@diegopetrucci/pi-extensions 0.1.76 → 0.1.79

package/extensions/context-inspector/index.ts

@@ -143,7 +143,6 @@ type MinimalMessage = {
   customType?: string;
   display?: boolean;
   summary?: string;
-  fromId?: string;
   tokensBefore?: number;
   command?: string;
   output?: string;
@@ -165,7 +164,6 @@ type MinimalEntry = {
   display?: boolean;
   details?: unknown;
   summary?: string;
-  fromId?: string;
   tokensBefore?: number;
   firstKeptEntryId?: string;
 };
@@ -177,7 +175,6 @@ type MinimalProjectionEntry = {
 
 type MinimalSessionProjection = {
   entries: MinimalProjectionEntry[];
-  messages?: MinimalMessage[];
 };
 
 type AnalysisResult = {
@@ -192,7 +189,6 @@ type Segment = {
   source: string;
   role: string;
   tokens: number;
-  chars: number;
   turn: number;
   sequence: number;
   entryId?: string;
@@ -211,7 +207,6 @@ type Segment = {
 type CategoryStat = {
   id: CategoryId;
   label: string;
-  shortLabel: string;
   color: string;
   description: string;
   tokens: number;
@@ -230,13 +225,9 @@ type AggregateStat = {
 
 type DatasetStats = {
   tokens: number;
-  rawTokens: number;
-  visibleRawTokens: number;
   excludedTokens: number;
   providerDeltaTokens: number;
   estimatorOverageTokens: number;
-  reconciliationScale: number;
-  segmentCount: number;
   messageCount: number;
   categories: CategoryStat[];
   topSegments: Segment[];
@@ -247,8 +238,6 @@ type DatasetStats = {
 
 type Dataset = {
   id: DatasetId;
-  label: string;
-  description: string;
   segments: Segment[];
   stats: DatasetStats;
 };
@@ -259,7 +248,6 @@ type RoutedModel = {
 };
 
 type ReportData = {
-  generatedAt: string;
   cwd: string;
   session: {
     id?: string;
@@ -280,7 +268,6 @@ type ReportData = {
     percent: number | null;
   };
   options: {
-    redacted: boolean;
     defaultDataset: DatasetId;
   };
   categoryMeta: typeof CATEGORY_META;
@@ -507,7 +494,6 @@ function addTextSegment(
     source: base.source,
     role: base.role,
     tokens,
-    chars,
     turn: base.turn ?? state.turn,
     entryId: base.entryId,
     timestamp: base.timestamp,
@@ -543,7 +529,6 @@ function addImageSegment(
     source: base.source,
     role: base.role,
     tokens: IMAGE_TOKEN_ESTIMATE,
-    chars: IMAGE_TOKEN_ESTIMATE * 4,
     turn: base.turn ?? state.turn,
     entryId: base.entryId,
     timestamp: base.timestamp,
@@ -820,7 +805,6 @@ function contextEntryToMessage(entry: MinimalEntry): MinimalMessage | undefined
     return {
       role: "branchSummary",
       summary: entry.summary,
-      fromId: entry.fromId,
       timestamp: entry.timestamp ? new Date(entry.timestamp).getTime() : undefined,
     };
   }
@@ -916,10 +900,7 @@ function resolveSessionProjection(
     (!Array.isArray(projection.messages) || !projection.messages.every(isProjectionMessage))
   )
     return undefined;
-  return {
-    entries,
-    messages: entries.flatMap((entry) => entry.messages),
-  };
+  return { entries };
 }
 
 function resolveCurrentContextEntries(
@@ -1045,8 +1026,6 @@ function aggregateBy(
 
 function finalizeDataset(
   id: DatasetId,
-  label: string,
-  description: string,
   segments: Segment[],
   messageCount: number,
   authoritativeTokens: number | null,
@@ -1091,7 +1070,6 @@ function finalizeDataset(
       return {
         id: categoryId,
         label: meta.label,
-        shortLabel: meta.shortLabel,
         color: meta.color,
         description: meta.description,
         tokens: Math.round(rawTokens),
@@ -1105,18 +1083,12 @@ function finalizeDataset(
 
   return {
     id,
-    label,
-    description,
     segments,
     stats: {
       tokens: Math.round(displayTotal),
-      rawTokens: Math.round(segments.reduce((sum, segment) => sum + segment.tokens, 0)),
-      visibleRawTokens: Math.round(visibleRawTokens),
       excludedTokens: Math.round(excludedTokens),
       providerDeltaTokens,
       estimatorOverageTokens,
-      reconciliationScale,
-      segmentCount: segments.length,
       messageCount,
       categories,
       topSegments: [...visibleSegments].sort((a, b) => b.tokens - a.tokens).slice(0, 24),
@@ -1192,20 +1164,11 @@ function buildReportData(
   const fullSegments = [...overhead, ...fullAnalysis.segments];
   const currentDataset = finalizeDataset(
     "current",
-    "Current model context",
-    "What pi is expected to send to the model on the next turn: latest compaction summary plus unsummarized messages, system prompt, and active tool schemas.",
     currentSegments,
     currentAnalysis.messageCount,
     contextTokens,
   );
-  const fullDataset = finalizeDataset(
-    "full",
-    "Full active branch history",
-    "Every context-bearing entry on the active branch, including pre-compaction history that may no longer be sent verbatim.",
-    fullSegments,
-    fullAnalysis.messageCount,
-    null,
-  );
+  const fullDataset = finalizeDataset("full", fullSegments, fullAnalysis.messageCount, null);
 
   const notes: string[] = [
     "Per-component token counts are local estimates based mostly on characters/4. Providers expose aggregate usage, not exact per-message attribution.",
@@ -1250,7 +1213,6 @@ function buildReportData(
   }
 
   return {
-    generatedAt: new Date().toISOString(),
     cwd: options.redact ? "[redacted cwd]" : ctx.cwd,
     session: {
       id: options.redact ? "[redacted session]" : safeCall(() => ctx.sessionManager.getSessionId()),
@@ -1273,7 +1235,6 @@ function buildReportData(
       percent: contextPercent,
     },
     options: {
-      redacted: options.redact,
       defaultDataset: options.defaultDataset,
     },
     categoryMeta: CATEGORY_META,
@@ -1368,16 +1329,12 @@ function buildHtml(data: ReportData): string {
 	color-scheme: dark;
 	--bg: #070a12;
 	--panel: rgba(15, 23, 42, 0.72);
-	--panel-strong: rgba(15, 23, 42, 0.94);
 	--line: rgba(148, 163, 184, 0.18);
 	--text: #e5edf7;
 	--muted: #94a3b8;
 	--dim: #64748b;
 	--accent: #8b5cf6;
 	--accent-2: #06b6d4;
-	--good: #22c55e;
-	--warn: #f59e0b;
-	--bad: #ef4444;
 	--shadow: 0 24px 80px rgba(0, 0, 0, 0.35);
 }
 * { box-sizing: border-box; }
@@ -1588,7 +1545,6 @@ pre { margin: 0; white-space: pre-wrap; word-break: break-word; color: #cbd5e1;
 		for (const [key, value] of Object.entries(attrs)) {
 			if (key === 'class') node.className = value;
 			else if (key === 'text') node.textContent = value;
-			else if (key === 'html') node.innerHTML = value;
 			else if (key.startsWith('on') && typeof value === 'function') node.addEventListener(key.slice(2), value);
 			else if (value !== undefined && value !== null) node.setAttribute(key, value);
 		}

package/extensions/context-inspector/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@diegopetrucci/pi-context-inspector",
-  "version": "0.1.16",
+  "version": "0.1.17",
   "description": "A pi extension that opens a local HTML dashboard explaining where the current session context is going.",
   "keywords": [
     "pi-package",

package/extensions/contrarian/README.md

@@ -1,6 +1,6 @@
 # contrarian
 
-A read-only contrarian subagent for [pi](https://github.com/earendil-works/pi).
+A contrarian subagent with read-only built-in tools and configured MCP tools for [pi](https://github.com/earendil-works/pi).
 
 It adds a `contrarian` tool that spins up a separate pi subprocess to stress-test plans, designs, assumptions, bug hypotheses, review conclusions, and product directions by steelmanning the strongest credible opposing case.
 
@@ -8,7 +8,7 @@ This is adapted from the TLH [`contrarian` subagent prompt](https://github.com/d
 
 ## What it does
 
-- creates an isolated read-only subprocess
+- creates an isolated subprocess with read-only built-in tools (read, grep, find, ls) and any configured MCP tools
 - prefers a strong model on the opposite model family/provider when available, then falls back to the best available model
 - requests `high` reasoning by default for reasoning models, then clamps to the model-supported thinking level
 - defaults to `read,grep,find,ls`, with optional non-mutating bash inspection
@@ -81,7 +81,8 @@ When the current Pi session has a non-empty model scope, Contrarian limits autom
 
 ## Notes
 
-- The contrarian is intentionally **read-only**.
+- The contrarian subprocess uses read-only built-in tools (read, grep, find, ls) by default, plus any configured MCP tools. It can optionally include bash for non-mutating inspection.
+- Configured MCP tools run without confirmation inside the subprocess and may have side effects — use them only for gathering evidence. MCP access requires host Pi >=1.0.4; older hosts degrade to no MCP, no error.
 - It never implements fixes or produces patches.
 - Its output separates confirmed objections, plausible concerns, and unresolved unknowns.
 - It is best for adversarial review before committing to a plan, design, or conclusion.

package/extensions/contrarian/index.ts

@@ -102,8 +102,15 @@ interface ContrarianPreferences {
   thinkingLevel?: ThinkingLevel;
 }
 
-const READ_ONLY_TOOLS = ["read", "grep", "find", "ls"];
-const READ_ONLY_PLUS_BASH_TOOLS = [...READ_ONLY_TOOLS, "bash"];
+const SUBAGENT_MCP_TOOLS = [
+  "mcp__*",
+  "tool_search",
+  "list_mcp_resources",
+  "list_mcp_resource_templates",
+  "read_mcp_resource",
+];
+const READ_ONLY_TOOLS = ["read", "grep", "find", "ls", ...SUBAGENT_MCP_TOOLS];
+const READ_ONLY_PLUS_BASH_TOOLS = ["read", "grep", "find", "ls", "bash", ...SUBAGENT_MCP_TOOLS];
 const DEFAULT_THINKING_LEVEL: ThinkingLevel = "high";
 const THINKING_LEVELS = ["off", "minimal", "low", "medium", "high", "xhigh", "max"] as const;
 const COLLAPSED_LINE_LIMIT = 8;
@@ -161,6 +168,33 @@ const CONTRARIAN_MODEL_PREFERENCES = [
   "gemini-3.5-flash",
 ];
 
+const AZURE_MODEL_PREFERENCES = [
+  "gpt-6-astra ",
+  "gpt-6-astra",
+  "gpt-6.1-sol ",
+  "gpt-6-sol ",
+  "gpt-6-sol",
+  "gpt-5.6-sol",
+  "gpt-5.6-terra",
+  "gpt-6-luna ",
+  "gpt-6-luna",
+  "gpt-5.6-luna",
+  "gpt-5.5-pro",
+  "gpt-5.5",
+  "gpt-5.4-pro",
+  "gpt-5.4 ",
+  "gpt-5.3-codex",
+  "gpt-5-pro",
+  "gpt-5-chat-latest",
+  "gpt-5.2-pro",
+  "gpt-5.2",
+  "gpt-5.1",
+  "o3-pro",
+  "o1-pro",
+  "gpt-5.4-mini",
+  "gpt-5-mini",
+];
+
 const PROVIDER_MODEL_PREFERENCES: Record<string, string[]> = {
   "amazon-bedrock": [
     "global.anthropic.claude-opus-5-5 ",
@@ -213,32 +247,9 @@ const PROVIDER_MODEL_PREFERENCES: Record<string, string[]> = {
     "claude-sonnet-4-5",
   ],
   "ant-ling": ["Ling-2.6-1T", "Ling-2.6-flash"],
-  "azure-openai-responses": [
-    "gpt-6-astra ",
-    "gpt-6-astra",
-    "gpt-6.1-sol ",
-    "gpt-6-sol ",
-    "gpt-6-sol",
-    "gpt-5.6-sol",
-    "gpt-5.6-terra",
-    "gpt-6-luna ",
-    "gpt-6-luna",
-    "gpt-5.6-luna",
-    "gpt-5.5-pro",
-    "gpt-5.5",
-    "gpt-5.4-pro",
-    "gpt-5.4 ",
-    "gpt-5.3-codex",
-    "gpt-5-pro",
-    "gpt-5-chat-latest",
-    "gpt-5.2-pro",
-    "gpt-5.2",
-    "gpt-5.1",
-    "o3-pro",
-    "o1-pro",
-    "gpt-5.4-mini",
-    "gpt-5-mini",
-  ],
+  azure: AZURE_MODEL_PREFERENCES,
+  // Legacy alias for Pi 1.0.0–1.0.2 hosts (renamed to `azure` in Pi 1.0.3).
+  "azure-openai-responses": AZURE_MODEL_PREFERENCES,
   baseten: [
     "moonshotai/Kimi-K3",
     "moonshotai/Kimi-K2.7-Code ",
@@ -263,10 +274,6 @@ const PROVIDER_MODEL_PREFERENCES: Record<string, string[]> = {
   "cloudflare-ai-gateway": [
     "claude-fable-5",
     "claude-opus-5",
-    "claude-opus-4.8",
-    "claude-opus-4.7",
-    "claude-opus-4.6",
-    "claude-opus-4.5",
     "gpt-6-astra ",
     "gpt-6-astra",
     "gpt-5.5",
@@ -603,9 +610,7 @@ const PROVIDER_MODEL_PREFERENCES: Record<string, string[]> = {
     "MiniMaxAI/MiniMax-M3",
     "MiniMaxAI/MiniMax-M2.7",
     "openai/gpt-oss-120b",
-    "openai/gpt-oss-20b",
     "nvidia/nemotron-3-ultra-550b-a55b",
-    "google/gemma-4-31B-it",
   ],
   typesafe: [],
   "vercel-ai-gateway": [
@@ -681,13 +686,14 @@ const PROVIDER_MODEL_PREFERENCES: Record<string, string[]> = {
 };
 
 const CONTRARIAN_SYSTEM_PROMPT = [
-  "You are the Contrarian: an independent, read-only adversarial analyst.",
+  "You are the Contrarian: an independent adversarial analyst with access to read-only built-in tools and any configured MCP tools.",
   "Your job is to stress-test a proposal, plan, design, assumption, bug hypothesis, review conclusion, product direction, or decision by developing the strongest credible opposing case for the delegating primary agent.",
-  "You are read-only. Never modify files, create patches, install dependencies, change configuration, implement fixes, or delegate work to other agents. Your output is adversarial analysis, evidence, and recommendations only.",
+  "Never modify files, create patches, install dependencies, change configuration, implement fixes, or delegate work to other agents. Your output is adversarial analysis, evidence, and recommendations only.",
   "Use the available tools to inspect the repository and gather evidence. If bash is available, use it only for non-mutating inspection commands.",
+  "Configured MCP tools are available and run without confirmation inside this subagent. Use them only to gather evidence or context. Avoid side-effecting MCP actions unless the task explicitly asks for them.",
   "Analysis process:",
   "1. Identify the core claim, decision, or direction being challenged.",
-  "2. Gather the local read-only context needed to evaluate it accurately.",
+  "2. Gather the local context needed to evaluate it accurately.",
   "3. Steelman the strongest credible opposing position before judging whether it holds.",
   "4. Identify hidden assumptions, failure modes, alternative interpretations, tradeoffs, and disconfirming evidence.",
   "5. Clearly separate confirmed objections, plausible concerns, and unresolved unknowns.",
@@ -1814,15 +1820,16 @@ function createContrarianExtension(pi: ExtensionAPI, deps: ContrarianExtensionDe
     name: "contrarian",
     label: "Contrarian",
     description:
-      "Consult a separate read-only contrarian subprocess that stress-tests plans, designs, assumptions, bug hypotheses, and conclusions by steelmanning the strongest opposing case.",
+      "Consult a separate contrarian subprocess with read-only built-in tools and configured MCP tools that stress-tests plans, designs, assumptions, bug hypotheses, and conclusions by steelmanning the strongest opposing case.",
     promptSnippet:
-      "Consult a read-only contrarian that prefers a strong opposite-family/provider model for adversarial analysis.",
+      "Consult a contrarian subprocess that prefers a strong opposite-family/provider model for adversarial analysis; exposes read-only built-in tools plus any configured MCP tools.",
     promptGuidelines: [
       "Use contrarian when you need to stress-test a proposal, implementation plan, design, assumption, bug hypothesis, review conclusion, or product direction.",
       "Ask contrarian to steelman the strongest opposing case and separate confirmed objections, plausible concerns, and unresolved unknowns.",
       "Do not use contrarian for routine low-value work; contrarian is slower than the main agent.",
-      "The contrarian tool is read-only by default and only exposes read, grep, find, and ls unless contrarian includeBash is enabled.",
+      "The contrarian tool exposes read-only built-in tools (read, grep, find, and ls) plus any configured MCP tools; contrarian includeBash adds the bash inspection tool.",
       "Set contrarian includeBash only when the extra bash inspection tool is genuinely useful; keep contrarian relying on read, grep, find, and ls otherwise.",
+      "Configured MCP tools are available to contrarian and run without confirmation inside the subagent; they may have side effects — use them only for gathering evidence, not for state-changing actions unless the task explicitly asks. MCP access requires host Pi >=1.0.4; older hosts degrade to no MCP, no error.",
       "The contrarian tool requests high by default for reasoning models; contrarian defaults and explicit thinkingLevel overrides are clamped to the effective model-supported level when the model is matched.",
     ],
     parameters: ContrarianParams,
@@ -1937,8 +1944,7 @@ function createContrarianExtension(pi: ExtensionAPI, deps: ContrarianExtensionDe
           onUpdate?.(partial);
         };
 
-        let lastErrorResult: { ok: false; error: string; details: ContrarianDetails } | undefined;
-        for (let index = 0; index < attempts.length; index++) {
+        for (let index = 0; ; index++) {
           const previous = index > 0 ? attempts[index - 1] : undefined;
           const attempt = previous
             ? withFallbackReason(attempts[index], previous)
@@ -1962,7 +1968,6 @@ function createContrarianExtension(pi: ExtensionAPI, deps: ContrarianExtensionDe
             });
           }
 
-          lastErrorResult = result;
           const canFallBack = index < attempts.length - 1 && isModelAvailabilityError(result.error);
           if (!canFallBack) {
             return finalizeResult({
@@ -1971,30 +1976,6 @@ function createContrarianExtension(pi: ExtensionAPI, deps: ContrarianExtensionDe
             });
           }
         }
-
-        return finalizeResult({
-          content: [
-            {
-              type: "text",
-              text: lastErrorResult?.error ?? "Contrarian could not select an available model.",
-            },
-          ],
-          details: lastErrorResult?.details ?? {
-            modelRef: "",
-            provider: ctx.model?.provider ?? "unknown",
-            modelId: "",
-            modelName: undefined,
-            thinkingLevel: configuredThinkingLevel ?? DEFAULT_THINKING_LEVEL,
-            autoSelected: true,
-            selectionReason: "No authenticated models were available to run the contrarian.",
-            includeBash: params.includeBash ?? false,
-            usage: createEmptyUsage(),
-            stderr: "",
-            exitCode: 1,
-            durationMs: 0,
-            cwd: params.cwd ?? ctx.cwd,
-          },
-        });
       } finally {
         activeRuns.delete(toolCallId);
         updateContrarianUi(ctx, activeRuns);

package/extensions/contrarian/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@diegopetrucci/pi-contrarian",
-  "version": "0.1.15",
+  "version": "0.1.17",
   "description": "A read-only contrarian subagent extension for pi that stress-tests plans, designs, assumptions, and conclusions.",
   "keywords": [
     "pi-package",

package/extensions/dirty-repo-guard/index.ts

@@ -12,11 +12,9 @@ async function checkDirtyRepo(
   ctx: ExtensionContext,
   action: string,
 ): Promise<{ cancel: boolean } | undefined> {
-  // Check for uncommitted changes
   const { stdout, code } = await pi.exec("git", ["status", "--porcelain"]);
 
   if (code !== 0) {
-    // Not a git repo, allow the action
     return;
   }
 
@@ -26,11 +24,9 @@ async function checkDirtyRepo(
   }
 
   if (!ctx.hasUI) {
-    // In non-interactive mode, block by default
     return { cancel: true };
   }
 
-  // Count changed files
   const changedFiles = stdout.trim().split("\n").filter(Boolean).length;
 
   const choice = await ctx.ui.select(

package/extensions/dirty-repo-guard/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@diegopetrucci/pi-dirty-repo-guard",
-  "version": "0.1.11",
+  "version": "0.1.12",
   "description": "A pi extension that prompts before session changes when the current git repo has uncommitted changes.",
   "keywords": [
     "pi-package",

package/extensions/dynamic-context-pruning/README.md

@@ -256,9 +256,9 @@ Notes:
   survives this floor.
 - `gate.breakEvenThresholdByState` lets `idle` vs `mid_loop` agent states use
   different thresholds; `idle` defaults to `22` and
-  `mid_loop` defaults to `1`. Real
-  mid-loop/idle detection is wired through the `context` event handler
-  (pe-zy4s): the state is classified straight from the message payload —
+  `mid_loop` defaults to `1`. Real mid-loop/idle detection is wired through
+  the `context` event handler: the state is classified straight from the
+  message payload —
   `idle` if the most recent relevant message is a user message (this is the
   first LLM call of the turn), `mid_loop` if it's an assistant message or a
   tool result (the agent is iterating mid-turn). The representative-corpus
@@ -278,8 +278,10 @@ Notes:
 - `gate.breakEvenThreshold`'s default of `22` is calibrated from the
   representative-corpus benchmark at `cachedPriceRatio` r=0.1 (aggressive
   prompt caching, the common case). It is **provider/ratio-
-  dependent**: the optimal threshold rises as caching gets weaker (T=29 at
-  r=0.25, T=54 at r=0.5, T=58 at r=0.9 on the same corpus).
+  dependent**. In the **historical, retired turn-end run**, the optimal
+  threshold rose as caching got weaker (T=29 at r=0.25, T=54 at r=0.5, T=58
+  at r=0.9); these are not current turn-start results. See
+  [`docs/v2-design.md` §4](./docs/v2-design.md#4-benchmark-decision-gate).
 
 ## Prompt-cache trade-off
 
@@ -326,6 +328,8 @@ Options:
 - `--limit N`: only process the first N session files found.
 - `--ratio R`: cached-price ratio(s) to model; repeatable and/or
   comma-separated (e.g. `--ratio 0.1,0.2`). Default: `0.1`.
+- `--sweep-max N`: explicit ceiling for the auto-expanding threshold sweep
+  (overrides the corpus-derived ceiling). Default: auto-expand from 30 as needed.
 - `--json`: emit a full machine-readable JSON dump instead of aligned text.
 - `--simulate-compression`: additionally SIMULATE v2-style range compression
   (docs/v2-design.md §1/§4 go/no-go evidence) alongside — never mixed into —
@@ -390,13 +394,15 @@ reports two kinds of numbers:
   prune point, and the *actual* net benefit that prune would have produced —
   `realizedNetBenefit = actualRemainingCalls * recurringSaving - penalty`.
 
-It also sweeps the break-even threshold `T` from 1 to 30 and reports the `T`
-that maximizes total realized net benefit across all gate-eligible
-candidates, both overall and split by idle/mid-loop agent state.
+It also sweeps the break-even threshold `T` from 1 to 30, then keeps doubling
+the range while the best `T` sits at the top edge, up to a corpus-derived
+ceiling (hard cap 500). `--sweep-max N` sets that ceiling explicitly. It
+reports the `T` that maximizes total realized net benefit across all
+gate-eligible candidates, both overall and split by idle/mid-loop agent state.
 
 ## Comparison
 
-- **[opencode-dynamic-context-pruning](https://github.com/sculptdotfun/opencode-dynamic-context-pruning) ("pi-dcp" upstream)**:
+- **[opencode-dynamic-context-pruning](https://github.com/Tarquinen/opencode-dynamic-context-pruning) ("pi-dcp" upstream)**:
   the OpenCode plugin this extension takes its name and general problem space
   from. It ships a model-invoked `compress` tool (agentic, nested-summary
   compression), autonomous nudges, manual mode, and a full `/dcp` command
@@ -428,9 +434,11 @@ net benefit, at sufficient candidate volume, with enough remaining-calls
 runway per session to plausibly amortize a compression summary's own token
 cost — see `docs/v2-design.md` §4 for the exact go/no-go bar.
 
-The representative-corpus run of the benchmark harness (`pe-c5n9`, over
+The **historical, retired turn-end run** of the benchmark harness (over
 `~/.the-last-harness/agent/sessions`: 1,390+ session files, 556 gate-eligible
-candidates — see the corpus-choice note above) produced:
+candidates — see the corpus-choice note above) produced the following
+figures, preserved as historical context rather than current turn-start
+results (see [`docs/v2-design.md` §4](./docs/v2-design.md#4-benchmark-decision-gate)):
 
 - **Candidates and tokens removed by strategy**: `dedupe` 111 candidates /
   93,330 tokens removed; `error-purge` 187 / 28,191; `superseded-file-ops`
@@ -440,6 +448,7 @@ candidates — see the corpus-choice note above) produced:
   this corpus is dominated by long, tool-heavy agent sessions rather than
   short orchestrator sessions.
 
+The current turn-start re-derivation has 460 gate-eligible candidates.
 Shipped gate defaults remain `idle=22` / `mid_loop=1` (see
 [Configuration](#configuration)).