@dotdrelle/wiki-manager 0.15.48 → 0.15.50

package/README.md

@@ -637,9 +637,15 @@ with `"replans": 1` in the `/run` body.
 Runtime approvals are bounded to a run, plan revision and approval class.
 Mutating orchestrated tasks **wait for approval by default**, including tasks
 created by a skill or a directly selected capability such as ingest or
-pipeline. Approve them by replying "valide tout", running `/approve`, or
-clicking Approve in either UI (Shell right-pane banner, or the `serve` banner
-above the composer). `POST /approve` also accepts an explicit run scope.
+pipeline. Approve them by running `/approve`, or clicking Approve in either UI:
+the Shell right-pane banner, or the `serve` banner.
+
+In `serve` that banner is a **fixed overlay, visible in every centre view**. It
+used to sit inside the composer, which the layout hides in the wiki, connectors
+and execution views — so a restore launched from `/history` waited on an
+approval nobody could see, and the Execution view, the one meant for monitoring,
+could not show it either. The Shell never had that gap because its plan pane is
+always on screen. `POST /approve` also accepts an explicit run scope.
 
 An explicitly launched skill is also approval-gated. For an orchestrated skill,
 the scheduler blocks each uncovered mutating task; for a `direct` skill, the
@@ -907,10 +913,35 @@ The shared `docker-compose.yml` starts one workspace stack:
 Use `wiki-workspace` whenever possible so Compose receives the right project
 name, env file, ports, and volume mounts.
 
+`PRODUCTION_ALLOWED_STEPS` gates what `production-mcp` will accept, and an
+omission from it is **silent**: `agent_plan` simply leaves the step's task out of
+the fragment instead of failing. `taxonomy` was missing from the shipped default
+for several releases, so every compose-deployed ingest ran without the taxonomy
+barrier and left the published map stale. Keep the variable in step with the
+in-code default of `production_mcp_server.py`: a test here asserts `taxonomy` is
+present, and one in `agent-wiki-production` compares the whole list against that
+in-code reference. Remember that an explicit value in your `.env` overrides the
+default entirely.
+
 Runtime split: the host manager/runtime uses Node.js 22+ for `node:sqlite`; the
 interactive OpenTUI shell uses Bun 1.2+; workspace Docker services run from the
 published images and do not depend on host `node_modules`.
 
+Two consequences worth knowing before debugging anything:
+
+- **The runtime is not a container.** `runtime/lifecycle.js` spawns it locally,
+  detached, from the manager sources — no Compose file declares it. Changing
+  runtime or shell code therefore needs a **restart**, never an image rebuild;
+  changing `llm-wiki` or an agent needs the image rebuilt.
+- **The runtime starts before the workspace containers.** Its first agent
+  discovery legitimately finds them absent. `agentRegistry` keeps a known
+  agent's capabilities when a probe fails — it only refreshes `lastSeenAt`, and
+  says so in the runtime log — and the periodic re-scan re-probes the MCP
+  endpoints instead of reusing a cached status. Without both, a capability the
+  agent really has stayed missing from the registry until the next successful
+  discovery, and the only symptom was a run failing much later with
+  `No agent provides capability …`.
+
 As of 0.11.4, the host runtime store carries a minimal format guard:
 `PRAGMA user_version = 1` in SQLite plus `.wiki/meta.json` with
 `schemaVersion: 1`. Unknown future versions stop startup with a clear error.

package/docker-compose.yml

@@ -124,7 +124,13 @@ services:
       - WORKSPACE_NAME=${WORKSPACE_NAME:-workspace}
       - WIKI_WORKSPACE_PATH=/workspace
       - WIKI_CONFIG_PATH=${WIKI_CONFIG_PATH:-}
-      - PRODUCTION_ALLOWED_STEPS=${PRODUCTION_ALLOWED_STEPS:-doctor,ingest,ingest_plan,ingest_apply,build,export,polish,restore,pipeline}
+      # `taxonomy` belongs here, and its absence is SILENT: agent_plan drops the
+      # taxonomy task from an ingest fragment when the step is not allowed
+      # (production_mcp_server.py, "taxonomy" in _ALLOWED_STEPS), without an
+      # error. Every compose-deployed ingest then ran without the Lot 4 barrier
+      # and left the published map stale — the very defect that work fixed.
+      # `copy` stays out on purpose: it is the legacy step, opt-in only.
+      - PRODUCTION_ALLOWED_STEPS=${PRODUCTION_ALLOWED_STEPS:-doctor,ingest,ingest_plan,ingest_apply,taxonomy,build,export,polish,restore,pipeline}
       - PRODUCTION_REQUIRE_CONFIRMATION=${PRODUCTION_REQUIRE_CONFIRMATION:-false}
       # Parallelism levers — effective concurrency ≈ recommendedConcurrency.
       # Intermediate defaults (4/8). Low profile 2/4, high profile 8/16.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dotdrelle/wiki-manager",
-  "version": "0.15.48",
+  "version": "0.15.50",
   "description": "Agentic shell and orchestration cockpit for llm-wiki workspaces.",
   "license": "PolyForm-Noncommercial-1.0.0",
   "author": "dotrelle",
@@ -11,7 +11,7 @@
   },
   "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/agentEvents.test.js src/core/skillChainView.test.js src/core/runtimeLog.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/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/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/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/approvalPolicy.test.js src/orchestrator/dispatcher.test.js src/orchestrator/objectiveResolver.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/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/auth.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/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/agentEvents.test.js src/core/skillChainView.test.js src/core/runtimeLog.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/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/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/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/approvalPolicy.test.js src/orchestrator/dispatcher.test.js src/orchestrator/objectiveResolver.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/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",
     "check-versions": "node scripts/check-versions.js",
     "prepack": "node scripts/check-versions.js",
     "prepublishOnly": "node scripts/check-versions.js",

package/src/agent/graph.js

@@ -26,7 +26,7 @@ import { createAgentEvent, dispatchAgentEvent } from '../core/agentEvents.js';
 import { enqueueProductionJob, ensureJobQueue, formatQueue, productionLockBusy } from '../core/jobQueue.js';
 import { loadWorkspaceProfile, updateWorkspaceProfilePreference } from '../core/profile.js';
 import { capabilityRegistryForSession } from '../orchestrator/capabilityRegistry.js';
-import { fetchRuntimeState, postRuntimeApprove, postRuntimeCancel, postRuntimeControl, postRuntimeDelegate, postRuntimeKill, postRuntimeSkill } from '../runtime/client.js';
+import { fetchRuntimeState, postRuntimeCancel, postRuntimeControl, postRuntimeDelegate, postRuntimeKill, postRuntimeSkill } from '../runtime/client.js';
 import { controlLanguage } from '../runtime/controlMessages.js';
 
 const MAX_TOOL_ITERATIONS = 80;
@@ -46,7 +46,7 @@ const MAX_SPINNER_ARG_LENGTH = 96;
 const INTERNAL_TOOL_SERVERS = {
   wiki: ['plan_set', 'plan_done'],
   shell: ['run_command', 'read_command', 'profile_update'],
-  runtime: ['kill', 'cancel', 'status', 'approve', 'enqueue', 'delegate', 'run_skill'],
+  runtime: ['kill', 'cancel', 'status', 'enqueue', 'delegate', 'run_skill'],
 };
 
 const AGENT_SLASH_COMMANDS = new Set([
@@ -164,15 +164,6 @@ const RUNTIME_STATUS_TOOL = {
   },
 };
 
-const RUNTIME_APPROVE_TOOL = {
-  type: 'function',
-  function: {
-    name: 'runtime__approve',
-    description: 'Grant the pending approval of the active runtime run (mutating tasks wait on it). Use when the user consents in ANY phrasing: "vas-y", "ok pour l\'export", "approuve", "valide". Confirm what was approved.',
-    parameters: { type: 'object', additionalProperties: false, properties: {} },
-  },
-};
-
 const RUNTIME_ENQUEUE_TOOL = {
   type: 'function',
   function: {
@@ -364,7 +355,6 @@ function toolDefinitionForCall(session, callName) {
     RUNTIME_STATUS_TOOL,
     RUNTIME_CANCEL_TOOL,
     RUNTIME_KILL_TOOL,
-    RUNTIME_APPROVE_TOOL,
     RUNTIME_ENQUEUE_TOOL,
     RUNTIME_DELEGATE_TOOL,
     RUNTIME_RUN_SKILL_TOOL,
@@ -852,28 +842,6 @@ export async function handleRuntimeControlTool(session, tool, args = {}) {
       const result = await postRuntimeCancel({ url, workspace });
       return result.cancelled ? 'Runtime run cancellation requested.' : `No active run to cancel${result.reason ? ` (${result.reason})` : ''}.`;
     }
-    if (tool === 'approve') {
-      const state = await fetchRuntimeState({ url, workspace });
-      const pending = (Array.isArray(state?.approvals) ? state.approvals : [])
-        .filter((approval) => approval.status === 'pending_approval');
-      const runId = state?.runId
-        ?? state?.runs?.find((run) => ['running', 'pending_approval'].includes(run.status))?.id
-        ?? null;
-      if (!runId || pending.length === 0) return 'No pending approval found.';
-      const approvalClasses = [...new Set(pending.flatMap((approval) => {
-        const value = approval.approvalClasses ?? approval.approvalClass ?? [];
-        return Array.isArray(value) ? value : [value];
-      }).map(String).filter(Boolean))];
-      const result = await postRuntimeApprove({
-        url,
-        workspace,
-        runId,
-        scope: 'run',
-        planRevision: state?.planRevision ?? null,
-        approvalClasses: approvalClasses.length > 0 ? approvalClasses : ['default'],
-      });
-      return result?.approved ? 'Current validated plan approved.' : 'No pending approval found.';
-    }
     if (tool === 'delegate') {
       const objective = String(args.objective ?? '').trim();
       if (!objective) return 'Delegation rejected: missing objective.';
@@ -1250,7 +1218,7 @@ export function buildAgentSystemPrompt(state) {
     workspaceProfile
       ? `Workspace profile (.wiki/profile.md) — durable user preferences, apply these to every reply (tone, tutoiement/vouvoiement, formatting, etc.):\n${workspaceProfile}`
       : null,
-    'Runtime control: you have runtime__status, runtime__cancel, runtime__kill, runtime__approve 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. When the user consents to a pending approval in any phrasing ("vas-y", "ok pour l\'export"), call runtime__approve. 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. 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.',
     '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.',
@@ -1285,7 +1253,7 @@ export function formatLlmUnavailableMessage(reason) {
 
 function toolsForClassification(classification, writeTools, session = null) {
   const controlTools = session?.runtime?.url
-    ? [RUNTIME_STATUS_TOOL, RUNTIME_CANCEL_TOOL, RUNTIME_KILL_TOOL, RUNTIME_APPROVE_TOOL, RUNTIME_ENQUEUE_TOOL]
+    ? [RUNTIME_STATUS_TOOL, RUNTIME_CANCEL_TOOL, RUNTIME_KILL_TOOL, RUNTIME_ENQUEUE_TOOL]
     : [];
   // Provider discovery and validation belong to the runtime. Hiding
   // delegation while the shell snapshot is temporarily empty forced Donna

package/src/agent/graph.test.js

@@ -782,67 +782,25 @@ test('runtime status does not manufacture a plan', async () => {
   }
 });
 
-test('one Donna approval grants the complete validated run revision', async () => {
-  const originalFetch = globalThis.fetch;
-  const requests = [];
-  globalThis.fetch = async (url, options = {}) => {
-    requests.push({ url: String(url), method: options.method ?? 'GET', body: options.body ? JSON.parse(options.body) : null });
-    if ((options.method ?? 'GET') === 'GET') {
-      return {
-        ok: true,
-        status: 200,
-        json: async () => ({
-          running: true,
-          runId: 'run-approval',
-          planRevision: 3,
-          approvals: [
-            { status: 'pending_approval', approvalClasses: ['workspace'] },
-            { status: 'pending_approval', approvalClasses: ['workspace'] },
-          ],
-        }),
-      };
-    }
-    return { ok: true, status: 202, json: async () => ({ approved: true, runId: 'run-approval' }) };
-  };
-  let calls = 0;
+test('Donna has no self-approval tool during an active run', async () => {
+  // Donna must never grant her own pending approval (the recurring "spontaneous
+  // approval" regression). Approval is the user's action, expressed through the
+  // banner button or /approve — never through an LLM tool call.
+  let seenTools = [];
   const session = sessionBase({
     runtime: { url: 'http://runtime.test' },
     agentProjection: { status: 'running', conversation: [], activities: [] },
     llm: {
-      async completeWithTools() {
-        calls += 1;
-        if (calls === 1) {
-          return {
-            content: null,
-            message: { role: 'assistant', content: null },
-            tool_calls: [{
-              id: 'approve-run',
-              type: 'function',
-              function: { name: 'runtime__approve', arguments: '{}' },
-            }],
-          };
-        }
-        return { content: 'Plan approuvé.', message: { role: 'assistant', content: 'Plan approuvé.' }, tool_calls: null };
+      async completeWithTools({ tools }) {
+        seenTools = tools.map((tool) => tool.function.name);
+        return { content: 'ok', message: { role: 'assistant', content: 'ok' }, tool_calls: null };
       },
     },
   });
 
-  try {
-    const result = await createAgentGraph().invoke({ input: 'oui', session });
-    assert.equal(result.response, 'Plan approuvé.');
-    const approval = requests.find((request) => request.url.includes('/approve'));
-    assert.deepEqual(approval.body, {
-      workspace: 'docs',
-      runId: 'run-approval',
-      itemId: null,
-      approvalId: null,
-      scope: 'run',
-      planRevision: 3,
-      approvalClasses: ['workspace'],
-    });
-  } finally {
-    globalThis.fetch = originalFetch;
-  }
+  await createAgentGraph().invoke({ input: 'oui', session });
+  assert.ok(seenTools.includes('runtime__status'), 'control tools still bound during an active run');
+  assert.ok(!seenTools.includes('runtime__approve'), 'no runtime__approve tool for Donna');
 });
 
 test('agent graph binds the full toolset for a "remember my preference" request, not just read-only tools', async () => {
@@ -1773,7 +1731,7 @@ test('agent graph lets Donna handle ambiguous input during a run with the contro
   assert.match(result.response, /mettre en file/);
   assert.ok(seenTools.includes('runtime__enqueue'));
   assert.ok(seenTools.includes('runtime__status'));
-  assert.ok(seenTools.includes('runtime__approve'));
+  assert.ok(!seenTools.includes('runtime__approve'), 'no self-approval tool for Donna');
   assert.ok(!seenTools.includes('production__production_start_job'), 'no write MCP tools during an active run for ambiguous intents');
 });
 

package/src/cli/wiki-manager.js

@@ -567,6 +567,7 @@ async function waitForRuntimeRun(session, log, { timeoutMs, pollMs = 1500, autoA
             scope: 'run',
             planRevision,
             approvalClasses: approvalClasses.length > 0 ? approvalClasses : ['default'],
+            caller: 'headless-auto-approve',
           });
           const line = `runtime-wait: auto-approved run ${currentRun.id} (revision ${planRevision})${result?.approved ? '' : ' [no pending approval matched]'}`;
           log.push(line); console.log(line);
@@ -845,6 +846,7 @@ export async function waitForRuntimeChain(session, log, {
             scope: 'run',
             planRevision,
             approvalClasses: approvalClasses.length > 0 ? approvalClasses : ['default'],
+            caller: 'chain-wait-auto-approve',
           });
           const line = `chain-wait: auto-approved run ${active.runId} (revision ${planRevision})`;
           log.push(line); console.log(line);
@@ -978,11 +980,12 @@ async function runRuntime(argv, agent) {
           payload: { message: `runtime: expired ${staleControl.length} stale queued control request(s) from a previous session` },
         }));
       }
-      session._onRuntimeError = (err) => {
+      session._onRuntimeError = (err, runId = null) => {
         const message = err instanceof Error ? err.message : String(err);
         dispatchAgentEvent(session, createAgentEvent('run_error', {
           origin: 'runtime',
-          payload: { message, workspace },
+          ...(runId ? { runId } : {}),
+          payload: { message, workspace, ...(runId ? { runId } : {}) },
         }));
       };
       context.approvalManager = createApprovalManager(session, {
@@ -991,7 +994,9 @@ async function runRuntime(argv, agent) {
           : undefined,
       });
       session._requestApproval = (request) => context.approvalManager.requestApproval(request);
-      context.supervisor = startActivitySupervisor(session);
+      context.supervisor = startActivitySupervisor(session, {
+        refreshMcp: () => refreshMcpRuntimeStatus(session),
+      });
       contexts.set(key, context);
       if (workspace && workspace !== key) contexts.set(workspace, context);
       return context;
@@ -1422,7 +1427,26 @@ async function runRuntime(argv, agent) {
         const provider = agents.find((item) => (item.description?.capabilities ?? [])
           .some((capability) => capability.id === body.capabilityPlan.capability));
         if (!provider?.serverName) {
-          throw new Error(`No agent provides capability ${body.capabilityPlan.capability}.`);
+          /*
+           Say what WAS seen, not only what was missing.
+
+           "No agent provides capability X." leaves three very different causes
+           indistinguishable: no agent registered at all, the right agent
+           registered but advertising a narrower set (an operation excluded by
+           PRODUCTION_ALLOWED_STEPS drops its whole capability from
+           agent_describe, silently), or a name mismatch. Listing the registry
+           turns the next occurrence into its own diagnosis instead of a guess.
+          */
+          const seen = agents.map((item) => {
+            const ids = (item.description?.capabilities ?? []).map((capability) => capability.id);
+            return `${item.serverName ?? '?'}[${ids.join(', ') || 'none'}]`;
+          });
+          throw new Error(
+            `No agent provides capability ${body.capabilityPlan.capability}. `
+            + (seen.length
+              ? `Registered agents: ${seen.join('; ')}.`
+              : 'No agent is registered: none answered agent_describe.'),
+          );
         }
         const fragment = parseJsonText(formatMcpToolResult(await callMcpTool(session.mcp, provider.serverName, 'agent_plan', {
           capability: body.capabilityPlan.capability,

package/src/commands/slash.js

@@ -1402,7 +1402,7 @@ export async function handleSlashCommand(line, context) {
           },
         });
         if (result?.runId) {
-          return { output: `▶ Run de capability accepté (${String(result.runId).slice(0, 8)}) — le plan de l'agent sera intégré et dispatché en parallèle ; approbation demandée avant les mutations (« valide tout » ou /approve).` };
+          return { output: `▶ Run de capability accepté (${String(result.runId).slice(0, 8)}) — le plan de l'agent sera intégré et dispatché en parallèle ; approbation demandée avant les mutations (/approve).` };
         }
         return { output: `Run non démarré: ${result?.explanation ?? result?.error ?? JSON.stringify(result)}` };
       }

package/src/core/agentEvents.js

@@ -561,7 +561,18 @@ function applyEvent(state, event) {
       return;
     case 'run_error':
       state.status = 'error';
-      state.logs.push(String(event.payload?.message ?? 'Agent run failed.'));
+      /*
+       "Run failed:" is part of the line, not decoration.
+
+       The serve journal keeps only the entries matching a keyword list
+       (failed, error, done, approval…). A run killed by a message that uses
+       none of those words — "No agent provides capability workspace.restore."
+       — was therefore filtered out as unimportant, and the panel showed "No
+       essential run event yet" over a run that had just died with its reason
+       already in hand. What ends a run is essential by construction; the
+       prefix states that instead of hoping the wording says so.
+      */
+      state.logs.push(`Run failed: ${String(event.payload?.message ?? 'Agent run failed.')}`);
       // A dead run must not leave "pending" plan steps and spinning
       // activities in the persisted projection: they reappeared as ghosts
       // at every relaunch ("des trucs dans le plan qui n'existent pas") and

package/src/core/agentEvents.test.js

@@ -675,3 +675,24 @@ test('a stream resumed after a discard carries only the final text', () => {
   assert.equal(projection.conversation[1].content, '12 pages.');
   assert.equal(projection.conversation[1].streaming, undefined, 'le message doit être figé');
 });
+
+test('run_error names the failure so the essential journal cannot filter it out', () => {
+  // La liste de mots-clés du journal serve (failed, error, done, approval…)
+  // rejetait un message qui n'en contient aucun — « No agent provides
+  // capability workspace.restore. » — et le panneau affichait « No essential
+  // run event yet » au-dessus d'un run mort avec sa raison déjà connue.
+  const state = reduceAgentEvents([
+    createAgentEvent('run_error', {
+      origin: 'runtime',
+      runId: 'run-1',
+      payload: { message: 'No agent provides capability workspace.restore.' },
+    }),
+  ]);
+
+  assert.equal(state.status, 'error');
+  const line = state.logs.at(-1);
+  assert.match(line, /^Run failed: /);
+  assert.match(line, /workspace\.restore/);
+  // Le mot qui rend l'entrée « essentielle » pour le journal serve.
+  assert.match(line, /failed/i);
+});

package/src/core/buildInfo.json

@@ -1,4 +1,4 @@
 {
-  "version": "0.15.48",
-  "commit": "7290b94"
+  "version": "0.15.50",
+  "commit": "844cf7e"
 }

package/src/core/dockerCompose.test.js

@@ -27,6 +27,24 @@ test('workspace production agent enables restore by default', async () => {
   assert.match(String(allowed), /(?:^|,)restore(?:,|})/);
 });
 
+test('every shipped default allows the taxonomy step', async () => {
+  /*
+   L'absence de `taxonomy` est SILENCIEUSE.
+
+   `agent_plan` retire la tâche taxonomique du fragment d'ingestion quand
+   l'étape n'est pas autorisée (`"taxonomy" in _ALLOWED_STEPS`), sans erreur ni
+   avertissement. Un déploiement par compose ingérait donc sans jamais publier
+   de taxonomie, et la carte restait périmée — exactement le défaut que le Lot 4
+   avait corrigé côté moteur. Le défaut de Python porte `taxonomy` ; un compose
+   POSITIONNE toujours la variable, donc ce défaut ne s'applique jamais là.
+  */
+  const raw = await readFile(new URL('../../docker-compose.yml', import.meta.url), 'utf8');
+  const compose = YAML.parse(raw);
+  const allowed = compose.services['production-mcp'].environment
+    .find((entry) => String(entry).startsWith('PRODUCTION_ALLOWED_STEPS='));
+  assert.match(String(allowed), /(?:^|,)taxonomy(?:,|})/);
+});
+
 test('shipped compose files never carry a build context', async () => {
   // Ces deux fichiers partent dans le paquet npm, où les dépôts frères
   // (`../agent-external/…`) n'existent pas : un `build:` y rend toute commande

package/src/core/mcp.js

@@ -1,7 +1,7 @@
 import { existsSync, readFileSync } from 'node:fs';
 import { managerEnvFile, managerMcpEndpointsFile, readEnvFile } from './env.js';
 
-const WIKI_MANAGER_VERSION = '0.15.48';
+const WIKI_MANAGER_VERSION = '0.15.50';
 
 function envValue(key) {
   const filePath = managerEnvFile();