π The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
Changelog
All notable changes to the Ruflo project (formerly Claude Flow) are documented here.
The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.
[3.65.0] - 2026-10-30
Minor release: @claude-flow/cli, claude-flow, ruflo 3.65.0, plus ruflo-mods plugin 0.12.0. This release ships the eight follow-ups to #4581 on how ruflo integrates with Claude Code mods (ADR-503 to ADR-510, #4585).
Added
Learning-loop gaming defences (ADR-503, #4590).
- A verified-success turn is tainted when it edited tests (
edited-tests), edited a verifier's config (edited-verifier-config), or ran a narrowed or weakened check (narrowed-verifier). - Tainted records never train unless
.claude-flow/config.jsonlearningLoop.allowTaintlists their classes. - An opt-in rotating re-verify step (
learningLoop.reverify, off by default) re-runs an allowlisted check in a temporary worktree. - An anomaly gate holds back lessons whose evidence looks gamed.
scripts/eval-learning-loop.mjsis a manual, paid evaluation harness.
- A verified-success turn is tainted when it edited tests (
Real-binary test tier (ADR-504, #4589). This runs the real
claudebinary with ruflo-mods against a scripted mock of the Anthropic Messages API, with no API key needed. The contracts check that:- every mod loads;
- the learning loop records outcomes;
- lessons reach the model;
- the guard denies risky commands;
- route context is added.
It runs as the CI job
mods-real-binaryand asnpm run test:real-binary.compactSnapshot(ruflo-mods option, default on; ADR-505, #4591).- Before compaction, the mod saves a screened transcript to rotating slots under
.claude-flow/sessions/compactions/. - After compaction, it adds one bounded
[RUFLO_AFTER_COMPACT]block with recent lessons and outcomes, open swarm and claim ids, and the snapshot path. ruflo doctor --component compact-snapshotswarns when snapshots exist but.claude-flow/is not gitignored.
- Before compaction, the mod saves a screened transcript to rotating slots under
Plugin secrets via
userConfigsensitive: true(ADR-506, #4587).- The cost-tracker, metaharness and x-gateway plugins read their keys through a shared
readSecrethelper. It checks the masked plugin option first and falls back to the environment. - A new
scripts/audit-plugin-secrets.mjsCI gate fails on secret-named options that are not marked sensitive, and on literal secrets in manifests.
- The cost-tracker, metaharness and x-gateway plugins read their keys through a shared
Plugin footprint cut (ADR-507, #4588).
- Operational skills are user-invoked only (
disable-model-invocation: true). - Long skill and command descriptions are trimmed to one sentence, with a ceiling enforced by an audit.
scripts/measure-plugin-footprint.mjsreports the result, with before and after numbers inv3/docs/validation/.
- Operational skills are user-invoked only (
Cost budgets that can feed a gateway (ADR-508, #4592).
.claude-flow/cost/budget.jsonholds the budget windows.ruflo cost budget setwrites it.ruflo cost budget export --gateway litellm [--apply]produces LiteLLMbudget_limitsand per-model budget fallbacks.--applyrequires HTTPS except on loopback, and reads the key from an environment variable only.ruflo cost reportbreaks spend down by agent and by tool.- New ruflo-mods option
costLedger(default on). --gateway meta-llmrefuses until meta-llm documents a budget API.
Plugin trust (ADR-509, #4593).
- Plugins carry Ed25519
signature.jsonmanifests, signed with the existing helpers key. ruflo plugins verifychecks a plugin against its manifest and reports verified, modified, unsigned or bad-signature.- A static scan runs before
mods install,initorplugins install. A high finding stops the install unless you pass--accept-risk <plugin>. ruflo initstops on risky project auto-approve keys (enableAllProjectMcpServers,bypassPermissions, wildcard Bash and others).--keep-risky-settingsand--drop-risky-settingslet you continue.- Under
modTrust: refuse-risky, a modified ruflo plugin is refused.
- Plugins carry Ed25519
Headless plugin staging (ADR-510, #4586).
resolvePluginPathsfinds ruflo plugins in a local checkout, the marketplace clone or the install cache.- Headless
claude -pworkers now get--plugin-dirfor ruflo-mods automatically. ruflo plugins paths [--json]prints the resolved folders for Agent SDK{ type: "local", path }callers.ruflo doctor --component headless-pluginschecks this setup.
Not yet
- Per-skill trimming (
skillTrim) was not built: Claude Code exposes no hook that hides a skill from the model's list. - The 30-task learning-loop evaluation from #4581 Β§7 has still not been run. The loop stays on by default, as the maintainers decided in ADR-502 Β§6.
[3.64.0] - 2026-10-29
Minor release: @claude-flow/cli, claude-flow, ruflo 3.64.0; ruflo-mods plugin 0.8.0.
Added
- Verified-outcome learning loop (#4581, #4583, ADR-502), on by default (ruflo-mods
learningLoop; set it tofalseto turn it off).- The mod records each main-loop turn. Its outcome is decided only by the exit status of a test, build, typecheck or lint command run after the turn's last edit.
|| true, pipes and;verify nothing, and the model's own words are never read.ruflo hooks verified-ingest, the MCP toolhooks_intelligence_verified-ingest, and automatic ingest inhooks_routeandhooks_session-endall train SONA on verified turns only. Unverified turns are counted and never trained.- Learned lessons at or above the 0.6 routing gate are written to
.claude-flow/data/learned-lessons.json. The mod cites at most three of them in a prompt they match. ruflo doctor --component learning-loopreports whether outcomes are reaching SONA.
Not yet
- #4581 stages 5β7 (console pane, autonomous
/loop-run, GEPA self-tuning) are not built. - The #4581 Β§7 evaluation (30+ tasks, baseline vs loop-enabled) has not been run; the default is on by maintainer decision ahead of it (ADR-502 Β§6).
[3.63.9] - 2026-10-29
Patch release: @claude-flow/cli, claude-flow, ruflo 3.63.9.
Security
github_repo_analyzeno longer returns credentials embedded in the origin remote. A remote such ashttps://x-access-token:<token>@github.com/...was echoed verbatim inremoteUrl, putting the token into tool output. URL userinfo is now stripped; scp-form remotes (git@host:o/r) are unchanged (#4577).
Fixed
migrate run --dry-runandmigrate fix --agents --dry-run(or-d) preview instead of migrating. Both read a flag key the parser never set, so a dry run overwrote.claude-flow/config.jsonand wrote restored agent files (#4568).github_repo_analyzereads repository names that contain a dot (vercel/next.js,socketio/socket.io) from the origin remote instead of filing them all underlocal/repo(#4571).hooks_worker-detectwithautoDispatchmarks its worker records and response assyntheticwith a note, instead of reporting a completed worker when no executor ran (#4576).
Changed
- The
agentic-qe-fleetmarketplace entry is pinned to Agentic QE v3.15.17 (#4574).
[3.63.8] - 2026-10-28
Patch release: @claude-flow/cli, claude-flow, ruflo 3.63.8, plus standalone @claude-flow/memory 3.1.0.
Fixed
@claude-flow/memory3.1.0 addsControllerRegistry.set(). The CLI's memory bridge already called it to register controllers it builds itself, but the method did not exist, so those registrations were silently dropped.set()closes any instance it replaces. The agentdb the CLI installs (3.0.0-alpha.20) does not ship the MutationGuard, AttestationLog or GuardedVectorBackend files, so those stay off. What does change: when agentdb's own ReasoningBank is unavailable, the intelligence module's local ReasoningBank is now really registered (#4549, #4550).agentdb_pattern-storeandagentdb_pattern-searchuse the same backend when that local ReasoningBank is registered. Before, the store wrote to the SQLpatternnamespace while search read the local bank, so a stored pattern could not be found once the bank held other patterns. Search also falls back to the SQL namespace when the bank has no match (#4567).
[3.63.7] - 2026-10-28
Patch release: @claude-flow/cli, claude-flow, ruflo 3.63.7.
Fixed
agent_spawnacceptstypeas an alias foragentType, so projects set up with older agent and command docs can spawn agents again.agentTypewins when both are sent. A call with neither now saysagentType is requiredinstead oftype: Required. 443agent_spawncalls in the shipped agent, command and skill docs now passagentType(#4542, #4543, #4564).coordination_synctriggerrefuses instead of waiting 50 ms and reporting a sync that never happened. The tool's description now says so (#4557, #4559, #4564).agent_poolfillis implemented instead of returning "Unknown action" (#4555, #4561).hooks_build-agentswrites under the project, validatesoutputDir, and no longer reports invented stats (#4558, #4562).performance_metricsreports recorded latency and throughput instead of fixed figures (#4554, #4556).- The daemon's local audit worker no longer reports a clean scan it did not run (#4563).
session_infoandsession_currentread the session records that hooks write (#4548).- The GitHub skills use registered MCP tools (#4141, #4541). A new check fails when a shipped skill runs a CLI command that does not exist (#4552, #4553).
[3.63.6] - 2026-10-27
Patch release: @claude-flow/cli, claude-flow, ruflo 3.63.6.
Fixed
- ruflo-console 0.49.7: Claude can clear Project Anatole alerts through the console again. Before,
console_run anatole-ack-<id>failed with "command.run: called from a tool.call hook, it would wait on the turn this hook is holding", so every alert stayed open. The console now sends/protectorwithout waiting inside Claude's tool call, and Claude's reply says the action is queued. The ack is applied when the turn ends (#4545). claims_rebalanceno longer overfills one agent (#4547).@claude-flow/aidefenceis declared as an optional dependency of the CLI (#4445).- The hooks-automation skill uses the
hookscommand and the real MCP tool names (#4538, #4540).
[3.63.5] - 2026-10-27
Patch release: @claude-flow/cli, claude-flow, ruflo 3.63.5, plus standalone @claude-flow/neural 3.0.1.
Security
http_fetchnow validates every redirect target, not only the first URL. A public URL can no longer redirect to loopback, private ranges, cloud metadata orfile://. Redirects are capped at 10 hops, andAuthorization/Cookieheaders are dropped on cross-origin hops (#4534).
Fixed
- SONA EWC++:
SONAManagernow records Fisher importance and anchor values after each learning cycle; before, they were always empty. The anchor is the parameter value at consolidation, so the penalty starts at zero (#4531, #4532,@claude-flow/neural3.0.1). The penalty is still not applied to the learning update. task_liststatus filter matches status aliases (#4536).- Workflow condition steps can read an earlier step's output via
step-N.output(#4530). task_summarycounts cancelled tasks, so the counts add up to the total (#4489).hooks_worker-statussummary counts pending workers (#4529).doctorrecognizes matching npm launchers on Windows (#4473).- ruflo-console 0.49.5 keeps a project copy of the mission ledger, so a mission survives a restart that loads the console from a different source (#4524).
[3.63.4] - 2026-10-26
Patch release: @claude-flow/cli, claude-flow, ruflo 3.63.4, plus standalone @claude-flow/mcp 3.1.4, @claude-flow/plugin-agent-federation 1.0.3 and @claude-flow/shared 3.0.6. Follow-ups to the 3.63.3 security fixes.
Security
mcp start -t http|websocketwith no token now uses a per-user token in~/.claude-flow/mcp-http-token(mode 0600, created on first use), so a loopback server is no longer open to every local process. SetRUFLO_MCP_ALLOW_UNAUTHENTICATED_HTTP=1to run without one. The standalone WebSocket transport now checks the token when the connection upgrades (Authorization: Beareror?token=), and--auth-tokenworks with it (#4516). Breaking: HTTP/WebSocket clients must send the token.- Federation:
VERIFIEDandATTESTEDtrust now require the key pinned for the peer in thetrustedPeerKeysconfig (node id to hex public key); a self-signed manifest or a handshake response only proves possession of the key it carries. A handshake response must come from the node that was challenged, the listener defaults to127.0.0.1, and nonces use crypto randomness (#4514). Breaking: peers that relied on self-asserted trust need atrustedPeerKeysentry. - A file in
.claude-flow/daemon-queue/no longer counts as a manual trigger, so it cannot skip the supervisor gate (#4515). @claude-flow/shared3.0.6 carries the same Origin/Host guard and WebSocket token check as@claude-flow/mcp(#4517).- AgentBBS: a
..in a room label can no longer move the room log outside its base path (#4518).
[3.63.3] - 2026-10-26
Patch release: @claude-flow/cli, claude-flow, ruflo 3.63.3, plus standalone @claude-flow/mcp 3.1.3 and @claude-flow/plugin-agent-federation 1.0.2. Security fixes from a private report.
Security
- A cloned project's config can no longer turn on AI workers:
daemon.aiWorkers.enabledis honoured only from the--headlessflag,RUFLO_DAEMON_AI_WORKERS=1or~/.claude-flow/config.json; a project file can only sayfalse. The audit worker no longer globs**/.env*(#4509). If you relied on a project file to enable AI workers, move the setting to your user config. @claude-flow/mcp3.1.3: HTTP and WebSocket transports with no token now refuse requests and upgrades from a web page (anyOriginnot listed by exact name;*does not count) or with a non-loopbackHost, and refuse to listen beyond loopback without a token unlessallowUnauthenticatedis set (#4510). A local process can still call a token-less loopback server; set a token.@claude-flow/plugin-agent-federation1.0.2: a policy denial oftask-assignment,memory-query,context-share,trust-change,topology-change,agent-spawn,claim-eventoragent-handoffis final inlegacyandobservemode too, and an evaluator error refuses them (#4511). Still open: the listener defaults to0.0.0.0, and discovery manifests are verified against their own key.
[3.63.2] - 2026-10-26
Patch release: @claude-flow/cli, claude-flow, ruflo 3.63.2. No leaf package published.
Security
http_fetchno longer lets IPv6 literals past its SSRF guard: brackets are stripped before matching, so thefc00::/7andfe80::/10checks now work, and IPv4-mapped forms such as::ffff:127.0.0.1and::ffff:169.254.169.254are decoded and checked against the IPv4 rules (#4503, #4504). Still open:::, multicast and NAT64 ranges, and redirect re-validation.
Fixed
- ruflo-console 0.49.4: the mission plan is submitted from a file, which fixes "command line is too long" on Windows (#4506).
- CI gates helper sync on import and manifest checks (#4507).
[3.63.1] - 2026-10-25
Patch release: @claude-flow/cli, claude-flow, ruflo 3.63.1. No leaf package published (the swarm fixes below live in @claude-flow/swarm, which the CLI does not depend on).
Fixed
- ruflo-console 0.49.3: the install record for a plugin installed more than once no longer comes from another project; with none applying the console says "not installed here", and the record with the longest
projectPathwins (#4496, #4497). Also in this train: Windows path and autopilot guard fixes (0.49.1/0.49.2, #4490, #4491, #4494). flywheel runskips the directory fsync on Windows so it can commit (#4492, #4493).federation_bbs_watchpages forward fromsinceEnvelopeId(#4480).- Managed agents refuse
.and..resource ids and the guard reads common argument wrappers (ruflo-agent0.4.1, #4478). ruflo-mods0.7.1 loads again (#4485, #4487).- Swarm (source only, unpublished): downgraded topology leaders wire only to the leader, consensus waits on events instead of polling, retries release the old domain slot (#4483, #4498, #4499).
[3.63.0] - 2026-10-25
Minor release: @claude-flow/cli, claude-flow, ruflo 3.63.0. Adds four MCP tools for Anthropic Claude Managed Agents (managed_agent_capabilities, managed_agent_control, managed_agent_learning, managed_agent_interrupt). No leaf package changed, so no standalone publishes.
Added
- Managed cloud agents: dynamic workflows, controls and verified advisory learning (
ruflo-agent0.4.0,ruflo-workflows0.7.0,ruflo-console0.49.0; ADR 0003 and 0004). Everything is opt-in. A cloud agent can run bounded dynamic workflows (dynamicWorkflows: truewith a mandatorybudgetCents, limited networking, optionalpredefinedAgents); provider completion is reported asverification: "unverified". Managed sessions show up in the console's Workflows page with their budget, blockers and worker counts, with zero provider traffic from the panels.managed_agent_controlinspects, interrupts or terminates one session only when the host setsRUFLO_MANAGED_CONTROL=1and lists the session inRUFLO_MANAGED_CONTROL_SESSIONS, after a fresh read of the provider's state. WithRUFLO_MANAGED_LEARNING=1and a host-configuredRUFLO_MANAGED_VERIFIER_PUBLIC_KEY,managed_agent_learningadmits signed outcomes (unsigned ones stay quarantined), recommends a strategy and evaluates paired holdouts; nothing changes routing on its own. - User guide: a "Managed cloud agents" section in
docs/USERGUIDE.md.
Fixed
memory_search_unifieddropped an entry whenever the same key existed in two namespaces; it now deduplicates on namespace and key (#4470, thanks @kevin9327).- Before this release shipped, an independent review of the managed-agent change found and fixed:
managed_agent_interrupthad no host policy gate (it now uses themanaged_agent_controlpath, andmanaged_agent_terminateobeys the same allowlist, deleting only the session's own environment, onceRUFLO_MANAGED_CONTROL=1);managed_agent_promptcould report the previous turn's answer as finished on a later poll; several valid base64 spellings of one verifier signature were stored as separate ledger records; and the console showed rescheduling, budget-stopped and action-blocked sessions as finished.
Notes
- Known limits (ADR 0004): the verifier signs with a key in the same user account as the checks it runs, so run only host-authored checks, never ones an agent produced; evidence is only partly bound to what ran; a crashed process can leave a lock file that needs removing by hand; a session in an
errorstate cannot be controlled. Live execution against the Anthropic Managed Agents beta has not been exercised (it needs an executor key), so treat the cloud runtime as unproven in production. ruflo-ai-team0.3.5 (#4475) also changed in this tree.RUFLO_MANAGED_AGENTS_BASE_URLis an operator-only override of the API origin (an HTTPS origin, no credentials, path or redirects); it is never a tool argument.
[3.62.0] - 2026-10-24
Minor release: @claude-flow/cli, claude-flow, ruflo 3.62.0. Adds one MCP tool (rure_cuda_launch_plan). No leaf package changed, so no standalone publishes.
Added
- ruRE: CUDA launch plan (ruflo-rure 0.6.0, ADR-501). New read-only MCP tool
rure_cuda_launch_planturns whatrure_analyze_cudafound into a plan for measuring: per architecture and kernel, how many blocks and warps fit on one multiprocessor at each block size, which resource sets the limit (registers, shared memory, warps, block slots, barriers), the best block sizes and flags. Eight architectures have a table (sm_70tosm_120); any other gets no figure rather than a guess. It still parses and never runs:includeHarnessonly returns the text of a small CUDA program for you to build and run yourself. The occupancy model matches NVIDIA'scuda_occupancy.hon 1.77 million combinations and the CUDA driver on five real GPUs (ADR-501, ADR-500).
Fixed
rure_analyze_cudareported the wrong architecture (sm_5) for a standalone.cubinbuilt with CUDA 12 forsm_70tosm_90; the SM number sits in a different byte depending on the cubin's OS ABI. Cubins inside a fatbin or host binary were not affected.- The worker daemon let a
maxConcurrentorworkerTimeoutMssaved indaemon-state.jsonoverride the same setting inconfig.json; an explicit value now wins (constructor, then config file, then saved state, then default) (#4462). - Cancelling a headless worker now settles its execution and escalates to SIGKILL when the child ignores SIGTERM; the fallback timer no longer keeps the host alive, and a timed-out child no longer gets a second SIGTERM (#4463).
Notes
- Plugins in this tree that changed:
ruflo-rure0.6.0,ruflo-ai-team0.3.4,ruflo-browser0.3.7 (doctor path from other working directories),ruflo-x-gateway0.10.2. The x.ruv.io gateway itself (0.9.1) is a separate deployment. @claude-flow/swarmchanges (#4456, #4457) are not a CLI dependency and are not part of this release.
[3.61.1] - 2026-10-24
Patch release: @claude-flow/cli, claude-flow, ruflo 3.61.1. No leaf package changed, so no standalone publishes.
Added
ruflo-console0.48.0 (ADR-499): the ruRE page can read a folder outside the project that you allow. New plugin optionrureFolders(name=/absolute/folderpairs); a sample then names it as@name/sub/file. Absolute paths,..and odd characters in a sample are still refused; the folder and every part below it must not be a link; and changing the option is an install-class write that always asks you first (Claude's console tools can only propose it at the full control level). ruRE still only parses. This is the version thatruflo tuiships, which is why it needs a CLI release.
Notes
- CI: plugin pull requests now fail unless the plugin's
plugin.jsonversion rises when shipped files change (#4447). @claude-flow/plugins(#4451, registry caches no longer re-run a surviving plugin'sregister*methods) is not a CLI dependency and is not part of this release.
[3.61.0] - 2026-10-23
Minor release: @claude-flow/cli, claude-flow, ruflo 3.61.0. Adds one MCP tool (rure_analyze_cuda). No leaf package changed, so no standalone publishes.
Added
- ruRE: CUDA static analysis (ruflo-rure 0.5.0, plugin ADR-0006). New read-only MCP tool
rure_analyze_cudareads the GPU code inside a CUDA fatbin, cubin or PTX file: fatbin entries (PTX or cubin, target architecture, LZ4/zstd compression), each kernel with register count, code size and constant/shared/local memory, and PTX version, target and kernel names. It parses only; nothing is executed, and strings taken from the file pass the AIDefence gate. Checked against real nvcc 12.8 and 13.0 output. Not included: SASS decoding,cuobjdumpshell-out. zstd entries need Node 22.15+. ruflo-console0.47.0: a CUDA row on the ruRE page, DETECT routes for.cubin,.fatbinand.ptx, and a hint on a native binary that holds a CUDA fatbin (ADR-495..497). The ruRE page refuses a sample that is a symlink, wraps long lists, and shows Claude a structuredrureblock inconsole_state.ruflo-mods0.7.0: opt-inrureOffer(off by default, ADR-498) adds one line under an answer when the model touched a binary, library,.asaror CUDA file, saying ruRE can inspect it without running it. Offer only; the mods plugin is not part of the CLI tarball.
Fixed
- Dependency audit:
zod4.6.7 inruflo-graph-intelligenceand@claude-flow/browser(GHSA-mrhx-c7xc-m3w5). Nested zod 3 copies stay under two time-boxed accepted findings that expire 2026-11-23 (#4442, #4444).
[3.60.1] - 2026-10-23
Patch release: @claude-flow/cli, claude-flow, ruflo 3.60.1. No leaf package changed, so no standalone publishes.
Fixed
ruflo cleanup --force --keep-configdeleted the config files it is meant to keep (claude-flow.config.json, thehooks/claudeFlowblocks of.claude/settings.json): the parser stores the flag askeepConfigand the action readkeep-config(#4435).ruflo-console0.45.1: ADR init, saves and exports work on Windows. The write layer shelled out todd,installandsh, which do not exist there; a newhost-fswrite flavor uses the host filesystem (#4433, #4434, carries #4429).ruflo-console: the autopilot envelope accepts Windows paths (C:\work\app, UNC) and compares them ignoring case and separator style; a POSIX-rooted path containing a backslash is still refused (#4436). Autopilot start stays gated off on Windows.
[3.60.0] - 2026-10-23
Minor release: @claude-flow/cli, claude-flow, ruflo 3.60.0, with @claude-flow/hooks 3.0.2 and @claude-flow/mcp 3.1.2 published standalone. Backward-compatible additions (six new reverse-engineering MCP tools, a ruRE page in the console) plus security and correctness fixes. The 3.59.x patch releases are not itemised here; this lists what changed since 3.59.17.
Added
- ruRE: static reverse engineering (ruflo-rure 0.4.2, ADR-0001..0005 of the plugin). Six read-only MCP tools β
rure_extract_asar,rure_analyze_js_app,rure_find_feature,rure_analyze_binary(ELF, PE, Mach-O),rure_analyze_dotnet(metadata only) andrure_analyze_evm(bytecode inventory) β that parse a file and never run it. Every string taken from the analyzed file passes the AIDefence gate; results sayscreened: falsewhen the scanner is unavailable. New dependency:@electron/asar. Known gaps: EVM and P/Invoke handling validated on built fixtures only; the.asarscratch folder is not cleaned up; the tools read any path the process can read. ruflo-console0.45.0: a ruRE page in the TOOLS group (ADR-494) with a row per tool, a project-relativesamplefield, and results cleaned of terminal escapes. The tools are needed in the CLI: on 3.59.17 and older the rows say to update.agentic-qe-fleetmarketplace pin moved to v3.15.12 (#4419).
Fixed
- Security:
agentbbsaccepted..in a room id, letting a room log path escape.agentbbs(#4416). - Security (
@claude-flow/mcp3.1.2): MCP session ids were a counter, a timestamp and sixMath.random()characters, so they were guessable; they now usecrypto.randomUUID()(#4418). neural_compresswithmethod: "prune"deleted every stored neural pattern and reported success, because pattern usage is never recorded; it now refuses and leaves the store untouched (#4428, #4423).github_pr_managemerge/close andgithub_issue_trackupdate/close reported success whenghfailed or no record matched, and PR actions ignoredowner/repo; they now fail withgh's error and pass--repo(#4425, #4424).daa_agent_adaptrecordedperformanceScore: 0, and an omitted score, as a 0.8 success; 0 is now a failure, an omitted score records none, and an out-of-range score is rejected (#4427, #4426).- The bridge search looked at only the newest 1000 rows, hiding older ones; the window is back to 10000 and configurable with
CLAUDE_FLOW_SEARCH_WINDOW(#4405, #4397). - The cost-tracker Stop hook stored its record in the session's project instead of
CLAUDE_FLOW_CWD(#4407), anddoctorcould not read the version of aclaude.cmdshim on Windows (#4408) (#4397). - The AgentDB embedder rescue applied only to the first store in a process (#4421).
transport use slim,agent publishandswarm joinnow report "not configured" when the installed AGNTCY SDK lacks the function they call, instead of failing at the call site (#4409, #4376).update allhonours--dry-run(#4398).WorkerDaemonkept two timers for a worker enabled twice (#4412). In@claude-flow/hooks3.0.2: an explicit stop is no longer undone by a pending auto-restart (#4413), a duplicate cluster is counted once (#4415), and fallback embeddings no longer block the event loop for up to 10 s (#4422).ruflo-ai-team'sannotation-review.mdmissed two tools (#4394); the federation peer's standalone TLS lock moved to patched rustls, webpki and quinn-proto (#4414).
[3.58.1] - 2026-10-20
Patch release: @claude-flow/cli, claude-flow, ruflo 3.58.1. Bug fixes only.
Fixed
- Security:
hooks_codemod'sresolveTargets()path-containment check used an unanchoredabs.startsWith(cwd)string-prefix comparison β a sibling directory merely starting with the project directory's name (e.g.cwd-staging) passed containment, letting this live, CLAUDE.md-recommended tool read and write outside the project root (CWE-22). Anchored both call sites at a path-separator boundary, matching the already-correctisWithinPrefix()pattern used elsewhere in this codebase (#4312/#4313). CommandParser.parse()flippedparsingFlags = falseon every literal--token it saw, not just the first β a later--supplied as an ordinary operand (e.g. a search query of exactly--) was silently swallowed instead of being kept as a positional (#4310/#4314).TokenOptimizer.optimizedEdit()calledagentBooster.editCode(), a method that has never existed onagentic-flow/agent-booster's real exports; a separate bug in the same file'sinitialize()picked the un-instantiatedAgentBoosterclass instead of calling the package'sgetAgentBooster()factory. Published standalone as@claude-flow/[email protected](#4311/#4315).- CI reliability:
scripts/__tests__/probe-mod-guards.test.mjs(anode:testfile) was never categorized into either the vitest-baseline or vitest-excluded list, causing an intermittent, unrelated "no test suite found" failure on PRs that don't touch it (#4316).
Added
ruflo-console0.44.6: the compact nav's menu, groups, search and page tabs are now separated by a dimβbar instead of just a gap (#3664).
[3.58.0] - 2026-10-20
Minor release: @claude-flow/cli, claude-flow, ruflo 3.58.0. Backward-compatible additions (new memory peek command, a Codex session activity viewer, configurable MCP rate limits) plus a critical install-breaking fix and a round of correctness fixes.
Fixed
- Critical:
@claude-flow/[email protected]pinned@claude-flow/[email protected], which had never been published, so every fresh install of the 3.57.x line failed withETARGET(#4303). Published@claude-flow/[email protected]standalone (verified against a clean install) and addedscripts/audit-leaf-versions-published.mjsβ a new, now-required CI check that queries the live npm registry for every non-bundled leaf pin so a PR that bumps a pin ahead of its own leaf's publish fails before merge, not after release (#4304). agent spawn --namewas accepted and echoed by the CLI but never persisted to theAgentRecord, soruflo-consolecould only label agents by type (#4300/#4301).agent list --allsent a literalstatus: 'all'thatagent_listdoesn't recognize, filtering out every agent;agent list --typewas accepted but never read (#4302).claims_claim'sexpiresAtwas declared but unsettable by any handler, and the claims-load utilization math duplicated the max-claims threshold as a second magic literal (#3180/#3927).- A stale docstring on
InMemoryClaimRepositoryclaimed SQLite-backed persistence that doesn't exist in this package (#3179/#3926). - The persistent embedding cache raced concurrent cold callers into redundant SQLite initializations, and its FNV-1a cache key could collide for distinct equal-length inputs, including the lone-surrogate case (#3685, #3681).
- The embedding model cache directory could be unwritable for non-root runtime users, permanently caching the mock fallback with no retry path (#3692/#3831).
ruflo-mods' risky-call/event classifier was missing several calls and hook events (agent.register,tool.call,session.append, read-side forgery vectors likefs.read/env.get, and more) that can rewrite or forge what other mods decide by β strictly widened, nothing un-flagged (#3802).ruflo-x-gatewaycould crash on a malformed relay frame instead of failing fast on the one relay connection affected (#3921).- The standalone terminal console (
npx ruflo tui) gave no on-screen indication that mouse support is off by default, silently dropped the0 menuhint from the footer below 80 columns, and dropped a mouse press+release pair when a terminal (tmux especially) coalesced them into one data chunk (#4305). - CI reliability, found while validating this release (#4307):
v3/@claude-flow/memory's FTS5 tests locatedsql-wasm.wasmvia asql.js/package.jsonsubpath resolve, which sql.js 1.14's tightenedexportsmap rejects even though the file exists β now resolvessql.js/dist/sql-wasm.wasmdirectly;plugins/ruflo-business-pods/scripts/pod-tick.test.mjs(anode:testfile) was miscategorized into the vitest-can-run-it baseline instead of the node:test-only exclusion list (#4308);plugins/ruflo-graph-intelligence's knowledge-graph adapter built its optional@claude-flow/cliimport specifier as a string literal that a bundler's static scan could eagerly (and wrongly) try to resolve during test collection, outside the method's own try/catch β now built from joined parts; the root test-suite ratchet's default vitest worker count intermittently crashed withERR_IPC_CHANNEL_CLOSEDpartway through the ~9500-test suite β capped at 2 workers as a stopgap pending real root-cause work;tests/all-plugins-smoke-workflow.test.mjs's hardcodedtimeout-minutes: 10expectation went stale when #3950 raised the real value to 15 for the new external-plugin smoke step.
Added
memory peekβ a read-only command that reports which store files the current directory resolves to, row counts, last-write time, and WAL state, without ever creating, initializing, or migrating anything (#3508/#3620).@claude-flow/codex activityβ inspect a Codex helper's assignment, recorded state, progress, and tool calls/results from an existing session, with symlink-safe, TOCTOU-resistant file access and best-effort credential masking in previews (#3930/#3962).mcp startnow supports configurable HTTP and per-session rate limits (#3694).
[3.57.2] - 2026-10-19
Patch release: @claude-flow/cli, claude-flow, ruflo 3.57.2. @claude-flow/cli-core 3.7.5 (#4295, #4296) published standalone and pinned exactly. A large community PR batch β mostly correctness fixes to error reporting, memory/graph boundary handling, and CLI diagnostics.
Fixed
agentdb_graph-query/agentdb_graph-pathfinderresolved the memory DB path ad hoc instead of through the sharedresolveDbPath()priority order, so an explicitCLAUDE_FLOW_DB_PATHbinding could silently read a different store than the one selected (#4291, #4298).backupMemoryDb()fell back to a raw file copy after any native online-backup failure β including ordinarySQLITE_BUSY/disk errors on an unencrypted source β and reportedbackedUp: trueregardless; the fallback is now gated on a confirmed-encrypted source with no active WAL, and the result records which method actually completed (#4287, #4297).JsonMemoryBackend.isExpired()used a falsy check onttlSeconds, so an explicitttl: 0never expired (cli-core3.7.5, #4295, #4296).RvfBackend's native@ruvector/rvfbranch could never succeed (wrongopen()call) and silently fell back to the pure-TS path; the dead native branch is removed rather than half-wired (#4278, #4279, #4280).ReasoningBank.consolidate()deleted pruned/deduplicated patterns from storage but never calledhnswIndex.removePoint(id), so the HNSW index filled with orphans untilstorePattern()started throwing permanently (#4270, #4271).TopologyManager's queen/coordinator failover never promoted a replacement after death despitefailoverEnabled: true, silently falling through to generic mesh-style ranking (#4285, #4286).memory export,memory import, andmemory_migratereported fabricated success text instead of the tool's real error or migrated count (#4204/#4207, #4194/#4209, and the memory-import tool failure fix).hooks worker dispatchprintedStatus: completedeven when no daemon ran anything (#4196/#4208).hooks_intelligence_attentionaccepted unimplemented modes instead of rejecting them (#4193).init upgradenow refreshesintelligence.cjs/session companions instead of leaving stale copies behind (#4198/#4220).- SmartRetrieval collapsed multi-result searches down to one (#4192); the ReasoningBank embedder used a hash fallback instead of the real embedder (#4191);
@claude-flow/cli's ownpackage.jsonwasn't exported for consumers resolving it directly (#4190). ruvocal's MCP server merge dropped a base server's credentials when the browser echoed the server back without headers (#4281).ruflo-x-gatewaynow advertisesoffline_accessso OAuth clients can refresh silently instead of re-authenticating on every access-token expiry (#4256).workflow list/status/stop/runnow read the actual shape and errors the workflow tools return (#4268); stale task owners are released on terminal states (#4272); unregistering a swarm agent now removes its domain-pool membership (#4269); orphaned atomic-write temp files under the policy lock are reaped (#3164/#4273); one-shot CLI runs now flush learned patterns beforeprocess.exit(0)instead of losing them whilestats.jsonkeeps counting (#4283/#4284).ruflo-console's mission ledger no longer leaks a mission from one project into every other project's cockpit (#4254/#4276).- Command examples in the user guide that failed when copied are fixed (#4275); a stray
npxinvocation referencing a nonexistent unscoped package is corrected to@claude-flow/codex(#4265). scripts/build-codex-mod-hooks.mjsnow rebuilds the codex-mod bundle the same way on every OS (#4292).
[3.57.1] - 2026-10-19
Patch release: @claude-flow/cli, claude-flow, ruflo 3.57.1. @claude-flow/cli-core 3.7.4 (#4229, #4230) was published earlier and is pinned exactly.
Security
DENYLISTED_ENV_NAMESnow blocks the credential/command-hijack and interpreter-flag-injection vars that 2026-09-21's env-sanitizer fix missed:GIT_ASKPASS,GIT_PROXY_COMMAND,GIT_CONFIG_NOSYSTEM,PERL5OPT,RUBYOPT,JAVA_TOOL_OPTIONS,_JAVA_OPTIONS,JDK_JAVA_OPTIONS,PYTHONHOME,CLASSPATH,PYTHONSTARTUP, plus git's numberedGIT_CONFIG_(COUNT|KEY_\d+|VALUE_\d+)family via a new regex (cli-core3.7.4, #4229, #4230).
Fixed
all-plugins-smoke: the adr/ddd/sparc codex-mod native hook bundles are rebuilt, clearing the native-bundle-drift failure that had been hitting every push tomain;scripts/build-codex-mod-hooks.mjsalso no longer hard-fails on Windows, whereexecFileSynccan't invoke a.cmdshim directly (#4266, #4277, @uyg7x).
[3.57.0] - 2026-10-17
Minor release: @claude-flow/cli, claude-flow, ruflo 3.57.0. A batch of community and dream-cycle fixes (#4210-#4252) plus a policy-ledger integrity fix (#3892). No standalone leaf republish needed this round.
Security
ruflo init/init --dual/@claude-flow/codex initno longer write a permissive.codex/config.tomlby default. The generated file now usessandbox_mode = "workspace-write",approval_policy = "on-request"andweb_search = "cached"(standard trust); the olddanger-full-access/never/livefile needs an explicit--codex-trust full(orRUFLO_CODEX_TRUST=full).initprints which level was written, an existing.codex/config.tomlis never rewritten (not even with--force), and a kept permissive file is flagged with how to tighten it. Found by the three-host e2e run (ADR-490).@claude-flow/codex3.0.4 (bundled, ships with the next train).- The MCP bridge on loopback no longer accepts tool-call requests from an arbitrary web page: local-only mode now requires the
Hostto be loopback (blocks DNS rebinding) and theOriginto be loopback-or-allowlisted when noMCP_AUTH_TOKENis set (#4226, #4243, @kevin9327). - The policy ledger now authenticates a retained/pruned ledger against its independent on-disk archive anchor instead of trusting the mutable
state.jsonalone, and streams full audits through fenced writers. Fixespolicy-ledger-verification-failedon any install whose ledger has already been pruned (#3892, #4216, @proffesor-for-testing).
Added
ruflo doctorreportsbetter-sqlite3copies compiled against Node 24.19+ headers when running Node 24.19β24.21, where they can abort the process (Assertion failed: (env) != nullptr, nodejs/node#65446), with the documented workarounds.doctor --component native-addonruns it alone (#4055, reported by @ronmikailov).ruflo doctorreports health checks as versioned JSON output (#4228, @bryanbans).
Fixed
@metaharness/darwinpin bumped from~0.10.2to~0.12.0, unblocking themetaharness-pin-driftCI gate repo-wide (#3219, #4252).- Swarm: subscriber delivery is cancelled cleanly at shutdown, acknowledgements are tracked per recipient, urgent messages survive queue overflow, and failed/cancelled task workers release their slot (#4233, #4234, #4235, #4236, @rudycelekli). Coordinator cancellation stays terminal (#4244, @rudycelekli); domain pools stop creating unregistered agents and release agents after terminal tasks (#4246, #4247, @Tong-bit-art); shared worker reservations are honored across agent pools, fixing a case where that bookkeeping change would have silently defeated the domain-pool-release fix above (#4245, @rudycelekli).
- Learning stats read-merge-write transactions are serialized across processes, fixing lost concurrent increments (#4217, #4240, @rudycelekli).
- Claude memories refresh across project and worktree roots (#4210, @rudycelekli); the secret screen no longer treats a hyphen-named memory key as a credential field (#4202, #4211, @akushonkamen).
- Hook lifecycle events are distinguished during deduplication, fixing the Stop hook claiming the session-end de-dup key so later hooks never ran (#4206, #4212, @dajiaohuang);
initomits unimplemented generated hooks (#4200, #4213, @dajiaohuang) and stops granting blanket execution of project helpers (#4221, @xiajiahang). - Intelligence: trajectories are forwarded using the ruvllm contract (#4222, @xiajiahang); neural quantization reports unchanged persisted storage instead of a false delta (#4224, @dajiaohuang).
- Console: Mission Control's prose wraps at 80 columns (#3899, #4225, console 0.44.3, @korjavin); the mission ledger is keyed per project instead of leaking one project's active mission into every other (#4237, #4242, @kevin9327).
daemon statusreports the running daemon's actual TTL instead of a stale--ttl 0default (#4238, #4241, @kevin9327).- CLI:
--no-coloris honored for help and version output, and--no-updateis honored before the startup update check runs (#4248, #4251, @shuaiyafei). dream(performance):RaftConsensus.awaitConsensus()is event-driven instead of busy-polling (#4214, #4215).- Docs: 25 user guide links repaired to v3 packages (#4223, @ketpatil77); ABP-governed SGit memory checkpoints and Ruflo CLI recovery validation documented (#4239, @mondweep).
Not included: #4220 (helper-execution-permission fix) touches a signed helper manifest and needs a re-sign with the private key before merge, held for a follow-up release. #4216's own large diff (ledger integrity) was reviewed in depth separately before merging, given it gates MCP tool-call authorization.
[Unreleased]
Fixed
@claude-flow/shared3.0.4: RVF event log no longer loses appends after a damaged tail (#4031, @rudycelekli). A partial length prefix, a partial payload or a record cut short is truncated on open, so later appends are readable; an unwritable log (EACCES/EROFS/EPERM) keeps its readable history and rejects appends explicitly. A tail whose length prefix is over the 100 MB limit is no longer accepted silently (found by @martinvlad): the bytes are kept andappend()rejects until repaired, or setrepairOversizedTail: trueto truncate it.append()also refuses a record the replay limit would never read back. A fresh install can now import the package root and its tarball ships no test files:zod,express,cors,helmet,wsand@types/wsare declared dependencies, Express 5 bodiless POST and malformed-JSON requests get JSON-RPC 400s (#4091/#4089, @proffesor-for-testing).
Batch of reviewed fixes from @rudycelekli (PRs #4020-#4026, #4028-#4030, #4032, #4033, #4035, #4037; #4027 and #4034 landed earlier):
@claude-flow/cli-core3.7.1: JSON memory backend owns its snapshots, rejects invalid entry containers, measures size in UTF-8 bytes, omits non-matching substring results and lets expired keys be reused; output formatter honours progress controls, empty totals, table padding, coloured header widths, and suppresses progress in quiet or redirected output.@claude-flow/shared3.0.3: RVF event log honours zero limits, rebuilds indexes on retried initialization and isolates persisted records from caller mutation; circuit breaker returns independent date snapshots and expires rolling requests on completion.ruflo-adr0.5.5 (see its changelog).- Not included: #4031 (RVF truncated-tail recovery) is waiting on the author for a read-only-log fallback.
[3.56.10] - 2026-10-14
Patch release: @claude-flow/cli, claude-flow, ruflo 3.56.10. @claude-flow/memory 3.0.5 (#4183) was published earlier and is pinned exactly. @claude-flow/cli-core 3.7.3 (#3987) was also published earlier and is pinned exactly.
Added
ruflo tuidraws the console's pictures, looks and motion (console 0.44.0, #4182). The bundledtui.mjsis rebuilt. The TUI draws at the default size when the terminal reports 0 columns (#4160, @martinvlad).
Changed
ruflo init: the v3 implementation skills are opt-in, likeagents.v3(#4140, #4139, @z4forzubair).
Fixed
- The MCP server starts in the session project,
CLAUDE_FLOW_CWDnaming the home directory or a filesystem root is ignored (cli-core3.7.3), user-level claims resolve throughos.homedir()on Windows, and duplicate or zombie daemons per workspace are stopped (#3987, @socrates6969). Trajectory counts persist across processes and daemon jobs feed learning (#3988, @socrates6969). One node process per MCP session: the launcher runs the resolved local CLI in-process and keepsnode:httpand the security barrel off the startup path (#3989, @socrates6969; ruflo-core plugin 0.2.10). - The console reads the project on Windows: a drive path is absolute, so the cwd is not prefixed twice (#4177, #4175). The neural data directory resolves through the project cwd, and stored worker error text is capped.
- Hooks and core: failed tool outcomes are recorded, hooks run from the project root, and telemetry stays off the tool path (#4178).
- Memory: calibrated default search threshold with near-miss hints (#4184, #4129);
memory_exportreads the CLI store as well as the AgentDB store (#4186, #4148); the AgentDB directory is created before the bridge opens it (#4155, #4132, @martinvlad);TieredMemoryStorerecalls multi-word and natural-language queries (memory3.0.5, #4183, #4152). - Intelligence: long memories are recalled for short route prompts (#4172, @rudycelekli); SONA reinforces the nearest pattern for near-identical tasks (#4185, #4134); AIDefence learns missed threats and autopilot learning loads (#4187, #4147, #4146).
- Workflow: invalid gates and unsupported steps fail closed (#4164, @rudycelekli). Agent health derives from observed execution outcomes (#4167, @rudycelekli).
- Analyze: diff content and commit subjects drive classification, safe Git revision ranges are accepted, encryption paths are flagged security-sensitive, and the reviewer-role docs are accurate (#4165, #4162, #4169, #4163, @rudycelekli).
- Security scan: interpolated
execcalls after imports are detected (#4166, @rudycelekli); a failednpm auditreports "could not scan" instead of a clean tree (#4154, #4124, @martinvlad). - A clear "not installed" message for
@ruvector/rvagent-wasm, and only new patterns are counted (#4179, #3979). - Skills, plugin contracts and ruvector/neural/kg surfaces (#4181; #4126, #4133, #4135, #4149, #4156-#4159, #4161).
[3.56.9] - 2026-10-12
Patch release: @claude-flow/cli, claude-flow, ruflo 3.56.9, with the bundled leaves @claude-flow/security 3.0.5 and @claude-flow/codex 3.0.5 published standalone and pinned exactly, and @claude-flow/shared 3.0.5 / @claude-flow/cli-core 3.7.2 (published earlier) pinned. Signed helpers changed (hook-handler.cjs, re-signed in #4096).
Fixed
- Policy envelope (
security3.0.5): explicit envelope ceilings require metering, and unmetered worker ceilings stay opt-in (#3503, @rudycelekli). - Codex MCP backend (
codex3.0.5, ruflo wrapper bridge): the opt-incodexbackend no longer launches@openai/codex mcp-server, which current Codex CLI dropped, so it is retired without a startup stall (#3522, @rudycelekli; reported in #3509 by @pacphi). - CLI parser, retry, rate-limit, monitoring and config fixes: a batch of review findings (#4107; #4000-#4003, #4006-#4011, #4013, #4016, #4017, @rudycelekli and others).
analyze diffrenders theanalyze_diffresult shape (#4102, #4101, @z4forzubair).shared3.0.5 andcli-core3.7.2: narrow-column truncation, rate-window retry times, structured network error codes (#4014, #4015, #4018, @rudycelekli).- ruflo-mods guards (#4096, @proffesor-for-testing): the root-delete guard no longer misses brace, nesting and fork-bomb forms, the screen
.envcheck is fixed, andinit-generated hooks carry the changes. - Console (marketplace plugin; its
tui.mjsis bundled into the CLI): worktree removal runs no fsmonitor or hooks and keeps untracked files (#4092), expired confirm cards are reported byconsole_state(#4067), pictures draw as SVG on Claude Code Desktop (#4112), all by @proffesor-for-testing; the Skills page pins its skills CLI (#4113, @martinvlad). - CI: the v3 root suite runs every test file and a crashed run is no longer accepted as a pass (#4099, @proffesor-for-testing); this does not ship in the package.
[3.56.8] - 2026-10-12
Patch release: @claude-flow/cli, claude-flow, ruflo 3.56.8. @claude-flow/shared 3.0.4 was published earlier and is now pinned exactly (3.56.7 still pinned 3.0.3). No other published package changed since 3.56.7; signed helpers did not change.
Fixed
@claude-flow/shared3.0.4: the RVF event log repairs an incomplete or damaged tail before appending (#4031 by @rudycelekli); a clean install can import.,./coreand./mcpbecause zod and express are now declared, and the tarball no longer ships compiled tests (#4089, #4091 by @proffesor-for-testing; the fresh-consumer finding came from @martinvlad, withscripts/smoke-shared-fresh-consumer.mjs).hooks coverage-gaps: files with no executable lines are no longer reported as 0%, overall coverage is line-weighted, the command readscoverage-final.json, and the report size is capped (#4082-#4084, reported by @ronmikailov; #4097). Plugin workflow contracts forruflo-testgenandruflo-sparcwere corrected in the same change (#4085-#4088; marketplace plugins, not npm).
[3.56.7] - 2026-10-11
Minor-feature patch: @claude-flow/codex 3.0.4 (published standalone; the CLI and root pin it exactly) and @claude-flow/cli, claude-flow, ruflo 3.56.7. Signed helpers changed (#4071, re-signed there).
Added
ruflo tui: the ruflo console in a terminal of its own (ADR-489, #4059, #4075, #4076). Six pages, opened with--viewor keys 1-6: Overview, Cost, Swarm, Events, Missions (read-only) and an AI page that asks Codex (pinned 0.162.0, read-only) or OpenRouter, each question after a confirm card;--no-aiturns it off and--max-spend-usd(default 0.5) stops sending to OpenRouter once the provider-reported spend reaches it.--plainprints one frame and exits 0; without a terminal it exits 2 with a hint.prepare-publishstages the bundledplugins/ruflo-console/dist/tui.mjs(about 1.27 MB, 446 KB gzipped) into the CLI tarball (about 0.4 MB larger, 3.9 MB packed). Console 0.42.7 is a marketplace plugin, not npm.- MCP resources capability on the CLI server, and a three-host integration loop (Claude Code, Codex, OpenRouter) with ADR-490/491 (#4068).
doctorflags a better-sqlite3 built against Node 24.19+ headers on Node 24.19-24.21 (#4069, #4074; reported in #4055 by @ronmikailov).ruflo mods: doctor repairs that repair,--strict, atomic install, startup self-check (#4073).
Fixed
- Auto-memory hooks deliver their subcommand, sync sees learnings, worktrees share the main memory dir, store writes are atomic (#4071, #4058 by @ronmikailov).
memory purgekeeps the on-disk format on sql.js write-back (#4057); command recording has an off switch and redaction, andmemory cleanup --expired-onlyreaches the bridge store (#4056, #4072; both reported by @ronmikailov).init --dualwrites a safe default.codex/config.toml; the previous permissive values now need--codex-trust full(#4070).
Known
- Nested better-sqlite3 11.x copies under agentdb and agentic-flow still ship; fixing that needs an agentdb release depending on better-sqlite3 ^12.
doctorwarns on affected Node versions.
[3.56.6] - 2026-10-11
Patch release: @claude-flow/cli, claude-flow, ruflo 3.56.6, with the already-published leaves @claude-flow/cli-core 3.7.1 and @claude-flow/shared 3.0.3 now pinned exactly. Signed helpers changed (hook-handler.cjs) and were re-signed in #4053.
Fixed
- Windows hooks launch through
node -einstead ofcmd /c, which started an interactive cmd that executed the hook JSON payload and created stray files (#4045). - The settings.json risk scanner now runs on the automatic SessionStart hook; both scanners match encoded PowerShell in linear time (#4052, #4053).
- Trajectory-end feedback is indexed even when steps were recorded (#3662, #4051).
@claude-flow/cli-core3.7.1 and@claude-flow/shared3.0.3: JSON memory backend, table/progress output, RVF event log and circuit breaker fixes from a reviewed batch by @rudycelekli (#4061; PRs #4020-#4026, #4028-#4030, #4032, #4033, #4035, #4037).
Plugins (no npm impact)
ruflo-cost-tracker: preserve Codex usage ownership (#4048, @Devesh-Maheshwari).ruflo-swarm: host-specific no-op manifest for native Codex hook discovery (#4050).ruflo-adr0.5.5.
[3.56.5] - 2026-10-11
Patch release: one fix in @claude-flow/cli. No leaf package changed since 3.56.4 (git diff v3.56.4..HEAD -- v3/@claude-flow touches only cli; the integration change drops publishConfig.tag and has no runtime effect).
Fixed
- The graph-edge-writer encryption-header sniff no longer closes the raw DB file descriptor, which dropped live SQLite POSIX locks; the header is read in a bounded child process (#4044, fixes #4040).
[3.56.4] - 2026-10-10
Patch release: security hardening of the plugin-trust and helper paths, MCP/task fixes, and the Node 24 sqlite fix. Standalone leaves published first: @claude-flow/memory 3.0.4, @claude-flow/integration 3.0.1, @claude-flow/plugins 3.0.2, @claude-flow/mcp 3.1.1 (bundled; republished standalone), @claude-flow/security 3.0.4 (bundled; republished standalone). Then @claude-flow/cli, claude-flow, ruflo 3.56.4.
Security
- Plugin registry trust: a placeholder (all-zero) trust anchor is a small-order Ed25519 key and no longer vouches for anything; signature verification uses strict (non-ZIP-215) rules; an unsigned registry no longer grants "official" trust (#3948).
plugins upgradeno longer runs the new version'spostinstallor registersshell:exechooks without the install-time gate (#3951).- Plugin and helper locators stop walking up at
node_modules, so a project-local install cannot reach projectplugins/ruflo-*code (#3986, replaces #3954). - Helpers: session ids and session paths are validated, module roots are no longer project-chosen (#3949); the statusline no longer interpolates
HELPER_ROOTinto a shell string, andnpxruns the helper CLI from the helper root (#3990, replaces #3953). - User-level settings pin home helpers (#3952).
task_assignrejects__proto__agent ids that wrote ontoObject.prototype(#3967).
Fixed
- Node 24:
better-sqlite3is routed to a build that does not abort at exit (#3955). - Importing the
agentic-flowroot entry no longer runs its CLI (#3958). - MCP resource cache: stale writes are fenced after invalidation (#4019).
security scancode patterns no longer report the false positives seen on a real tree (#3999). - MCP, task and CLI fixes: #3975, #3974, #3970, #3969, #3968, #3980, #3947; graph path search ~94x faster at n=3200 (#3977).
[3.56.3] - 2026-10-09
Patch release: AgentDB 3.0.0-alpha.20 update. @claude-flow/memory 3.0.3 (published standalone; the CLI pins it exactly) and @claude-flow/cli 3.56.3. neural, shared and cli-core are unchanged.
Changed
- AgentDB floor raised from
^3.0.0-alpha.17to^3.0.0-alpha.20(memory,cli, root,ruflo). alpha.20 swaps@xenova/transformersfor@huggingface/transformers,@ruvector/rvf0.1.9 -> 0.2.3 andruvector-> 0.2.x. Semantic search scores shift slightly (same-text cosine ~0.992 against alpha.17 embeddings); an entry scored just under the 0.3 default threshold under alpha.17 can now appear just above it, ranked far below the exact-keyword hit.
Fixed
- Dev tree only:
v3/package.jsongains apnpm.packageExtensionsentry declaringonnxruntime-common1.24.3 for@huggingface/[email protected](it imports it without declaring it, so pnpm's hoist gave it 1.14.0 and AgentDB silently fell back to mock embeddings in the pnpm workspace; npm installs were not affected). - AgentDB alpha.20 refuses a sql.js save when the database changed under an open handle or when
<db>.agentdb.lockexists, and a lock left by a crashed writer was never cleared, so every later save failed and the database file was never created.@claude-flow/memorynow removes that lock only when it is provably stale (regular file named exactly<db>.agentdb.lock, numeric PID that is dead, older than 10 s, re-verified after an atomic rename-aside), retries once, and otherwise throwsAgentdbLockError. A failed persist on shutdown is surfaced (ControllerRegistry,AgentDBBackend,HybridBackend,shutdownBridge), never swallowed. This only affects the sql.js path (no native better-sqlite3).
[3.56.2] - 2026-10-09
Patch release: one bug fix in @claude-flow/cli. No leaf package changed since 3.56.1 (git diff v3.56.1..HEAD -- v3/@claude-flow touches only cli). The ruflo-console fixes merged since 3.56.1 (console 0.40.4-0.40.8, #3933, #3935, #3938, #3941, #3942) and ruflo-swarm 0.3.6 ship through the plugin marketplace, not this npm release.
Fixed
task_updateaccepts only the five task statuses; a task stored ascompleteis no longer shown as pending or re-dispatched (#3932, fixes #3931).
[3.56.1] - 2026-10-08
Patch release: five bug fixes (four in @claude-flow/cli, one console). No leaf package changed since 3.56.0 (git diff v3.56.0..HEAD -- v3/@claude-flow touches only cli), and no behaviour is removed. The ruflo-console plugin fixes merged since 3.56.0 (console 0.40.2/0.40.3, #3917, #3924) ship through the plugin marketplace, not this npm release.
Fixed
- The policy trust mirror and key now honour
HOMEandXDG_CONFIG_HOMEinstead of a fixed home directory, so a throwawayHOMEno longer yieldspolicy-anchor-log-missing(#3922, fixes #3919, reported by @HF-teamdev). ensureSchemaColumnsrewrites the store only when a column was actually added; an up-to-date encrypted memory store is no longer rewritten by a read such asmemory stats(#3923, fixes #3918, reported by @HF-teamdev).config reset --section <name>resets only that section and removes the matching flat overrides, preserving unrelated configuration (#3901, @rudycelekli).- MCP config tools (
config_get/set/list/export/import/reset) now work on the plain nested JSON document written by theconfigCLI command instead of assuming a{values, scopes}envelope; existing envelope stores keep their format (#3902, @rudycelekli). This does not change how the daemon readsconfig.json(see #3192, #3449, #3239). - Console approvals badge counts only rows that have an approve or deny action (#3924, fixes #3920, reported by @HF-teamdev); delivered as
ruflo-console0.40.3 via the plugin marketplace.
[3.56.0] - 2026-10-08
Minor release: a hardening and correctness batch from @rudycelekli, @nicholas-ruest, @martinvlad, @HF-teamdev, @drakeo338 and @stuinfla. Several fixes change behaviour on purpose (see Behaviour changes). Integrated by merging each PR head, so authorship is preserved.
Behaviour changes
- Unknown ids and value-less options are now errors (#3911, @martinvlad) β
agent status/task statuswith an unknown id printnot foundand exit 1; a declared string/number option with no value (hooks route --task) now reportsOption --task needs a valueinstead of silently becomingtrue. Applies to every command. session_savecaptures the persisted task/agent/memory stores by default, andsession_importrejects malformed stores (#3489, #3490); explicitfalsestill excludes a store.- Default project memory backup snapshots both
memory.dbandagentdb-memory.db(#3903); explicit--dband a non-blankCLAUDE_FLOW_DB_PATHstill select one file. - Agent pool scale-to-zero is honored and invalid targets fail before writing (#3488); a zero fallback retry budget is honored (#3496).
- Project path validators block sensitive directory names anywhere under the project root, not only at the leaf (#3504,
@claude-flow/security3.0.3). - The OAuth callback server retains an early callback and ignores unrelated requests (#3507,
@claude-flow/security3.0.3). - Provider selection honors an explicit execution provider (#3493) and Anthropic gateway credentials/endpoint (
ANTHROPIC_AUTH_TOKEN,ANTHROPIC_BASE_URL; #3494, fixes #3100); provider timeouts now cover reading the response body (#3495). hooks_routeconsults persisted SONA outcome patterns before local routing (#3904) β only a supported agent at confidence >= 0.6 is used;useSemanticRouter: falsekeeps keyword routing. With the opt-in neural router, selection uses calibrated quality predictions (#3498) so calibration can move the pick across the quality bar.- The WAL-sidecar safety gate is liveness-based (#3894, fixes #3161) β leftover
-wal/-shmfiles from a closed connection no longer blockmemory store/list; a genuinely open holder is still refused. ruflo initno longer pins a Claude model in generated.claude/settings.json(#3528, fixes #3527).- The
ruflowrapper pins the matching@claude-flow/cliand warns when the resolved CLI version differs (#3438, fixes #3306); it warns and continues, it does not refuse. - Claims are stored in one lossless format shared by CLI and MCP; unreadable stores are refused rather than replaced (#3502).
- Native immutable conditional append for
memory store(--append-only,--append-conditions,--require-native,--no-embedding; #3750, @stuinfla) β additive, CLI only.
Fixed
- Shipped
.claude/settings.jsonhad unescaped quotes in hook commands and did not parse (#3895, fixes #3888, @nicholas-ruest); a new test checks that every shipped settings/plugin JSON parses. - Remaining
better-sqlite3opens go through the shared loader (#3896, fixes #3883's sites, @nicholas-ruest); an RFE1-encrypted store is reported as encrypted, not broken (#3906, fixes #2889, @nicholas-ruest); the demo-registry fallback is labelled as unverified (#3897, @martinvlad). analyzepreserves machine-readable JSON (#3910);doctorrejects unsupported component selections (#3907) and loads MetaHarness modules via file URLs on Windows (#3523, fixes #3186);daemon starthonors explicit scheduled workers (#3586).- Unreadable task/agent records are preserved, not overwritten (#3484); async agent-registry updates do not revive retired agents (#3492); fallback learning is attributed to each attempted model (#3497).
- Workflow step boundaries are validated (#3491), persisted workflows resume without duplicate executors (#3500) and keep control signals and sibling state across awaits (#3505).
- Memory JSON import reports rejected writes accurately (#3499); export reads all pages only after successful reads (#3501); concurrent rotating-token refreshes coalesce (#3506).
- Test fixtures no longer break under pnpm's
NODE_PATH(#3543, fixes #3529) and the vitest runner is resolved viapackage.json. @claude-flow/memory3.0.2: cache admission accounts for replacement sizes (#3679, fixes #3678) and invalidation matches each key independently (#3687, fixes #3686).- Removed an empty
.claude/agents/tmp.jsonthat shipped in the CLI tarball.
Not included
#3503 (fail-closed policy envelope: it would deny MCP dispatch for any envelope with a ceiling because dispatch does not supply the metering), #3892, #3831, #3879, #3835, #3751, #3694, #3909 and the draft security PRs: see each PR for the status.
Leaf packages published with this release
@claude-flow/memory 3.0.2, @claude-flow/security 3.0.3.
[3.55.0] - 2026-10-07
Minor release. Contains a security behaviour change: ruflo mcp start -t http on a non-loopback host now refuses to start without a token (see Security and Breaking changes).
Security
- MCP HTTP transport: refuse unauthenticated non-loopback binds, optional bearer-token auth (#3598, #3599) β
ruflo mcp start -t http --host 0.0.0.0(or any non-loopback host) now exits 1 unless a token is set (RUFLO_MCP_HTTP_TOKEN,--auth-token-file,--auth-token) or the operator opts out withRUFLO_MCP_ALLOW_UNAUTHENTICATED_HTTP=1. With a token,/rpc,/mcpand/infoneedAuthorization: Bearer <token>(constant-time compare);/healthstays public and minimal. Also fixes a fail-open in@claude-flow/mcpwhere auth enabled with an empty token list accepted any bearer value. Not closed: loopback with no token is still unauthenticated (any local process can call every tool, and DNS rebinding is not blocked because there is noHostallow-list); set a token whenever the HTTP transport is used. A valid token grants every tool (no per-tool authorization). - Hive-mind gating that keeps local users working (ADR-476, #3338, #3339) β
hive-mind_initno longer returnshiveToken, andspawn,consensuspropose,broadcast,shutdown,memoryset/deleteandoptimize-memory(plus the existingjoin/leave/vote) now require an operator credential from remote callers (HTTP/WebSocket MCP). Local stdio MCP,ruflo mcp exec, thehive-mindCLI and in-process callers need no credential and no extra step. Remote clients setRUFLO_HIVE_BOOTSTRAP_SECRET(or read.claude-flow/hive-mind/bootstrap.secret, created 0600 by the first localhive-mind init) and send it asbootstrapSecret;RUFLO_HIVE_REQUIRE_AUTH=1applies the same rule to stdio on bridged servers.state.jsonis now written 0600 atomically and the hive directory 0700. Migration: clients that readhiveTokenfrom theinitresponse must drop that (it isundefined); over stdio they omit it everywhere. This is a speed bump, not a boundary, against a local prompt-injected agent that can read the files. - Policy ledger anchor can no longer be deleted and silently re-established (#3602, #3886, ADR-475) β
policy verifyno longer trusts an anchor stored only instate.json. A second, hash-chained anchor log plus a mirror under~/.config/ruflo/policy-trust/now detect truncation, andpolicy verify --establish-anchor(interactive TTY only) is the logged repair path. Limit: this does not defend against an attacker who can rewrite both the project directory and~/.config/ruflo; stronger evidence needs an external witness. A pre-#3568 ledger with receipts and no anchor now fails policy transactions withpolicy-ledger-anchor-missinguntil repaired. Ships in@claude-flow/security3.0.2.
Fixed
- Resilience batch (#3885; fixes #3676 #3674 #3668 #3670 #3682 #3672 #3593 #3592; thanks @rudycelekli) β
@claude-flow/shared3.0.2: event-store pagination, bulkhead sync-throw handling, retry timeout cleanup, connection-pool reservation (also in@claude-flow/mcp); cli: production retry on non-Error rejections, worker-queue terminal outcomes and cancelled head, bounded worker pool with an already-aborted input. @claude-flow/memory3.0.1 βmmrRerank()caches embedding-cosine max-similarity across rounds (#3516, #3517).@claude-flow/swarm3.0.1 β Byzantine quorums now intersect for every cluster size (#3560, #3587, @rudycelekli);spawnAgent()auto-domain branch registers the agent in its pool (#3538, #3539);MessageBusis event-driven instead of a 10 ms unconditional poll (#3563).@claude-flow/plugin-agent-federation1.0.1 β outbound sends are authorized with the peer's current trust, not a stale snapshot (#3561, #3588, @rudycelekli).- CLI string options keep their declared value (#3594, #3601), config cache is isolated per project directory (#3590, #3591), nested config defaults are cloned (#3595, #3600) β all @rudycelekli.
- Hive consensus commands fail when the hive rejects the operation (#3609, #3654, @rudycelekli).
daemon start --workersselection reaches the WorkerDaemon (#3547, #3875).- Plugins run the installed ruflo CLI before
npx @claude-flow/cli@latest(#3558, #3559, @HF-teamdev) β ruflo-cost-tracker 0.27.2, ruflo-adr 0.5.4, ruflo-metaharness 0.2.4, ruflo-goals 0.4.3. - ruflo-adr edge keys use a separator the memory validator accepts (#3633, #3636, @drakeo338).
- ruflo-console 0.36.1: Settings lists Claude control and spending first (#3887).
Breaking changes
- Binding the MCP HTTP transport off loopback (for example Docker with
--host 0.0.0.0) now requiresRUFLO_MCP_HTTP_TOKEN(and clients must send the header) orRUFLO_MCP_ALLOW_UNAUTHENTICATED_HTTP=1; otherwisemcp startexits with an error. Loopback users are unaffected. hive-mind_initno longer returnshiveToken(ADR-476).
Leaf packages published with this release
@claude-flow/shared 3.0.2, @claude-flow/memory 3.0.1, @claude-flow/swarm 3.0.1, @claude-flow/security 3.0.2, @claude-flow/mcp 3.1.0, @claude-flow/plugin-agent-federation 1.0.1.
[3.34.0] - 2026-07-31
Added
- AGNTCY/Outshift runtime integration (ADR-378/379/380) β optional, removable augmentation per ADR-150's pattern; the kernel stays fully operational with these packages absent.
ruflo transport use slim,ruflo agent publish, andruflo swarm join <namespace>CLI verbs; CASA (Continuous Agentic Semantic Authorization) envelope schema, deterministic compiler, and deny-by-default enforcement gate with Ed25519-signed decision receipts (.swarm/casa-receipts.jsonl); AGNTCY OTel span attributes (coordination.episode,authorization.decision). All verbs exit 0 with a clear message whenRUFLO_AGNTCY_SLIM_ENDPOINTis unset β no fake success paths. Companion Rust cratev3/crates/ruflo-agntcymirrors the TS enforcement logic, with a real in-processLocalTransportand aSlimTransportstub behind a non-defaultslimCargo feature. Companion package@metaharness/agntcy(build-time half β identity, OASF export, Directory publish, semantic observability) ships from the siblingmetaharnessrepo. (v3/@claude-flow/cli/src/commands/agntcy/,plugins/ruflo-agntcy/,v3/crates/ruflo-agntcy/,v3/docs/adr/ADR-378-*.md,ADR-379-*.md,ADR-380-*.md) - npm Trusted Publishing (OIDC) release workflow (ADR-378) β
.github/workflows/stable-npm-release.ymlpublishes the three-package stable train (@claude-flow/cli,claude-flow,ruflo) from an immutable, tag-pinned checkout with a full test/build/pack/install smoke-test gate before any registry write, verifies published-package integrity against the locally-built archive, and rolls outlatest/alpha/v3alphadist-tags together. Manualworkflow_dispatchonly, gated to theruvnetactor. - Configurable statusline cost segment via two environment variables (defaults unchanged):
RUFLO_STATUSLINE_COST_SYMBOLβ override the leading$(e.g.β‘,β¬,π±); empty string shows the number alone.RUFLO_STATUSLINE_HIDE_COSTβ1/true/yes/onhides the segment.cost.total_cost_usdis a client-side estimate that may differ from the actual bill and is misleading on subscription plans.
Fixed
@agntcy/slim-bindingspinned to the confirmed-working alpha β the SLIM maintainers moved offuniffi-bindgen-react-native(which shipped raw, uncompiled TypeScript incompatible with plain Noderequire/importβ filed as agntcy/slim#1916, reproduced and confirmed) onto@ubjs/core/@ubjs/nodein the2.0.0-alpha.4+dist-tag, not yet promoted tolatest.v3/@claude-flow/cli/package.jsonnow pins the exact working alpha version (deliberate exact-pin for a pre-release channel, not a caret range) rather than the still-brokenlatest; verified live end-to-end (server bring-up, client connect, graceful shutdown, zeroERR_UNSUPPORTED_NODE_MODULES_TYPE_STRIPPING).detectAgntcyRuntime()required zero logic changes β its existing graceful-degradation design already returnedconfigured: trueonce the upstream package resolved.security scanfailed open on unvalidated--depth/--type/--targetβ an unrecognised--typematched none of the three phase guards, so no phase ran at all and the command still printed "No security issues found!" and exited 0; an unrecognised--depthfell through chained ternaries to the shallowest traversal (so--depth fullscanned less than the defaultstandard), which on a tree whose only HIGH finding sat below that budget also flipped the critical/high exit-code gate from 1 to 0; and a non-existent or non-directory--targetread nothing, which the swallowed dir-read catches turned into a clean banner, exit 0, and a persisted CLEAN report thatgetSecurityStatusthen surfaced as CLEAN. All three now fail closed before anything is scanned or written. Chained ternaries replaced with exhaustiveRecord<ScanDepth, number>maps behind real type predicates, and the recursion guard now positive-tests its budget so a bad value stops traversal instead of disabling the limiter. (v3/@claude-flow/cli/src/commands/security.ts)
Changed
security scan --depth fullis deprecated, not removed βfullwas never a supported depth, but the CLI itself emitted it (statusline insight, announcement, the CLAUDE.mdinitgenerates, two shipped agent definitions). It is now normalised todeepwith a warning rather than rejected, and those emitters have been updated to--depth deep. (v3/@claude-flow/cli/src/funnel/insights.ts,funnel/messages.ts,commands/announcements.ts,init/claudemd-generator.ts,.claude/agents/v3/security-architect*.md)
Removed
security scan --type containeris rejected instead of silently reporting clean β it had been advertised in--type's help since the command was written, but no phase ever implemented it, so it scanned nothing and printed a clean bill of health. It now exits 1 with an explicit "not implemented" message, and the help text no longer offers it. Breaking for any pipeline passing--type container, which previously exited 0. (v3/@claude-flow/cli/src/commands/security.ts)
[3.32.10] - 2026-07-26
Bug-fix release closing the tracker-fire sweep from 2026-07-24 β 2026-07-26. All fixes are surgical and additive; no API surface change, no schema change, patch semver.
Fixed
- Statusline promo row blank on new installs β restored the cold-start local seed pool that ADR-311 had emptied. The remote Cognitum-served pool (
funnel.ruv.io/v1/messages) remains authoritative viaeligibleMessagesFromPools(remote wins byid); the seed only renders during the β€ 60s cold-start window before the first successful remote fetch lands. Nine seed messages ship: one disclosure, seven educational tips (all pointing atcognitum.onedocs), one promotional. (v3/@claude-flow/cli/src/funnel/messages.ts) - #2777 β
ruflo initno longer imports the entire ruvnet/ruflo repo (97 MB, 384SKILL.md) into.agents/skills/ruflo/. Materialization now writes only the single platformSKILL.md, and the idempotency gate detects a "bloated" prior install (Cargo.toml/crates/ present or size > 1 MB) and wipes+re-materializes. (v3/@claude-flow/cli/src/commands/init.ts) - #2781 β
ruflo-adr adr-indexno longer silently drops ADR data: status regex accepts the- **Status**: proposedformadr-createemits, single-line and wrapped relation lines parse fully, andCLI_CORE=1no longer forks writes into an alternate store the default reader can't see (warns and unifies on@claude-flow/cli@latest). Dry-run over the repo's 531 ADRs now captures 608 edges cleanly. (plugins/ruflo-adr/scripts/lib/parse-adrs.mjs,import.mjs,reindex.mjs) - #2770 β Windows:
browser-sessionMCP tools and two init-flowexecFileSync('npx', β¦)sites now setshell: process.platform === 'win32'so cmd.exe can resolvenpx.cmd. POSIX behavior unchanged. Each site carries a shell-injection note for future editors. (v3/@claude-flow/cli/src/mcp-tools/browser-tools.ts,v3/@claude-flow/cli/src/commands/init.ts) - #2782 β
WorkerDaemon.saveState(),autopilot-state.saveState(), andautopilot-state.appendLog()no longer race on a shared.tmpfilename under in-process concurrent workers. All three sites now call the existingwriteFileAtomichelper (pid + timestamp + random-suffixed temp) so concurrent callers cannot collide; each write is wrapped in try/catch so a losing racer can't crash the daemon. (v3/@claude-flow/cli/src/services/worker-daemon.ts,v3/@claude-flow/cli/src/autopilot-state.ts) - #2785 β
ruflo hooks post-tasknow accepts--task(short-t) and--store-resultsflags, matching the CLAUDE.md-documented usage. Routing outcomes finally persist to the namespacehooks_metricsreads, so the "Pattern Learning" / "Agent Routing" numbers become non-zero. (v3/@claude-flow/cli/src/commands/hooks.ts) - #2786 β AgentDB no longer silently fails to initialize when
CLAUDE_FLOW_ENCRYPT_AT_REST=1is set. AddedgetAgentDbPath()which returns the same directory asgetDbPath()but with basenameagentdb-memory.db, so the ControllerRegistry (native better-sqlite3) opens a distinct file from the sql.js CRUD writer'smemory.db(which stays encrypted).learningSystem/reasoningBankpopulate correctly with encryption enabled. (v3/@claude-flow/cli/src/memory/memory-bridge.ts) - #2776 β Statusline security segment:
STALEandIN_PROGRESSstates are now reachable via a local overlay (getLocalSecurity) that recomputes freshness on every render from.claude/security-scans/scan-*.json. Configurable viaRUFLO_SCAN_STALE_HOURS(default 24) andRUFLO_SCAN_PENDING_CAP_MIN(default 30); STALE renders dim gray so it stops shouting for attention once escalated. (.claude/helpers/statusline.cjs) - #2775 β Memory store to an existing key no longer dead-ends:
bridgeStoreEntryusesINSERT ... ON CONFLICTwith tombstone auto-resurrect; UNIQUE returns a typed error instead ofnull(no more misleading demotion into the #2735 guard);bridgeDeleteEntryrunswal_checkpoint(PASSIVE); CLImemory store --upsertdefault now applies at the parser layer (root cause:parser.ts:applyDefaultsonly walkedglobalOptions); thememory_storeMCP tool schema defaultsupsert: truefor CLI parity. (v3/@claude-flow/cli/src/memory/memory-bridge.ts,commands/memory.ts,parser.ts,mcp-tools/memory-tools.ts)
Investigated
- #2774 β Reporter's diagnosis "Codex MCP generator registers management CLI instead of stdio server" appears incorrect:
ruflo mcp startIS a working stdio server perv3/@claude-flow/cli/src/commands/mcp.ts(MCP Server started on stdio), and upstreamd20f1323b fix(codex): ship stable Windows-safe Ruflo integrationcleanly centralized the Codex MCP config inmcp-config.tsusing the same command shape. A proposed swap toclaude-flow-mcpwas reverted during rebase. Recommend closing the issue unless the reporter can produce a fresh repro against 3.32.9+.
[3.5.0] - 2026-02-27
Ruflo v3.5 β First Major Stable Release
This release marks the official rebranding from Claude Flow to Ruflo and represents the first major stable release after 5,800+ commits, 55 alpha iterations, and 10 months of development.
Highlights
- Rebranding: Claude Flow β Ruflo across all packages (
@claude-flow/cli,claude-flow,ruflo) - agentic-flow v3.0.0-alpha.1 Integration: Full deep integration with 10 subpath exports (ReasoningBank, Router, Orchestration, Agent Booster, SDK, Security, QUIC transport)
- AgentDB v3.0.0-alpha.9: 8 new controllers (HierarchicalMemory, MemoryConsolidation, SemanticRouter, GNNService, RVFOptimizer, MutationGuard, AttestationLog, GuardedVectorBackend) + 6 MCP tools
- 215 MCP Tools: Full Model Context Protocol server with vector memory, neural training, swarm coordination
- Security Hardening: Command injection fix, TOCTOU race fix, eliminated hardcoded HMAC keys, timing attack fixes
- Doctor Health Check: New
agentic-flowdiagnostic (filesystem-based, ESM-compatible) - 0 Production Vulnerabilities: Clean
npm auditacross all packages
Added
agentic-flow-bridge.tsβ Unified lazy-loading bridge for all agentic-flow v3 modules- Tiered embedding resolution: ReasoningBank WASM (Tier 1) β @claude-flow/embeddings (Tier 2) β mock fallback (Tier 3)
- Agent Booster local import with npx fallback
checkAgenticFlow()doctor health check- 7 TypeScript module declarations for agentic-flow subpath exports
- ADR-056: agentic-flow v3 Integration Architecture
Fixed
- Command injection vulnerability in enhanced-model-router.ts (SAFE_LANGUAGES whitelist)
- TOCTOU race condition in bridge singleton initialization (Promise-based caching)
- 22 agent/skill files updated from stale v1.5.11/v2.0.0-alpha to v3.0.0-alpha.1
- ESM compatibility for doctor checks (filesystem-based instead of
require.resolve) - @ruvector/gnn pinned to 0.1.25 to fix fatal process crash (issue #216)
Changed
- All 3 packages bumped from
3.1.0-alpha.55to3.5.0 - Publish tags changed from
alpha/v3alphatolatest - agentic-flow minimum version:
0.1.0β3.0.0-alpha.1 - agentdb minimum version:
2.0.0-alpha.3.4β3.0.0-alpha.10
[3.1.0-alpha.55] - 2026-02-27
AgentDB 3.0.0-alpha.9 Integration (ADR-053/ADR-055)
- Activated 8 AgentDB v3 controllers with MutationGuard proof engine
- Added 6 new MCP tools:
agentdb_hierarchical_*,agentdb_consolidation_*,agentdb_semantic_* - Fixed controller registry activation bugs (ADR-055)
- Statusline fixes for real-time controller status
- Pinned @ruvector/[email protected] to fix fatal process crash
[3.1.0-alpha.43] - 2026-02-15
Ruflo Branding Fix
- Fixed CLI branding: show 'ruflo' instead of 'claude-flow' when run via
npx ruflo - Fixed Windows ESM import crash with
pathToFileURL - Fixed init hook prompt overflow and description field
[3.1.0-alpha.36] - 2026-02-10
Stability & Compatibility
- Fixed hooks backward compatibility:
--successand--filemade optional - Fixed Windows npm install crash (404 optional dependencies)
- Bumped agentdb to 2.0.0-alpha.3.6
- Fixed V3 build errors (missing helmet, VERSION type, vitest spy)
[3.1.0-alpha.29] - 2026-02-01
Security & Agent Teams
- Security fixes, backward compatibility, and Agent Teams hooks
- Added
--settingsflag to upgrade command for Agent Teams - Fixed npm 11 install crash by pinning agentdb
v3.0.0-alpha Series (2025-10 to 2026-02)
v3.0.0-alpha.184 β CLI Help & Categorization (2025-12)
- Fixed CLI help categorization across 26 commands
- Published install optimizations
- curl-style installer script
- SEO-optimized npm packages for discovery
v3.0.0-alpha.170 β Plugins & Marketplace (2025-12)
- Plugin Marketplace: 8 official plugins + IPFS registry via Pinata
- Gas Town Bridge Plugin: WASM-accelerated orchestrator integration
- 10 RuVector WASM Plugins: 50 MCP tools for neural computation
- @claude-flow/teammate-plugin: MCP tools for Agent Teams coordination
v3.0.0-alpha.150 β SONA & SemanticRouter (2025-11)
- SemanticRouter: SONA WASM integration with verified benchmarks
- Fixed phantom Claude popups on Windows
- Fixed statusline safe multi-line output for Claude Desktop
- Fixed MCP tool naming (
/β_) for Claude Desktop compatibility - Memory namespace support in delete command
v3.0.0-alpha.100 β @claude-flow/guidance (2025-11)
- @claude-flow/guidance Control Plane: Governance, compliance, and policy enforcement
- Wave 1: Proof, gateway, memory-gate, coherence, hooks, persistence primitives
- Wave 2: Conformance kit, capability algebra, evolution pipeline, artifact ledger
- Wave 3: Civilization-grade primitives (trust, truth, uncertainty, time, authority)
- Rust WASM Policy Kernel: SIMD128-accelerated policy evaluation
- ContinueGate: Safety gate for agent continuation decisions
- 22-benchmark suite with before/after performance reporting
- CLAUDE.md generators, analyzer, and auto-optimizer
- Content-aware executor with statistical validation (Spearman Ο, Cohen's d)
v3.0.0-alpha.50 β Core V3 Implementation (2025-10)
- Complete V3 implementation across all ADRs
- ADR-003: Coordinator consolidation + security tests
- Complete hooks system with AgentDB, HNSW, tests
- ReasoningBank guidance system with CLI
- V2βV3 migration documentation
- MCP memory tools upgraded to sql.js + HNSW backend
- Claims-based authorization (ADR-016)
- Node.js worker daemon system
- Auto-update system for @claude-flow packages (ADR-025)
- Replaced all mock implementations with real functionality
v3.0.0-alpha.1 β Foundation (2025-10)
- Complete V3 monorepo structure (
@claude-flow/cli,shared,memory,hooks,security) - 26 CLI commands with 140+ subcommands
- 215 MCP tools via FastMCP 3.x
- RuVector intelligence system (SONA, MoE, HNSW, EWC++, Flash Attention)
- Hive-Mind consensus (Byzantine, Raft, Gossip, CRDT, Quorum)
- 17 hooks + 12 background workers
- 60+ specialized agent types
- Cross-platform helper system
v2.7.x Series (2025-08 to 2025-10)
v2.7.34 β PostgreSQL & Neural Persistence
- PostgreSQL Bridge with attention, GNN, hyperbolic embeddings
- Neural pattern persistence to disk
- Hive-mind
--claudeflag for spawn command - Real statusline data, hive-mind shutdown fixes, daemon persistence
- Multi-platform builds (Linux, macOS, Windows) in CI/CD
v2.7.0 β agentic-flow Integration
- Deep integration with agentic-flow coordination engine
- SDK architecture analysis and hooks & learning integration
- Modular installation strategy
- Optimized v3 migration plan
v2.0.0-alpha Series (2025-05 to 2025-08)
v2.0.0-alpha.128 β Maturity
- Comprehensive hive-mind optimization
- Database schema robustness (missing columns, optimization errors)
- Auto-rebuild better-sqlite3 on NODE_MODULE_VERSION mismatch
- InMemoryStore interval cleanup for clean process exit
v2.0.0-alpha.53 β Hook Safety
- Critical hook safety system
- Hive-mind optimization command
- Safety & security features documentation
- Neural Link System with safety protocols
v2.0.0-alpha.33 β Windows & WSL
- Windows/WSL compatibility fixes
- Module import error resolution
- README restructure for v2.0.0 features
- Comprehensive test suite
v1.x Series (2025-01 to 2025-05)
v1.0.71 β Final v1 Release
- npm publishing compatibility
- Full CLI command functionality
- SPARC integration with full prompt loading
- Cross-platform support
v1.0.50 β Swarm & SPARC
- Parallel execution for swarm tasks
- Background task management
- Swarm command with improved error handling
- Claude Code slash commands integration
v1.0.28 β Project Management
- CLI project management commands
- System monitoring and SPARC commands
- Orchestration templates (monitoring, optimization, security review)
v1.0.1 β Initial Release (2025-01-01)
- Complete Claude-Flow AI Agent Orchestration System
- Configuration guide and comprehensive tests
- Initial commit
Milestone Summary
| Milestone | Version | Date | Key Feature |
|---|---|---|---|
| Initial Release | v1.0.1 | 2025-01 | AI agent orchestration system |
| SPARC Integration | v1.0.50 | 2025-03 | Swarm + SPARC methodology |
| Alpha Foundation | v2.0.0-alpha.33 | 2025-05 | V2 alpha with hook safety |
| agentic-flow | v2.7.0 | 2025-08 | agentic-flow coordination engine |
| V3 Foundation | v3.0.0-alpha.1 | 2025-10 | V3 monorepo, 215 MCP tools |
| Plugin Marketplace | v3.0.0-alpha.170 | 2025-12 | 8 plugins + IPFS registry |
| Guidance Control Plane | v3.0.0-alpha.100 | 2026-01 | WASM policy kernel, ContinueGate |
| AgentDB v3 | v3.1.0-alpha.55 | 2026-02 | 8 controllers, MutationGuard |
| Ruflo v3.5 | v3.5.0 | 2026-02-27 | First stable release, rebranding |