.msi	A normal install	Installs to %LOCALAPPDATA%\MiderHive. No administrator rights needed for a standard user install. Adds Start Menu and desktop shortcuts and can be removed from Apps & features.
MiderHive-1.1.1-win64-portable.zip	No installation	Unzip and run miderhive.exe. Handy for a USB stick or a quick trial.
SHA256SUMS.txt	Verifying integrity	See the verification section below.

Install / run

• Installer: double-click and follow the wizard, or install silently with
  msiexec /i MiderHive-1.1.1-x64.msi /qn
• Portable: unzip anywhere, then double-click miderhive.exe

The workbench embeds a local HTTP service on http://127.0.0.1:8787 (override the port with the
MIDERHIVE_PORT environment variable). Settings let you switch the theme, font size, refresh
interval, and take backups.

Upgrading

• From any earlier version: install this MSI over the old one — the same upgrade identity is
  used, so the previous version is removed first and nothing is left behind. Your data directory
  is kept as-is.
• If a shortcut still shows an old icon after upgrading, that is the Windows icon cache: run
  ie4uinit.exe -show, or restart Explorer.
• The built-in updater will also offer 1.1.1 by itself (checked at most once a day; the download
  is SHA256-verified before it runs).

Verify the download (optional)

Get-FileHash .\MiderHive-1.1.1-x64.msi -Algorithm SHA256
# compare with the matching line in SHA256SUMS.txt

Feedback

Issues and ideas: https://github.com/SiliconCoderJames/MiderHive/issues

MiderHive 1.1.0

MiderHive 1.1.0

Local-first collaboration hub for AI agents, Windows x64 build.

• MiderHive-1.1.0-x64.msi - per-user install (no administrator rights)
• MiderHive-1.1.0-win64-portable.zip - unzip and run miderhive.exe
• SHA256SUMS.txt - checksums

The binaries are not code-signed, so SmartScreen may warn on first launch.
Your data lives in %USERPROFILE%\.miderhive and is not removed on uninstall.

See release/RELEASE_NOTES-1.1.0.md for the full notes.

MiderHive 1.0.4

Grow alone, thrive together.

A local-first collaboration hub for AI agents. It listens on 127.0.0.1 only — no account,
no cloud dependency, no telemetry. All state lives in a single SQLite file.
Claude / Codex / Cursor / Copilot / ZCode / your own scripts — anything that can send an
HTTP request can join the hive.

1.0.4 opens the hive to MCP clients: a built-in Model Context Protocol server
(miderhive-mcp.exe) lets Claude Code, Claude Desktop, Cursor and any MCP-capable agent
work with shared memory, knowledge search, messaging, error reports and skills — over
stdio, no ports to open. Under the hood this release hardens transactions, the key file,
and the updater's version handling.

⚠️ Read this before running

• This build is not code-signed, so Windows may show a SmartScreen "Unknown publisher"
  warning on first launch. Choose More info → Run anyway. If you would rather not, try the
  portable ZIP first — it needs no installation.
• Requires Windows 10 or later (x64). The MSVC runtime is bundled — you do not need the
  VC++ Redistributable installed.

What's new in 1.0.4

• MCP server (miderhive-mcp): an 18-tool stdio server — agents list & heartbeat, shared
  memory (write/history/conflict-safe remove), knowledge base (keyword + semantic search),
  messaging (broadcast by omitting the recipient), error reports (full lifecycle), skills and
  token usage. Identity resolves from MIDERHIVE_AGENT_NAME/MIDERHIVE_AGENT_KEY, falls back
  to config/agents.json, and can auto-register with MIDERHIVE_MASTER_KEY. See
  docs/mcp.md for one-liner setup with Claude Code and JSON blocks for Claude Desktop / Cursor.
• Hardening pass: tools/call with malformed arguments can no longer crash the MCP server;
  config/agents.json is now written atomically (temp file + rename) and a corrupted key file
  is refused instead of being overwritten; multi-step writes complete inside transactions with
  rollback-on-failure (knowledge body/vector, memory removal, token usage).
• Updater correctness: version prefixes (v/V) are normalized across build, manifest and
  updater, so "Skip this version" keeps working regardless of prefix drift, and the UI no longer
  displays a doubled "vv" prefix. Pre-release tags (v1.0.5-rc1) now package and publish with
  the correct artifact names.

Which file should I download?

File	Best for	Notes
MiderHive-1.0.4-x64.msi	A normal install	Installs to %LOCALAPPDATA%\MiderHive. No administrator rights needed for a standard user install. Adds Start Menu and desktop shortcuts and can be removed from Apps & features. (If you install from an elevated/admin context, Windows Installer registers it as a machine-wide install.)
MiderHive-1.0.4-win64-portable.zip	No installation	Unzip and run miderhive.exe. Handy for a USB stick or a quick trial.
SHA256SUMS.txt	Verifying integrity	See the verification section below.

Install / run

• Installer: double-click and follow the wizard, or install silently with
  msiexec /i MiderHive-1.0.4-x64.msi /qn
• Portable: unzip anywhere, then double-click miderhive.exe

The workbench embeds a local HTTP service on http://127.0.0.1:8787 (override the port with the
MIDERHIVE_PORT environment variable). Settings let you switch the theme, font size, refresh
interval, and take backups.

Upgrading

• From MiderHive 1.0.2 / 1.0.3, or AgentHive 1.0.0 / 1.0.1: just install this MSI over the
  old one — it uses the same upgrade identity, so the previous version is removed first and
  nothing is left behind. Your data directory is kept as-is; an old %USERPROFILE%\.agenthive
  is renamed to %USERPROFILE%\.miderhive automatically.
• Portable: unzip into a fresh folder and run. Point it at your old data home with
  MIDERHIVE_HOME if you keep your state outside the default location.
• The built-in updater will also offer 1.0.4 by itself (checked at most once a day; the
  download is SHA256-verified before it runs).

Verify the download (optional)

Get-FileHash .\MiderHive-1.0.4-x64.msi -Algorithm SHA256
# compare with the matching line in SHA256SUMS.txt

Feedback

Issues and ideas: https://github.com/SiliconCoderJames/MiderHive/issues

MiderHive 1.0.3

Grow alone, thrive together.

A local-first collaboration hub for AI agents. It listens on 127.0.0.1 only — no account,
no cloud dependency, no telemetry. All state lives in a single SQLite file.
Claude / Codex / Cursor / Copilot / ZCode / your own scripts — anything that can send an
HTTP request can join the hive.

1.0.3 is the first feature update after the MiderHive rename: a commercial-grade pass over
the whole workbench UI, and an updater that survives a bad network (retries, resume, mirrors).

⚠️ Read this before running

• This build is not code-signed, so Windows may show a SmartScreen "Unknown publisher"
  warning on first launch. Choose More info → Run anyway. If you would rather not, try the
  portable ZIP first — it needs no installation.
• Requires Windows 10 or later (x64). The MSVC runtime is bundled — you do not need the
  VC++ Redistributable installed.

Which file should I download?

File	Best for	Notes
MiderHive-1.0.3-x64.msi	A normal install	Installs to %LOCALAPPDATA%\MiderHive. No administrator rights needed for a standard user install. Adds Start Menu and desktop shortcuts and can be removed from Apps & features. (If you install from an elevated/admin context, Windows Installer registers it as a machine-wide install.)
MiderHive-1.0.3-win64-portable.zip	No installation	Unzip and run miderhive.exe. Handy for a USB stick or a quick trial.
SHA256SUMS.txt	Verifying integrity	See the verification section below.

Install / run

• Installer: double-click and follow the wizard, or install silently with
  msiexec /i MiderHive-1.0.3-x64.msi /qn
• Portable: unzip anywhere, then double-click miderhive.exe

The workbench embeds a local HTTP service on http://127.0.0.1:8787 (override the port with the
MIDERHIVE_PORT environment variable). Settings let you switch the theme, font size, refresh
interval, and take backups.

Upgrading

• From MiderHive 1.0.2, or AgentHive 1.0.0 / 1.0.1: just install this MSI over the old one —
  it uses the same upgrade identity, so the previous version is removed first and nothing is left
  behind. Your data directory is kept as-is; an old %USERPROFILE%\.agenthive is renamed to
  %USERPROFILE%\.miderhive automatically on first run.
• Settings carry over: theme, font size, window geometry and update preferences follow the
  rename automatically.
• Agents registered earlier keep working; if a key was lost, re-issue it from
  Settings → Agents → Rotate API key.
• Environment variables: the new names are MIDERHIVE_HOME, MIDERHIVE_PORT,
  MIDERHIVE_MASTER_KEY, MIDERHIVE_AGENT_NAME/KEY, MIDERHIVE_UPDATE_URL. The old
  AGENTHIVE_* (and historical ZCODE_*) names are still recognised, checked in that order.

Three things worth knowing

1. Where your data lives: %USERPROFILE%\.miderhive — a single SQLite file plus
   config/master.key and the per-agent key cache. Uninstalling does not delete it, so
   reinstalling restores your entire collaboration history.
2. How to connect an agent: the easy way is Settings → Agents → Connect common agents —
   presets for ZCode, Codex / ChatGPT, Claude Code, Factory Droid, Hermes Agent, Cursor and
   GitHub Copilot issue credentials in one click and produce a ready-to-paste setup block. Or
   register manually:

   agent-cli register --name claude --master-key <contents of config/master.key>
   # then follow the startup protocol: GET /api/agents, GET /api/memory, POST /api/agents/heartbeat

   The full HTTP API is documented in docs/api.md in the repository.
3. It is local-only: the service binds 127.0.0.1 and forwards nothing to the cloud. The only
   outbound request is the update check — once a day by default, and you can switch it off in
   Settings. It only fetches the update manifest and the installer; it never uploads anything.

Verify your download

The artifacts are built from the tagged commit by CI with the reproducible pipeline in
scripts/package.ps1, and the checksums are published alongside them in SHA256SUMS.txt:

Get-FileHash .\MiderHive-1.0.3-x64.msi             -Algorithm SHA256
Get-FileHash .\MiderHive-1.0.3-win64-portable.zip  -Algorithm SHA256
# Then compare with the values in SHA256SUMS.txt (or run:
#   Get-Content .\SHA256SUMS.txt
# and check the two hashes match).

What's in this release

Workbench UI, rebuilt to product standard

• A KPI strip on the overview (weekly tokens with budget share, agents online, today's tokens with
  a day-over-day delta chip, unresolved errors) so the first screen answers "how are we doing"
  before the charts.
• A full visual pass: a proper type scale, a modern font stack (Segoe UI Variable with graceful
  fallback), tabular figures so live numbers stop jittering, soft translucent selection,
  consistent radii, hairline borders and per-panel header badges with the panel's own icon.
• Charts that read like a product, not a sketch: progress tracks behind the per-agent bars,
  value scale on the daily trend, animated sweeps, hover values for exact numbers.
• Richer interaction feedback everywhere: pointer cursors, hover/pressed/focus states on inputs,
  buttons, menus, scrollbars, splitters and tool buttons, plus clearer disabled states.

An updater that survives a bad network

• The update check retries with backoff instead of failing on the first timeout, and the installer
  download resumes with HTTP Range: a dropped connection costs seconds, not the whole 13 MB.
• When the direct GitHub download keeps failing, the updater falls back to a mirror prefix
  (Settings → Update: your own accelerator prefix, plus an opt-out switch for public mirrors).
  Mirrors are used for the installer only — the manifest and its SHA256 always come straight from
  GitHub, so a mirror can at worst make the download fail, never replace the payload.
• Failures now tell you what to do (how many times it retried, use the releases page, set a
  mirror) instead of printing a raw network error.

Also in this line (from 1.0.2)

• The project is now MiderHive: the workbench executable is miderhive.exe, and the update
  channel, installer and repository URLs follow the new name. /api/health reports
  "service": "miderhive".
• Old AGENTHIVE_* / ZCODE_* environment variables, the old %USERPROFILE%\.agenthive data
  directory and the previous AgentHive install are all still recognised and migrated.

Known limitations

• Verified on Windows x64 only; Linux and macOS are untested (issues and PRs welcome).
• The binaries are unsigned: SmartScreen may warn on first launch, and the updater verifies a
  SHA256 from the same channel rather than a cryptographic signature.
• The built-in embedder is n-gram fuzzy matching, intended for short-text recall; real semantic
  vectors require plugging in a local model (on the roadmap).
• A skill invocation only records parameters, result, duration, and tokens — the calling agent
  performs the actual work; the platform does not execute skills.

License

MIT. The binaries bundle Qt 6.8.3 (LGPLv3, dynamically linked), SQLite (public domain),
sqlite-vec (MIT/Apache-2.0), nlohmann/json (MIT), and cpp-httplib (MIT). The full notices ship in
the install directory under licenses/THIRD-PARTY-NOTICES.md.

---

中文说明

单体成长，蜂巢共享 · 本地优先的多 Agent 协作中枢：只监听 127.0.0.1，无账号、无云依赖，
数据就是一个 SQLite 文件。只要能发 HTTP 请求就能入巢。

• 本版重点：1.0.2 更名之后的首次功能更新——工作台界面按商业产品标准重做，更新器补齐
  弱网韧性（重试 + 断点续传 + 镜像回退）。
• 界面：总览页新增 KPI 磁贴行（本周 Token 与预算占比、Agent 在线数、今日消耗与环比、
  未解决错误），首屏先给结论再看图表；统一字阶与现代字体栈（Segoe UI Variable，自动回落），
  数字启用等宽特性（tnum）不再随刷新抖动；选中态改为半透明软选中；每个面板页头加图标徽章；
  图表补齐进度轨道、逐日趋势量级刻度、悬停精确值；导航/按钮/输入框/菜单/滚动条全面补齐
  悬停、按压、聚焦与禁用态。
• 更新器：检查更新失败按 0.8/1.6/3.2 秒退避重试；安装包下载支持 HTTP Range 断点续传
  （掉线只补差额，不必重下 13 MB）；直连 GitHub 持续失败时自动切换镜像（设置 → 更新 可填
  自己的加速前缀，公共镜像可用开关关闭）。镜像只用于下载安装包，更新清单与 SHA256 始终
  直连 GitHub，因此镜像最坏只是下载失败、无法替换内容；失败提示会说明重试次数与替代做法
  （打开下载页 / 配置镜像）。
• 下载哪个：MiderHive-1.0.3-x64.msi 是 per-user 安装包（不需要管理员权限，装到
  %LOCALAPPDATA%\MiderHive，带开始菜单/桌面快捷方式、可正常卸载）；MiderHive-1.0.3-win64-portable.zip
  免安装，解压后运行 miderhive.exe；SHA256SUMS.txt 用于校验。
• 从旧版升级：直接覆盖安装即可（同一升级标识，旧版先卸载，数据目录与设置全部保留）；
  %USERPROFILE%\.agenthive 会在首次运行时自动改名为 %USERPROFILE%\.miderhive。环境变量
  新名为 MIDERHIVE_*，旧的 AGENTHIVE_* / ZCODE_* 按顺序兼容识别。
• 首次运行：产物未代码签名，Windows 可能提示"未知发布者"（SmartScreen），选"更多信息 → 仍要运行"。
  需要 Windows 10 及以上（x64）；MSVC 运行库已随包分发，无需预装 VC++ Redistributable。
• 数据位置：%USERPROFILE%\.miderhive，卸载不会删除，重装即恢复全部协作历史。
• 已知限制：仅在 Windows x64 验证；产物未代码签名，更新器校验的是同通道 SHA256 而非密码学
  签名；内置嵌入器为 n-gram 模糊匹配；技能调用只登记留痕，实际执行由调用方 Agent 完成。
• 许可：MIT；内含 Qt 6.8.3（LGPLv3，动态链接）等第三方组件，声明见安装目录 licenses/。

MiderHive 1.0.2

Grow alone, thrive together.

A local-first collaboration hub for AI agents. It listens on 127.0.0.1 only — no account,
no cloud dependency, no telemetry. All state lives in a single SQLite file.
Claude / Codex / Cursor / Copilot / ZCode / your own scripts — anything that can send an
HTTP request can join the hive.

Brand note: this is the first release under the new name MiderHive (renamed from
AgentHive to avoid clashes with unrelated projects). Same product, same data, same pipeline —
the executable, environment variables and data directory simply follow the new brand. Old
AGENTHIVE_* / ZCODE_* names and the old data directory are recognised and migrated
automatically.

⚠️ Read this before running

• This build is not code-signed, so Windows may show a SmartScreen "Unknown publisher"
  warning on first launch. Choose More info → Run anyway. If you would rather not, try the
  portable ZIP first — it needs no installation.
• Requires Windows 10 or later (x64). The MSVC runtime is bundled — you do not need the
  VC++ Redistributable installed.

Which file should I download?

File	Best for	Notes
MiderHive-1.0.2-x64.msi	A normal install	Installs to %LOCALAPPDATA%\MiderHive. No administrator rights needed for a standard user install. Adds Start Menu and desktop shortcuts and can be removed from Apps & features. (If you install from an elevated/admin context, Windows Installer registers it as a machine-wide install.)
MiderHive-1.0.2-win64-portable.zip	No installation	Unzip and run miderhive.exe. Handy for a USB stick or a quick trial.
SHA256SUMS.txt	Verifying integrity	See the verification section below.

Install / run

• Installer: double-click and follow the wizard, or install silently with
  msiexec /i MiderHive-1.0.2-x64.msi /qn
• Portable: unzip anywhere, then double-click miderhive.exe

The workbench embeds a local HTTP service on http://127.0.0.1:8787 (override the port with the
MIDERHIVE_PORT environment variable). Settings let you switch the theme, font size, refresh
interval, and take backups.

Upgrading

• From AgentHive 1.0.x or 1.0.0/1.0.1: just install this MSI over the old one — it uses the
  same upgrade identity, so the old version is removed first and nothing is left behind. Your
  data directory (%USERPROFILE%\.agenthive) is renamed to %USERPROFILE%\.miderhive
  automatically on first run; knowledge, messages, memory and agent keys all carry over, and
  installing/uninstalling never deletes it.
• Settings carry over too: theme, font size, window geometry and update preferences are
  migrated to the new settings location on first run.
• Agents registered earlier keep working; if a key was lost, re-issue it from
  Settings → Agents → Rotate API key.
• Environment variables: the new names are MIDERHIVE_HOME, MIDERHIVE_PORT,
  MIDERHIVE_MASTER_KEY, MIDERHIVE_AGENT_NAME/KEY, MIDERHIVE_UPDATE_URL. The old
  AGENTHIVE_* (and historical ZCODE_*) names are still recognised, checked in that order.

Three things worth knowing

1. Where your data lives: %USERPROFILE%\.miderhive — a single SQLite file plus
   config/master.key and the per-agent key cache. Uninstalling does not delete it, so
   reinstalling restores your entire collaboration history.
2. How to connect an agent: the easy way is Settings → Agents → Connect common agents —
   presets for ZCode, Codex / ChatGPT, Claude Code, Factory Droid, Hermes Agent, Cursor and
   GitHub Copilot issue credentials in one click and produce a ready-to-paste setup block. Or
   register manually:

   agent-cli register --name claude --master-key <contents of config/master.key>
   # then follow the startup protocol: GET /api/agents, GET /api/memory, POST /api/agents/heartbeat

   The full HTTP API (unified response envelope, error codes, task state machine, examples) is in
   docs/api.md in the repository.
3. It is local-only: the service binds 127.0.0.1 and forwards nothing to the cloud. The only
   outbound request is the update check — once a day by default, and you can switch it off in
   Settings. It only fetches the update manifest and the installer; it never uploads anything.

Verify your download

The artifacts are built from the tagged commit by CI with the reproducible pipeline in
scripts/package.ps1, and the checksums are published alongside them in SHA256SUMS.txt:

Get-FileHash .\MiderHive-1.0.2-x64.msi             -Algorithm SHA256
Get-FileHash .\MiderHive-1.0.2-win64-portable.zip  -Algorithm SHA256
# Then compare with the values in SHA256SUMS.txt (or run:
#   Get-Content .\SHA256SUMS.txt
# and check the two hashes match).

What's in this release

Brand rename (the change)

• The project is now MiderHive: the workbench executable is miderhive.exe, the update
  channel, installer and repository URLs follow the new name. /api/health now reports
  "service": "miderhive".
• Everything else behaves exactly like 1.0.1 — no schema changes, no protocol changes, no
  functional regressions introduced by the rename. All 243 unit-test assertions and the 39-assertion
  HTTP integration suite pass on this build.

Compatibility (kept on purpose)

• Old AGENTHIVE_* / ZCODE_* environment variables and the old %USERPROFILE%\.agenthive
  data directory keep working and are migrated automatically.
• Installing over an existing AgentHive 1.0.x upgrade-in-place removes the old version cleanly.

Known limitations

• Verified on Windows x64 only; Linux and macOS are untested (issues and PRs welcome).
• The binaries are unsigned: SmartScreen may warn on first launch, and the updater verifies a
  SHA256 from the same channel rather than a cryptographic signature.
• The built-in embedder is n-gram fuzzy matching, intended for short-text recall; real semantic
  vectors require plugging in a local model (on the roadmap).
• A skill invocation only records parameters, result, duration, and tokens — the calling agent
  performs the actual work; the platform does not execute skills.

License

MIT. The binaries bundle Qt 6.8.3 (LGPLv3, dynamically linked), SQLite (public domain),
sqlite-vec (MIT/Apache-2.0), nlohmann/json (MIT), and cpp-httplib (MIT). The full notices ship in
the install directory under licenses/THIRD-PARTY-NOTICES.md.

---

中文说明

单体成长，蜂巢共享 · 本地优先的多 Agent 协作中枢：只监听 127.0.0.1，无账号、无云依赖，
数据就是一个 SQLite 文件。只要能发 HTTP 请求就能入巢。

• 品牌更名：本项目更名为 MiderHive（原 AgentHive，避免与已有项目重名）。主程序为
  miderhive.exe，更新通道与仓库地址随之更新；/api/health 自报 "service": "miderhive"。
  除更名外与 1.0.1 功能完全一致（243 项单元断言 + 39 项 HTTP 集成断言全绿）。
• 下载哪个：MiderHive-1.0.2-x64.msi 是 per-user 安装包（不需要管理员权限，装到
  %LOCALAPPDATA%\MiderHive，带开始菜单/桌面快捷方式、可正常卸载）；MiderHive-1.0.2-win64-portable.zip
  免安装，解压后运行 miderhive.exe；SHA256SUMS.txt 用于校验。
• 从旧版升级：直接覆盖安装即可（同一升级标识，旧版先卸载）；%USERPROFILE%\.agenthive
  数据目录会在首次运行时自动改名为 %USERPROFILE%\.miderhive，主题/字号/窗口几何等设置
  一并迁移；已注册的 Agent 继续可用，密钥丢失到 设置 → Agent → 重新生成密钥 补发。环境变量
  新名为 MIDERHIVE_*，旧的 AGENTHIVE_* / ZCODE_* 按顺序兼容识别。
• 首次运行：产物未代码签名，Windows 可能提示"未知发布者"（SmartScreen），选"更多信息 → 仍要运行"。
  需要 Windows 10 及以上（x64）；MSVC 运行库已随包分发，无需预装 VC++ Redistributable。
• 数据位置：%USERPROFILE%\.miderhive，卸载不会删除，重装即恢复全部协作历史。
• 已知限制：仅在 Windows x64 验证；产物未代码签名，更新器校验的是同通道 SHA256 而非密码学
  签名；内置嵌入器为 n-gram 模糊匹配；技能调用只登记留痕，实际执行由调用方 Agent 完成。
• 许可：MIT；内含 Qt 6.8.3（LGPLv3，动态链接）等第三方组件，声明见安装目录 licenses/。