Reported issues for claude-flow
Pod holds 24 of 362 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.
Back to claude-flow.
Most discussed
hooks statusline --json writes [WARN]/[INFO] to stdout, breaking JSON parsers
@claude-flow/cli v3.34.0, macOS, Node v26.1.0.
hooks statusline --json emits diagnostics on stdout, ahead of the JSON. With stderr fully
discarded, stdout still starts with a log line:
$ npx @claude-flow/cli hooks statusline --json 2>/dev/null | head -c 60
[WARN] Failed to load config from <project>/claude-flow.config.json: ...
JSON.parse → Unexpected token 'W', "[WARN] Fail"... is not valid JSON
**It's intermittent, which is worse than a hard fail
Read the thread · 2026-08-03 · closed · external user · 7 comments
MCP memory_search defaults to namespace='default' instead of 'all' — returns 0 results for entries in other namespaces
Bug Description
The MCP memory_search tool uses a default similarity threshold of 0.3 which causes valid semantic searches to return 0 results, while the identical query via claude-flow memory search CLI returns the expected results.
Steps to Reproduce
- Store an entry with a 768-dim embedding:
claude-flow memory store --key "test-entry" --value "ADR-048 memory pipeline with curateIndex and topic files" --namespace patterns
- Search via CLI — works:
[Read the thread](https://github.com/ruvnet/ruflo/issues/1131) · 2026-02-13 · closed · external user · 9 comments
### [Bug] Missing hive-mind/spawn MCP tool in V3
## Description
The CLI command `npx claude-flow@alpha hive-mind spawn` exists but fails because the corresponding MCP tool `hive-mind/spawn` is not implemented. I should mention I'm migrating a project currently using 2.7.x
## Steps to Reproduce
1. Install claude-flow V3: `npx claude-flow@alpha`
2. Initialize hive-mind: `npx claude-flow@alpha hive-mind init`
3. Attempt to spawn workers: `npx claude-flow@alpha hive-mind spawn -n 5`
## Expected Behavior
Workers should be spawned
[Read the thread](https://github.com/ruvnet/ruflo/issues/943) · 2026-01-14 · closed · outside contributor · 7 comments
### 🚨 Critical: 85% of MCP Tools Are Mock/Stub Implementations - Deep Validation Analysis
# 🚨 Critical Issue: Extensive Mock/Stub Implementation Detected in Claude-Flow MCP Server
## Summary
Through systematic deep validation analysis using a Hive Mind collective intelligence approach, we discovered that approximately **85% of claude-flow MCP tools are mock/stub implementations** that return success responses without performing actual functionality.
## Impact
- **High Severity**: Production workflows using claude-flow tools will fail silently
- **Silent Failures**: Tools return suc
[Read the thread](https://github.com/ruvnet/ruflo/issues/653) · 2025-08-14 · closed · external user · 11 comments
### [Bug] Memory is not working at all
Still have no working memory in `v2.0.0-alpha.78`
It's pretty unusable right now
After trying with ~5 hive-mind spawns i still get this output:
✅ Memory Bank Statistics: Total Entries: 0 Namespaces: 0 Size: 0.00 KB
Furthermore all my hive-mind spawns show the same status basically regardless of their progress:
Status: active Created: 7/30/2025, 6:18:10 AM
Agents: 👑 Queen: Queen Coordinator (active) 🐝 Workers: - Researcher Worker 1 (researcher) idle - Coder Wor
Read the thread · 2025-07-30 · open · external user · 12 comments
Could not create Claude Code slash commands: Deno is not defined
run npx claude-flow@alpha init --force
📝 Creating Claude Code slash commands...
✓ Created slash command: /sparc-architect
✓ Created slash command: /sparc-code
✓ Created slash command: /sparc-tdd
✓ Created slash command: /sparc-debug
✓ Created slash command: /sparc-security-review
✓ Created slash command: /sparc-docs-writer
✓ Created slash command: /sparc-integration
✓ Created slash command: /sparc-post-deployment-monitoring-mode
✓ Created slash command: /sparc-refinement-o
[Read the thread](https://github.com/ruvnet/ruflo/issues/480) · 2025-07-27 · closed · external user · 7 comments
### mcpServers in wrong place
mcpServers is not a claude code settings.json directive. The mcp servers belong in a new file .mcp.json coupled with a `enabledMcpjsonServers` directive in settings.local.json.
https://github.com/ruvnet/claude-flow/pull/238#issuecomment-3067448714
[Read the thread](https://github.com/ruvnet/ruflo/issues/251) · 2025-07-14 · closed · external user · 12 comments
### orchestrator stars with no access to claude
if claude is not in the path, ocestrator starts succesfully while not really working i guess. monitor command does not see it.
after fixing the path it is looking the same but actually started working.
this is the output with no claude in path available:
```bash
✘ vasiliybondarenko@MBP16 ~/code/AntiSotka ↱ main ±✚ npx claude-flow start
✅ Starting Claude-Flow Orchestration System...
🚀 System Configuration:
Mode: Interactive
MCP Port: 3000
Working Directory: /Users/vasiliybondar
[Read the thread](https://github.com/ruvnet/ruflo/issues/21) · 2025-06-12 · closed · external user · 8 comments
## Most recent
### feat(adr): expose ADR creation via first-class CLI/MCP (currently Claude Code plugin-only)
## Summary
Ruflo has ADR creation support, but the supported automated creation path is only exposed through the Claude Code `ruflo-adr` plugin (`/adr-create` or the plugin's `adr create` prompt command). There is no first-class Ruflo CLI command or MCP tool, so Codex, other MCP clients, CI, and ordinary shells cannot use the same lifecycle operation.
Manual creation of a Markdown file is possible, but it bypasses the plugin's sequential numbering, standard template, AgentDB registration, rela
[Read the thread](https://github.com/ruvnet/ruflo/issues/3112) · 2026-08-27 · open · external user · 1 comment
### Optional Parallel Search MCP setup for the self-hosted Web UI
Hey! Would you be open to a small Parallel Search MCP integration for RuVocal?
I was reading through the [bridge's web research path](https://github.com/ruvnet/ruflo/blob/e21aa352fdc80fd2d3cc4e83404a76a18d118b96/ruflo/src/ruvocal/mcp-bridge/index.js#L979-L997). When there's no research Cloud Function configured, it falls back to Google, and [that path needs `GOOGLE_API_KEY`](https://github.com/ruvnet/ruflo/blob/e21aa352fdc80fd2d3cc4e83404a76a18d118b96/ruflo/src/ruvocal/mcp-bridge/index.js#L653-
[Read the thread](https://github.com/ruvnet/ruflo/issues/3106) · 2026-08-26 · open · external user · 0 comments
### YAML parse error in plugins/ruflo-metaharness/skills/harness-gepa/SKILL.md — unquoted colon in description breaks `npx skills add`
YAML parse error in plugins/ruflo-metaharness/skills/harness-gepa/SKILL.md — unquoted colon in description breaks `npx skills add`
Corpo:
## Summary
`npx skills add ruvnet/ruflo --all` fails to install the `harness-gepa` skill (labeled as `Eve` in the plugin manifest) because its YAML frontmatter contains an unquoted colon inside the `description` field, which the parser interprets as a nested mapping.
## Environment
- OS: Windows 11
- Shell: PowerShell
- Node: v22+ (via `npx`)
- Command: `
[Read the thread](https://github.com/ruvnet/ruflo/issues/3065) · 2026-08-19 · closed · external user · 0 comments
### CLAUDE_FLOW_MCP_TOOLS env var ignored by 'mcp tools' (v3.38.12)
## Description
`claude-flow doctor` suggests setting `CLAUDE_FLOW_MCP_TOOLS` to restrict advertised MCP tool categories and reduce schema overhead:
MCP Schema Overhead: 333 advertised tools ≈ 61550 schema tokens (all tools) Suggested fix: Set CLAUDE_FLOW_MCP_TOOLS to required categories/tool names (for example: memory,swarm,agent,hooks)
Setting the variable has no effect — `mcp tools` still lists all 333 tools across all 26 categories as `Enabled`.
## Steps to reproduce
```bash
CLAU
[Read the thread](https://github.com/ruvnet/ruflo/issues/3055) · 2026-08-18 · closed · external user · 0 comments
### memory_store (MCP tool): tags parameter accepted but never persisted
## Summary
The `memory_store` MCP tool accepts a `tags` array parameter (per its own tool schema: `"tags": {"description": "Optional tags for filtering", "items": {"type": "string"}, "type": "array"}`) and returns `success: true`, but the tags are never written to storage. Every subsequent read (`memory_retrieve`, `memory_list`) shows an empty/absent tags value, and the raw SQLite `tags` column is empty for the row.
## Environment
- Package: `@claude-flow/memory` v3.0.0-alpha.22 (resolved via `
[Read the thread](https://github.com/ruvnet/ruflo/issues/3051) · 2026-08-17 · closed · external user · 1 comment
### agentTeams teammateIdle autoAssign: no scope/ownership check, no refusal back-off — pressed write-heavy tasks onto read-only agents 20+ times in one session
## Summary
The `agentTeams.coordination.autoAssignOnIdle` / `hooks.teammateIdle.autoAssign` mechanism (v3 CLI, observed on 4.x this week) pairs "idle agent" with "unclaimed task" purely on liveness. In one working session we observed it:
- fire **20+ times across 5 different subagents**, byte-identical instruction each time, with **no back-off after an explicit reasoned refusal** (one agent declined the same assignment 8 times; the hook re-fired immediately each time);
- assign **implementatio
[Read the thread](https://github.com/ruvnet/ruflo/issues/3031) · 2026-08-14 · closed · outside contributor · 1 comment
### mcp start -t http (3.38.2, post-#2984): initialize returns protocolVersion as object — spec-validating clients reject the handshake; http tools/list serves only 4 system tools
## Summary
Follow-up to #2984 (fixed in 3.38.2 via #2987 — thank you, the process now stays alive and the port binds). Testing 3.38.2 with a **real MCP client** (a chat-ui fork using the MCP SDK) rather than curl surfaces two remaining defects that make the HTTP transport unusable in practice, plus one minor bind quirk:
1. **`initialize` returns `protocolVersion` as an object; the MCP spec requires a string.** Spec-validating clients reject the handshake outright.
2. **`tools/list` over HTTP s
[Read the thread](https://github.com/ruvnet/ruflo/issues/2990) · 2026-08-12 · closed · external user · 2 comments
### migrate status (ADR-382 Part B) suggests `ruflo plugins install <plugin>` — wrong subsystem and wrong syntax; the printed fix cannot succeed
## Summary
v3.38.0's new removed-agent detection (#2973, ADR-382 Part B) works exactly as promised — but the remediation command it prints is broken twice over. Verified on a fresh v3.38.0 install (macOS, Node 26):
$ ruflo migrate status Removed Agents (ADR-128) Agent Owning Plugin Install coder.md ruflo-core ruflo plugins install ruflo-core …
$ ruflo plugins install ruflo-core [ERROR] Required option missing: --name
1. **Wrong syntax** — `plugins install` takes `
[Read the thread](https://github.com/ruvnet/ruflo/issues/2985) · 2026-08-12 · closed · outside contributor · 0 comments
### mcp start -t http prints a full success table (Status: Running, health/RPC URLs) then exits 0 without ever binding the port
## Version / environment
- ruflo v3.38.0 (npm global install, pinned)
- macOS (Darwin 25.6.0), node v22.23.0
## Reproduce
$ ruflo mcp start -t http --port 39217
[INFO] Starting MCP Server...
Initializing server... Loading tool registry... Server started in 62.35ms
+------------+------------+ | Property | Value | +------------+------------+ | Server PID | 77164 | | Transport | http | | Host | localhost | | Port | 39217 | | Tools | 27 enabled
Read the thread · 2026-08-12 · closed · external user · 0 comments
memory-bridge.js bridgeSearchEntries(): LIMIT 1000 with no ORDER BY silently truncates search to oldest rows
Summary
bridgeSearchEntries() in @claude-flow/cli's memory-bridge.js (part of the ruflo package published from this repo, ruflo/ directory) runs a query of the shape:
SELECT id, key, namespace, content, embedding
FROM memory_entries WHERE status='active' ${nsFilter}
LIMIT 1000
with no ORDER BY, before any BM25/embedding ranking is applied. SQLite returns rows in whatever engine-chosen order it happens to store/scan them in — not recency. Once a namespace's `memor
Read the thread · 2026-08-12 · closed · external user · 1 comment
@claude-flow/memory registry references 8 agentdb exports that no installable agentdb provides — native controllers silently dead on every fresh install
Found while verifying the #2887 fix (#2966) on a fresh 3.38.0 install — the fix works, the tiered fallback stores durably and reports honestly (fallbackFrom metadata is genuinely useful). But the investigation showed the native controller paths are permanently unreachable, at a larger scale than #2887 suggested.
The problem
@claude-flow/memory (3.0.0-alpha.22) controller-registry.js references 18 exports from agentdb:
AgentDB, AttestationLog, BatchOperations, CausalMemory
[Read the thread](https://github.com/ruvnet/ruflo/issues/2977) · 2026-08-11 · open · external user · 2 comments
### [BUG] memory_search_unified candidate scan uses unordered LIMIT 1000, missing recently written rows on larger DBs
## Summary
`memory_search_unified` (the unified memory search MCP tool in `ruflo` v3.33.0) silently fails to return newly written rows when operating on databases containing more than 1,000 entries. Because `bridgeSearchEntries` collects candidates using an unordered SQL query capped at `LIMIT 1000`, the candidate window saturates with older rows based on SQLite insertion order.
As a result, queries return plausible-looking historical entries without raising errors, creating a false-sight
[Read the thread](https://github.com/ruvnet/ruflo/issues/2976) · 2026-08-11 · closed · external user · 2 comments
### `github_issue_track` `create` silently falls back to local-only storage when `gh` is unavailable, and its `gh`-backed real path ignores the supplied `owner`/`repo`
## Environment
- `ruflo` (npm) v3.34.0, wrapping `@claude-flow/cli` v3.34.0
- No `gh` CLI installed, no `GITHUB_TOKEN`/`GH_TOKEN` set
## Reproduction
Call the github_issue_track MCP tool:
{ "action": "create", "owner": "ruvnet", "repo": "claude-flow",
"title": "...", "body": "..." }
With no `gh` CLI present, this returns:
```json
{
"success": true,
"source": "local-store",
"action": "created",
"issue": {
"id": "issue-1786319660352",
"title": "...",
"status":
[Read the thread](https://github.com/ruvnet/ruflo/issues/2963) · 2026-08-10 · closed · external user · 1 comment
### Explicit provider/model selection (providers configure, agent spawn --provider/--model) does not propagate into actual agent execution
## Environment
- `ruflo` (npm) v3.34.0, wrapping `@claude-flow/cli` v3.34.0
- Ollama 0.32.6, local model `qwen3.6:27b`, reachable at `http://127.0.0.1:11434`
- macOS; no `ANTHROPIC_API_KEY`/`OLLAMA_API_KEY`/`OPENROUTER_API_KEY` set
## Summary
`ruflo providers configure` and `ruflo agent spawn --provider/--model` both succeed and look like they've set up local-Ollama execution — but neither the persisted provider config nor a non-alias `--model` value (a literal Ollama tag) reaches the code th
[Read the thread](https://github.com/ruvnet/ruflo/issues/2962) · 2026-08-10 · closed · external user · 2 comments
### `ruflo hooks post-task` (CLI) drops the ADR-130 "reinforced-by" graph edge that the MCP `hooks_post-task` tool writes — root cause is an unawaited fire-and-forget write racing the CLI's explicit `process.exit(0)`
**Package / versions checked:** `ruflo` (latest published: 3.36.0) → depends on `@claude-flow/cli@^3.33.0`, which resolves to `@claude-flow/cli` latest published 3.36.0. Verified against a fresh `npm pack` of both tarballs (not the locally installed copy, not vendored source).
## Summary
`ruflo hooks post-task` and the MCP tool `hooks_post-task` are **the same code path** — the CLI command literally calls `callMCPTool('hooks_post-task', …)`, which looks the tool up in the shared `TOOL_REGISTRY
[Read the thread](https://github.com/ruvnet/ruflo/issues/2961) · 2026-08-10 · closed · external user · 1 comment
### Codex DualModeOrchestrator leaves worker stdin open, causing `codex exec` to wait for EOF
## Summary
`@claude-flow/codex` DualModeOrchestrator launches Codex workers with stdin
configured as a pipe but never writes to or closes that pipe.
The prompt is already passed positionally to `codex exec`.
On a real Codex execution this leaves the worker blocked waiting for stdin EOF.
## Environment
- Ruflo: 3.34.0
- bundled `@claude-flow/codex`: 3.0.3
- Codex CLI: 0.146.0
- Node: 22.23.1
- macOS
- DualModeOrchestrator
- real Codex model/provider routes
## Source
In:
`v3/@claude-flow/co
[Read the thread](https://github.com/ruvnet/ruflo/issues/2947) · 2026-08-08 · closed · external user · 1 comment
The remaining reports are on [the project's issue tracker](https://github.com/ruvnet/claude-flow/issues).