agentic-flow 3.0.0-alpha.2 → 3.1.0

package/.claude/statusline.mjs

@@ -0,0 +1,126 @@
+/**
+ * Agentic Flow Statusline for Claude Code
+ * Shows model, tokens, cost, swarm status, and memory usage
+ */
+
+import { execFileSync } from 'child_process';
+import { existsSync } from 'fs';
+import { join } from 'path';
+
+// Cache for expensive operations
+let lastSwarmCheck = 0;
+let cachedSwarmStatus = null;
+const CACHE_TTL = 5000; // 5 seconds
+
+/**
+ * Get swarm status (cached)
+ */
+function getSwarmStatus() {
+  const now = Date.now();
+  if (cachedSwarmStatus && (now - lastSwarmCheck) < CACHE_TTL) {
+    return cachedSwarmStatus;
+  }
+
+  // Ask the project's own install only. `npx agentic-flow@<tag>` here ran
+  // whatever the tag pointed at (the deprecated alpha) on every refresh.
+  const bin = join(process.cwd(), 'node_modules', '.bin',
+    process.platform === 'win32' ? 'agentic-flow.cmd' : 'agentic-flow');
+  if (!existsSync(bin)) {
+    cachedSwarmStatus = '⚡';
+    lastSwarmCheck = now;
+    return cachedSwarmStatus;
+  }
+
+  try {
+    const result = execFileSync(bin, ['mcp', 'status'], {
+      encoding: 'utf-8',
+      timeout: 2000,
+      stdio: ['ignore', 'pipe', 'ignore']
+    }).trim();
+
+    cachedSwarmStatus = result.includes('running') ? '🐝' : '⚡';
+    lastSwarmCheck = now;
+    return cachedSwarmStatus;
+  } catch {
+    cachedSwarmStatus = '⚡';
+    lastSwarmCheck = now;
+    return cachedSwarmStatus;
+  }
+}
+
+/**
+ * Format token count
+ */
+function formatTokens(tokens) {
+  if (tokens >= 1000000) {
+    return `${(tokens / 1000000).toFixed(1)}M`;
+  }
+  if (tokens >= 1000) {
+    return `${(tokens / 1000).toFixed(1)}K`;
+  }
+  return String(tokens);
+}
+
+/**
+ * Format cost
+ */
+function formatCost(cost) {
+  if (cost >= 1) {
+    return `$${cost.toFixed(2)}`;
+  }
+  return `$${cost.toFixed(4)}`;
+}
+
+/**
+ * Main statusline export
+ */
+export default function statusline(context) {
+  const parts = [];
+
+  // Agentic Flow indicator
+  parts.push('🤖');
+
+  // Model name (shortened)
+  // Claude Code passes `model` as { id, display_name }; older callers pass a string.
+  const modelName = typeof context.model === 'string'
+    ? context.model
+    : context.model?.display_name ?? context.model?.id;
+  if (typeof modelName === 'string' && modelName) {
+    const model = modelName
+      .replace('claude-', '')
+      .replace('-20250514', '')
+      .replace('sonnet-4', 'S4')
+      .replace('opus-4', 'O4')
+      .replace('haiku-3.5', 'H3.5');
+    parts.push(model);
+  }
+
+  // Token usage
+  if (context.inputTokens !== undefined || context.outputTokens !== undefined) {
+    const input = formatTokens(context.inputTokens || 0);
+    const output = formatTokens(context.outputTokens || 0);
+    parts.push(`↑${input} ↓${output}`);
+  }
+
+  // Cost
+  if (context.totalCost !== undefined && context.totalCost > 0) {
+    parts.push(formatCost(context.totalCost));
+  }
+
+  // Swarm/MCP status indicator
+  parts.push(getSwarmStatus());
+
+  // Session time
+  if (context.sessionStartTime) {
+    const elapsed = Math.floor((Date.now() - context.sessionStartTime) / 1000);
+    const mins = Math.floor(elapsed / 60);
+    const secs = elapsed % 60;
+    if (mins > 0) {
+      parts.push(`${mins}m${secs}s`);
+    } else {
+      parts.push(`${secs}s`);
+    }
+  }
+
+  return parts.join(' │ ');
+}

package/.claude/statusline.sh

@@ -0,0 +1,74 @@
+#!/bin/bash
+# Agentic Flow Intelligence Status Line
+# Shows real-time learning metrics from SONA + HNSW + ReasoningBank
+
+INTEL_FILE=".agentic-flow/intelligence.json"
+INTEL_DB=".agentic-flow/intelligence.db"
+
+# Default values
+PATTERNS=0
+MEMORIES=0
+ROUTES=0
+TRAJECTORIES=0
+LEARNING_RATE=10
+EPSILON=10
+
+# Get current git branch
+GIT_BRANCH=""
+if git rev-parse --is-inside-work-tree >/dev/null 2>&1; then
+  GIT_BRANCH=$(git branch --show-current 2>/dev/null || echo "")
+fi
+
+# Read intelligence metrics
+if [ -f "$INTEL_FILE" ]; then
+  PATTERNS=$(jq -r '.patterns | length // 0' "$INTEL_FILE" 2>/dev/null || echo "0")
+  MEMORIES=$(jq -r '.memories | length // 0' "$INTEL_FILE" 2>/dev/null || echo "0")
+  ROUTES=$(jq -r '.metrics.totalRoutes // 0' "$INTEL_FILE" 2>/dev/null || echo "0")
+  TRAJECTORIES=$(jq -r '.trajectories | length // 0' "$INTEL_FILE" 2>/dev/null || echo "0")
+
+  # Get learning parameters
+  LR=$(jq -r '.config.learningRate // 0.1' "$INTEL_FILE" 2>/dev/null || echo "0.1")
+  EPS=$(jq -r '.config.epsilon // 0.1' "$INTEL_FILE" 2>/dev/null || echo "0.1")
+  # Values come from a file in the working tree: accept plain decimals only, and
+  # pass them to awk as data (-v), never as program text.
+  NUM_RE='^[0-9]*\.?[0-9]+$'
+  [[ "$LR" =~ $NUM_RE ]] || LR="0.1"
+  [[ "$EPS" =~ $NUM_RE ]] || EPS="0.1"
+  LEARNING_RATE=$(awk -v v="$LR" 'BEGIN {printf "%.0f", v * 100}' 2>/dev/null || echo "10")
+  EPSILON=$(awk -v v="$EPS" 'BEGIN {printf "%.0f", v * 100}' 2>/dev/null || echo "10")
+
+  # Get active count
+  ACTIVE=$(jq -r '[.trajectories // {} | to_entries[] | select(.value.status == "active")] | length' "$INTEL_FILE" 2>/dev/null || echo "0")
+fi
+
+# Attention mechanism status (check which are enabled)
+NEURAL="●Neural"
+DAG="◐DAG"
+GRAPH="○Graph"
+SSM="●SSM"
+
+# Check DB for vector count
+VECTOR_COUNT=0
+if [ -f "$INTEL_DB" ]; then
+  VECTOR_COUNT=$(sqlite3 "$INTEL_DB" "SELECT COUNT(*) FROM vectors;" 2>/dev/null || echo "0")
+fi
+
+# Build output
+OUTPUT=""
+
+# Line 1: Model + Project + Branch
+OUTPUT="Opus 4.5 in agentic-flow"
+if [ -n "$GIT_BRANCH" ]; then
+  OUTPUT="${OUTPUT} on ⎇ ${GIT_BRANCH}"
+fi
+
+# Line 2: Agentic Flow metrics
+OUTPUT="${OUTPUT}\n🧠 Agentic Flow ◆ ${PATTERNS} patterns ⬡ ${MEMORIES} mem ↝${ROUTES} #${TRAJECTORIES}"
+
+# Line 3: Agent routing/learning metrics
+OUTPUT="${OUTPUT}\n🎯 Agent Routing q-learning lr:${LEARNING_RATE}% ε:${EPSILON}%"
+
+# Line 4: Swarm coordination mechanisms
+OUTPUT="${OUTPUT}\n⚡ Swarm: ●Neural ◐DAG ○Graph ●SSM"
+
+printf "%b\n" "$OUTPUT"

package/CHANGELOG.md

@@ -2,9 +2,178 @@
 
 All notable changes to this project will be documented in this file.
 
+## [3.1.0] - 2026-10-31
+
+Fixes two correctness bugs left after 3.0.0 and brings the docs in line with its proxy auth. See ADR-080.
+
+### Changed
+- **One embedding space per process (#246).** A failed model load used to be retried on the next call, so one process could hold hash-fallback vectors and model vectors side by side and compare them (cosine of the same text across the two was ~0.03). The first embedding now fixes the space:
+  - a failed load is **not retried automatically**; the process stays on hash vectors and `isSemanticEmbedding()` stays false;
+  - if the model fails on a text **after model vectors were produced**, `computeEmbedding`/`computeEmbeddingBatch` **throw** instead of returning a hash vector. `retrieveMemories` and `distillMemories` propagate that error;
+  - new `resetEmbeddingBackend()` clears the choice and the cache so the model is tried again (re-embed anything you kept); new `getEmbeddingKind()` returns `'local'`, `'hash'` or `null`.
+
+### Security
+- **`.claude/statusline.mjs` ran `npx agentic-flow@alpha mcp status` on every refresh**, which executes the deprecated March 3.0.0-alpha line (and its critical `protobufjs` advisory). It now asks only the project's own `node_modules/.bin/agentic-flow`, without a shell, and shows the idle marker when there is none.
+- **`hooks init` pre-approved `Bash(npx:*)`, `Bash(npm run:*)` and the whole `agentic-flow`, `claude-flow` and `ruv-swarm` MCP servers**, so any package or tool ran without a prompt. It now allows only `Bash(agentic-flow:*)`, as `init` has since 3.0.0. Re-run `hooks init --force` in existing projects.
+- **`hooks init` wrote unversioned `npx agentic-flow …` hooks**, which resolve to whatever `@latest` is that day. They now pin the installed version, as `init` has since 3.0.0. The `swarm` commands print a pinned `swarm connect` line instead of `@alpha`, and `scripts/validate-providers.sh`, the versioned `docs/docker-tests` Dockerfiles and the worker skills pin too. The version comes from a shared `src/utils/package-version.ts`, which throws instead of falling back to a tag.
+
+### Fixed
+- `.claude/statusline.mjs` no longer crashes when Claude Code passes `model` as `{ id, display_name }`; it shows `display_name ?? id` and still accepts a string.
+- The Gemini and Requesty proxies' streams no longer drop an SSE line split across reads or corrupt a multibyte character split across chunks (3.0.0 fixed only OpenRouter). All three now share `src/proxy/sse-lines.ts`; the Gemini proxy also strips CR from its CRLF lines.
+
+### Docs
+- The proxy guides (`STANDALONE_PROXY_GUIDE`, `CLAUDE-CODE-INTEGRATION`, `PROXY-ARCHITECTURE-AND-EXTENSION`, both copies) and the shipped Requesty plans sent a dummy key that the 3.0.0 proxies reject. They now use the token the proxy prints and document `AGENTIC_FLOW_PROXY_TOKEN` and `AGENTIC_FLOW_PROXY_HOST`; the extension guide's template uses the `proxy-auth` helpers.
+- The READMEs and the skills in `.claude/skills` ran `npx agentic-flow@alpha`, which resolves to the March 3.0.0-alpha line. They pin `agentic-flow@3.1.0`, and a test keeps that pin equal to the package version.
+
+## [3.0.0] - 2026-10-31
+
+First stable 3.x. It is the main line (2.1.5) plus a security and correctness review; it supersedes 3.0.0-alpha.1/.2, which were built from a different tree. Published without an alpha dist-tag (ruflo ADR-513/515).
+
+### Breaking
+- **Agents no longer bypass permissions.** By default only Read/Glob/Grep/TodoWrite are auto-approved; Bash/Write/Edit/WebFetch are denied in headless runs, including the `agentic_flow_agent` MCP tool. The old default let a prompt injection reach an unprompted Bash agent through that tool, which `init` also auto-approved. Opt back in with `AGENTIC_FLOW_ALLOW_BYPASS_PERMISSIONS=1` (logs a warning).
+- **Translation proxies (OpenRouter, Gemini, Requesty, Ollama) bind `127.0.0.1` and require a per-run token** as `x-api-key` or `Authorization: Bearer`. They used to listen on every interface with no auth, so anyone on the network could spend your provider key. The CLI hands the token to the agent automatically; the standalone `proxy` command prints it. New: `AGENTIC_FLOW_PROXY_TOKEN`, `AGENTIC_FLOW_PROXY_HOST`. Code that constructs `AnthropicTo*Proxy` directly must send `getAuthToken()`. `/health` stays open.
+- **The HTTP/SSE MCP server binds `127.0.0.1` and requires a token.** New: `AGENTIC_FLOW_MCP_TOKEN` (generated and printed when unset), `AGENTIC_FLOW_MCP_HOST`, `AGENTIC_FLOW_MCP_PORT`, `AGENTIC_FLOW_MCP_ALLOWED_ORIGINS` (no browser origins by default).
+- **`init`** pins hooks and the MCP server to the installed version (it ran `npx agentic-flow@alpha`), no longer pre-approves `npx`, `npm run` or whole MCP servers, and registers the server as `agentic-flow` (tool ids become `mcp__agentic-flow__*`). Re-run `init --force` in existing projects. It no longer copies `.claude/mcp.json`, which pointed at a developer-machine path.
+- **Only `./.env` is read**, never a parent directory's. Endpoint keys (`*_BASE_URL`, `*_PROXY_*`, `HTTP(S)/ALL_PROXY`) and `AGENTIC_FLOW_ALLOW_*`/`AGENTIC_FLOW_MCP_*` are ignored from it with a warning: a cloned repo's `.env` could otherwise send your API key to its own endpoint. Export them, or pass `--env-file <path>`.
+- **Router config** is read only from an explicit path, `AGENTIC_FLOW_ROUTER_CONFIG` or `~/.agentic-flow/router.config.json`, not from the working directory.
+- Agent and command names are `[a-z0-9-]`; categories are such segments joined by `/`. The agent-booster file tools only touch files inside the working directory. The `readonly` MCP profile no longer includes `agent_booster_parse_markdown`.
+- QUIC/WebSocket transport defaults to `127.0.0.1`, caps frames at 1 MiB, and `receive()` no longer returns messages an `onMessage` handler already consumed.
+- Embeddings under `npx` are real: the npx heuristic that silently returned hash vectors is gone, so the first use downloads the MiniLM model (~23 MB). With no network it falls back to hash vectors as before; `isSemanticEmbedding()` says which you got.
+
+### Security
+- `statusline.sh` (copied by `init`) executed code from a repo's `.agentic-flow/intelligence.json` through awk; values are now validated. `github-safe.js` passed arguments through a shell; it now uses `execFileSync`.
+- Path traversal in the agent and command creation tools, and arbitrary file writes in the agent-booster edit tools, are closed.
+- `certs` and `.claude/settings*.json` (which set `defaultMode: bypassPermissions`) are no longer published.
+
+### Fixed
+- **Self-learning that actually learns.** ReasoningBank moves the confidence of the memories a task used toward that task's verdict and decays them on failure; retrieval no longer counts as a reward, and pruning considers low confidence regardless of use. agent-booster counts success only from `recordOutcome`, not from cache hits, and no longer returns another file's cached edit as a success.
+- **AutopilotLearning is back** (`dist/coordination/autopilot-learning.js`). It shipped in 3.0.0-alpha.2 from an uncommitted tree and was missing from main; it now loads the real `agentdb` and reports unavailable honestly when it cannot.
+- `computeEmbedding` no longer keeps the process alive for up to an hour (an unref'd LRU, no per-entry timers); `computeEmbeddingBatch` batches (256 texts: ~1.3 s → ~0.5 s).
+- The OpenRouter proxy's stream no longer drops a line split across reads or corrupts multibyte characters.
+- The SONA service imports without `@ruvector/sona` and without crashing in ESM.
+- `ModelRouter.stream()` records metrics and falls back on a provider error before the first chunk; cost-optimal routing no longer mutates the caller's params.
+- ReasoningBank logs go to stderr, not stdout.
+
+### Dependencies
+- `agentdb` ^3.0.0 (stable), `@huggingface/transformers` ^4.2 (3.8 pinned onnxruntime-node 1.21, whose install script fails on CUDA 13 hosts), `ruvector` ^0.3.3, `fastmcp` ^3.35 (3.21 crashed on every session with MCP SDK 1.30), `sharp` ^0.35.5.
+
+## [2.1.5] - 2026-10-26
+
+### Security
+
+- `FederationHubServer` now verifies the agent token (HMAC JWT, secret from `jwtSecret` or `JWT_SECRET`) and takes agent and tenant identity from it; it used to accept any non-empty token. The hub also listens on `127.0.0.1` unless `host` is set. **Breaking for hub users:** agents need a token signed with the hub's secret, so set `JWT_SECRET` on both sides (#242).
+
+## [2.1.0] - 2026-06-23
+
+### Added
+
+- **Cost-optimal model routing (ADR-073).** New `CostOptimalRouter`
+  (`src/router/cost-optimal-router.ts`) wrapping
+  [`@metaharness/router`](https://www.npmjs.com/package/@metaharness/router):
+  routes each query to the _cheapest model predicted to clear a quality bar_,
+  learned from eval logs via k-NN / kernel-ridge (optional native FastGRNN via
+  the already-bundled `@ruvector/tiny-dancer`). Build from a flat
+  `(embedding → per-model quality)` dataset or explicit candidates; resolves
+  `"<provider>/<model>"` ids to concrete provider/model bindings.
+- **`ModelRouter.enableCostOptimalRouting({ router, embed })`** adds a new
+  `'cost-optimal'` routing mode that embeds the query, picks the cost-optimal
+  model, and steers `params.model` — with a graceful fallback to the existing
+  cost heuristic when unconfigured or on error (routing never hard-fails). The
+  previous rule-based `'cost-optimized'` mode is unchanged.
+- Bounded LRU embedding cache on the cost-optimal hot path (embedding dominates
+  latency vs the µs-scale k-NN; recurring prompts are common).
+- Benchmark (`benchmarks/cost-optimal-router-benchmark.mjs`) and tests
+  (`tests/router/cost-optimal-router.test.ts`). Measured on a 3-tier lineup
+  (1000-query held-out test, bar=0.8): **28.5% cheaper than always-opus** while
+  holding **98.1%** of queries at/above the bar; routing latency **p50 73µs /
+  p99 125µs**.
+- **Autonomous repair via Darwin Mode (ADR-074).** New `repair()` wrapper
+  (`src/repair/darwin-repair.ts`) and `agentic-flow-repair` CLI over
+  [`@metaharness/darwin`](https://www.npmjs.com/package/@metaharness/darwin):
+  freeze the model and evolve the harness (planner/context/reviewer/retry/tool/
+  memory/score policy), keeping only variants that measurably improve under a
+  frozen scorer + safety gate. Defaults to Test-Driven Repair (`'real'` sandbox —
+  the repo's own tests gate promotion, run shell-free with a scrubbed env);
+  `'mock'` is a deterministic, Docker-free substrate for hermetic tests. The full
+  SWE-bench-Lite TDR product (Docker grading) is run via Darwin's own CLI — see
+  ADR-074. New exports `agentic-flow/repair` and `agentic-flow/router/cost-optimal`.
+- **Harness evolution MCP tools + provenance (ADR-075).** New `harness_repair`,
+  `harness_manifest`, and `harness_verify` MCP tools (`src/mcp/tools/harness-tools.ts`,
+  registered in the stdio server), mirroring claude-flow's `metaharness_*`. Plus an
+  Ed25519 **witness manifest** module (`src/harness/provenance.ts`, exported as
+  `agentic-flow/harness/provenance`): sign a sha256 manifest over harness/agent
+  config files and verify it + on-disk drift — built on Node's `crypto`, no deps.
+- **Repositioned as the agentic meta-harness (ADR-076)** — README hero +
+  four-pillar framing (route / evolve / orchestrate / verify), `package.json`
+  description and keywords. Docs/positioning only; no behavior change.
+- ADR-073/074/075/076 documenting the metaharness integration (all implemented).
+
+### Dependencies
+
+- Added `@metaharness/router` and `@metaharness/darwin` (both dependency-free;
+  optional `@ruvector/tiny-dancer` peer already present).
+
+## [2.0.15] - 2026-06-23
+
+### Fixed
+
+- **#167 — `agentic-flow/agent-booster` subpath export pointed at a missing file.**
+  The `./agent-booster` export resolved to `dist/agent-booster/index.js`, which
+  was never emitted (the booster code lives under `intelligence/` and
+  `optimizations/`), so `import 'agentic-flow/agent-booster'` failed with
+  `ERR_MODULE_NOT_FOUND`. Added the missing barrel entrypoint
+  (`src/agent-booster/index.ts`) which re-exports the booster API and provides
+  the documented `AgentBooster` name (alias of `EnhancedAgentBooster`).
+- **#162 — WebSocket fallback transport: stale-route `EHOSTUNREACH` after idle
+  on a direct point-to-point link.** `WebSocketFallbackTransport` now reuses an
+  already-open inbound (server-accepted) socket for replies before dialing a
+  fresh outbound connection — the full-duplex inbound direction stays healthy
+  when a new outbound dial would hit a per-process stale scoped route. Added a
+  liveness ping/pong heartbeat that terminates stale sockets, driven by
+  `maxIdleTimeoutMs` (previously accepted but unused — ping cadence is 1/6 of
+  it, default 30000ms → 5000ms).
+
+### Notes
+
+- **#146 (Ollama provider in config-wizard)** was already resolved in the 2.0.x
+  line: `OLLAMA_API_KEY`, `OLLAMA_BASE_URL`, and `'ollama'` are accepted by the
+  config wizard. The remaining asks in that issue (agentdb controller
+  prerequisites) are tracked in the `agentdb` package.
+
+## [2.0.1-alpha.4] - 2025-12-03
+
+### Added
+
+- **SONA v0.1.4 Federated Learning Integration**: Complete integration with AgentDB
+  - Updated AgentDB dependency to 2.0.0-alpha.2.16
+  - Full support for `EphemeralLearningAgent`, `FederatedLearningCoordinator`, and `FederatedLearningManager`
+  - Quality-based filtering and weighted aggregation
+  - Large-scale federation (50+ agents with configurable limits)
+
+### Changed
+
+- **Dependencies Updated**:
+  - `agentdb`: 2.0.0-alpha.2.15 → 2.0.0-alpha.2.16 (SONA v0.1.4 federated learning)
+
+### Documentation
+
+- Comprehensive federated learning guide available in AgentDB package
+- 5 detailed use cases for distributed learning
+- API documentation and performance tuning recommendations
+
+### Tested
+
+- ✅ Complete federated learning workflow with 50+ agents
+- ✅ Quality filtering and weighted consolidation
+- ✅ Multi-agent coordination and automatic aggregation
+- ✅ All test suites passing
+
+---
+
 ## [1.10.2] - 2025-01-10
 
 ### Fixed
+
 - **Critical Bug**: Fixed ANTHROPIC_API_KEY overriding `--provider` CLI argument ([#60](https://github.com/ruvnet/agentic-flow/issues/60))
   - CLI arguments (`--provider`, `--openrouter-key`, `--anthropic-key`, `--model`) now properly propagate to environment variables
   - Removed silent fallback to `ANTHROPIC_API_KEY` for non-Anthropic providers (OpenRouter, Gemini, Requesty)
@@ -12,11 +181,13 @@ All notable changes to this project will be documented in this file.
   - CLI arguments now correctly override environment variables as expected
 
 ### Added
+
 - Comprehensive test suite for provider CLI argument handling
 - Clear error messages when provider-specific API keys are missing
 - Logging for provider selection from CLI arguments
 
 ### Changed
+
 - **Breaking Change**: Providers now require their specific API keys (no fallback to ANTHROPIC_API_KEY)
   - OpenRouter requires `OPENROUTER_API_KEY` or `--openrouter-key`
   - Gemini requires `GOOGLE_GEMINI_API_KEY`
@@ -24,7 +195,9 @@ All notable changes to this project will be documented in this file.
   - Anthropic continues to require `ANTHROPIC_API_KEY` or `--anthropic-key`
 
 ### Migration Guide
+
 If you were relying on the fallback behavior:
+
 ```bash
 # Before (relied on ANTHROPIC_API_KEY fallback)
 export ANTHROPIC_API_KEY=sk-ant-...
@@ -39,4 +212,5 @@ npx agentic-flow --provider openrouter --openrouter-key sk-or-... --task "test"
 ```
 
 ## [1.10.0] - Previous Release
+
 - See git history for previous changes