@staix/agent-hub 0.12.19 → 0.12.21

package/src/models/route/relay-selector.ts

@@ -1,17 +1,24 @@
 import type { ModelBackend, ModelRelayOptions, RelayRequest } from "../relay.ts";
 import { normalizeConversation } from "./normalize.ts";
-import { extractToolSignals } from "./signals.ts";
-import { selectStage, type StageState } from "./stage.ts";
+import { extractToolSignals, turnKind } from "./signals.ts";
+import { selectStage, stayOrSwitch, type StageState, type SwitchTrace, type Tier } from "./stage.ts";
 import { SessionState } from "./state.ts";
 
-type RelaySelectorOptions = Pick<ModelRelayOptions, "mlx" | "selectBackend" | "routeSessionKey" | "onRoute" | "allowedDGXmodels">;
-type RouteSource = "override" | "dimensions" | "hold" | "classifier" | "default";
-type RelayRouteEvent = { route: "hub/auto"; tier: string; source: RouteSource; score: number; ms: number };
+type RelaySelectorOptions = Pick<ModelRelayOptions, "mlx" | "selectBackend" | "routeSessionKey" | "allowedDGXmodels" | "staySwitch"> & { dgxMaxInputTokens: number };
+type RouteSource = "override" | "dimensions" | "hold" | "classifier" | "default" | "load" | "cooldown" | "pin";
+/** `stateless`: the request carried no session key, so the planner saw a new session on every call. */
+export type RelayRouteEvent = { route: "hub/auto"; tier: string; source: RouteSource; score: number; ms: number; stateless?: boolean } & Partial<SwitchTrace>;
 type EstimateInputTokens = (messages: RelayRequest["messages"], tools?: unknown[]) => number;
+type AutoState = { stage: StageState; pin?: { tier: Tier; alias: string } };
+/**
+ * A `hub/auto` decision (#199). `backend` is the stage backend; the relay only reorders it with its fallback, trying
+ * `prefer` (an enforced tool loop's pinned backend) first when that is the fallback, and reports a reorder through `moved`.
+ */
+export interface AutoChoice { backend: ModelBackend; route: RelayRouteEvent; movable: boolean; prefer?: string; moved: (alias: string, source: "load" | "cooldown" | "pin") => void }
 
 /** Stage selection for the relay's virtual `hub/auto` model. */
 export class AutoRouteSelector {
-  private readonly states = new SessionState<StageState>(512, 60 * 60_000);
+  private readonly states = new SessionState<AutoState>(512, 60 * 60_000);
 
   constructor(
     private readonly options: RelaySelectorOptions,
@@ -20,33 +27,38 @@ export class AutoRouteSelector {
     private readonly estimateInputTokens: EstimateInputTokens,
   ) {}
 
-  async select(body: RelayRequest): Promise<ModelBackend> {
+  async select(body: RelayRequest): Promise<AutoChoice> {
     const started = performance.now();
     let sessionKey: string | undefined;
     try {
       const key = this.options.routeSessionKey?.(body);
       if (typeof key === "string" && key.length > 0 && key.length <= 512) sessionKey = key;
     } catch { /* host identity lookup is fail-open */ }
-    const priorState = sessionKey ? this.states.get(sessionKey) ?? { capableHoldTurnsRemaining: 0 } : { capableHoldTurnsRemaining: 0 };
+    const prior = sessionKey ? this.states.get(sessionKey) : undefined;
 
     let backend = this.defaultBackend;
-    let tier = this.aliasOf(backend);
     let source: RouteSource = "default";
     let score = 0;
+    let trace: SwitchTrace | undefined;
+    let next: AutoState | undefined;
+    let movable = true;
+    let prefer: string | undefined;
     try {
-      const signals = extractToolSignals(normalizeConversation(body));
-      const decision = selectStage(signals, { mode: "efficient_first", confidenceThreshold: 0.5, capableHoldTurns: 2 }, priorState);
-      if (sessionKey) this.states.set(sessionKey, decision.state);
+      const conversation = normalizeConversation(body);
+      const decision = selectStage(extractToolSignals(conversation), { mode: "efficient_first", confidenceThreshold: 0.5, capableHoldTurns: 2 }, prior?.stage ?? { capableHoldTurnsRemaining: 0 });
+      const inputTokens = this.estimateInputTokens(body.messages, body.tools);
+      const fits = (candidate: Tier) => { const staged = this.stageBackend(candidate, body, inputTokens); return !!staged && (staged.kind === "mlx" || inputTokens <= this.options.dgxMaxInputTokens); };
+      const staged = stayOrSwitch(this.options.staySwitch?.(), prior?.pin?.tier, decision, turnKind(conversation), { inputTokens, fits });
+      next = { stage: decision.state, ...(staged.pin ? { pin: { tier: staged.pin, alias: "" } } : {}) };
+      trace = staged.trace;
       score = decision.score;
       source = this.sourceOf(decision.source);
-      const selectedTier = decision.tier ?? decision.defaultTier;
-      const stagedBackend = this.stageBackend(selectedTier, body);
-      if (stagedBackend) {
-        backend = stagedBackend;
-        tier = this.aliasOf(stagedBackend);
-      } else {
-        source = "default";
-      }
+      // Enforced, a tool loop that keeps its tier keeps its backend too: a move inside the tier costs the same prefill (#199).
+      movable = !(trace.staySwitch === "enforce" && trace.turnType === "tool_result" && trace.plan === "stay" && trace.reason !== "new_pin");
+      if (!movable) prefer = prior?.pin?.alias;
+      const stagedBackend = this.stageBackend(staged.tier, body, inputTokens);
+      if (stagedBackend) backend = stagedBackend;
+      else source = "default";
     } catch {
       source = "default";
     }
@@ -56,27 +68,31 @@ export class AutoRouteSelector {
         const override = await this.options.selectBackend(body);
         if (this.isAllowed(override)) {
           backend = override;
-          tier = this.aliasOf(override);
           source = "override";
         }
       } catch { /* deterministic stage choice is the fail-open route */ }
     }
 
-    this.emit({ route: "hub/auto", tier, source, score, ms: Math.max(0, performance.now() - started) });
-    return backend;
+    if (next?.pin) next.pin.alias = this.aliasOf(backend);
+    if (sessionKey && next) this.states.set(sessionKey, next);
+    const route: RelayRouteEvent = { route: "hub/auto", tier: this.aliasOf(backend), source, score, ms: Math.max(0, performance.now() - started), ...trace, ...(sessionKey ? {} : { stateless: true }) };
+    return { backend, route, movable, ...(prefer ? { prefer } : {}), moved: (alias, moveSource) => {
+      route.tier = alias;
+      route.source = moveSource;
+      if (next?.pin) next.pin.alias = alias;
+    } };
   }
 
-  private stageBackend(tier: "capable" | "efficient", body: RelayRequest): ModelBackend | undefined {
+  private stageBackend(tier: "capable" | "efficient", body: RelayRequest, input: number): ModelBackend | undefined {
     if (tier === "capable") return "dgx/coding" in this.options.allowedDGXmodels ? { kind: "dgx", alias: "dgx/coding" } : undefined;
-    if (this.options.mlx && this.mlxFitsBudget(body)) return { kind: "mlx", alias: this.mlxAlias };
+    if (this.options.mlx && this.mlxFitsBudget(body, input)) return { kind: "mlx", alias: this.mlxAlias };
     return "dgx/fast" in this.options.allowedDGXmodels ? { kind: "dgx", alias: "dgx/fast" } : undefined;
   }
 
-  private mlxFitsBudget(body: RelayRequest): boolean {
+  private mlxFitsBudget(body: RelayRequest, input: number): boolean {
     const context = Math.min(8192, this.options.mlx?.contextWindow ?? 8192);
     const maxOutput = this.options.mlx?.maxTokens ?? 2048;
     const output = body.max_tokens ?? maxOutput;
-    const input = this.estimateInputTokens(body.messages, body.tools);
     const inputLimit = this.options.mlx?.maxInputTokens ?? (this.options.mlx?.provider === "ollama" ? 6000 : 16_000);
     return Number.isInteger(output) && output > 0 && output <= maxOutput && input <= inputLimit && input + output <= context;
   }
@@ -95,8 +111,4 @@ export class AutoRouteSelector {
     if (source === "llm-classifier") return "classifier";
     return "default";
   }
-
-  private emit(event: RelayRouteEvent): void {
-    try { this.options.onRoute?.(event); } catch { /* observation cannot fail routing */ }
-  }
 }

package/src/models/route/runtime.ts

@@ -3,22 +3,24 @@ import { AdvisorGate, buildAdvisorJudgeRequest, redoFeedback } from "./advisor.t
 import type { HubRoute } from "./config.ts";
 import { buildEscalationJudgeRequest, parseEscalationVerdict } from "./escalation.ts";
 import { normalizeConversation } from "./normalize.ts";
-import { extractToolSignals } from "./signals.ts";
+import { extractToolSignals, turnKind } from "./signals.ts";
 import { RouteLabelTracker, type RouteLabelEvent, type RouteTurnContext, type RouteTurnOutcome } from "./labels.ts";
-import { dimensionsFromSignal, selectStage, type StageState } from "./stage.ts";
+import { dimensionsFromSignal, estimateInputTokens, selectStage, stayOrSwitch, type StageState, type StaySwitchPolicy, type SwitchTrace, type Tier } from "./stage.ts";
 import { EscalationState, SessionState } from "./state.ts";
 import { parseAdvisorVerdict } from "./text.ts";
 import { planExecutePhase, type PlanExecuteState } from "./plan-execute.ts";
 
-export interface RouteEvent {
+export interface RouteEvent extends Partial<SwitchTrace> {
   route: string; tier: string; source: "override" | "dimensions" | "hold" | "classifier" | "default"; score: number; ms: number;
   decision?: string; turn?: string; task?: number; pii?: boolean; severity?: number; spinning?: number; exploring?: number; production?: number;
 }
 export interface AdvisorEvent { route: string; trigger: string; verdict: "approve" | "redo" | "failed"; discardedChars: number }
-interface RouteState { fingerprint: string; stage: StageState; plan: PlanExecuteState; escalation: EscalationState; advisor: AdvisorGate }
+interface RouteState { fingerprint: string; stage: StageState; pin?: Tier; plan: PlanExecuteState; escalation: EscalationState; advisor: AdvisorGate }
 export interface RouteHost {
   execute: (model: string, messages: ChatMessage[], judge: boolean, signal: AbortSignal, maxTokens?: number) => Promise<ChatResult>;
   onCampus: () => Promise<boolean>;
+  /** Read on every stage call (#197); absent is the shadow default. */
+  staySwitch?: () => StaySwitchPolicy | undefined;
   onRoute?: (event: RouteEvent) => void;
   onRouteOutcome?: (event: RouteLabelEvent) => void;
   onAdvisor?: (event: AdvisorEvent) => void;
@@ -59,11 +61,19 @@ export class HubRouteRuntime {
     const dimensions = dimensionsFromSignal(signals);
     const efficient = config.efficient ?? "fast", capable = config.capable ?? "coding";
     let model = efficient, source: RouteEvent["source"] = "default", score = 0;
-    let sent = messages;
+    let sent = messages, trace: Partial<SwitchTrace> = {};
     if (config.type === "stage") {
       const decision = selectStage(signals, { confidenceThreshold: config.confidence_threshold, capableHoldTurns: config.hold_turns }, state.stage);
       state.stage = decision.state;
-      model = (decision.tier ?? decision.defaultTier) === "capable" ? capable : efficient;
+      // OmniRoute models publish no context limit to the hub, so every tier fits here; the estimate leaves out the tool
+      // schemas the host adds to each call, as the relay counts what Pi sends. A PII conversation's size says
+      // something about its text, as a private envelope's does: the planner never sees it (no prefill bound, so no
+      // plan or reason can depend on it) and the event leaves it out.
+      const staged = stayOrSwitch(this.host.staySwitch?.(), state.pin, decision, turnKind(conversation), { inputTokens: pii ? 0 : estimateInputTokens(messages), fits: () => true });
+      state.pin = staged.pin;
+      const { prefillTokens, ...shape } = staged.trace;
+      trace = pii ? shape : staged.trace;
+      model = staged.tier === "capable" ? capable : efficient;
       score = decision.score;
       source = decision.source === "capable_hold" ? "hold" : decision.source === "override" || decision.source === "dimensions" ? decision.source : "default";
     } else if (config.type === "plan_execute") {
@@ -72,7 +82,7 @@ export class HubRouteRuntime {
       state.plan = decision.state;
       if (decision.planningPrompt) sent = [{ role: "system", content: decision.planningPrompt }, ...messages];
     } else if (config.type === "escalation" && state.escalation.snapshot().latched) model = capable;
-    let decisionId = this.emitRoute({ route, tier: model, source, score, ms: performance.now() - started }, labelTurn, dimensions, pii, messages);
+    let decisionId = this.emitRoute({ route, tier: model, source, score, ms: performance.now() - started, ...trace }, labelTurn, dimensions, pii, messages);
     const result = await this.host.execute(model, sent, false, signal);
     if (config.type !== "escalation" || state.escalation.snapshot().latched) return this.withDecision(result, decisionId);
     const request = buildEscalationJudgeRequest(normalizeConversation([...messages, result.message]), signals.assistantTurnCount + 1);

package/src/models/route/signals.ts

@@ -217,16 +217,27 @@ function buildSignals(calls: Array<{name:string;command?:string;source?:string}>
   return signal;
 }
 
+const COMPACTED = 'session is being continued';
+
 export function extractToolSignals(conversation: Conversation, recentWindow = DEFAULT_RECENT_WINDOW): ToolSignals {
   const calls: Array<{name:string;command?:string}> = []; const results: Result[]=[]; const ids=new Map<string,boolean>(); let assistants=0; let compacted=false;
   for(const message of conversation.messages){if(message.role==='assistant')assistants++;
-    if(message.content.toLowerCase().includes('session is being continued'))compacted=true;
+    if(message.content.toLowerCase().includes(COMPACTED))compacted=true;
     for(const call of message.toolCalls){const command=argumentCommand(call.arguments);calls.push({name:call.name,...(command!==undefined?{command}:{})}); ids.set(call.id,semantic(call.name,command)==='read');}
     for(const result of message.toolResults){const retrieval=ids.get(result.toolCallId)===true&&!result.isError;results.push({text:retrieval?'':result.content,isError:result.isError===true});}
   }
   return buildSignals(calls,results,conversation.messages.length,assistants,compacted,recentWindow);
 }
 
+export type TurnKind = 'tool_result' | 'user' | 'compaction';
+
+/** What a call answers (#197, not ported): its last message is a tool result, a compaction summary or a user message. */
+export function turnKind(conversation: Conversation): TurnKind {
+  const last = conversation.messages.at(-1);
+  if (last?.toolResults.length) return 'tool_result';
+  return last?.content.toLowerCase().includes(COMPACTED) ? 'compaction' : 'user';
+}
+
 export function extractToolSignalsFromObservations(observations: readonly ToolObservation[], turnDepth = observations.length, recentWindow = DEFAULT_RECENT_WINDOW): ToolSignals {
   const calls=observations.map((o)=>({name:o.name, ...(o.command!==undefined?{command:o.command}:{}), ...(o.source!==undefined?{source:o.source}:{})}));
   const results=observations.filter((o)=>o.resultText!==undefined||o.isError===true).map((o)=>({text:!o.isError && semantic(o.name,o.command,o.source)==='read' ? '' : o.resultText??'',isError:o.isError===true}));

package/src/models/route/stage.ts

@@ -2,7 +2,7 @@
 // SPDX-License-Identifier: Apache-2.0
 // Ported to TypeScript from NVIDIA NeMo Switchyard crates/libsy/src/algorithms/util/stage.rs at c8848511, modified.
 
-import type { ToolSignals } from './signals.ts';
+import type { ToolSignals, TurnKind } from './signals.ts';
 
 export type Tier = 'capable' | 'efficient';
 export type PickerMode = 'capable_first' | 'efficient_first';
@@ -19,7 +19,8 @@ export interface HandoffNoteConfig {
   onlyOnWrongSignalEscalation?: boolean;
 }
 export interface StageOptions { mode?:PickerMode; confidenceThreshold?:number; capableHoldTurns?:number; handoffNotes?:HandoffNoteConfig }
-export interface StageDecision { tier:Tier|undefined; defaultTier:Tier; source:DecisionSource; probability:number; confidence:number; score:number; dimensions:CodingAgentDimensions; note?:string; state:StageState }
+/** `hardOverride` (#197, not ported): pickTier's override predicate holds, even when a capable hold answered first. */
+export interface StageDecision { tier:Tier|undefined; defaultTier:Tier; source:DecisionSource; probability:number; confidence:number; score:number; dimensions:CodingAgentDimensions; note?:string; state:StageState; hardOverride:boolean }
 
 const STALL_MIN_TURN_DEPTH=8;
 const SCORE_GAIN=5;
@@ -46,9 +47,14 @@ export function scoreSignal(signal:ToolSignals):ScoreResult {
   return {score,confidence:Math.abs(score)};
 }
 
+/** Compaction, critical severity or a repeated failure: capable whatever the scores say. */
+export function hardOverride(signal:ToolSignals):boolean {
+  return signal.compacted||signal.severity>=SEVERITY_CRITICAL||signal.repeatedFailure;
+}
+
 export function pickTier(signal:ToolSignals, mode:PickerMode='efficient_first', confidenceThreshold=0.5):PickOutcome {
   const defaultTier:Tier=mode==='capable_first'?'capable':'efficient';
-  if(signal.compacted||signal.severity>=SEVERITY_CRITICAL||signal.repeatedFailure)
+  if(hardOverride(signal))
     return {kind:'resolved',tier:'capable',source:'override',probability:0.5,confidence:1};
   const scored=scoreSignal(signal); const probability=(scored.score+1)/2; const half=confidenceThreshold/2;
   if(probability>0.5+half||probability<0.5-half)
@@ -89,5 +95,53 @@ export function selectStage(signal:ToolSignals, options:StageOptions={}, state:S
     dimensions,
     ...(note === undefined ? {} : { note }),
     state: { capableHoldTurnsRemaining: remaining },
+    hardOverride: hardOverride(signal),
   };
 }
+
+// Session-aware stay/switch (#197): not ported from Switchyard; the rules follow Weave Router's planner, without its
+// history summary on a switch.
+export type StaySwitchMode = 'off' | 'shadow' | 'enforce';
+/** The routing.toml keys the planner reads. */
+export interface StaySwitchPolicy { stay_switch: StaySwitchMode; max_switch_prefill_tokens: number }
+// ponytail: an unmeasured prefill bound; #197 AC6 measures prefill before enforce becomes the default.
+export const DEFAULT_STAY_SWITCH: StaySwitchPolicy = { stay_switch: 'shadow', max_switch_prefill_tokens: 32_000 };
+export type SwitchReason = 'new_pin' | 'compaction' | 'context_fit' | 'same_tier' | 'override' | 'tool_loop' | 'prefill_bound' | 'user_turn';
+export interface SwitchCost { inputTokens: number; maxSwitchPrefillTokens: number; fits: (tier: Tier) => boolean }
+/** Input tokens a backend would prefill for a conversation, 4 characters each; the relay and local routes share it. */
+export function estimateInputTokens(messages: unknown, tools?: unknown[]): number {
+  return Math.ceil(JSON.stringify({ messages, tools }).length / 4);
+}
+export interface SwitchPlan { plan: 'stay' | 'switch'; tier: Tier; reason: SwitchReason }
+/** What a route event records about the planner; never prompt text. */
+export interface SwitchTrace { turnType: TurnKind; prefillTokens: number; staySwitch?: 'shadow' | 'enforce'; plan?: 'stay' | 'switch'; reason?: SwitchReason }
+
+/**
+ * Stay on the session's pinned tier or switch to this call's stage decision. A new session or a compaction starts a fresh
+ * pin, a hard override escalates on any turn, a tool loop keeps its tier, and other changes wait for a user turn; a
+ * de-escalation that makes the efficient backend prefill more than the bound stays. A tier whose backend cannot hold the
+ * conversation is never chosen: nothing here summarizes or trims the history to make a switch fit.
+ */
+export function planSwitch(pin: Tier | undefined, fresh: Pick<StageDecision, 'tier' | 'defaultTier' | 'source'> & { hardOverride?: boolean }, turn: TurnKind, cost: SwitchCost): SwitchPlan {
+  const wanted = fresh.tier ?? fresh.defaultTier, other: Tier = wanted === 'capable' ? 'efficient' : 'capable';
+  const want = cost.fits(wanted) || !cost.fits(other) ? wanted : other;
+  const plan = (tier: Tier, reason: SwitchReason): SwitchPlan => ({ plan: pin === undefined || pin === tier ? 'stay' : 'switch', tier, reason });
+  if (pin === undefined) return plan(want, 'new_pin');
+  if (turn === 'compaction') return plan(want, 'compaction');
+  if (want !== wanted || !cost.fits(pin)) return plan(want, 'context_fit');
+  if (want === pin) return plan(pin, 'same_tier');
+  // A hold answers before pickTier's overrides, so the override is read on its own: it escalates even during a hold.
+  if (fresh.source === 'override' || fresh.hardOverride) return plan(want, 'override');
+  if (turn === 'tool_result') return plan(pin, 'tool_loop');
+  if (want === 'efficient' && cost.inputTokens > cost.maxSwitchPrefillTokens) return plan(pin, 'prefill_bound');
+  return plan(want, 'user_turn');
+}
+
+/** The one place hub/auto and local stage routes apply the mode, so both record the same trace. `shadow` and `off` keep the stage tier. */
+export function stayOrSwitch(policy: StaySwitchPolicy | undefined, pin: Tier | undefined, fresh: StageDecision, turn: TurnKind, cost: Omit<SwitchCost, 'maxSwitchPrefillTokens'>): { tier: Tier; pin?: Tier; trace: SwitchTrace } {
+  const { stay_switch: mode, max_switch_prefill_tokens: maxSwitchPrefillTokens } = policy ?? DEFAULT_STAY_SWITCH;
+  const tier = fresh.tier ?? fresh.defaultTier, trace: SwitchTrace = { turnType: turn, prefillTokens: cost.inputTokens };
+  if (mode === 'off') return { tier, trace };
+  const planned = planSwitch(pin, fresh, turn, { ...cost, maxSwitchPrefillTokens });
+  return { tier: mode === 'enforce' ? planned.tier : tier, pin: planned.tier, trace: { ...trace, staySwitch: mode, plan: planned.plan, reason: planned.reason } };
+}

package/src/omniroute/client.ts

@@ -158,7 +158,7 @@ export class OmniRoute {
     }
   }
 
-  async chat(body: { model: string; messages: ChatMessage[]; tools?: unknown[]; max_tokens?: number }, opts: ChatOptions = {}): Promise<ChatResult> {
+  async chat(body: { model: string; messages: ChatMessage[]; tools?: unknown[]; max_tokens?: number; temperature?: number; reasoning_effort?: string }, opts: ChatOptions = {}): Promise<ChatResult> {
     const base = opts.via ?? (await this.base());
     if (!base) throw new Error("no model gateway is configured or reachable (omniroute.urls in .agenthub/config.json; see ahub doctor)");
     if (opts.onCampusOnly && (opts.via || this.isAccessHost(base))) throw new Error("PII model calls require the confirmed campus gateway");

package/src/pi/process-signature.ts

@@ -17,3 +17,26 @@ export function processSignature(pid: number): string | undefined {
     return text ? createHash("sha256").update(text).digest("hex") : undefined;
   } catch { return undefined; }
 }
+
+export type Liveness = "live" | "gone" | "unknown";
+
+/**
+ * #226: the one ownership check behind every recorded pid (hub manifest, registry claim, recovery runner, manager
+ * owner, launcher records). With a recorded `signature` the identity decides: equal is live, different is gone (the
+ * pid was reused, after a reboot for example, also by a process we may not signal), and a pid that exists but whose
+ * identity cannot be read is unknown. Without one (records written by 0.12.20 and older) the pid probe answers as it
+ * always did: live, gone on ESRCH, unknown otherwise. Unknown is never gone.
+ */
+// ponytail: `ps lstart` is derived from boot time plus the process's age; a wall-clock step on Linux can change it for a
+// live process, which then reads as gone (a second hub could take its claim). Upgrade path: a boot-id or start-tick
+// (Linux /proc/<pid>/stat field 22) component in the signature.
+export function processLiveness(pid: unknown, signature?: string | null, identity: (pid: number) => string | undefined = processSignature): Liveness {
+  if (typeof pid !== "number" || !Number.isSafeInteger(pid) || pid <= 0) return "unknown";
+  if (signature) {
+    let current: string | undefined;
+    try { current = identity(pid); } catch { /* unreadable: the pid probe below decides between gone and unknown */ }
+    if (current) return current === signature ? "live" : "gone";
+  }
+  try { process.kill(pid, 0); return signature ? "unknown" : "live"; }
+  catch (error) { return (error as NodeJS.ErrnoException).code === "ESRCH" ? "gone" : "unknown"; }
+}

package/src/pi/tool-receipts.ts

@@ -1,6 +1,7 @@
 import { chmodSync } from "node:fs";
 import { Database } from "bun:sqlite";
 import { createHash } from "node:crypto";
+import { PreEffectToolRefusal } from "../hub/tool-refusal.ts";
 
 /** Private, session-scoped dedupe. An interrupted effect is never automatically repeated. */
 export class PiToolReceipts {
@@ -31,10 +32,15 @@ export class PiToolReceipts {
     }
     const claim = this.db.query("INSERT OR IGNORE INTO pi_tool_receipts(session,call_id,fingerprint,state) VALUES(?,?,?,'pending')").run(session, callId, fingerprint);
     if (!claim.changes) return this.execute(session, callId, name, args, run);
-    const work = Promise.resolve().then(run).then((result) => {
+    const settle = (result: string): string => {
       const text = result.slice(0, 100_000);
       this.db.query("UPDATE pi_tool_receipts SET state='done',result=? WHERE session=? AND call_id=?").run(text, session, callId);
       return text;
+    };
+    // Legacy pending rows have no refusal proof: neither the current board nor an error's text can settle them.
+    const work = Promise.resolve().then(run).then(settle, (error: unknown) => {
+      if (error instanceof PreEffectToolRefusal) return settle(`error: ${error.message}`);
+      throw error;
     }).catch(() => "error: tool outcome is uncertain; stop and reconcile it, do not repeat the effect")
       .finally(() => this.pending.delete(key));
     this.pending.set(key, work);

package/src/ui/index.html

@@ -4,17 +4,27 @@
 <meta charset="utf-8">
 <meta name="viewport" content="width=device-width,initial-scale=1">
 <title>agent-hub | Local dashboard</title>
+<script>
+const dashboardTheme = (() => {
+  let choice = 'system';
+  try { const saved = document.cookie.split(';').map(value => value.trim()).find(value => value.startsWith('agent-hub-theme='))?.slice('agent-hub-theme='.length); if (['system','light','dark'].includes(saved)) choice = saved; } catch {}
+  const apply = () => { document.documentElement.dataset.theme = choice; };
+  apply();
+  return { get: () => choice, set: (value) => { choice = ['system','light','dark'].includes(value) ? value : 'system'; apply(); try { document.cookie = `agent-hub-theme=${choice}; Path=/; Max-Age=31536000; SameSite=Strict`; } catch {} } };
+})();
+</script>
 <style>
-:root{color-scheme:light;--paper:#f5f4ef;--ink:#252b29;--muted:#68716c;--line:#d9ddd5;--panel:#ffffff;--panel-soft:#f4f6f2;--control:#ffffff;--control-border:#bbc4bb;--hover:#f0f4ef;--green:#245f45;--green-soft:#e4efe9;--amber:#8a5b12;--amber-soft:#f8eed6;--blue:#245a93;--blue-soft:#e6eef8;--red:#a2392f;--red-soft:#f6e5e2;--dot:#98a29b;--dot-ok:#2f7a52;--dot-warn:#b07a16;--dot-err:#b24539;--focus:#245f45;--on-accent:#ffffff}@media(prefers-color-scheme:dark){:root{color-scheme:dark;--paper:#161a18;--ink:#e5e8e3;--muted:#97a099;--line:#2c342f;--panel:#1d2320;--panel-soft:#17201b;--control:#1d2320;--control-border:#3a443d;--hover:#263029;--green:#6fc2a0;--green-soft:#1e3a2e;--amber:#d9a64f;--amber-soft:#3a2f14;--blue:#8ab8ea;--blue-soft:#1d2f44;--red:#e28b80;--red-soft:#3d2320;--dot:#5b6660;--dot-ok:#4cb77f;--dot-warn:#d9a64f;--dot-err:#e28b80;--focus:#6fc2a0;--on-accent:#161a18}}*{box-sizing:border-box}body{margin:0;background:var(--paper);color:var(--ink);font:14px/1.5 system-ui,sans-serif}header,main{max-width:1440px;margin:auto;padding:24px 32px}header{display:flex;align-items:center;justify-content:space-between;border-bottom:1px solid var(--line);gap:20px}h1{font-size:24px;letter-spacing:-1px;margin:0}h1 span{font-weight:400;color:var(--muted)}h2{font-size:15px;margin:0}p{margin:6px 0}.eyebrow{overflow-wrap:anywhere;font-size:12px;color:var(--muted)}.eyebrow{letter-spacing:1.5px;text-transform:uppercase}.meta{overflow-wrap:anywhere;font-size:12px;color:var(--muted);font-variant-numeric:tabular-nums}.connection{font-size:12px;color:var(--muted)}main{display:grid;grid-template-columns:minmax(0,1fr) 340px;gap:24px;align-items:start}.stack{display:flex;flex-direction:column;gap:24px;min-width:0}section{border:1px solid var(--line);background:var(--panel);border-radius:8px;overflow:hidden}.section-head{display:flex;justify-content:space-between;align-items:center;padding:14px 18px;border-bottom:1px solid var(--line);gap:12px}.section-head .filters{display:flex;align-items:center;gap:8px;flex-wrap:wrap}.content{padding:16px 18px}.empty{color:var(--muted);padding:20px 18px}.row{padding:14px 18px;border-bottom:1px solid var(--line)}.row:last-child{border-bottom:0}.row-head{display:flex;align-items:center;justify-content:space-between;gap:12px}.peername{display:flex;align-items:center;gap:8px;min-width:0}.label{font-weight:600;overflow-wrap:anywhere}.badge{display:inline-block;font:11px ui-monospace,monospace;padding:2px 7px;background:var(--panel-soft);color:var(--muted);border-radius:4px;white-space:nowrap}.badge.ok{background:var(--green-soft);color:var(--green)}.badge.warn{color:var(--amber);background:var(--amber-soft)}.badge.info{color:var(--blue);background:var(--blue-soft)}.badge.err{color:var(--red);background:var(--red-soft)}.dot{flex:none;width:9px;height:9px;border-radius:50%;background:var(--dot)}.dot.ok{background:var(--dot-ok)}.dot.warn{background:var(--dot-warn)}.dot.err{background:var(--dot-err)}.chips{display:flex;gap:6px;flex-wrap:wrap}.chip{background:var(--control);color:var(--muted);border:1px solid var(--control-border);border-radius:999px;padding:4px 10px;font-size:12px;cursor:pointer;line-height:1.2}.chip:hover{background:var(--hover)}.chip[aria-pressed="true"]{background:var(--green);color:var(--on-accent);border-color:var(--green);font-weight:600}.actions{display:flex;gap:8px;align-items:center;flex-wrap:wrap;margin-top:10px}button,input,select,textarea{font:inherit}button{background:var(--control);color:var(--ink);border:1px solid var(--control-border);border-radius:5px;padding:6px 10px;cursor:pointer}button:hover{background:var(--hover)}button.primary{background:var(--green);color:var(--on-accent);border-color:var(--green)}button.danger{color:var(--red)}button:disabled{opacity:.5;cursor:default}:focus-visible{outline:2px solid var(--focus);outline-offset:3px}input,select,textarea{width:100%;border:1px solid var(--control-border);border-radius:5px;padding:8px;background:var(--control);color:var(--ink)}input[type="checkbox"]{width:auto;margin-right:6px}textarea{resize:vertical;min-height:78px}label{display:block;font-size:12px;color:var(--muted);margin-bottom:5px}form .field{margin-bottom:12px}.actions select{width:auto;max-width:220px}.section-head .filters select{width:auto;max-width:220px}.body{white-space:pre-wrap;overflow-wrap:anywhere;margin-top:7px}.stream{height:360px;overflow:auto}.event{padding:12px 18px;border-bottom:1px solid var(--line)}.event[hidden]{display:none}.event .body{font:13px/1.55 ui-monospace,monospace}.notice{grid-column:1/-1;margin:0;border:1px solid var(--line);border-radius:6px;padding:10px 14px;background:var(--panel);overflow-wrap:anywhere}.notice.error{color:var(--red);border-color:var(--red)}.notice:empty{display:none}progress{width:100%;height:8px;accent-color:var(--green);display:block;margin:8px 0}progress.warn{accent-color:var(--amber)}progress.warn::-webkit-progress-value{background:var(--amber)}progress.critical{accent-color:var(--red)}progress.critical::-webkit-progress-value{background:var(--red)}progress::-moz-progress-bar{background:var(--green)}progress.warn::-moz-progress-bar{background:var(--amber)}progress.critical::-moz-progress-bar{background:var(--red)}.budget-window+.budget-window{margin-top:14px}.split{display:grid;grid-template-columns:1fr 1fr;gap:12px}details summary{cursor:pointer;padding:14px 18px;color:var(--green);font-weight:600}details form{padding:0 18px 18px}.task-list{max-height:460px;overflow:auto}footer{max-width:1440px;margin:auto;padding:0 32px 24px;color:var(--muted);font-size:12px}@media(max-width:900px){main{grid-template-columns:1fr}.stack{gap:18px}header,main{padding:18px}.stream{height:300px}footer{padding:0 18px 18px}}@media(prefers-reduced-motion:reduce){*{scroll-behavior:auto}}
+:root{color-scheme:light;--waiting:#795399;--paper:#f5f4ef;--ink:#252b29;--muted:#68716c;--line:#d9ddd5;--panel:#ffffff;--panel-soft:#f4f6f2;--control:#ffffff;--control-border:#bbc4bb;--hover:#f0f4ef;--green:#245f45;--green-soft:#e4efe9;--amber:#8a5b12;--amber-soft:#f8eed6;--blue:#245a93;--blue-soft:#e6eef8;--red:#a2392f;--red-soft:#f6e5e2;--dot:#98a29b;--dot-ok:#2f7a52;--dot-warn:#b07a16;--dot-err:#b24539;--focus:#245f45;--on-accent:#ffffff}:root[data-theme="dark"]{color-scheme:dark;--waiting:#b9a0df;--paper:#161a18;--ink:#e5e8e3;--muted:#97a099;--line:#2c342f;--panel:#1d2320;--panel-soft:#17201b;--control:#1d2320;--control-border:#3a443d;--hover:#263029;--green:#6fc2a0;--green-soft:#1e3a2e;--amber:#d9a64f;--amber-soft:#3a2f14;--blue:#8ab8ea;--blue-soft:#1d2f44;--red:#e28b80;--red-soft:#3d2320;--dot:#5b6660;--dot-ok:#4cb77f;--dot-warn:#d9a64f;--dot-err:#e28b80;--focus:#6fc2a0;--on-accent:#161a18}@media(prefers-color-scheme:dark){:root:not([data-theme="light"]):not([data-theme="dark"]){color-scheme:dark;--waiting:#b9a0df;--paper:#161a18;--ink:#e5e8e3;--muted:#97a099;--line:#2c342f;--panel:#1d2320;--panel-soft:#17201b;--control:#1d2320;--control-border:#3a443d;--hover:#263029;--green:#6fc2a0;--green-soft:#1e3a2e;--amber:#d9a64f;--amber-soft:#3a2f14;--blue:#8ab8ea;--blue-soft:#1d2f44;--red:#e28b80;--red-soft:#3d2320;--dot:#5b6660;--dot-ok:#4cb77f;--dot-warn:#d9a64f;--dot-err:#e28b80;--focus:#6fc2a0;--on-accent:#161a18}}*{box-sizing:border-box}body{margin:0;background:var(--paper);color:var(--ink);font:14px/1.5 system-ui,sans-serif}header,main{max-width:1440px;margin:auto;padding:24px 32px}header{display:flex;align-items:center;justify-content:space-between;border-bottom:1px solid var(--line);gap:20px}h1{font-size:24px;letter-spacing:-1px;margin:0}h1 span{font-weight:400;color:var(--muted)}h2{font-size:15px;margin:0}p{margin:6px 0}.eyebrow{overflow-wrap:anywhere;font-size:12px;color:var(--muted)}.eyebrow{letter-spacing:1.5px;text-transform:uppercase}.meta{overflow-wrap:anywhere;font-size:12px;color:var(--muted);font-variant-numeric:tabular-nums}.connection{font-size:12px;color:var(--muted)}main{display:grid;grid-template-columns:minmax(0,1fr) 340px;gap:24px;align-items:start}.stack{display:flex;flex-direction:column;gap:24px;min-width:0}section{border:1px solid var(--line);background:var(--panel);border-radius:8px;overflow:hidden}.section-head{display:flex;justify-content:space-between;align-items:center;padding:14px 18px;border-bottom:1px solid var(--line);gap:12px}.section-head .filters{display:flex;align-items:center;gap:8px;flex-wrap:wrap}.content{padding:16px 18px}.empty{color:var(--muted);padding:20px 18px}.row{padding:14px 18px;border-bottom:1px solid var(--line)}.row:last-child{border-bottom:0}.row-head{display:flex;align-items:center;justify-content:space-between;gap:12px}.peername{display:flex;align-items:center;gap:8px;min-width:0}.label{font-weight:600;overflow-wrap:anywhere}.badge{display:inline-block;font:11px ui-monospace,monospace;padding:2px 7px;background:var(--panel-soft);color:var(--muted);border-radius:4px;white-space:nowrap}.badge.ok{background:var(--green-soft);color:var(--green)}.badge.warn{color:var(--amber);background:var(--amber-soft)}.badge.info{color:var(--blue);background:var(--blue-soft)}.badge.err{color:var(--red);background:var(--red-soft)}.dot{flex:none;width:9px;height:9px;border-radius:50%;background:var(--dot)}.dot.ok{background:var(--dot-ok)}.dot.warn{background:var(--dot-warn)}.dot.err{background:var(--dot-err)}.chips{display:flex;gap:6px;flex-wrap:wrap}.chip{background:var(--control);color:var(--muted);border:1px solid var(--control-border);border-radius:999px;padding:4px 10px;font-size:12px;cursor:pointer;line-height:1.2}.chip:hover{background:var(--hover)}.chip[aria-pressed="true"]{background:var(--green);color:var(--on-accent);border-color:var(--green);font-weight:600}.actions{display:flex;gap:8px;align-items:center;flex-wrap:wrap;margin-top:10px}button,input,select,textarea{font:inherit}button{background:var(--control);color:var(--ink);border:1px solid var(--control-border);border-radius:5px;padding:6px 10px;cursor:pointer}button:hover{background:var(--hover)}button.primary{background:var(--green);color:var(--on-accent);border-color:var(--green)}button.danger{color:var(--red)}button:disabled{opacity:.5;cursor:default}:focus-visible{outline:2px solid var(--focus);outline-offset:3px}input,select,textarea{width:100%;border:1px solid var(--control-border);border-radius:5px;padding:8px;background:var(--control);color:var(--ink)}input[type="checkbox"]{width:auto;margin-right:6px}textarea{resize:vertical;min-height:78px}label{display:block;font-size:12px;color:var(--muted);margin-bottom:5px}form .field{margin-bottom:12px}.actions select{width:auto;max-width:220px}.section-head .filters select{width:auto;max-width:220px}.body{white-space:pre-wrap;overflow-wrap:anywhere;margin-top:7px}.stream{height:360px;overflow:auto}.event{padding:12px 18px;border-bottom:1px solid var(--line)}.event[hidden]{display:none}.event .body{font:13px/1.55 ui-monospace,monospace}.notice{grid-column:1/-1;margin:0;border:1px solid var(--line);border-radius:6px;padding:10px 14px;background:var(--panel);overflow-wrap:anywhere}.notice.error{color:var(--red);border-color:var(--red)}.notice:empty{display:none}progress{width:100%;height:8px;accent-color:var(--green);display:block;margin:8px 0}progress.warn{accent-color:var(--amber)}progress.warn::-webkit-progress-value{background:var(--amber)}progress.critical{accent-color:var(--red)}progress.critical::-webkit-progress-value{background:var(--red)}progress::-moz-progress-bar{background:var(--green)}progress.warn::-moz-progress-bar{background:var(--amber)}progress.critical::-moz-progress-bar{background:var(--red)}.budget-window+.budget-window{margin-top:14px}.split{display:grid;grid-template-columns:1fr 1fr;gap:12px}details summary{cursor:pointer;padding:14px 18px;color:var(--green);font-weight:600}details form{padding:0 18px 18px}.task-list{max-height:460px;overflow:auto}footer{max-width:1440px;margin:auto;padding:0 32px 24px;color:var(--muted);font-size:12px}@media(max-width:900px){main{grid-template-columns:1fr}.stack{gap:18px}header,main{padding:18px}.stream{height:300px}footer{padding:0 18px 18px}}@media(prefers-reduced-motion:reduce){*{scroll-behavior:auto}}
+.task-progress{padding:12px 18px;border-bottom:1px solid var(--line)}.task-progress svg{display:block;width:100%;height:12px;margin:8px 0}.progress-proposed{fill:var(--dot)}.progress-unknown{fill:var(--dot)}.progress-waiting{fill:var(--waiting)}.progress-in_progress{fill:var(--blue)}.progress-in_review{fill:var(--amber)}.progress-changes_requested{fill:var(--red)}.progress-approved{fill:var(--green)}.task-stage{font:12px ui-monospace,monospace;margin:8px 0;color:var(--muted)}.task-stage .reached{color:var(--green)}.task-stage.changes{color:var(--red)}.theme-choice{width:auto}
 </style>
 </head>
 <body>
-<header><div><div class="eyebrow">Local control room</div><h1>agent-hub <span>/ dashboard</span></h1><div id="project-context" class="meta"></div></div><div id="connection" class="connection" role="status">Connecting...</div></header>
+<header><div><div class="eyebrow">Local control room</div><h1>agent-hub <span>/ dashboard</span></h1><div id="project-context" class="meta"></div></div><div><label for="theme">Theme</label><select id="theme" class="theme-choice"><option value="system">System</option><option value="light">Light</option><option value="dark">Dark</option></select><div id="connection" class="connection" role="status">Connecting...</div></div></header>
 <main>
 <p id="notice" class="notice" role="status" aria-live="polite"></p>
 <div class="stack">
 <section aria-labelledby="stream-title"><div class="section-head"><h2 id="stream-title">Live stream</h2><div class="filters"><div id="stream-filters" class="chips" role="group" aria-label="Filter events by kind"></div><label><input id="follow" type="checkbox" checked> Follow latest</label></div></div><div id="stream" class="stream" tabindex="0" aria-label="Recent hub events"><p class="empty">Waiting for hub activity...</p></div><p id="stream-empty" class="empty" hidden>No events match this filter.</p></section>
-<section aria-labelledby="tasks-title"><div class="section-head"><h2 id="tasks-title">Task board</h2><div class="filters"><select id="task-filter" aria-label="Filter tasks by state"><option value="">All states</option><option value="proposed">Proposed</option><option value="in_progress">In progress</option><option value="in_review">In review</option><option value="approved">Approved</option><option value="changes_requested">Changes requested</option></select><span id="task-count" class="meta"></span></div></div><div id="tasks" class="task-list"></div><details><summary>Propose a task</summary><form id="propose"><div class="field"><label for="title">Title</label><input id="title" name="title" required maxlength="300" placeholder="What should the team do?"></div><div class="field"><label for="detail">Details (optional)</label><textarea id="detail" name="detail" maxlength="8000"></textarea></div><div class="split"><div class="field"><label for="class">Class</label><select id="class" name="class"><option>implement</option><option>plan</option><option>bulk_edit</option><option>test</option><option>review</option><option>summarize</option><option>triage</option></select></div><div class="field"><label for="owner">Owner</label><select id="owner" name="owner"><option value="">Automatic</option></select></div></div><button class="primary" data-mutation="project" type="submit">Propose task</button></form></details></section>
+<section aria-labelledby="tasks-title"><div class="section-head"><h2 id="tasks-title">Task board</h2><div class="filters"><select id="task-filter" aria-label="Filter tasks by state"><option value="">All states</option><option value="proposed">Proposed</option><option value="in_progress">In progress</option><option value="in_review">In review</option><option value="approved">Approved</option><option value="changes_requested">Changes requested</option></select><span id="task-count" class="meta"></span></div></div><div id="task-progress" class="task-progress" aria-live="polite"></div><div id="tasks" class="task-list"></div><details><summary>Propose a task</summary><form id="propose"><div class="field"><label for="title">Title</label><input id="title" name="title" required maxlength="300" placeholder="What should the team do?"></div><div class="field"><label for="detail">Details (optional)</label><textarea id="detail" name="detail" maxlength="8000"></textarea></div><div class="split"><div class="field"><label for="class">Class</label><select id="class" name="class"><option>implement</option><option>plan</option><option>bulk_edit</option><option>test</option><option>review</option><option>summarize</option><option>triage</option></select></div><div class="field"><label for="owner">Owner</label><select id="owner" name="owner"><option value="">Automatic</option></select></div></div><button class="primary" data-mutation="project" type="submit">Propose task</button></form></details></section>
 <section aria-labelledby="compose-title"><div class="section-head"><h2 id="compose-title">Console message</h2></div><form id="send" class="content"><div class="field"><label for="recipient">Recipient</label><select id="recipient" name="recipient"><option value="">All peers</option></select></div><div class="field"><label for="message">Message</label><textarea id="message" name="body" required maxlength="8000" placeholder="Send an instruction to the team..."></textarea></div><button class="primary" data-mutation="project" type="submit">Send message</button></form></section>
 </div>
 <div class="stack">
@@ -22,11 +32,14 @@
 <section aria-labelledby="peers-title"><div class="section-head"><h2 id="peers-title">Peers &amp; queues</h2><span id="peer-count" class="meta"></span></div><div id="peers"></div></section>
 <section aria-labelledby="permissions-title"><div class="section-head"><h2 id="permissions-title">Pending approvals</h2><span id="permission-count" class="meta"></span></div><div id="permissions"></div></section>
 <section aria-labelledby="budget-title"><div class="section-head"><h2 id="budget-title">Usage budgets</h2></div><div id="budget"></div></section>
+<section aria-labelledby="bench-title"><div class="section-head"><h2 id="bench-title">Benchmarks</h2><span id="bench-count" class="meta"></span></div><div id="bench"></div></section>
 </div>
 </main>
 <footer>Loopback session. Private task content stays in the terminal. Reopen with <code>ahub ui</code> when your session expires.</footer>
 <script>
 'use strict';
+/* TASK_PROGRESS_MODEL */
+const themeControl = document.getElementById('theme'); themeControl.value = dashboardTheme.get(); themeControl.addEventListener('change', () => dashboardTheme.set(themeControl.value));
 let ticket = location.hash.slice(1);
 history.replaceState(null, '', location.pathname);
 const $ = (id) => document.getElementById(id);
@@ -98,7 +111,7 @@ function resetView(id, instance) {
   signatures.clear(); $('stream').replaceChildren(); empty($('stream'), managerMode ? 'Select a running project...' : 'Waiting for hub activity...');
   if (streamFilter) { streamFilter = ''; for (const child of $('stream-filters').children) child.setAttribute('aria-pressed', 'false'); }
   $('stream-empty').hidden = true; $('task-filter').value = '';
-  for (const [id, text] of [['peers','No project selected.'],['tasks','No project selected.'],['permissions','No project selected.'],['budget','No project selected.']]) { $(id).replaceChildren(); empty($(id), text); }
+  for (const [id, text] of [['peers','No project selected.'],['tasks','No project selected.'],['task-progress','No project selected.'],['permissions','No project selected.'],['budget','No project selected.']]) { $(id).replaceChildren(); empty($(id), text); }
   $('peer-count').textContent = ''; $('task-count').textContent = ''; $('permission-count').textContent = ''; $('project-context').textContent = '';
   notice('');
   draftChoicesPending = managerMode && Boolean(selectedProjectId);
@@ -146,6 +159,25 @@ async function refreshProjects() {
   if (!stopped) projectsTimer = setTimeout(refreshProjects, 3000);
 }
 function peerOptions(select, ids, first) { const selected = select.value; select.replaceChildren(); const initial = el('option', first); initial.value = ''; select.append(initial); for (const id of ids) { const option = el('option', id); option.value = id; select.append(option); } if (ids.includes(selected)) select.value = selected; }
+function renderTaskProgress(progress) {
+  update('task-progress', progress, (target) => {
+  const percent = progress.total ? Math.floor(progress.counts.approved * 100 / progress.total) : 0;
+  target.append(el('strong', `${progress.counts.approved}/${progress.total} approved (${percent}%)`));
+  const svg = document.createElementNS('http://www.w3.org/2000/svg', 'svg'); svg.setAttribute('viewBox', '0 0 100 12'); svg.setAttribute('preserveAspectRatio', 'none'); svg.setAttribute('aria-hidden', 'true');
+  let x = 0;
+  for (const [state, count] of Object.entries(progress.counts)) {
+    if (!count) continue;
+    const width = progress.total ? count / progress.total * 100 : 0;
+    const rect = document.createElementNS('http://www.w3.org/2000/svg', 'rect'); rect.setAttribute('x', String(x)); rect.setAttribute('y', '0'); rect.setAttribute('width', String(width)); rect.setAttribute('height', '12'); rect.setAttribute('class', 'progress-' + state); svg.append(rect); x += width;
+  }
+  target.append(svg, el('p', Object.entries(progress.counts).map(([state, count]) => state.replaceAll('_', ' ') + ' ' + count).join(' / '), 'meta'));
+  });
+}
+function renderTaskStage(stage) {
+  const row = el('div', undefined, 'task-stage' + (stage.changes ? ' changes' : ''));
+  for (const [index, label] of ['Proposed', 'In progress', 'In review', 'Approved'].entries()) row.append(el('span', `${index < stage.stage ? '[x]' : '[ ]'} ${label}${index < 3 ? ' > ' : ''}`, index < stage.stage ? 'reached' : ''));
+  row.append(el('span', ` / Stage ${stage.stage}/4: ${stage.label}`)); return row;
+}
 function render(snapshot) {
   const peers = snapshot.status.peers || {}, ids = Object.keys(peers);
   $('peer-count').textContent = ids.length + ' peers';
@@ -159,6 +191,7 @@ function render(snapshot) {
       if (peer.needsReview) row.append(el('p', peer.needsReview + ' need review: use ahub queue list', 'meta'));
       if (peer.oldestQueuedAt !== undefined) row.append(el('p', 'Oldest pending: ' + Math.max(0, Math.floor((Date.now() - peer.oldestQueuedAt) / 1000)) + 's', 'meta'));
       if (peer.attached === false) row.append(el('p', 'Disconnected recipient; work is retained.', 'meta'));
+      if (peer.toolsOnly) row.append(el('p', peer.toolsOnly, 'meta'));
       if (peer.paused) row.append(el('p', peer.paused, 'meta'));
       if (peer.requestedModel) row.append(el('p', 'Requested: ' + peer.requestedModel, 'meta'));
       for (const backend of peer.backends || []) {
@@ -171,16 +204,20 @@ function render(snapshot) {
   const assignable = ids.filter((id) => peers[id].state !== 'offline' && peers[id].attached !== false);
   const peerSignature = JSON.stringify([ids, assignable]);
   if (signatures.get('peer-options') !== peerSignature) { signatures.set('peer-options', peerSignature); peerOptions($('owner'), assignable, 'Automatic'); peerOptions($('recipient'), ids, 'All peers'); if (draftChoicesPending) { restoreDraft(true); draftChoicesPending = false; } }
+  const progress = taskProgress(snapshot.tasks);
+  renderTaskProgress(progress);
+  const progressStages = new Map(progress.stages.map(stage => [stage.id, stage]));
   $('task-count').textContent = snapshot.tasks.length + ' tasks';
   const taskFilter = $('task-filter').value;
   const visibleTasks = taskFilter ? snapshot.tasks.filter((task) => task.state === taskFilter) : snapshot.tasks;
-  update('tasks', [visibleTasks, ids, taskFilter, snapshot.tasks.length], (target) => {
+  update('tasks', [visibleTasks, ids, taskFilter, progress], (target) => {
     if (!snapshot.tasks.length) return empty(target, 'No public tasks yet. Propose a task below to get started.');
     if (!visibleTasks.length) return empty(target, 'No tasks in this state.');
     for (const task of visibleTasks) {
       const row = el('div', undefined, 'row'), head = el('div', undefined, 'row-head');
       head.append(el('span', '#' + task.id + ' ' + task.title, 'label'), badge(task.state.replaceAll('_', ' '), taskBadge[task.state] || ''));
-      row.append(head, el('p', task.class + ' / ' + (task.owner || 'unassigned') + (task.reviewer ? ' / review: ' + task.reviewer : ''), 'meta'));
+      const stage = progressStages.get(task.id);
+      row.append(head, renderTaskStage(stage), el('p', task.class + ' / ' + (task.owner || 'unassigned') + (task.reviewer ? ' / review: ' + task.reviewer : ''), 'meta'));
       if (task.detail) row.append(el('p', task.detail, 'body'));
       if (task.state !== 'approved' && ids.length) {
         const actions = el('div', undefined, 'actions'), select = el('select'); select.setAttribute('aria-label', 'Assign task ' + task.id); peerOptions(select, ids, 'Select peer'); if (ids.includes(task.owner)) select.value = task.owner;
@@ -225,6 +262,7 @@ function render(snapshot) {
       target.append(row);
     }
   });
+  renderBench(snapshot.bench);
   const stream = $('stream');
   if (!eventCount) { stream.replaceChildren(); empty(stream, 'No activity yet.'); }
   for (const entry of snapshot.events || []) {
@@ -262,6 +300,46 @@ const STREAM_KINDS = ['chat', 'task', 'review', 'status', 'budget', 'presence',
 let streamFilter = '';
 function applyStreamFilter(node) { node.hidden = streamFilter !== '' && node.dataset.kind !== streamFilter; }
 function refreshStreamEmpty() { const has = $('stream').querySelector('.event:not([hidden])') !== null, any = $('stream').querySelector('.event') !== null; $('stream-empty').hidden = has || !any; }
+// #251: benchmark runs on this machine and arm comparisons per suite. Every bar carries its number as text.
+function renderBench(bench) {
+  const pct = (v) => v === null || v === undefined ? '-' : Math.round(v * 100) + '%';
+  const num = (v) => v === null || v === undefined ? '-' : String(Math.round(v));
+  const dur = (ms) => ms === null || ms === undefined ? '-' : ms < 120000 ? Math.round(ms / 1000) + 's' : Math.round(ms / 60000) + 'm';
+  $('bench-count').textContent = bench ? bench.runs.length + ' runs' : '';
+  update('bench', bench || null, (target) => {
+    if (!bench) return empty(target, 'No benchmark runs. In a project kept for benchmarks, run ahub bench run <suite.json> --arm <label>.');
+    for (const run of bench.runs) {
+      const row = el('div', undefined, 'row'), head = el('div', undefined, 'row-head');
+      const tone = run.state === 'running' ? 'info' : run.state === 'finished' ? 'ok' : run.state === 'unknown' ? '' : 'warn';
+      head.append(el('span', run.suite + ' [' + run.arm + ']', 'label'), el('span', run.state, 'badge ' + tone));
+      const bar = el('progress'); bar.max = run.total || 1; bar.value = run.done; bar.setAttribute('aria-label', run.suite + ' ' + run.arm + ': ' + run.done + ' of ' + run.total + ' attempts');
+      row.append(head);
+      if (run.state === 'running') row.append(el('p', 'now: ' + (run.current || '-') + ' / running ' + dur(Date.now() - Date.parse(run.startedAt)) + ' / ' + (run.total - run.done) + ' attempts left', 'meta'));
+      row.append(bar, el('p', run.done + '/' + run.total + ' attempts / pass ' + pct(run.passRate) + ' (' + run.pass + ' pass, ' + run.fail + ' fail, ' + run.timeout + ' timeout, ' + run.error + ' error) / tokens p50 ' + num(run.tokensMedian) + ' / wall p50 ' + dur(run.wallMsMedian), 'meta'));
+      target.append(row);
+    }
+    const suites = new Map();
+    for (const arm of bench.arms) { if (!suites.has(arm.suite)) suites.set(arm.suite, []); suites.get(arm.suite).push(arm); }
+    for (const [suite, arms] of suites) {
+      if (arms.length < 2) continue;
+      const block = el('div', undefined, 'row');
+      block.append(el('div', 'Compare arms: ' + suite, 'label'));
+      for (const [key, label, rate, fmt] of [['passRate', 'Pass rate', true, pct], ['firstPassRate', 'First pass', true, pct], ['tokensMedian', 'Tokens per attempt (p50)', false, num], ['wallMsMedian', 'Wall time per attempt (p50)', false, dur]]) {
+        const values = arms.map((a) => a.measures[key]).filter((v) => v !== null && v !== undefined);
+        const max = rate ? 1 : Math.max(1, ...values);
+        block.append(el('p', label, 'meta'));
+        for (const arm of arms) {
+          const value = arm.measures[key];
+          const counted = key === 'passRate' || key === 'wallMsMedian' ? arm.measures.scored : arm.measures.measured;
+          const line = el('p', arm.arm + ': ' + fmt(value) + ' (' + counted + ' counted of ' + arm.measures.attempts + ' attempts, ' + arm.runs + ' runs)' + (counted < bench.minAttempts ? ', too few to compare' : ''), 'meta');
+          const bar = el('progress'); bar.max = max; bar.value = value === null || value === undefined ? 0 : value; bar.setAttribute('aria-label', label + ', ' + arm.arm + ': ' + fmt(value));
+          block.append(line, bar);
+        }
+      }
+      target.append(block);
+    }
+  });
+}
 function renderStreamChips() {
   const group = $('stream-filters');
   if (group.children.length) return;