@dotdrelle/wiki-manager 0.15.101 → 0.16.0

package/mcp.endpoints.example.json

@@ -29,7 +29,7 @@
   "chatAccess": {
     "maxToolIterations": 8,
     "servers": {
-      "llm-wiki":         { "allow": ["help_list", "help_read", "help_search", "wiki_workspace_status", "wiki_list_pages", "wiki_read_page", "wiki_read_pages", "wiki_search_context", "wiki_collect_context", "wiki_read_ingested_source", "wiki_outline", "template_read", "template_write", "build_context_write", "wiki_read_deliverable", "wiki_graph_query", "wiki_graph_path"] },
+      "llm-wiki":         { "allow": ["help_list", "help_read", "help_search", "wiki_workspace_status", "wiki_list_pages", "wiki_read_page", "wiki_read_pages", "wiki_search_context", "wiki_collect_context", "wiki_read_ingested_source", "wiki_outline", "template_read", "wiki_read_deliverable", "wiki_graph_query", "wiki_graph_path"] },
       "wiki-production":  { "allow": ["production_job_status", "production_jobs_list"] },
       "cme":              { "allow": ["cme_status", "cme_sources_list", "cme_export_status", "cme_confluence_search", "cme_wiki_search"] }
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dotdrelle/wiki-manager",
-  "version": "0.15.101",
+  "version": "0.16.00",
   "description": "Agentic shell and orchestration cockpit for llm-wiki workspaces.",
   "repository": {
     "type": "git",
@@ -19,7 +19,7 @@
   },
   "scripts": {
     "start": "bun ./bin/wiki-manager.js",
-    "test": "node --test src/core/skillInvocation.test.js src/core/skillCompiler.test.js src/runtime/skillRun.test.js src/runtime/controlDrain.test.js src/runtime/controlCancellation.test.js src/cli/runtimeStartup.test.js src/cli/wiki-manager.test.js src/agent/graph.test.js src/agent/skillRecursion.test.js src/contracts/schemas.test.js src/core/activity.test.js src/core/env.test.js src/core/agentsCompose.test.js src/core/profileServiceStatus.test.js src/core/workspaceProfile.test.js src/core/buildInfo.test.js src/core/testGate.test.js src/core/agentEvents.test.js src/core/skillChainView.test.js src/core/runtimeLog.test.js src/core/runtimeEventAdapter.test.js src/activity/activityAggregator.test.js src/graph/runGraphProjector.test.js src/core/workflow.test.js src/core/planPatch.test.js src/core/agentLoop.test.js src/core/plan.test.js src/core/mcp.test.js src/core/mcpEndpoints.test.js src/core/toolLoop.test.js src/core/documentIntake.test.js src/core/dockerCompose.test.js src/core/otherWorkspacesRunning.test.js src/core/wikiSetup.test.js src/core/wikiWorkspace.test.js src/core/wikiWorkspaceStart.test.js src/core/wikirc.test.js src/core/workspaceInherit.test.js src/core/cacert.test.js src/core/composeOverrides.test.js src/core/setEnvValue.test.js src/core/commandFailure.test.js src/core/currentArtifact.test.js src/core/googleGrants.test.js src/core/modelFetch.test.js src/core/startupCheck.test.js src/core/queueStore.test.js src/orchestrator/agentRegistry.test.js src/orchestrator/capabilityRegistry.test.js src/orchestrator/capabilityResolver.test.js src/orchestrator/planValidator.test.js src/orchestrator/planIntegrator.test.js src/orchestrator/taskStatuses.test.js src/orchestrator/scheduler.test.js src/orchestrator/attemptManager.test.js src/orchestrator/resultAggregator.test.js src/orchestrator/proactiveReviewScheduler.test.js src/orchestrator/knowledgeSignals.test.js src/orchestrator/approvalPolicy.test.js src/orchestrator/dispatcher.test.js src/orchestrator/objectiveResolver.test.js src/orchestrator/providers/fakeRuntimeProvider.test.js src/orchestrator/providers/runtimeProviders.test.js src/orchestrator/providers/dispatcherExternalRuntime.test.js src/orchestrator/providers/deepAgentsProvider.test.js src/commands/slash.test.js src/shell/repl.test.js src/shell/setupWizardModality.test.js src/shell/setupWizardPlaceholders.test.js src/shell/setupWizardSuggestions.test.js src/shell/setupWizardDiscovery.test.js src/shell/wrapText.test.js src/runtime/lifecycle.test.js src/runtime/store.test.js src/runtime/workspaceIsolation.test.js src/runtime/controlMessages.test.js src/runtime/controlClassify.test.js src/runtime/deltaCoalescer.test.js src/runtime/recoveryManager.test.js src/runtime/server.test.js src/runtime/supervisor.test.js src/runtime/delegation.test.js src/runtime/runner.test.js src/runtime/runner.e2e.test.js src/runtime/skillChain.e2e.test.js src/runtime/donna-contract.test.js src/runtime/approvals.test.js src/runtime/auth.test.js src/runtime/totp.test.js src/runtime/loginSession.test.js src/runtime/loginRoutes.test.js",
+    "test": "node --test src/core/skillInvocation.test.js src/core/skillCompiler.test.js src/runtime/skillRun.test.js src/runtime/controlDrain.test.js src/runtime/controlCancellation.test.js src/cli/runtimeStartup.test.js src/cli/wiki-manager.test.js src/agent/graph.test.js src/agent/llm.test.js src/agent/skillRecursion.test.js src/contracts/schemas.test.js src/core/activity.test.js src/core/env.test.js src/core/agentsCompose.test.js src/core/profileServiceStatus.test.js src/core/llmCapabilities.test.js src/core/workspaceProfile.test.js src/core/buildInfo.test.js src/core/testGate.test.js src/core/agentEvents.test.js src/core/skillChainView.test.js src/core/runtimeLog.test.js src/core/runtimeEventAdapter.test.js src/activity/activityAggregator.test.js src/graph/runGraphProjector.test.js src/core/workflow.test.js src/core/planPatch.test.js src/core/agentLoop.test.js src/core/plan.test.js src/core/mcp.test.js src/core/mcpEndpoints.test.js src/core/toolLoop.test.js src/core/logLabel.test.js src/runtime/conversationCompaction.test.js src/core/documentIntake.test.js src/core/dockerCompose.test.js src/core/otherWorkspacesRunning.test.js src/core/wikiSetup.test.js src/core/wikiWorkspace.test.js src/core/wikiWorkspaceStart.test.js src/core/wikirc.test.js src/core/workspaceInherit.test.js src/core/cacert.test.js src/core/composeOverrides.test.js src/core/setEnvValue.test.js src/core/commandFailure.test.js src/core/currentArtifact.test.js src/core/googleGrants.test.js src/core/modelFetch.test.js src/core/startupCheck.test.js src/core/queueStore.test.js src/orchestrator/agentRegistry.test.js src/orchestrator/capabilityRegistry.test.js src/orchestrator/capabilityResolver.test.js src/orchestrator/planValidator.test.js src/orchestrator/planIntegrator.test.js src/orchestrator/taskStatuses.test.js src/orchestrator/scheduler.test.js src/orchestrator/attemptManager.test.js src/orchestrator/resultAggregator.test.js src/orchestrator/proactiveReviewScheduler.test.js src/orchestrator/knowledgeSignals.test.js src/orchestrator/approvalPolicy.test.js src/orchestrator/dispatcher.test.js src/orchestrator/objectiveResolver.test.js src/orchestrator/providers/fakeRuntimeProvider.test.js src/orchestrator/providers/runtimeProviders.test.js src/orchestrator/providers/dispatcherExternalRuntime.test.js src/orchestrator/providers/deepAgentsProvider.test.js src/commands/slash.test.js src/shell/repl.test.js src/shell/setupWizardModality.test.js src/shell/setupWizardPlaceholders.test.js src/shell/setupWizardSuggestions.test.js src/shell/setupWizardDiscovery.test.js src/shell/wrapText.test.js src/runtime/lifecycle.test.js src/runtime/store.test.js src/runtime/workspaceIsolation.test.js src/runtime/controlMessages.test.js src/runtime/controlClassify.test.js src/runtime/deltaCoalescer.test.js src/runtime/recoveryManager.test.js src/runtime/server.test.js src/runtime/supervisor.test.js src/runtime/delegation.test.js src/runtime/runner.test.js src/runtime/runner.e2e.test.js src/runtime/skillChain.e2e.test.js src/runtime/donna-contract.test.js src/runtime/approvals.test.js src/runtime/auth.test.js src/runtime/totp.test.js src/runtime/loginSession.test.js src/runtime/loginRoutes.test.js",
     "check-versions": "node scripts/check-versions.js",
     "prepack": "node scripts/check-versions.js",
     "prepublishOnly": "node scripts/check-versions.js",

package/src/agent/graph.js

@@ -27,6 +27,8 @@ import { toolResultNote, toolStartNote } from '../core/progressNotes.js';
 import { openWikiPagesPromptLine } from '../core/openWikiPages.js';
 import { enqueueProductionJob, ensureJobQueue, formatQueue, productionLockBusy } from '../core/jobQueue.js';
 import { loadWorkspaceProfile, updateWorkspaceProfilePreference } from '../core/profile.js';
+import { formatLlmConfigFact } from '../core/wikirc.js';
+import { wikiSearchContextMessages } from '../core/wikiPresearch.js';
 import { artifactFromToolCall, currentArtifactFor, currentArtifactPromptLine, rememberArtifact } from '../core/currentArtifact.js';
 import { capabilityRegistryForSession } from '../orchestrator/capabilityRegistry.js';
 import { objectiveForResolution } from '../orchestrator/objectiveResolver.js';
@@ -445,7 +447,15 @@ export function bareToolCallJson(content, tools = []) {
  * reached the user and nothing ran.
  */
 export function narratedToolCallText(content) {
-  const match = String(content ?? '').trim().match(/^([a-z][a-z0-9_-]*__[a-z][a-z0-9_-]*)\s*\{/i);
+  // Accept EITHER `name{...}` (bare JSON body) OR `name({...})` (function-call
+  // style). Some providers/gateways write the call with parentheses around the
+  // arguments object — `runtime__run_skill({...})` — and the old brace-only
+  // regex missed it, so the raw call surfaced to the user AND executed nothing.
+  // The parenthesis alone is NOT enough: `runtime__delegate (the delegation
+  // tool) hands…` is a legitimate answer about the tool, and treating it as a
+  // call discarded it and ended on "Action not executed". Require the `{` of an
+  // arguments object right after the parenthesis.
+  const match = String(content ?? '').trim().match(/^([a-z][a-z0-9_-]*__[a-z][a-z0-9_-]*)\s*(?:\{|\(\s*\{)/i);
   return match ? match[1] : null;
 }
 
@@ -1311,6 +1321,7 @@ export function buildAgentSystemPrompt(state) {
     `Reply language: ${language}.`,
     `Current workspace: ${workspace}.`,
     `Current wikirc profile: ${wikirc}.`,
+    formatLlmConfigFact(state.session.wikircConfig, state.session.wikirc),
     `Available primitives: ${commandList(state.session)}.`,
     'Only announce or call slash commands that appear exactly in Available primitives. Do not invent command names, subcommands, or arguments.',
     'Connected MCP tools you may call directly (server__tool naming convention). Everything listed below is directly callable. When the requested action has no matching direct tool, call runtime__delegate with the original objective: the runtime resolves it against the discovered agent capability contracts, including executor-only single-task capabilities.',
@@ -1333,6 +1344,20 @@ export function buildAgentSystemPrompt(state) {
     'Write the way a thoughtful colleague speaks: warm, plain, and to the point. For a simple factual question, 1 to 3 sentences is the sweet spot. Stay synthetic and information-dense — use only the lines needed, and never exceed roughly 15 to 20 short lines even for a detailed answer. Never expose internal reasoning, repeated checks, tool-selection commentary, or a chronological diary. Prioritize the result, essential facts, concrete errors, and actual outputs — but say them in human language, not as a field dump.',
     'Call a matching direct tool when one is offered. Otherwise, for an action backed by a discovered agent capability, call runtime__delegate with the original objective. This applies to both planner agents and executor-only single-task agents. Never call an agent orchestration-contract or plan tool directly.',
     'Configuration is not a business run. When a connected server offers a setup or configuration tool, use it directly; never delegate configuration to an export, collect, send, build, or ingest capability. Read that server status first when existing non-secret values are needed, then ask only for required values that are still missing.',
+    // Observed: « récupère le numéro de ticket depuis le wiki » delegated to
+    // the external runtime's agent.answer, which burned 519k tokens and failed
+    // on its budget, while one wiki search answers it in a second.
+    'A question about the subject matter of this workspace (its projects, documents, tickets, people, decisions, figures, dates) is answered from the wiki: search it FIRST with the wiki search/read tools and answer from what they return. Never runtime__delegate a question those read tools can answer; delegation is for actions, or for an analysis the user explicitly asks an agent to perform.',
+    // Observed (acpi): "cherche sur internet" was delegated to the external
+    // runtime, which answered it could not search the internet — while a
+    // connected web-search read tool was available and offered right here. The
+    // wiki stays first for workspace facts; the web tools cover the rest and
+    // are never denied.
+    'A wiki-first rule covers workspace facts only. For an internet/web search the wiki cannot answer, use the connected web-search read tool directly when one is offered; otherwise delegate the objective to a research capability. Never answer that you cannot search the internet when a web-search tool is available, and never claim a search you did not run.',
+    // Observed: « compare les options A et B » answered from the previous
+    // answers alone — the history carries Donna's text, not the pages — and
+    // option A was invented, the opposite of what the wiki says.
+    'Your earlier answers in this conversation are not evidence: they keep your text, not the pages. For each new question, search the wiki again for every fact you have not quoted from a tool result in this very turn. Never fill a gap (an acronym expansion, a missing option, a figure) from general knowledge.',
     'For any question about the current workspace inventory or what is waiting there, call wiki__wiki_workspace_status first and answer only from its result. This is the canonical read-only workspace state; do not reconstruct it from upload, connector, or production tools.',
     'Tool identifiers are private implementation details. Never print MCP tool names such as server__tool in a user-facing answer. Describe the human result instead.',
     'Internal data shapes are private too. Never quote raw JSON field names (e.g. pendingSources.files), internal directory paths (e.g. raw/untracked/), or config keys in a user-facing answer — translate them into plain language. Say "36 pages sources sont en attente d\'ingestion", not the field or path they came from.',
@@ -1382,7 +1407,7 @@ export function buildAgentSystemPrompt(state) {
       : null,
     openWikiPagesPromptLine(state.session.openWikiPages),
     currentArtifactPromptLine(currentArtifactFor(state.session)),
-    'Runtime control: you have runtime__status, runtime__cancel, runtime__kill and runtime__enqueue. When the user asks to stop, remove, clean or kill the current run, its jobs or the queue ("supprime le job et la queue", "arr\u00eate tout"), call runtime__kill (or runtime__cancel for a soft stop of just the run) and confirm what was stopped. When the user explicitly asks to delete, reset, abandon or replace the current plan, call runtime__kill with purge=true; never set purge=true for a simple stop. For questions about what is running or queued, call runtime__status and answer from its data. You have no approval tool: a pending approval is granted only by the user through the approval button or the /approve command. Never grant, claim or report an approval yourself; when the user asks to proceed with pending mutations, tell them to use those controls. A request to approve, validate, confirm or accept a pending run is a human control action: NEVER call runtime__delegate (or any capability) for it — answer with the control to use and nothing else. When the user asks for a NEW action while a run is active, do not execute it: propose runtime__enqueue (run it after) or, if they insist it replaces the current work, runtime__kill then the new action.',
+    'Runtime control: you have runtime__status, runtime__cancel, runtime__kill and runtime__enqueue. When the user asks to stop, remove, clean or kill the current run, its jobs or the queue ("supprime le job et la queue", "arr\u00eate tout"), call runtime__kill (or runtime__cancel for a soft stop of just the run) and confirm what was stopped. When the user explicitly asks to delete, reset, abandon or replace the current plan, call runtime__kill with purge=true; never set purge=true for a simple stop. For questions about what is running or queued, call runtime__status and answer from its data. You have no approval tool: a pending approval is granted only by the user through the approval button or the /approve command. Never grant, claim or report an approval yourself; when the user asks to proceed with pending mutations, tell them to use those controls. A request to approve, validate, confirm or accept a pending run is a human control action: NEVER call runtime__delegate (or any capability) for it — answer with the control to use and nothing else. When the user asks for a NEW action while a run is active: if one of your direct tools performs it (write or edit a template, a build-context file or a wiki page; a read or a check), do it now — a write may be refused because the running job is using that part of the workspace, and then say so plainly and offer runtime__enqueue. Heavier work (ingest, build, export, curate, rebuilding an index) is queued with runtime__enqueue (it runs after), or, if the user insists it replaces the current work, runtime__kill then the new action.',
     'When the user asks to refresh, show, or update the displayed plan or status, call runtime__status. This is a state refresh request, not a new business capability, and must never be delegated.',
     'Report every runtime control outcome exactly as the tool returned it \u2014 never embellish. If runtime__kill reports 0 run(s)/0 task(s)/0 purged, say there was nothing active to stop or purge; do NOT claim a run, plan, pending approval or queue item was removed. If runtime__status returns an error or could not be read, say the runtime state could not be retrieved and do not describe a state you never obtained. Never assert that something was cleaned, cancelled, approved or purged unless that specific tool result confirms it.',
     'Durable profile updates are actions in this stabilized version: delegate them instead of writing directly.',
@@ -1433,7 +1458,14 @@ function toolsForClassification(classification, writeTools, session = null) {
     // suite: she can answer, approve, enqueue for later, soft-cancel or
     // kill — but she must not fire new MCP jobs alongside the run (that is
     // what runtime__enqueue is for). No canned regex answers anywhere.
-    return [SHELL_READ_COMMAND_TOOL, ...controlTools, ...capabilityRunTools];
+    // The direct tools stay offered — reads, and unitary writes such as a
+    // template or a page: without them a question sent as a run could only be
+    // delegated (observed: a wiki question handed to the external runtime,
+    // which spent 519k tokens on it), and a template asked for during an
+    // ingest waited for the whole ingest. ordinaryDirectTools never carries a
+    // job starter, and the engine refuses a write that would race with the
+    // running job (plan-demandes-pendant-run.md, lot 3).
+    return [SHELL_READ_COMMAND_TOOL, ...controlTools, ...capabilityRunTools, ...ordinaryDirectTools(writeTools)];
   }
   if (session?.runtime?.url) {
     // Offer every connected tool directly EXCEPT orchestration-bypass tools
@@ -1476,7 +1508,7 @@ export function isDonnaReadTool(item) {
   if (tool === 'wiki_workspace_status' || tool === 'agent_describe' || tool === 'agent_status') return true;
   // Match a read verb anywhere in the underscore-tokenized name, not just as
   // a trailing suffix — third-party MCPs don't all name tools verb-last
-  // (e.g. exa's "web_search_exa"/"web_fetch_exa" put the verb in the middle).
+  // (e.g. a connector whose `web_search_x` puts the verb in the middle).
   return tool.split('_').some((segment) => DONNA_READ_VERBS.has(segment));
 }
 
@@ -1585,7 +1617,8 @@ export function createAgentGraph(options = {}) {
             kind: 'agent_turn',
             confidence: 1,
             reason: 'agent_mode_llm_decision',
-            activeRun: state.session?.agentProjection?.status === 'running'
+            activeRun: state.session?._runActive === true
+              || ['running', 'pending_approval'].includes(state.session?.agentProjection?.status)
               || sessionActivities(state.session).some((activity) => !activity.terminal),
           })
       : (state.inputClassification ?? { kind: 'modify_run', confidence: 1, reason: 'tool_iteration' });
@@ -1614,11 +1647,28 @@ export function createAgentGraph(options = {}) {
       : toolsForClassification(classification, writeTools, state.session);
     const system = buildAgentSystemPrompt(state);
 
+    // The wiki is searched before the first model call of a turn, as in chat
+    // mode (core/wikiPresearch.js): a question about the workspace was
+    // otherwise answered from the previous answers, or delegated to an agent
+    // that burned its whole token budget on it. Only on the user's own words —
+    // never on a compiled skill objective (private material) nor on the
+    // continuation prompts of a run's later turns. It lives for this turn
+    // only: the next turn's history is rebuilt from the conversation.
+    const runTurnId = state.session._currentRunIdentity?.turnId;
+    const userWords = iterations === 0
+      && normalizedSkillStack(state.session).length === 0
+      && !state.session._responseSynthesisOnly
+      && (!state.session._currentRunIdentity || !runTurnId || String(runTurnId).endsWith(':turn-1'));
+    const userTurn = [
+      ...(userWords ? await wikiSearchContextMessages(state.input, state.session, tools, state.session._onStep) : []),
+      { role: 'user', content: state.input },
+    ];
+
     // On iteration 0: prior history is in state.messages, user input must be appended.
     // On subsequent iterations: user message was already stored in state.messages by the
     // iteration-0 return below, so use state.messages as-is.
     const conversationMessages = iterations === 0
-      ? [...(state.messages ?? []), { role: 'user', content: state.input }]
+      ? [...(state.messages ?? []), ...userTurn]
       : (state.messages ?? []);
 
     try {
@@ -1655,7 +1705,7 @@ export function createAgentGraph(options = {}) {
             return {
               pendingToolCalls: null,
               messages: [
-                ...(iterations === 0 ? [{ role: 'user', content: state.input }] : []),
+                ...(iterations === 0 ? userTurn : []),
                 {
                   role: 'user',
                   content: 'Your previous tool call was incomplete or contained invalid JSON arguments. Call the appropriate available tool again with one complete valid JSON object. Do not narrate or reproduce the broken call.',
@@ -1689,7 +1739,7 @@ export function createAgentGraph(options = {}) {
           tool_calls: result.tool_calls,
         };
         const newMessages = iterations === 0
-          ? [{ role: 'user', content: state.input }, assistantToolMessage]
+          ? [...userTurn, assistantToolMessage]
           : [assistantToolMessage];
         return {
           pendingToolCalls: result.tool_calls,
@@ -1722,7 +1772,7 @@ export function createAgentGraph(options = {}) {
           return {
             pendingToolCalls: null,
             messages: [
-              ...(iterations === 0 ? [{ role: 'user', content: state.input }] : []),
+              ...(iterations === 0 ? userTurn : []),
               {
                 role: 'user',
                 content: `You described a call to ${bareCall} as JSON text instead of calling it. Issue a real tool call now, or answer in plain language. Never print the call as text.`,
@@ -1746,7 +1796,7 @@ export function createAgentGraph(options = {}) {
         return {
           pendingToolCalls: null,
           messages: [
-            { role: 'user', content: state.input },
+            ...userTurn,
             result.message ?? { role: 'assistant', content: result.content ?? '' },
             {
               role: 'user',
@@ -1790,7 +1840,7 @@ export function createAgentGraph(options = {}) {
         return {
           pendingToolCalls: null,
           messages: [
-            { role: 'user', content: state.input },
+            ...userTurn,
             result.message ?? { role: 'assistant', content: result.content ?? '' },
             { role: 'user', content: 'This is an action request. Call runtime__delegate now with the original objective only. Do not provide instructions or narration.' },
           ],
@@ -1834,7 +1884,7 @@ export function createAgentGraph(options = {}) {
           return {
             pendingToolCalls: null,
             messages: [
-              ...(iterations === 0 ? [{ role: 'user', content: state.input }] : []),
+              ...(iterations === 0 ? userTurn : []),
               result.message ?? { role: 'assistant', content: result.content ?? '' },
               {
                 role: 'user',
@@ -1863,7 +1913,7 @@ export function createAgentGraph(options = {}) {
         emitAgentEvent(state.session, 'assistant_message', 'llm', { content: result.content ?? '' });
         // Text was streamed inline via session._onStream — no second LLM call needed.
         const newMessages = iterations === 0
-          ? [{ role: 'user', content: state.input }, result.message]
+          ? [...userTurn, result.message]
           : [result.message];
         return {
           response: null,

package/src/agent/graph.test.js

@@ -3,7 +3,7 @@ import test from 'node:test';
 import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs';
 import { tmpdir } from 'node:os';
 import { join } from 'node:path';
-import { bareToolCallJson, buildAgentSystemPrompt, connectorConfigurationTarget, createAgentGraph, invalidSuggestedSlashCommands, invalidUserFacingToolNames, isOrchestrationBypassTool, knownCapabilityIds, normalizeToolArgumentsFromSchema, nothingToDoForDonna } from './graph.js';
+import { bareToolCallJson, buildAgentSystemPrompt, connectorConfigurationTarget, createAgentGraph, invalidSuggestedSlashCommands, invalidUserFacingToolNames, isOrchestrationBypassTool, knownCapabilityIds, narratedToolCallText, normalizeToolArgumentsFromSchema, nothingToDoForDonna } from './graph.js';
 
 test('user-facing response guard hides MCP identifiers generically', () => {
   const session = sessionBase();
@@ -1122,6 +1122,81 @@ test('buildAgentSystemPrompt includes .wiki/profile.md content so preferences ap
   }
 });
 
+test('buildAgentSystemPrompt answers workspace questions from the wiki instead of delegating them', () => {
+  // Observed: a wiki question delegated to the external runtime's agent.answer
+  // burned 519k tokens and failed on its budget.
+  const prompt = buildAgentSystemPrompt({ session: sessionBase({}) });
+  assert.match(prompt, /search it FIRST with the wiki search\/read tools/);
+  assert.match(prompt, /Never runtime__delegate a question those read tools can answer/);
+});
+
+async function agentFirstCallMessages(sessionOverrides, input = 'compare les options A et B') {
+  const searched = [];
+  const originalFetch = globalThis.fetch;
+  globalThis.fetch = async (url, init) => {
+    const body = JSON.parse(init?.body ?? '{}');
+    let result = {};
+    if (body.method === 'tools/call') {
+      searched.push(body.params.arguments.question);
+      result = { content: [{ type: 'text', text: 'wiki/concepts/a.md: option A, sans tracé manuel' }] };
+    }
+    const text = JSON.stringify({ jsonrpc: '2.0', id: body.id ?? 1, result });
+    return { ok: true, status: 200, headers: { get: () => null }, text: async () => text, json: async () => JSON.parse(text) };
+  };
+  const seen = [];
+  const session = sessionBase({
+    mcp: { wiki: { status: 'connected', url: `http://agent-presearch-${Math.random()}.test/mcp`, tools: [
+      { name: 'wiki_search_context', inputSchema: { type: 'object', properties: {} } },
+    ] } },
+    llm: { async completeWithTools({ messages }) { seen.push(messages); return { tool_calls: [], content: 'ok' }; } },
+    ...sessionOverrides,
+  });
+  try {
+    await createAgentGraph().invoke({ input, session });
+  } finally {
+    globalThis.fetch = originalFetch;
+  }
+  return { searched, first: seen[0] ?? [] };
+}
+
+test('an agent turn searches the wiki before its first model call', async () => {
+  const { searched, first } = await agentFirstCallMessages({});
+  assert.deepEqual(searched, ['compare les options A et B']);
+  const index = first.findIndex((m) => /WIKI SEARCH RESULTS/.test(m.content));
+  assert.ok(index >= 0);
+  assert.equal(first[index + 1].content, 'compare les options A et B');
+});
+
+test('an agent turn never pre-searches a skill objective or a later run turn', async () => {
+  const skill = await agentFirstCallMessages({ _skillStack: ['pipeline'] });
+  assert.deepEqual(skill.searched, []);
+  const laterTurn = await agentFirstCallMessages({ _currentRunIdentity: { runId: 'r1', turnId: 'r1:turn-2' } });
+  assert.deepEqual(laterTurn.searched, []);
+  const firstTurn = await agentFirstCallMessages({ _currentRunIdentity: { runId: 'r1', turnId: 'r1:turn-1' } });
+  assert.deepEqual(firstTurn.searched, ['compare les options A et B']);
+});
+
+test('an agent turn during a pending run keeps its direct tools and cannot start a second run', async () => {
+  // plan-demandes-pendant-run.md, lot 3. The runtime says a run is active
+  // (_runActive): a projection in `pending_approval` used to read as idle.
+  let offered = [];
+  const session = sessionBase({
+    runtime: { url: 'http://127.0.0.1:7788' },
+    _runActive: true,
+    agentProjection: { status: 'pending_approval' },
+    mcp: { wiki: { status: 'connected', url: 'http://x.test/mcp', tools: [
+      { name: 'template_write', inputSchema: { type: 'object', properties: {} } },
+      { name: 'wiki_search_context', inputSchema: { type: 'object', properties: {} } },
+    ] }, production: sessionBase().mcp.production },
+    llm: { async completeWithTools({ tools }) { offered = tools.map((item) => item.function.name); return { tool_calls: [], content: 'ok' }; } },
+  });
+  await createAgentGraph().invoke({ input: 'salut', session });
+  assert.ok(offered.includes('wiki__template_write'), offered.join(','));
+  assert.ok(offered.includes('runtime__enqueue'), offered.join(','));
+  assert.ok(!offered.includes('runtime__delegate'), 'no second run beside the active one');
+  assert.ok(!offered.includes('production__production_start_job'), 'never a job starter');
+});
+
 test('buildAgentSystemPrompt omits the profile section when profile.md is missing or empty', () => {
   const workspacePath = mkdtempSync(join(tmpdir(), 'donna-profile-empty-'));
   try {
@@ -2612,3 +2687,52 @@ test('LOT F: the guard message stays English-only for every session language', a
   assert.match(await answer(undefined), /repeatedly printed an internal tool request/);
   assert.match(await answer('fr'), /repeatedly printed an internal tool request/);
 });
+
+test('narratedToolCallText recognises the brace and the function-call forms', () => {
+  assert.equal(narratedToolCallText('runtime__delegate{"objective":"x"}'), 'runtime__delegate');
+  // Some gateways write the call with parentheses around the arguments object;
+  // the brace-only regex missed this and executed nothing.
+  assert.equal(narratedToolCallText('runtime__run_skill({"skillName":"deliver"})'), 'runtime__run_skill');
+  assert.equal(
+    narratedToolCallText('  production__production_start_job({ "type": "build" })  '),
+    'production__production_start_job',
+  );
+  assert.equal(narratedToolCallText('Bonjour, je peux vous aider.'), null);
+  // Prose ABOUT a tool, opening on its name and a parenthesis, is an answer.
+  assert.equal(
+    narratedToolCallText('runtime__delegate (the delegation tool) hands an objective to the runtime.'),
+    null,
+  );
+  assert.equal(narratedToolCallText('runtime__delegate() takes an objective.'), null);
+  assert.equal(narratedToolCallText('{"name":"runtime__run_skill"}'), null);
+});
+
+test('a skill call narrated with parentheses is neither executed nor shown', async () => {
+  // Observed: `/deliver TechSections.export.md polish` answered with the raw
+  // `runtime__run_skill({…})` text — nothing ran and the call leaked, because
+  // the guard only matched `name{`.
+  const raw =
+    'runtime__run_skill({"skillName":"deliver","arguments":{"deliverable":"TechSections.export.md","polish":"true"},"selectionKind":"explicit_name"})';
+  const ran = [];
+  let calls = 0;
+  const session = sessionBase({
+    language: 'fr-FR',
+    runtime: { url: 'http://runtime.test' },
+    _runSkillWithinRun: async (name) => { ran.push(name); return { ok: true }; },
+    _onStreamReset: () => {},
+    llm: {
+      async completeWithTools() {
+        calls += 1;
+        return { content: raw, message: { role: 'assistant', content: raw }, tool_calls: null };
+      },
+    },
+  });
+
+  const result = await createAgentGraph().invoke({ input: 'Lance le polish.', session });
+
+  assert.deepEqual(ran, [], 'a call written as text must never reach the skill runner');
+  assert.doesNotMatch(result.response, /runtime__run_skill/);
+  assert.doesNotMatch(result.response, /[{}]/, 'no JSON fragment may survive');
+  assert.match(result.response, /repeatedly printed an internal tool request/);
+  assert.ok(calls >= 2, 'the first occurrence is retried, not surfaced');
+});

package/src/agent/llm.js

@@ -1,3 +1,5 @@
+import { supportsTemperature } from '../core/llmCapabilities.js';
+
 function trimTrailingSlash(value) {
   return value.replace(/\/+$/, '');
 }
@@ -25,6 +27,13 @@ export function createLlmClientFromWikiConfig(config) {
     return null;
   }
 
+  // A gpt-5-class model rejects `temperature` outright (HTTP 400, "Only the
+  // default (1) value is supported"), whether the profile sets one or not. When
+  // it is refused, the field is omitted whole rather than sent and rejected.
+  const temperatureBody = supportsTemperature(llmConfig)
+    ? { temperature: typeof llmConfig.temperature === 'number' ? llmConfig.temperature : 0.2 }
+    : {};
+
   return {
     async complete({ system, input, signal }) {
       const response = await fetch(`${baseUrl}/chat/completions`, {
@@ -40,7 +49,7 @@ export function createLlmClientFromWikiConfig(config) {
             { role: 'system', content: system },
             { role: 'user', content: input },
           ],
-          temperature: typeof llmConfig.temperature === 'number' ? llmConfig.temperature : 0.2,
+          ...temperatureBody,
         }),
       });
 
@@ -64,7 +73,7 @@ export function createLlmClientFromWikiConfig(config) {
       const body = {
         model,
         messages: allMessages,
-        temperature: typeof llmConfig.temperature === 'number' ? llmConfig.temperature : 0.2,
+        ...temperatureBody,
       };
       if (tools.length > 0) {
         body.tools = tools;
@@ -99,7 +108,7 @@ export function createLlmClientFromWikiConfig(config) {
       const body = {
         model,
         messages: allMessages,
-        temperature: typeof llmConfig.temperature === 'number' ? llmConfig.temperature : 0.2,
+        ...temperatureBody,
         stream: true,
       };
       if (tools.length > 0) {
@@ -195,7 +204,7 @@ export function createLlmClientFromWikiConfig(config) {
         body: JSON.stringify({
           model,
           messages: allMessages,
-          temperature: typeof llmConfig.temperature === 'number' ? llmConfig.temperature : 0.2,
+          ...temperatureBody,
           stream: true,
         }),
       });

package/src/agent/llm.test.js

@@ -0,0 +1,59 @@
+import assert from 'node:assert/strict';
+import test from 'node:test';
+import { createLlmClientFromWikiConfig } from './llm.js';
+
+function captureFetch(reply) {
+  const calls = [];
+  const original = globalThis.fetch;
+  globalThis.fetch = async (url, init) => {
+    calls.push({ url, body: JSON.parse(init.body) });
+    return { ok: true, json: async () => reply };
+  };
+  return { calls, restore: () => { globalThis.fetch = original; } };
+}
+
+const reply = { choices: [{ message: { content: 'ok' } }] };
+const gateway = {
+  provider: 'ai-gateway',
+  baseUrl: 'https://gw.example.com/v1',
+  apiKey: 'k',
+};
+
+test('omits temperature for a gpt-5-class model even when the profile sets one', async () => {
+  const { calls, restore } = captureFetch(reply);
+  try {
+    const client = createLlmClientFromWikiConfig({
+      llm: { ...gateway, model: 'openai/gpt-5-mini', temperature: 0.2 },
+    });
+    await client.complete({ system: 's', input: 'i' });
+    assert.equal('temperature' in calls[0].body, false);
+  } finally {
+    restore();
+  }
+});
+
+test('keeps the configured temperature for a model that accepts it', async () => {
+  const { calls, restore } = captureFetch(reply);
+  try {
+    const client = createLlmClientFromWikiConfig({
+      llm: { ...gateway, model: 'openai/gpt-4.1', temperature: 0.3 },
+    });
+    await client.complete({ system: 's', input: 'i' });
+    assert.equal(calls[0].body.temperature, 0.3);
+  } finally {
+    restore();
+  }
+});
+
+test('keeps the 0.2 default when the profile declares no temperature', async () => {
+  const { calls, restore } = captureFetch(reply);
+  try {
+    const client = createLlmClientFromWikiConfig({
+      llm: { ...gateway, model: 'openai/gpt-4.1' },
+    });
+    await client.complete({ system: 's', input: 'i' });
+    assert.equal(calls[0].body.temperature, 0.2);
+  } finally {
+    restore();
+  }
+});

package/src/cli/wiki-manager.js

@@ -18,7 +18,7 @@ import { ensureManagerScaffold, loadManagerEnv } from '../core/env.js';
 loadManagerEnv();
 import { createAgentGraph } from '../agent/graph.js';
 import { handleSlashCommand, printHelp, printVersion, refreshMcpRuntimeStatus } from '../commands/slash.js';
-import { runShell, runHeadlessChatTurn, sanitizeOpenWikiPages } from '../shell/repl.js';
+import { runShell, runHeadlessChatTurn, sanitizeOpenWikiPages, chatInputBudgetChars } from '../shell/repl.js';
 import { runPreflightChecks, withRuntimePreflight } from '../core/startupCheck.js';
 import { refreshRunningContainers } from '../core/wikiSetup.js';
 import { applySessionWikircProfile } from '../core/sessionConfig.js';
@@ -32,6 +32,7 @@ import { runAgentTurn, runAgenticLoop } from '../core/agentLoop.js';
 import { createDeltaCoalescer } from '../runtime/deltaCoalescer.js';
 import { TERMINAL_STATUS_SET } from '../orchestrator/taskStatuses.js';
 import { resolveCapabilityConcurrency } from '../orchestrator/scheduler.js';
+import { workspaceLockRegistry } from '../orchestrator/lockManager.js';
 import { capabilityRegistryForSession } from '../orchestrator/capabilityRegistry.js';
 import { CapabilityUnavailableError, resolve as resolveCapability } from '../orchestrator/capabilityResolver.js';
 import { listWorkspaces } from '../core/workspaces.js';
@@ -358,6 +359,9 @@ export function createInteractiveSession(context, { runtimeUrl, turnId, signal =
   session.headlessPlan = null;
   session.turnId = turnId ?? null;
   session._abortSignal = signal;
+  // The SAME lock registry as the workspace's runs, never a copy: a direct
+  // write from this turn must see what a run holds (plan-demandes-pendant-run.md).
+  session._workspaceLocks = workspaceLockRegistry(source);
   return session;
 }
 
@@ -956,7 +960,8 @@ async function runRuntime(argv, agent) {
   const { resolveRuntimeAuthToken } = await import('../runtime/auth.js');
   const { createSqliteQueueStore } = await import('../runtime/queueStore.js');
   const { createApprovalManager } = await import('../runtime/approvals.js');
-  const { conversationSeed, runRuntimeAgenticWorkflow } = await import('../runtime/runner.js');
+  const { conversationSeed, conversationCompactionPlan, compactionNoteForDonna, runRuntimeAgenticWorkflow } = await import('../runtime/runner.js');
+  const { summarizeCompactedConversation } = await import('../runtime/conversationCompact.js');
 
   // Same default as wiki-workspace runtime up and ensureRuntime: a loopback
   // bind is invisible to the serve container.
@@ -1741,13 +1746,17 @@ async function runRuntime(argv, agent) {
       await refreshMcpRuntimeStatus(context.session);
     }
     const ephemeral = createInteractiveSession(context, { runtimeUrl: selfRuntimeUrl, turnId, signal });
+    // Whether a run is active in this workspace, from the runtime itself: the
+    // projection says `pending_approval` for an ingest waiting on its approval,
+    // and reading only `running` made such a turn believe the workspace idle.
+    ephemeral._runActive = Boolean(context.running);
     // Seed from a freshly reduced COPY of persisted events. Interactive turn
     // events deliberately do not mutate the canonical run projection, so the
     // canonical session alone is not a reliable conversation-history source.
     const persistedProjection = reduceAgentEvents(store.listEvents({
       workspace: context.workspace ?? ephemeral.workspace ?? null,
     }));
-    const messages = conversationSeed({ agentProjection: persistedProjection }, input);
+    let messages = conversationSeed({ agentProjection: persistedProjection }, input);
     // Streaming fragments are coalesced before they are persisted and pushed —
     // one synchronous SQLite insert (plus one SSE write) per token stalled the
     // event loop, freezing both chats (serve and ShellUI) while a long answer
@@ -1779,6 +1788,37 @@ async function runRuntime(argv, agent) {
       workspace: context.workspace ?? null,
       payload: { message },
     }));
+    // Automatic memory compaction, before this turn's own message enters the
+    // conversation. Not during a run (the manual gauge refuses it too: the
+    // boundary must not move under a run's own messages). A compaction whose
+    // summary could not be produced moves nothing — dropping messages without
+    // their summary is the silent loss this exists to end. Donna, not the
+    // system, tells the user it happened.
+    const compaction = context.running ? null : conversationCompactionPlan(persistedProjection, {
+      budgetChars: chatInputBudgetChars(context.session?.wikircConfig),
+    });
+    if (compaction) {
+      ephemeral._onStep?.('Condensing the earlier conversation…');
+      const summary = await summarizeCompactedConversation(context.session, compaction);
+      if (summary && summary !== compaction.previousSummary) {
+        dispatchAgentEvent(ephemeral, createAgentEvent('conversation_reset', {
+          origin: 'runtime_turn',
+          turnId,
+          workspace: context.workspace ?? null,
+          payload: { summary, keepLast: compaction.keepLast, automatic: true, reason: compaction.reason },
+        }));
+        const compactedProjection = reduceAgentEvents(store.listEvents({
+          workspace: context.workspace ?? ephemeral.workspace ?? null,
+        }));
+        const count = compaction.segment.filter((message) => ['user', 'assistant'].includes(message?.role)).length;
+        // Right after the summary it refers to — never last: the chat branch
+        // drops a trailing user message as the current input.
+        const seed = conversationSeed({ agentProjection: compactedProjection }, input);
+        messages = [seed[0], { role: 'user', content: compactionNoteForDonna(count) }, ...seed.slice(1)];
+      } else {
+        emitRuntimeLog(context.session, 'conversation-compact: automatic compaction skipped — no new summary; the conversation window is unchanged');
+      }
+    }
     dispatchAgentEvent(ephemeral, createAgentEvent('user_message', {
       origin: 'runtime_turn',
       turnId,

package/src/commands/slash.js

@@ -144,8 +144,10 @@ function wikircSummaryText(summary) {
     `profile=${summary.profile}`,
     `file=${summary.fileName}`,
     `provider=${summary.provider ?? '-'}`,
+    `engine=${summary.engine ?? '-'}`,
     `model=${summary.model ?? '-'}`,
     `baseUrl=${summary.baseUrl ?? '-'}`,
+    `temperature=${summary.temperature ?? '-'}`,
     `language=${summary.language ?? '-'}`,
     `apiKey=${summary.hasApiKey ? 'configured' : 'missing'}`,
     `vector=${summary.vectorEnabled ? 'enabled' : 'disabled'}`,

package/src/core/agentEvents.js

@@ -5,6 +5,7 @@ import { applyPlanPatch, normalizePlanPatch, normalizePlanRevision, rebasePlanPa
 import { formatRuntimeLogPayload, isDispatchPlumbingLine, normalizeRuntimeLog, shortTaskLabel } from './runtimeLog.js';
 import { projectSkillChains, TERMINAL as CONTROL_TERMINAL_STATUSES } from './skillChainView.js';
 import { projectWorkflow } from './workflow.js';
+import { compactLogLabel } from './logLabel.js';
 import { validateContractInDev } from '../contracts/schemas.js';
 import { isActive, isTerminal, isSuccessful, isUnknownStatus, normalizeTaskStatus } from '../orchestrator/taskStatuses.js';
 
@@ -362,7 +363,16 @@ function applyEvent(state, event) {
       // (length at reset time) instead of clearing state.conversation is what
       // keeps the two concerns apart: the gauge/seed read conversationSeedStart,
       // the display reads conversation.
-      state.conversationSeedStart = state.conversation.length;
+      // `keepLast` (automatic compaction): the last exchanges stay verbatim in
+      // the seed — a follow-up such as « rajoute la liste » needs the previous
+      // answer word for word, not its summary. The manual gauge sends none.
+      {
+        const keepLast = Math.max(0, Math.floor(Number(event.payload?.keepLast) || 0));
+        state.conversationSeedStart = Math.max(
+          Number(state.conversationSeedStart) || 0,
+          state.conversation.length - keepLast,
+        );
+      }
       // The summary is best-effort (an LLM call the compact route makes before
       // dispatching this event): when it succeeds it REPLACES the previous one
       // — it is a rolling summary of "everything before this point", not an
@@ -1237,7 +1247,7 @@ function planTaskById(state, taskId) {
 function taskLabelFor(state, taskId) {
   const step = planTaskById(state, taskId);
   const label = step?.label ?? step?.description ?? null;
-  if (label && !/^Step \d+$/.test(label)) return label;
+  if (label && !/^Step \d+$/.test(label)) return compactLogLabel(label);
   return shortTaskLabel(taskId) || String(taskId ?? '') || 'task';
 }