pantheon-opencode 1.3.5-beta.60.efa533f → 1.3.6

package/src/plugins/tui/dist/tui.js

@@ -49,6 +49,20 @@ import { For, Show, createEffect, createMemo, createResource, createSignal, onCl
 *   (the old ContextBar was removed) and we deliberately do NOT hack the
 *   native footer. Track upstream: opencode statusline config.
 */
+const __tuiPluginOnceKey = "__pantheonPluginsLoaded";
+function pantheonPluginOnce(key) {
+	try {
+		const g = globalThis;
+		if (g[__tuiPluginOnceKey] === void 0) g[__tuiPluginOnceKey] = /* @__PURE__ */ new Set();
+		const set = g[__tuiPluginOnceKey];
+		if (!(set instanceof Set)) return false;
+		if (set.has(key)) return true;
+		set.add(key);
+		return false;
+	} catch {
+		return false;
+	}
+}
 /** How often the sidebar re-reads .pantheon/active-preset.json so `set-tier`
 *  changes made while opencode is open show up within ~30s. */
 const PRESET_REFRESH_MS = 3e4;
@@ -441,10 +455,10 @@ function parseConfig(raw) {
 	if (!toml?.parse) return defaultConfig();
 	const root = asTable(toml.parse(raw));
 	const cfg = defaultConfig();
-	const ui = asTable(root["ui"]);
-	cfg.showBars = bool(ui["show_bars"], cfg.showBars);
-	cfg.showStatus = bool(ui["show_status"], cfg.showStatus);
-	const rawWidth = ui["bar_width"];
+	const ui = asTable(root.ui);
+	cfg.showBars = bool(ui.show_bars, cfg.showBars);
+	cfg.showStatus = bool(ui.show_status, cfg.showStatus);
+	const rawWidth = ui.bar_width;
 	if (typeof rawWidth === "number" && Number.isFinite(rawWidth) && rawWidth >= 1) cfg.barWidth = Math.min(40, Math.floor(rawWidth));
 	for (const id of [
 		"anthropic",
@@ -453,14 +467,14 @@ function parseConfig(raw) {
 	]) {
 		const t = asTable(root[id]);
 		const p = cfg.providers[id];
-		p.enabled = bool(t["enabled"], p.enabled);
-		p.show["5h"] = bool(t["show_5h"], p.show["5h"]);
-		p.show["7d"] = bool(t["show_7d"], p.show["7d"]);
-		p.show["1m"] = bool(t["show_1m"], p.show["1m"]);
-		p.show.model = bool(t["show_model"], p.show.model);
-		const credentialsPath = str(t["credentials_path"]);
+		p.enabled = bool(t.enabled, p.enabled);
+		p.show["5h"] = bool(t.show_5h, p.show["5h"]);
+		p.show["7d"] = bool(t.show_7d, p.show["7d"]);
+		p.show["1m"] = bool(t.show_1m, p.show["1m"]);
+		p.show.model = bool(t.show_model, p.show.model);
+		const credentialsPath = str(t.credentials_path);
 		if (credentialsPath !== void 0) p.credentialsPath = credentialsPath;
-		const codexAuthPath = str(t["codex_auth_path"]);
+		const codexAuthPath = str(t.codex_auth_path);
 		if (codexAuthPath !== void 0) p.codexAuthPath = codexAuthPath;
 	}
 	return cfg;
@@ -751,7 +765,7 @@ function parseDelegationMarkdown(raw, fileAlias, sessionID = "") {
 	const startedAt = started !== void 0 ? Date.parse(started) : NaN;
 	if (agent === void 0 || state === void 0 || Number.isNaN(startedAt)) return null;
 	const normalized = state.toLowerCase();
-	if (normalized !== "running" && normalized !== "completed" && normalized !== "error" && normalized !== "cancelled") return null;
+	if (normalized !== "running" && normalized !== "completed" && normalized !== "error" && normalized !== "startup_failed" && normalized !== "startup_unknown" && normalized !== "cancelled") return null;
 	const finalizedAt = finalized !== void 0 ? Date.parse(finalized) : NaN;
 	return {
 		alias: title ?? (fileAlias !== void 0 ? stripMdSuffix(fileAlias) : "unknown"),
@@ -884,7 +898,7 @@ function delegationElapsed(entry, now) {
 * state machine testable without booting OpenCode's renderer. */
 function delegationActivity(entry) {
 	if (entry.state === "completed") return "completed";
-	if (entry.state === "error") return "error";
+	if (entry.state === "error" || entry.state === "startup_failed" || entry.state === "startup_unknown") return "error";
 	if (entry.state === "cancelled") return "cancelled";
 	if (entry.read) return "reading";
 	if (entry.alias.startsWith("live-") && entry.taskID === void 0) return "delegating";
@@ -933,6 +947,7 @@ function mergeChildDelegationSources(children, live) {
 		bySessionAlias.set(key(entry.sessionID, entry.alias), index);
 	});
 	for (const liveEntry of live) {
+		if (liveEntry.alias === null && liveEntry.taskID === null && Date.now() - liveEntry.startedAt > 3e4) continue;
 		const incoming = toDelegationEntry(liveEntry);
 		const index = (incoming.taskID !== void 0 ? byTask.get(incoming.taskID) : void 0) ?? (incoming.alias !== "" ? bySessionAlias.get(key(incoming.sessionID, incoming.alias)) : void 0);
 		if (index === void 0) {
@@ -944,7 +959,12 @@ function mergeChildDelegationSources(children, live) {
 		}
 		const existing = result[index];
 		if (existing === void 0) continue;
-		if (existing.source === "md" && existing.state !== "running") continue;
+		if (existing.state !== "running") {
+			if (liveEntry.alias !== null && existing.alias !== incoming.alias) existing.alias = incoming.alias;
+			if (incoming.agent !== "agent" && existing.agent !== incoming.agent) existing.agent = incoming.agent;
+			if (incoming.read && !existing.read) existing.read = true;
+			continue;
+		}
 		result[index] = {
 			...existing,
 			alias: liveEntry.alias !== null ? incoming.alias : existing.alias,
@@ -1530,6 +1550,7 @@ function View(props) {
 					...c,
 					status: resolveStatus(c.id)
 				})), md, now());
+				for (const [k, v] of props.liveStore.map) if (v.alias === null && v.taskID === null && Date.now() - v.startedAt > 3e4) props.liveStore.map.delete(k);
 				const liveEntries = [...props.liveStore.map.values()].filter((entry) => entry.sessionID === sessionID);
 				setChildDelegations(mergeChildDelegationSources(childEntries, liveEntries));
 				panelLog.info(`panel: children=${children.length} md=${md.length} events=${eventRefreshCount}`);
@@ -1768,6 +1789,7 @@ function View(props) {
 	})();
 }
 const tui = (api, _options, _meta) => {
+	if (pantheonPluginOnce("pantheon:tui")) return;
 	const [version, setVersion] = createSignal(null);
 	detectVersion(api).then((detected) => setVersion(detected)).catch(() => setVersion(null));
 	setupUsageBar(api);

package/src/plugins/tui/dist/tui.tsx

@@ -63,6 +63,24 @@ import {
   Show,
 } from 'solid-js'
 
+// ─── Double-load guard (mirrors src/pantheon/plugin-once.ts) ───────────────
+// Opencode merges global + project plugin paths, so this module can load twice
+// in one process. The second factory invocation becomes a no-op.
+const __tuiPluginOnceKey = '__pantheonPluginsLoaded'
+function pantheonPluginOnce(key: string): boolean {
+  try {
+    const g = globalThis as unknown as Record<string, unknown>
+    if (g[__tuiPluginOnceKey] === undefined) g[__tuiPluginOnceKey] = new Set<string>()
+    const set = g[__tuiPluginOnceKey]
+    if (!(set instanceof Set)) return false
+    if ((set as Set<string>).has(key)) return true
+    ;(set as Set<string>).add(key)
+    return false
+  } catch {
+    return false
+  }
+}
+
 /* ─── Constants ─────────────────────────────────────────── */
 
 /** How often the sidebar re-reads .pantheon/active-preset.json so `set-tier`
@@ -659,25 +677,25 @@ function parseConfig(raw: string): UsageBarConfig {
   const root = asTable(toml.parse(raw))
   const cfg = defaultConfig()
 
-  const ui = asTable(root['ui'])
-  cfg.showBars = bool(ui['show_bars'], cfg.showBars)
-  cfg.showStatus = bool(ui['show_status'], cfg.showStatus)
-  const rawWidth = ui['bar_width']
+  const ui = asTable(root.ui)
+  cfg.showBars = bool(ui.show_bars, cfg.showBars)
+  cfg.showStatus = bool(ui.show_status, cfg.showStatus)
+  const rawWidth = ui.bar_width
   if (typeof rawWidth === 'number' && Number.isFinite(rawWidth) && rawWidth >= 1)
     cfg.barWidth = Math.min(40, Math.floor(rawWidth))
 
   for (const id of ['anthropic', 'openai', 'opencodego'] as ProviderId[]) {
     const t = asTable(root[id])
     const p = cfg.providers[id]
-    p.enabled = bool(t['enabled'], p.enabled)
-    p.show['5h'] = bool(t['show_5h'], p.show['5h'])
-    p.show['7d'] = bool(t['show_7d'], p.show['7d'])
-    p.show['1m'] = bool(t['show_1m'], p.show['1m'])
-    p.show.model = bool(t['show_model'], p.show.model)
+    p.enabled = bool(t.enabled, p.enabled)
+    p.show['5h'] = bool(t.show_5h, p.show['5h'])
+    p.show['7d'] = bool(t.show_7d, p.show['7d'])
+    p.show['1m'] = bool(t.show_1m, p.show['1m'])
+    p.show.model = bool(t.show_model, p.show.model)
     // Assign only when present — keeps `exactOptionalPropertyTypes` happy.
-    const credentialsPath = str(t['credentials_path'])
+    const credentialsPath = str(t.credentials_path)
     if (credentialsPath !== undefined) p.credentialsPath = credentialsPath
-    const codexAuthPath = str(t['codex_auth_path'])
+    const codexAuthPath = str(t.codex_auth_path)
     if (codexAuthPath !== undefined) p.codexAuthPath = codexAuthPath
   }
   return cfg
@@ -924,7 +942,7 @@ export type DelegationEntry = {
   taskID?: string
   /** Agent name, e.g. "apollo". */
   agent: string
-  state: 'running' | 'completed' | 'error' | 'cancelled'
+  state: 'running' | 'completed' | 'error' | 'startup_failed' | 'startup_unknown' | 'cancelled'
   /** Epoch ms of the `Started` header. */
   startedAt: number
   /** Epoch ms of the `Finalized` header — null while still running. */
@@ -1034,6 +1052,8 @@ export function parseDelegationMarkdown(
     normalized !== 'running' &&
     normalized !== 'completed' &&
     normalized !== 'error' &&
+    normalized !== 'startup_failed' &&
+    normalized !== 'startup_unknown' &&
     normalized !== 'cancelled'
   ) {
     return null
@@ -1203,7 +1223,12 @@ export type DelegationActivity =
 
 export function delegationActivity(entry: DelegationEntry): DelegationActivity {
   if (entry.state === 'completed') return 'completed'
-  if (entry.state === 'error') return 'error'
+  if (
+    entry.state === 'error' ||
+    entry.state === 'startup_failed' ||
+    entry.state === 'startup_unknown'
+  )
+    return 'error'
   if (entry.state === 'cancelled') return 'cancelled'
   if (entry.read) return 'reading'
   if (entry.alias.startsWith('live-') && entry.taskID === undefined) return 'delegating'
@@ -1256,6 +1281,12 @@ export function mergeChildDelegationSources(
   })
 
   for (const liveEntry of live) {
+    if (
+      liveEntry.alias === null &&
+      liveEntry.taskID === null &&
+      Date.now() - liveEntry.startedAt > 30_000
+    )
+      continue
     const incoming = toDelegationEntry(liveEntry)
     const index =
       (incoming.taskID !== undefined ? byTask.get(incoming.taskID) : undefined) ??
@@ -1273,8 +1304,26 @@ export function mergeChildDelegationSources(
 
     const existing = result[index]
     if (existing === undefined) continue
-    const finalized = existing.source === 'md' && existing.state !== 'running'
-    if (finalized) continue
+    // Any terminal state is authoritative — a child session that went idle
+    // correctly derives 'completed' via childStatusToState before the MD
+    // report is written. The live tool-part channel stays 'running' until
+    // pantheon_delegation_read fires, so allowing it to overwrite a
+    // terminal state would flip completed → running (the bug this fixes).
+    //
+    // When the existing entry IS terminal, we still absorb live metadata
+    // (alias, agent, read) but preserve the terminal state.
+    if (existing.state !== 'running') {
+      if (liveEntry.alias !== null && existing.alias !== incoming.alias) {
+        existing.alias = incoming.alias
+      }
+      if (incoming.agent !== 'agent' && existing.agent !== incoming.agent) {
+        existing.agent = incoming.agent
+      }
+      if (incoming.read && !existing.read) {
+        existing.read = true
+      }
+      continue
+    }
     result[index] = {
       ...existing,
       alias: liveEntry.alias !== null ? incoming.alias : existing.alias,
@@ -2201,6 +2250,9 @@ function View(props: {
         }
         const withStatus = children.map((c) => ({ ...c, status: resolveStatus(c.id) }))
         const childEntries = childrenToDelegationEntries(withStatus, md, now())
+        for (const [k, v] of props.liveStore.map)
+          if (v.alias === null && v.taskID === null && Date.now() - v.startedAt > 30_000)
+            props.liveStore.map.delete(k)
         // The event channel is optimistic: it renders a job even while the
         // child session/report is still being created. Filter by parent so a
         // different focused session never leaks rows into this sidebar.
@@ -2397,6 +2449,7 @@ function View(props: {
 /* ─── Plugin Registration ────────────────────────────────── */
 
 const tui: TuiPlugin = (api, _options, _meta) => {
+  if (pantheonPluginOnce('pantheon:tui')) return
   // Do not await project file/process probes here. OpenCode waits for the TUI
   // factory before it finishes startup, and a Windows host opening a WSL
   // project can leave those probes pending while the project bridge comes up.
@@ -2467,7 +2520,9 @@ const tui: TuiPlugin = (api, _options, _meta) => {
     order: 900,
     slots: {
       sidebar_content(_ctx, props) {
-        return <View api={api} sessionID={props.session_id} version={version()} liveStore={liveStore} />
+        return (
+          <View api={api} sessionID={props.session_id} version={version()} liveStore={liveStore} />
+        )
       },
     },
   })

package/src/plugins/tui/package.json

@@ -1,7 +1,7 @@
 {
   "name": "pantheon-tui",
   "type": "module",
-  "version": "1.3.5-beta.60.efa533f",
+  "version": "1.3.6",
   "description": "Pantheon TUI plugin — sidebar (version, preset, sessions, real-time delegations) + AI subscription usage gauges incl. OpenCode Go/Zen dollar limits (vendored, MIT)",
   "license": "MIT",
   "exports": {

package/src/plugins/tui/src/index.tsx

@@ -63,6 +63,24 @@ import {
   Show,
 } from 'solid-js'
 
+// ─── Double-load guard (mirrors src/pantheon/plugin-once.ts) ───────────────
+// Opencode merges global + project plugin paths, so this module can load twice
+// in one process. The second factory invocation becomes a no-op.
+const __tuiPluginOnceKey = '__pantheonPluginsLoaded'
+function pantheonPluginOnce(key: string): boolean {
+  try {
+    const g = globalThis as unknown as Record<string, unknown>
+    if (g[__tuiPluginOnceKey] === undefined) g[__tuiPluginOnceKey] = new Set<string>()
+    const set = g[__tuiPluginOnceKey]
+    if (!(set instanceof Set)) return false
+    if ((set as Set<string>).has(key)) return true
+    ;(set as Set<string>).add(key)
+    return false
+  } catch {
+    return false
+  }
+}
+
 /* ─── Constants ─────────────────────────────────────────── */
 
 /** How often the sidebar re-reads .pantheon/active-preset.json so `set-tier`
@@ -659,25 +677,25 @@ function parseConfig(raw: string): UsageBarConfig {
   const root = asTable(toml.parse(raw))
   const cfg = defaultConfig()
 
-  const ui = asTable(root['ui'])
-  cfg.showBars = bool(ui['show_bars'], cfg.showBars)
-  cfg.showStatus = bool(ui['show_status'], cfg.showStatus)
-  const rawWidth = ui['bar_width']
+  const ui = asTable(root.ui)
+  cfg.showBars = bool(ui.show_bars, cfg.showBars)
+  cfg.showStatus = bool(ui.show_status, cfg.showStatus)
+  const rawWidth = ui.bar_width
   if (typeof rawWidth === 'number' && Number.isFinite(rawWidth) && rawWidth >= 1)
     cfg.barWidth = Math.min(40, Math.floor(rawWidth))
 
   for (const id of ['anthropic', 'openai', 'opencodego'] as ProviderId[]) {
     const t = asTable(root[id])
     const p = cfg.providers[id]
-    p.enabled = bool(t['enabled'], p.enabled)
-    p.show['5h'] = bool(t['show_5h'], p.show['5h'])
-    p.show['7d'] = bool(t['show_7d'], p.show['7d'])
-    p.show['1m'] = bool(t['show_1m'], p.show['1m'])
-    p.show.model = bool(t['show_model'], p.show.model)
+    p.enabled = bool(t.enabled, p.enabled)
+    p.show['5h'] = bool(t.show_5h, p.show['5h'])
+    p.show['7d'] = bool(t.show_7d, p.show['7d'])
+    p.show['1m'] = bool(t.show_1m, p.show['1m'])
+    p.show.model = bool(t.show_model, p.show.model)
     // Assign only when present — keeps `exactOptionalPropertyTypes` happy.
-    const credentialsPath = str(t['credentials_path'])
+    const credentialsPath = str(t.credentials_path)
     if (credentialsPath !== undefined) p.credentialsPath = credentialsPath
-    const codexAuthPath = str(t['codex_auth_path'])
+    const codexAuthPath = str(t.codex_auth_path)
     if (codexAuthPath !== undefined) p.codexAuthPath = codexAuthPath
   }
   return cfg
@@ -924,7 +942,7 @@ export type DelegationEntry = {
   taskID?: string
   /** Agent name, e.g. "apollo". */
   agent: string
-  state: 'running' | 'completed' | 'error' | 'cancelled'
+  state: 'running' | 'completed' | 'error' | 'startup_failed' | 'startup_unknown' | 'cancelled'
   /** Epoch ms of the `Started` header. */
   startedAt: number
   /** Epoch ms of the `Finalized` header — null while still running. */
@@ -1034,6 +1052,8 @@ export function parseDelegationMarkdown(
     normalized !== 'running' &&
     normalized !== 'completed' &&
     normalized !== 'error' &&
+    normalized !== 'startup_failed' &&
+    normalized !== 'startup_unknown' &&
     normalized !== 'cancelled'
   ) {
     return null
@@ -1203,7 +1223,12 @@ export type DelegationActivity =
 
 export function delegationActivity(entry: DelegationEntry): DelegationActivity {
   if (entry.state === 'completed') return 'completed'
-  if (entry.state === 'error') return 'error'
+  if (
+    entry.state === 'error' ||
+    entry.state === 'startup_failed' ||
+    entry.state === 'startup_unknown'
+  )
+    return 'error'
   if (entry.state === 'cancelled') return 'cancelled'
   if (entry.read) return 'reading'
   if (entry.alias.startsWith('live-') && entry.taskID === undefined) return 'delegating'
@@ -1256,6 +1281,12 @@ export function mergeChildDelegationSources(
   })
 
   for (const liveEntry of live) {
+    if (
+      liveEntry.alias === null &&
+      liveEntry.taskID === null &&
+      Date.now() - liveEntry.startedAt > 30_000
+    )
+      continue
     const incoming = toDelegationEntry(liveEntry)
     const index =
       (incoming.taskID !== undefined ? byTask.get(incoming.taskID) : undefined) ??
@@ -1273,8 +1304,26 @@ export function mergeChildDelegationSources(
 
     const existing = result[index]
     if (existing === undefined) continue
-    const finalized = existing.source === 'md' && existing.state !== 'running'
-    if (finalized) continue
+    // Any terminal state is authoritative — a child session that went idle
+    // correctly derives 'completed' via childStatusToState before the MD
+    // report is written. The live tool-part channel stays 'running' until
+    // pantheon_delegation_read fires, so allowing it to overwrite a
+    // terminal state would flip completed → running (the bug this fixes).
+    //
+    // When the existing entry IS terminal, we still absorb live metadata
+    // (alias, agent, read) but preserve the terminal state.
+    if (existing.state !== 'running') {
+      if (liveEntry.alias !== null && existing.alias !== incoming.alias) {
+        existing.alias = incoming.alias
+      }
+      if (incoming.agent !== 'agent' && existing.agent !== incoming.agent) {
+        existing.agent = incoming.agent
+      }
+      if (incoming.read && !existing.read) {
+        existing.read = true
+      }
+      continue
+    }
     result[index] = {
       ...existing,
       alias: liveEntry.alias !== null ? incoming.alias : existing.alias,
@@ -2201,6 +2250,9 @@ function View(props: {
         }
         const withStatus = children.map((c) => ({ ...c, status: resolveStatus(c.id) }))
         const childEntries = childrenToDelegationEntries(withStatus, md, now())
+        for (const [k, v] of props.liveStore.map)
+          if (v.alias === null && v.taskID === null && Date.now() - v.startedAt > 30_000)
+            props.liveStore.map.delete(k)
         // The event channel is optimistic: it renders a job even while the
         // child session/report is still being created. Filter by parent so a
         // different focused session never leaks rows into this sidebar.
@@ -2397,6 +2449,7 @@ function View(props: {
 /* ─── Plugin Registration ────────────────────────────────── */
 
 const tui: TuiPlugin = (api, _options, _meta) => {
+  if (pantheonPluginOnce('pantheon:tui')) return
   // Do not await project file/process probes here. OpenCode waits for the TUI
   // factory before it finishes startup, and a Windows host opening a WSL
   // project can leave those probes pending while the project bridge comes up.
@@ -2467,7 +2520,9 @@ const tui: TuiPlugin = (api, _options, _meta) => {
     order: 900,
     slots: {
       sidebar_content(_ctx, props) {
-        return <View api={api} sessionID={props.session_id} version={version()} liveStore={liveStore} />
+        return (
+          <View api={api} sessionID={props.session_id} version={version()} liveStore={liveStore} />
+        )
       },
     },
   })

package/src/plugins/tui/tsdown.config.ts

@@ -1,5 +1,5 @@
-import { defineConfig } from 'tsdown'
 import solid from 'rolldown-plugin-solid'
+import { defineConfig } from 'tsdown'
 
 export default defineConfig({
   entry: { tui: 'src/index.tsx', server: 'src/server.ts' },

package/src/routing.yml

@@ -151,6 +151,7 @@ presets:
 agents:
   zeus:
     role: Central orchestrator
+    max_steps: 40   # R4: step budget — at max_steps the agent is forced to summarize-and-stop. Tracker is permanent-per-process (no reset wiring).
     reasoning_effort: medium
     user_invocable: true
     can_delegate: true
@@ -189,6 +190,7 @@ agents:
     - incremental-implementation
   athena:
     role: Strategic planner
+    max_steps: 15   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: high
     user_invocable: true
     can_delegate: true
@@ -208,6 +210,7 @@ agents:
     - memory-bank
   apollo:
     role: Investigation scout
+    max_steps: 25   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: low
     user_invocable: false
     can_delegate: false
@@ -225,6 +228,7 @@ agents:
     - auto-continue
   hermes:
     role: Backend specialist
+    max_steps: 60   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: medium
     user_invocable: true
     can_delegate: true
@@ -247,6 +251,7 @@ agents:
     - incremental-implementation
   aphrodite:
     role: Frontend specialist
+    max_steps: 60   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: medium
     user_invocable: true
     can_delegate: false
@@ -267,6 +272,7 @@ agents:
     - file-prompts
   demeter:
     role: Database specialist
+    max_steps: 50   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: medium
     user_invocable: true
     can_delegate: false
@@ -286,6 +292,7 @@ agents:
     - incremental-implementation
   themis:
     role: Quality gate
+    max_steps: 25   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: high
     user_invocable: true
     can_delegate: false
@@ -307,6 +314,7 @@ agents:
     - context-compression
   prometheus:
     role: Infrastructure specialist
+    max_steps: 30   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: medium
     user_invocable: true
     can_delegate: false
@@ -325,6 +333,7 @@ agents:
     - incremental-implementation
   hephaestus:
     role: AI pipelines
+    max_steps: 50   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: medium
     user_invocable: true
     can_delegate: false
@@ -343,6 +352,7 @@ agents:
     - auto-continue
   nyx:
     role: Observability
+    max_steps: 25   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: low
     user_invocable: true
     can_delegate: false
@@ -361,6 +371,7 @@ agents:
     - auto-continue
   gaia:
     role: Remote sensing domain specialist
+    max_steps: 25   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: low
     user_invocable: false
     can_delegate: false
@@ -375,6 +386,7 @@ agents:
     - auto-continue
   iris:
     role: GitHub operations
+    max_steps: 20   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: low
     user_invocable: true
     can_delegate: false
@@ -392,6 +404,7 @@ agents:
     - artifact-management
   mnemosyne:
     role: Memory bank
+    max_steps: 20   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: low
     user_invocable: true
     can_delegate: false
@@ -415,6 +428,7 @@ agents:
     - session-goal
   talos:
     role: Hotfix agent
+    max_steps: 15   # R4: step budget — at max_steps the agent is forced to summarize-and-stop
     reasoning_effort: low
     user_invocable: true
     can_delegate: false
@@ -457,6 +471,11 @@ intent_gate:
     - search
     - lookup
     - documentation
+    - procure
+    - busque
+    - encontre
+    - localize
+    - pesquise
     target: apollo
   investigate:
     patterns:
@@ -568,6 +587,34 @@ background_delegation:
   read_only_agents: [apollo, gaia]
   session_max: 3
   retry_count: 1
+# ── R1: Per-error-type retry + provider cooldown (LiteLLM pattern) ──────
+# retry_policy: max retries per error class before escalating. Exponential
+# backoff (base 1s, doubling, capped 30s) applies between retries. auth
+# errors are NEVER retried (0) — a credential problem will not fix itself.
+# Enforcement: zeus-delegate-with-retry (manual waves) + RetryPolicyEngine.
+retry_policy:
+  auth: 0
+  rate_limit: 3
+  timeout: 2
+  other: 1
+# cooldown: after `allowed_fails` consecutive failures for a provider, skip
+# that provider for `cooldown_time_seconds`. A success resets the counter.
+# In-memory tracker (no Redis) — process-wide, shared across retry helpers.
+cooldown:
+  allowed_fails: 3
+  cooldown_time_seconds: 60
+# ── O5: permission.task — glob rules for which subagents an agent may
+# invoke (last matching rule wins; deny removes the agent from the delegate
+# tool description entirely — prevents circular delegation at the
+# permission layer). Default "*": allow keeps the existing runtime matrix
+# (zeus → anyone; athena/hermes → apollo only). Example deny-first config:
+#   permission:
+#     task:
+#       "*": deny
+#       "orchestrator-*": allow
+permission:
+  task:
+    "*": allow
 todo_enforcer:
   enabled: true
   cooldown_base_ms: 5000            # base backoff between injections

package/src/agents/README.md

@@ -1,26 +0,0 @@
-# Agent Reference — Pantheon
-
-This directory contains the definitions for all 14 Pantheon agents. Each agent is defined in a `.agent.md` file with YAML frontmatter specifying tools, model tier, capabilities, and delegation rules.
-
-## Agents
-
-| Agent | File | Role |
-|-------|------|------|
-| @zeus | [zeus.agent.md](zeus.agent.md) | Central orchestrator |
-| @athena | [athena.agent.md](athena.agent.md) | Strategic planner |
-| @apollo | [apollo.agent.md](apollo.agent.md) | Investigation scout |
-| @hermes | [hermes.agent.md](hermes.agent.md) | Backend specialist |
-| @aphrodite | [aphrodite.agent.md](aphrodite.agent.md) | Frontend specialist |
-| @demeter | [demeter.agent.md](demeter.agent.md) | Database specialist |
-| @themis | [themis.agent.md](themis.agent.md) | Quality & security gate |
-| @prometheus | [prometheus.agent.md](prometheus.agent.md) | Infrastructure specialist |
-| @hephaestus | [hephaestus.agent.md](hephaestus.agent.md) | AI pipelines |
-| @nyx | [nyx.agent.md](nyx.agent.md) | Observability |
-| @gaia | [gaia.agent.md](gaia.agent.md) | Remote sensing |
-| @iris | [iris.agent.md](iris.agent.md) | GitHub operations |
-| @mnemosyne | [mnemosyne.agent.md](mnemosyne.agent.md) | Memory bank |
-| @talos | [talos.agent.md](talos.agent.md) | Hotfix agent |
-
-## Full Reference
-
-See [AGENTS.md](../../AGENTS.md) for the complete agent reference with roles, capabilities, and invocation patterns.