@swarmclawai/swarmclaw 0.8.1 → 0.8.2

package/README.md

@@ -148,7 +148,7 @@ curl -fsSL https://raw.githubusercontent.com/swarmclawai/swarmclaw/main/install.
 ```
 
 The installer resolves the latest stable release tag and installs that version by default.
-To pin a version: `SWARMCLAW_VERSION=v0.8.1 curl ... | bash`
+To pin a version: `SWARMCLAW_VERSION=v0.8.2 curl ... | bash`
 
 Or run locally from the repo (friendly for non-technical users):
 
@@ -701,7 +701,7 @@ npm run update:easy     # safe update helper for local installs
 SwarmClaw uses tag-based releases (`vX.Y.Z`) as the stable channel.
 
 ```bash
-# example minor release (v0.8.1 style)
+# example minor release (v0.8.2 style)
 npm version minor
 git push origin main --follow-tags
 ```
@@ -711,13 +711,14 @@ On `v*` tags, GitHub Actions will:
 2. Create a GitHub Release
 3. Build and publish Docker images to `ghcr.io/swarmclawai/swarmclaw` (`:vX.Y.Z`, `:latest`, `:sha-*`)
 
-#### v0.8.1 Release Readiness Notes
+#### v0.8.2 Release Readiness Notes
 
-Before shipping `v0.8.1`, confirm the following user-facing changes are reflected in docs:
+Before shipping `v0.8.2`, confirm the following user-facing changes are reflected in docs:
 
-1. Release notes call out notification deduping and occurrence counts so operators know repeated health/approval events now refresh one notification instead of piling up duplicates.
-2. Memory/tooling docs mention the new current-thread recall rule: same-thread recall should stay local to the chat, while direct memory writes should go straight to `memory_store` or `memory_update` rather than detouring through agent/task/delegation flows.
-3. Site and README install/version strings are updated to `v0.8.1`, including install snippets, release notes index text, and sidebar/footer labels.
+1. Runtime/defaults docs mention the higher default agent recursion limit so long-running bounded turns get more headroom without custom tuning.
+2. Memory/tooling docs mention the narrower direct-memory-write routing: remember-and-confirm turns stay on `memory_store`/`memory_update`, bundled related facts should be stored as one canonical write, and same-thread recall should not steal those turns.
+3. File-output guidance notes that exact bullet-count and titled-section constraints are now treated as hard structure requirements during deliverable follow-through.
+4. Site and README install/version strings are updated to `v0.8.2`, including install snippets, release notes index text, and sidebar/footer labels.
 
 ## CLI
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@swarmclawai/swarmclaw",
-  "version": "0.8.1",
+  "version": "0.8.2",
   "description": "Self-hosted AI agent orchestration dashboard — manage LLM providers, orchestrate agent swarms, schedule tasks, and bridge agents to chat platforms.",
   "license": "MIT",
   "publishConfig": {

package/src/lib/runtime-loop.ts

@@ -22,7 +22,7 @@ export const CLAUDE_CODE_TIMEOUT_SEC_MAX = 7200
 export const CLI_PROCESS_TIMEOUT_SEC_MIN = 10
 export const CLI_PROCESS_TIMEOUT_SEC_MAX = 7200
 
-export const DEFAULT_AGENT_LOOP_RECURSION_LIMIT = 60
+export const DEFAULT_AGENT_LOOP_RECURSION_LIMIT = 120
 export const DEFAULT_ORCHESTRATOR_LOOP_RECURSION_LIMIT = 80
 export const DEFAULT_LEGACY_ORCHESTRATOR_MAX_TURNS = 16
 export const DEFAULT_ONGOING_LOOP_MAX_ITERATIONS = 250

package/src/lib/server/memory-policy.test.ts

@@ -36,6 +36,10 @@ test('isCurrentThreadRecallRequest detects same-thread recall without matching s
     isCurrentThreadRecallRequest('Remember that my favorite programming language is Rust and I prefer functional programming patterns.'),
     false,
   )
+  assert.equal(
+    isCurrentThreadRecallRequest('Remember that my favorite programming language is Rust and I prefer functional programming patterns. Then confirm what you just stored.'),
+    false,
+  )
 })
 
 test('isDirectMemoryWriteRequest detects remember-and-confirm turns without matching recall questions', () => {

package/src/lib/server/memory-policy.ts

@@ -29,6 +29,7 @@ export function isCurrentThreadRecallRequest(message: string): boolean {
   const trimmed = normalizeWhitespace(message)
   if (!trimmed) return false
   if (!CURRENT_THREAD_RECALL_MARKER_RE.test(trimmed)) return false
+  if (DIRECT_MEMORY_WRITE_MARKER_RE.test(trimmed) && DIRECT_MEMORY_WRITE_FOLLOWUP_RE.test(trimmed)) return false
   if (/\b(?:remember|store|save)\b/i.test(trimmed) && !/\?\s*$/.test(trimmed) && !/\b(?:what|which|who|when|where|did|confirm|recap|summarize|repeat|list|tell me|answer|recall)\b/i.test(trimmed)) {
     return false
   }

package/src/lib/server/runtime-settings.test.ts

@@ -46,7 +46,7 @@ describe('runtime settings defaults', () => {
     `)
 
     assert.equal(output.settings.loopMode, 'bounded')
-    assert.equal(output.settings.agentLoopRecursionLimit, 60)
+    assert.equal(output.settings.agentLoopRecursionLimit, 120)
     assert.equal(output.settings.orchestratorLoopRecursionLimit, 80)
     assert.equal(output.settings.legacyOrchestratorMaxTurns, 16)
     assert.equal(output.settings.ongoingLoopMaxIterations, 250)
@@ -61,7 +61,7 @@ describe('runtime settings defaults', () => {
     assert.equal(output.settings.heartbeatShowAlerts, true)
     assert.equal(output.settings.heartbeatTarget, null)
     assert.equal(output.settings.heartbeatPrompt, null)
-    assert.equal(output.runtime.agentLoopRecursionLimit, 60)
+    assert.equal(output.runtime.agentLoopRecursionLimit, 120)
     assert.equal(output.runtime.orchestratorLoopRecursionLimit, 80)
     assert.equal(output.runtime.legacyOrchestratorMaxTurns, 16)
   })

package/src/lib/server/session-tools/crud.test.ts

@@ -3,6 +3,7 @@ import assert from 'node:assert/strict'
 import fs from 'node:fs'
 import os from 'node:os'
 import path from 'node:path'
+import type { Agent } from '@/types'
 
 const originalEnv = {
   DATA_DIR: process.env.DATA_DIR,
@@ -30,11 +31,12 @@ before(async () => {
   loadAgents = storageMod.loadAgents
   saveAgents = storageMod.saveAgents
 
-  const agents = loadAgents({ includeTrashed: true })
+  const agents = loadAgents({ includeTrashed: true }) as Record<string, Agent>
   agents['agent-soul-test'] = {
     id: 'agent-soul-test',
     name: 'Soul Test Agent',
     description: 'Agent used for CRUD soul validation tests',
+    systemPrompt: '',
     provider: 'ollama',
     model: 'glm-5:cloud',
     plugins: ['manage_agents'],
@@ -42,7 +44,7 @@ before(async () => {
     platformAssignScope: 'self',
     createdAt: Date.now(),
     updatedAt: Date.now(),
-  } as any
+  }
   saveAgents(agents)
 })
 
@@ -124,8 +126,8 @@ describe('manage_agents soul validation', () => {
 
     const first = JSON.parse(String(firstRaw)) as Record<string, unknown>
     const second = JSON.parse(String(secondRaw)) as Record<string, unknown>
-    const created = Object.values(loadAgents({ includeTrashed: true }))
-      .filter((agent: any) => agent.createdInSessionId === 'agent-dedupe-session')
+    const created = Object.values(loadAgents({ includeTrashed: true }) as Record<string, Agent & { createdInSessionId?: string }>)
+      .filter((agent) => agent.createdInSessionId === 'agent-dedupe-session')
 
     assert.equal(created.length, 1)
     assert.equal(second.id, first.id)

package/src/lib/server/session-tools/memory.ts

@@ -788,6 +788,7 @@ const MemoryPlugin: Plugin = {
       'For info already in the current conversation, respond directly without calling any memory tool.',
       'For questions about prior work, decisions, dates, people, preferences, or todos from earlier conversations: start with one durable `memory_search`, then use `memory_get` only if you need a more targeted read. Only use archive/session history when the user explicitly needs transcript-level detail or the durable search is insufficient.',
       'When the user directly says to remember, store, or correct a fact, do one `memory_store` or `memory_update` call immediately. Treat the newest direct user statement as authoritative.',
+      'When one user message contains multiple related facts to remember, prefer one canonical `memory_store` write that captures the full set instead of many near-duplicate store calls.',
       'If someone says "remember this", write it down; do not rely on RAM alone.',
       'Memory writes merge canonical memories and retire superseded variants. After a successful store/update, do not keep re-searching unless the user explicitly asked you to verify.',
       'By default, memory searches focus on durable memories. Only include archives or working execution notes when you explicitly need transcript or run-history context.',
@@ -867,7 +868,7 @@ const MemoryPlugin: Plugin = {
     },
     {
       name: 'memory_store',
-      description: 'Store a durable fact, preference, decision, or correction from the user. Use this immediately when the user says to remember something.',
+      description: 'Store a durable fact, preference, decision, or correction from the user. Use this immediately when the user says to remember something. If several related facts arrive in one request, prefer one canonical write over many near-duplicate calls.',
       parameters: {
         type: 'object',
         properties: {
@@ -884,6 +885,7 @@ const MemoryPlugin: Plugin = {
         capabilities: ['memory.write'],
         disciplineGuidance: [
           'When the user says to remember or store a fact, call `memory_store` immediately. Do not delegate or use platform-management tools first.',
+          'If the user bundled multiple related facts into one remember request, store them together in one canonical write unless they asked for separate memories.',
         ],
       },
       execute: async (args, context) => executeNamedMemoryAction('store', args, context),

package/src/lib/server/stream-agent-chat.test.ts

@@ -7,6 +7,7 @@ import {
   buildExternalWalletExecutionBlock,
   buildToolDisciplineLines,
   getExplicitRequiredToolNames,
+  isNarrowDirectMemoryWriteTurn,
   isWalletSimulationResult,
   looksLikeOpenEndedDeliverableTask,
   resolveContinuationAssistantText,
@@ -39,6 +40,8 @@ describe('buildToolDisciplineLines', () => {
     const lines = buildToolDisciplineLines(['files'])
 
     assert.ok(lines.some((line) => line.includes('{"action":"read","filePath":"path/to/file.md"}')))
+    assert.ok(lines.some((line) => line.includes('exactly N bullet points')))
+    assert.ok(lines.some((line) => line.includes('Lower-priority logistics belong in FYI')))
   })
 
   it('adds schedule reuse and stop guidance when schedule tools are enabled', () => {
@@ -137,10 +140,12 @@ describe('buildToolDisciplineLines', () => {
     assert.ok(streamAgentChatSource.includes('call `memory_store` or `memory_update` immediately before any planning, delegation, task creation, or agent management'))
     assert.ok(streamAgentChatSource.includes('Do not inspect skills, browse the workspace, request capabilities, manage tasks, manage agents, or delegate before the direct memory write is complete.'))
     assert.ok(streamAgentChatSource.includes('Do NOT call memory tools, web search, or session-history tools'))
-    assert.ok(streamAgentChatSource.includes('const currentThreadRecallRequest = isCurrentThreadRecallRequest(message)'))
-    assert.ok(streamAgentChatSource.includes('const directMemoryWriteOnlyTurn = isDirectMemoryWriteRequest(message)'))
+    assert.ok(streamAgentChatSource.includes('const currentThreadRecallRequest = !directMemoryWriteOnlyTurn && isCurrentThreadRecallRequest(message)'))
+    assert.ok(streamAgentChatSource.includes('const directMemoryWriteOnlyTurn = isNarrowDirectMemoryWriteTurn(message)'))
     assert.ok(streamAgentChatSource.includes('shouldAllowToolForDirectMemoryWrite(toolName)'))
     assert.ok(streamAgentChatSource.includes('shouldAllowToolForCurrentThreadRecall(toolName)'))
+    assert.ok(streamAgentChatSource.includes('Preserve hard structural constraints from the original request'))
+    assert.ok(streamAgentChatSource.includes('## Exact Structural Constraints'))
   })
 
   it('blocks memory, session-history, web, and context tools during same-thread recall turns', () => {
@@ -164,6 +169,21 @@ describe('buildToolDisciplineLines', () => {
     assert.equal(shouldAllowToolForDirectMemoryWrite('files'), false)
   })
 
+  it('treats long remember-and-confirm turns as narrow direct memory writes', () => {
+    assert.equal(
+      isNarrowDirectMemoryWriteTurn('Remember that my favorite programming language is Rust and I prefer functional programming patterns. Then confirm what you just stored.'),
+      true,
+    )
+    assert.equal(
+      isNarrowDirectMemoryWriteTurn('Remember these facts for future conversations: My favorite programming language is Rust. My deploy target is Fly.io. My team size is 7 people. The project is codenamed "Neptune".'),
+      true,
+    )
+    assert.equal(
+      isNarrowDirectMemoryWriteTurn('Remember that my favorite programming language is Rust, then write a file summarizing it and send it to me.'),
+      false,
+    )
+  })
+
   it('canonicalizes required tool names when checking completion', () => {
     // The requiredToolsPending filter must canonicalize tool names so that
     // alias names (e.g. ask_human) match canonical names from LangGraph events.

package/src/lib/server/stream-agent-chat.ts

@@ -133,6 +133,11 @@ export function buildToolDisciplineLines(enabledPlugins: string[]): string[] {
     lines.push('Use `manage_capabilities` only when a needed tool is actually unavailable. If a direct tool for the job is already enabled in this session, call that tool immediately instead of requesting access or re-running discovery.')
   }
 
+  if (uniqueTools.includes('files') || uniqueTools.includes('edit_file')) {
+    lines.push('When the user specifies exact counts or exact section titles for file content, treat those as hard constraints. If a file must have exactly N bullet points, keep the total bullet count at N and put extra required detail into short prose under titled sections unless the user explicitly asked for more bullets.')
+    lines.push('When summarizing or restructuring a source document into named sections, make sure each top-level source section is represented somewhere in the output. Lower-priority logistics belong in FYI rather than being dropped.')
+  }
+
   if (uniqueTools.includes('delegate') && (uniqueTools.includes('shell') || uniqueTools.includes('files') || uniqueTools.includes('edit_file'))) {
     lines.push('When local workspace tools like `shell`, `files`, or `edit_file` are already enabled, prefer using them directly for straightforward coding and verification. Use `delegate` when you need a specialist backend, a second implementation pass, or parallel work.')
   }
@@ -524,6 +529,7 @@ function buildDeliverableFollowthroughPrompt(params: {
     'Do not stop after one partial batch. Finish every requested deliverable that is still outstanding before concluding.',
     'If a requested artifact cannot be produced, say exactly which artifact is missing, what blocked it, and what you already completed.',
     'Use the existing files, screenshots, and generated outputs first. Inspect them if needed, then complete the remaining work.',
+    'Preserve hard structural constraints from the original request: exact counts stay exact, required titled sections stay present, and source coverage gaps should be filled instead of skipped.',
     'End with a concise grouped completion summary that lists exact file paths, upload URLs, localhost URLs/ports, and screenshots you produced.',
   ]
 
@@ -553,6 +559,18 @@ function buildDeliverableFollowthroughPrompt(params: {
   return lines.join('\n')
 }
 
+function buildExactStructureBlock(userMessage: string): string {
+  const exactBulletMatch = userMessage.match(/\bexactly\s+(\d+)\s+bullet points?\b/i)
+  if (!exactBulletMatch) return ''
+  const bulletCount = exactBulletMatch[1]
+  return [
+    '## Exact Structural Constraints',
+    `The user required exactly ${bulletCount} bullet points.`,
+    'Treat that as a hard file-wide constraint unless the user explicitly says later sections get their own separate bullets.',
+    'If the file also needs titled sections such as Owners or Risks, use short prose under those headings instead of adding more bullet lines.',
+  ].join('\n')
+}
+
 /** Detect whether a user message is a broad, high-level goal that benefits from decomposition. */
 function isBroadGoal(text: string): boolean {
   if (text.length < 50) return false
@@ -591,10 +609,7 @@ function buildAgenticExecutionPolicy(opts: {
   const pluginLines = buildPluginCapabilityLines(opts.enabledPlugins, { platformAssignScope: opts.platformAssignScope })
   const toolDisciplineLines = buildToolDisciplineLines(opts.enabledPlugins)
   const hasMemoryTools = opts.enabledPlugins.some((toolId) => (canonicalizePluginId(toolId) || toolId) === 'memory')
-  const directMemoryWriteRequest = Boolean(opts.userMessage && isDirectMemoryWriteRequest(opts.userMessage))
-  const directMemoryWriteOnlyTurn = directMemoryWriteRequest
-    && !isBroadGoal(opts.userMessage || '')
-    && !looksLikeOpenEndedDeliverableTask(opts.userMessage || '')
+  const directMemoryWriteOnlyTurn = Boolean(opts.userMessage && isNarrowDirectMemoryWriteTurn(opts.userMessage))
 
   const parts: string[] = []
 
@@ -657,6 +672,10 @@ function buildAgenticExecutionPolicy(opts: {
   if (opts.userMessage && looksLikeOpenEndedDeliverableTask(opts.userMessage) && opts.enabledPlugins.some((toolId) => toolId === 'files' || toolId === 'edit_file')) {
     parts.push(OPEN_ENDED_REVISION_BLOCK)
   }
+  if (opts.userMessage) {
+    const exactStructureBlock = buildExactStructureBlock(opts.userMessage)
+    if (exactStructureBlock) parts.push(exactStructureBlock)
+  }
   if (opts.userMessage && isCurrentThreadRecallRequest(opts.userMessage)) {
     parts.push(buildCurrentThreadRecallBlock(opts.history || []))
   }
@@ -711,10 +730,24 @@ function buildDirectMemoryWriteBlock(): string {
     '## Direct Memory Write',
     'This turn is a direct request to remember, store, or correct a durable fact.',
     'Call `memory_store` or `memory_update` immediately, then confirm the stored value succinctly.',
+    'If the user bundled several related facts into one remember request, store them together in one canonical memory write unless they explicitly asked for separate entries.',
     'Do not inspect skills, browse the workspace, request capabilities, manage tasks, manage agents, or delegate before the direct memory write is complete.',
   ].join('\n')
 }
 
+const DIRECT_MEMORY_WRITE_CONFIRMATION_ONLY_RE = /\b(?:then|and then|after that)?\s*(?:confirm|recap|repeat|summarize|tell me|say)\b[\s\S]{0,120}\b(?:stored|saved|updated|remembered|wrote|write)\b/i
+const DIRECT_MEMORY_WRITE_EXTRA_ACTION_RE = /\b(?:then|and then|after that|also)\b[\s\S]{0,160}\b(?:write|create|send|email|message|delegate|research|search|browse|open|edit|build|schedule|plan|review|analy[sz]e)\b/i
+
+export function isNarrowDirectMemoryWriteTurn(message: string): boolean {
+  const trimmed = String(message || '').trim()
+  if (!trimmed || !isDirectMemoryWriteRequest(trimmed)) return false
+  if (looksLikeOpenEndedDeliverableTask(trimmed)) return false
+  if (DIRECT_MEMORY_WRITE_EXTRA_ACTION_RE.test(trimmed) && !DIRECT_MEMORY_WRITE_CONFIRMATION_ONLY_RE.test(trimmed)) {
+    return false
+  }
+  return !isBroadGoal(trimmed) || DIRECT_MEMORY_WRITE_CONFIRMATION_ONLY_RE.test(trimmed) || !/[?]$/.test(trimmed)
+}
+
 const CURRENT_THREAD_RECALL_BLOCKED_TOOL_IDS = new Set([
   'memory',
   'manage_sessions',
@@ -840,7 +873,8 @@ export async function streamAgentChat(opts: StreamAgentChatOpts): Promise<Stream
 
   const stateModifierParts: string[] = []
   const hasProvidedSystemPrompt = typeof systemPrompt === 'string' && systemPrompt.trim().length > 0
-  const currentThreadRecallRequest = isCurrentThreadRecallRequest(message)
+  const directMemoryWriteOnlyTurn = isNarrowDirectMemoryWriteTurn(message)
+  const currentThreadRecallRequest = !directMemoryWriteOnlyTurn && isCurrentThreadRecallRequest(message)
 
   if (hasProvidedSystemPrompt) {
     stateModifierParts.push(systemPrompt!.trim())
@@ -1054,9 +1088,6 @@ export async function streamAgentChat(opts: StreamAgentChatOpts): Promise<Stream
     projectDescription: activeProjectContext.project?.description || null,
     memoryScopeMode: agentMemoryScopeMode,
   })
-  const directMemoryWriteOnlyTurn = isDirectMemoryWriteRequest(message)
-    && !isBroadGoal(message)
-    && !looksLikeOpenEndedDeliverableTask(message)
   const toolsForTurn = currentThreadRecallRequest
     ? tools.filter((tool) => {
         const toolName = typeof (tool as { name?: unknown }).name === 'string'
@@ -1268,7 +1299,7 @@ export async function streamAgentChat(opts: StreamAgentChatOpts): Promise<Stream
   const MAX_REQUIRED_TOOL_CONTINUES = 2
   const MAX_EXECUTION_FOLLOWTHROUGHS = 1
   const MAX_DELIVERABLE_FOLLOWTHROUGHS = 2
-  const MAX_TOOL_SUMMARY_RETRIES = 1
+  const MAX_TOOL_SUMMARY_RETRIES = 2
   let autoContinueCount = 0
   let transientRetryCount = 0
   let requiredToolContinueCount = 0