release: v3.9.0

The auto-commit section instructed Hephaestus to automatically commit after
implementation work. Users who didn't know about this behavior would get
surprise commits — a trust-breaking behavioral change flagged by 5 Oracle
reviews as the sole publish blocker for 3.9.0.

assets/oh-my-opencode.schema.json

@@ -24,19 +24,7 @@
     "disabled_agents": {
       "type": "array",
       "items": {
-        "type": "string",
-        "enum": [
-          "sisyphus",
-          "hephaestus",
-          "prometheus",
-          "oracle",
-          "librarian",
-          "explore",
-          "multimodal-looker",
-          "metis",
-          "momus",
-          "atlas"
-        ]
+        "type": "string"
       }
     },
     "disabled_skills": {

benchmarks/bun.lock

@@ -0,0 +1,62 @@
+{
+  "lockfileVersion": 1,
+  "configVersion": 1,
+  "workspaces": {
+    "": {
+      "name": "hashline-edit-benchmark",
+      "dependencies": {
+        "@ai-sdk/openai": "^1.3.0",
+        "@friendliai/ai-provider": "^1.0.9",
+        "ai": "^6.0.94",
+        "zod": "^4.1.0",
+      },
+    },
+  },
+  "packages": {
+    "@ai-sdk/gateway": ["@ai-sdk/gateway@3.0.55", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.15", "@vercel/oidc": "3.1.0" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-7xMeTJnCjwRwXKVCiv4Ly4qzWvDuW3+W1WIV0X1EFu6W83d4mEhV9bFArto10MeTw40ewuDjrbrZd21mXKohkw=="],
+
+    "@ai-sdk/openai": ["@ai-sdk/openai@1.3.24", "", { "dependencies": { "@ai-sdk/provider": "1.1.3", "@ai-sdk/provider-utils": "2.2.8" }, "peerDependencies": { "zod": "^3.0.0" } }, "sha512-GYXnGJTHRTZc4gJMSmFRgEQudjqd4PUN0ZjQhPwOAYH1yOAvQoG/Ikqs+HyISRbLPCrhbZnPKCNHuRU4OfpW0Q=="],
+
+    "@ai-sdk/openai-compatible": ["@ai-sdk/openai-compatible@2.0.30", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.15" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-iTjumHf1/u4NhjXYFn/aONM2GId3/o7J1Lp5ql8FCbgIMyRwrmanR5xy1S3aaVkfTscuDvLTzWiy1mAbGzK3nQ=="],
+
+    "@ai-sdk/provider": ["@ai-sdk/provider@1.1.3", "", { "dependencies": { "json-schema": "^0.4.0" } }, "sha512-qZMxYJ0qqX/RfnuIaab+zp8UAeJn/ygXXAffR5I4N0n1IrvA6qBsjc8hXLmBiMV2zoXlifkacF7sEFnYnjBcqg=="],
+
+    "@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@2.2.8", "", { "dependencies": { "@ai-sdk/provider": "1.1.3", "nanoid": "^3.3.8", "secure-json-parse": "^2.7.0" }, "peerDependencies": { "zod": "^3.23.8" } }, "sha512-fqhG+4sCVv8x7nFzYnFo19ryhAa3w096Kmc3hWxMQfW/TubPOmt3A6tYZhl4mUfQWWQMsuSkLrtjlWuXBVSGQA=="],
+
+    "@friendliai/ai-provider": ["@friendliai/ai-provider@1.1.4", "", { "dependencies": { "@ai-sdk/openai-compatible": "2.0.30", "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.15" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.12" } }, "sha512-9TU4B1QFqPhbkONjI5afCF7Ox4jOqtGg1xw8mA9QHZdtlEbZxU+mBNvMPlI5pU5kPoN6s7wkXmFmxpID+own1A=="],
+
+    "@opentelemetry/api": ["@opentelemetry/api@1.9.0", "", {}, "sha512-3giAOQvZiH5F9bMlMiv8+GSPMeqg0dbaeo58/0SlA9sxSqZhnUtxzX9/2FzyhS9sWQf5S0GJE0AKBrFqjpeYcg=="],
+
+    "@standard-schema/spec": ["@standard-schema/spec@1.1.0", "", {}, "sha512-l2aFy5jALhniG5HgqrD6jXLi/rUWrKvqN/qJx6yoJsgKhblVd+iqqU4RCXavm/jPityDo5TCvKMnpjKnOriy0w=="],
+
+    "@vercel/oidc": ["@vercel/oidc@3.1.0", "", {}, "sha512-Fw28YZpRnA3cAHHDlkt7xQHiJ0fcL+NRcIqsocZQUSmbzeIKRpwttJjik5ZGanXP+vlA4SbTg+AbA3bP363l+w=="],
+
+    "ai": ["ai@6.0.101", "", { "dependencies": { "@ai-sdk/gateway": "3.0.55", "@ai-sdk/provider": "3.0.8", "@ai-sdk/provider-utils": "4.0.15", "@opentelemetry/api": "1.9.0" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-Ur/NgbgOp1rdhyDiKDk6EOpSgd1g5ADlbcD1cjQJtQsnmhEngz3Rf8nK5JetDh0vnbLy2aEBpaQeL+zvLRWuaA=="],
+
+    "eventsource-parser": ["eventsource-parser@3.0.6", "", {}, "sha512-Vo1ab+QXPzZ4tCa8SwIHJFaSzy4R6SHf7BY79rFBDf0idraZWAkYrDjDj8uWaSm3S2TK+hJ7/t1CEmZ7jXw+pg=="],
+
+    "json-schema": ["json-schema@0.4.0", "", {}, "sha512-es94M3nTIfsEPisRafak+HDLfHXnKBhV3vU5eqPcS3flIWqcxJWgXHXiey3YrpaNsanY5ei1VoYEbOzijuq9BA=="],
+
+    "nanoid": ["nanoid@3.3.11", "", { "bin": { "nanoid": "bin/nanoid.cjs" } }, "sha512-N8SpfPUnUp1bK+PMYW8qSWdl9U+wwNWI4QKxOYDy9JAro3WMX7p2OeVRF9v+347pnakNevPmiHhNmZ2HbFA76w=="],
+
+    "secure-json-parse": ["secure-json-parse@2.7.0", "", {}, "sha512-6aU+Rwsezw7VR8/nyvKTx8QpWH9FrcYiXXlqC4z5d5XQBDRqtbfsRjnwGyqbi3gddNtWHuEk9OANUotL26qKUw=="],
+
+    "zod": ["zod@4.3.6", "", {}, "sha512-rftlrkhHZOcjDwkGlnUtZZkvaPHCsDATp4pGpuOOMDaTdDDXF91wuVDJoWoPsKX/3YPQ5fHuF3STjcYyKr+Qhg=="],
+
+    "@ai-sdk/gateway/@ai-sdk/provider": ["@ai-sdk/provider@3.0.8", "", { "dependencies": { "json-schema": "^0.4.0" } }, "sha512-oGMAgGoQdBXbZqNG0Ze56CHjDZ1IDYOwGYxYjO5KLSlz5HiNQ9udIXsPZ61VWaHGZ5XW/jyjmr6t2xz2jGVwbQ=="],
+
+    "@ai-sdk/gateway/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.15", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-8XiKWbemmCbvNN0CLR9u3PQiet4gtEVIrX4zzLxnCj06AwsEDJwJVBbKrEI4t6qE8XRSIvU2irka0dcpziKW6w=="],
+
+    "@ai-sdk/openai-compatible/@ai-sdk/provider": ["@ai-sdk/provider@3.0.8", "", { "dependencies": { "json-schema": "^0.4.0" } }, "sha512-oGMAgGoQdBXbZqNG0Ze56CHjDZ1IDYOwGYxYjO5KLSlz5HiNQ9udIXsPZ61VWaHGZ5XW/jyjmr6t2xz2jGVwbQ=="],
+
+    "@ai-sdk/openai-compatible/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.15", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-8XiKWbemmCbvNN0CLR9u3PQiet4gtEVIrX4zzLxnCj06AwsEDJwJVBbKrEI4t6qE8XRSIvU2irka0dcpziKW6w=="],
+
+    "@friendliai/ai-provider/@ai-sdk/provider": ["@ai-sdk/provider@3.0.8", "", { "dependencies": { "json-schema": "^0.4.0" } }, "sha512-oGMAgGoQdBXbZqNG0Ze56CHjDZ1IDYOwGYxYjO5KLSlz5HiNQ9udIXsPZ61VWaHGZ5XW/jyjmr6t2xz2jGVwbQ=="],
+
+    "@friendliai/ai-provider/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.15", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-8XiKWbemmCbvNN0CLR9u3PQiet4gtEVIrX4zzLxnCj06AwsEDJwJVBbKrEI4t6qE8XRSIvU2irka0dcpziKW6w=="],
+
+    "ai/@ai-sdk/provider": ["@ai-sdk/provider@3.0.8", "", { "dependencies": { "json-schema": "^0.4.0" } }, "sha512-oGMAgGoQdBXbZqNG0Ze56CHjDZ1IDYOwGYxYjO5KLSlz5HiNQ9udIXsPZ61VWaHGZ5XW/jyjmr6t2xz2jGVwbQ=="],
+
+    "ai/@ai-sdk/provider-utils": ["@ai-sdk/provider-utils@4.0.15", "", { "dependencies": { "@ai-sdk/provider": "3.0.8", "@standard-schema/spec": "^1.1.0", "eventsource-parser": "^3.0.6" }, "peerDependencies": { "zod": "^3.25.76 || ^4.1.8" } }, "sha512-8XiKWbemmCbvNN0CLR9u3PQiet4gtEVIrX4zzLxnCj06AwsEDJwJVBbKrEI4t6qE8XRSIvU2irka0dcpziKW6w=="],
+  }
+}

benchmarks/headless.ts

@@ -0,0 +1,193 @@
+#!/usr/bin/env bun
+import { readFile, writeFile, mkdir } from "node:fs/promises"
+import { join, dirname } from "node:path"
+import { stepCountIs, streamText, type CoreMessage } from "ai"
+import { tool } from "ai"
+import { createFriendli } from "@friendliai/ai-provider"
+import { z } from "zod"
+import { formatHashLines } from "../src/tools/hashline-edit/hash-computation"
+import { normalizeHashlineEdits } from "../src/tools/hashline-edit/normalize-edits"
+import { applyHashlineEditsWithReport } from "../src/tools/hashline-edit/edit-operations"
+import { canonicalizeFileText, restoreFileText } from "../src/tools/hashline-edit/file-text-canonicalization"
+
+const DEFAULT_MODEL = "MiniMaxAI/MiniMax-M2.5"
+const MAX_STEPS = 50
+const sessionId = `bench-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`
+
+const emit = (event: Record<string, unknown>) =>
+  console.log(JSON.stringify({ sessionId, timestamp: new Date().toISOString(), ...event }))
+
+// ── CLI ──────────────────────────────────────────────────────
+function parseArgs(): { prompt: string; modelId: string } {
+  const args = process.argv.slice(2)
+  let prompt = ""
+  let modelId = DEFAULT_MODEL
+  for (let i = 0; i < args.length; i++) {
+    if ((args[i] === "-p" || args[i] === "--prompt") && args[i + 1]) {
+      prompt = args[++i]
+    } else if ((args[i] === "-m" || args[i] === "--model") && args[i + 1]) {
+      modelId = args[++i]
+    } else if (args[i] === "--reasoning-mode" && args[i + 1]) {
+      i++ // consume
+    }
+    // --no-translate, --think consumed silently
+  }
+  if (!prompt) {
+    console.error("Usage: bun run benchmarks/headless.ts -p <prompt> [-m <model>]")
+    process.exit(1)
+  }
+  return { prompt, modelId }
+}
+
+// ── Tools ────────────────────────────────────────────────────
+const readFileTool = tool({
+  description: "Read a file with hashline-tagged content (LINE#ID format)",
+  inputSchema: z.object({ path: z.string().describe("File path") }),
+  execute: async ({ path }) => {
+    const fullPath = join(process.cwd(), path)
+    try {
+      const content = await readFile(fullPath, "utf-8")
+      const lines = content.split("\n")
+      const tagged = formatHashLines(content)
+      return `OK - read file\npath: ${path}\nlines: ${lines.length}\n\n${tagged}`
+    } catch {
+      return `Error: File not found: ${path}`
+    }
+  },
+})
+
+const editFileTool = tool({
+  description: "Edit a file using hashline anchors (LINE#ID format)",
+  inputSchema: z.object({
+    path: z.string(),
+    edits: z.array(
+      z.object({
+        op: z.enum(["replace", "append", "prepend"]),
+        pos: z.string().optional(),
+        end: z.string().optional(),
+        lines: z.union([z.array(z.string()), z.string(), z.null()]),
+      })
+    ).min(1),
+  }),
+  execute: async ({ path, edits }) => {
+    const fullPath = join(process.cwd(), path)
+    try {
+      let rawContent = ""
+      let exists = true
+      try {
+        rawContent = await readFile(fullPath, "utf-8")
+      } catch {
+        exists = false
+      }
+
+      const normalized = normalizeHashlineEdits(edits)
+
+      if (!exists) {
+        const canCreate = normalized.every(
+          (e) => (e.op === "append" || e.op === "prepend") && !e.pos
+        )
+        if (!canCreate) return `Error: File not found: ${path}`
+      }
+
+      const envelope = canonicalizeFileText(rawContent)
+      const result = applyHashlineEditsWithReport(envelope.content, normalized)
+
+      if (result.content === envelope.content) {
+        return `Error: No changes made to ${path}. The edits produced identical content.`
+      }
+
+      const writeContent = restoreFileText(result.content, envelope)
+      await mkdir(dirname(fullPath), { recursive: true })
+      await writeFile(fullPath, writeContent, "utf-8")
+
+      const oldLineCount = rawContent.split("\n").length
+      const newLineCount = writeContent.split("\n").length
+      const delta = newLineCount - oldLineCount
+      const sign = delta > 0 ? "+" : ""
+      const action = exists ? "Updated" : "Created"
+      return `${action} ${path}\n${edits.length} edit(s) applied, ${sign}${delta} line(s)`
+    } catch (error) {
+      return `Error: ${error instanceof Error ? error.message : String(error)}`
+    }
+  },
+})
+
+// ── Agent Loop ───────────────────────────────────────────────
+async function run() {
+  const { prompt, modelId } = parseArgs()
+
+  const friendli = createFriendli({ apiKey: process.env.FRIENDLI_TOKEN! })
+  const model = friendli(modelId)
+  const tools = { read_file: readFileTool, edit_file: editFileTool }
+
+  emit({ type: "user", content: prompt })
+
+  const messages: CoreMessage[] = [{ role: "user", content: prompt }]
+  const system =
+    "You are a code editing assistant. Use read_file to read files and edit_file to edit them. " +
+    "Always read a file before editing it to get fresh LINE#ID anchors."
+
+  for (let step = 0; step < MAX_STEPS; step++) {
+    const stream = streamText({
+      model,
+      tools,
+      messages,
+      system,
+      stopWhen: stepCountIs(1),
+    })
+
+    let currentText = ""
+    for await (const part of stream.fullStream) {
+      switch (part.type) {
+        case "text-delta":
+          currentText += part.text
+          break
+        case "tool-call":
+          emit({
+            type: "tool_call",
+            tool_call_id: part.toolCallId,
+            tool_name: part.toolName,
+            tool_input: part.args,
+            model: modelId,
+          })
+          break
+        case "tool-result": {
+          const output = typeof part.result === "string" ? part.result : JSON.stringify(part.result)
+          const isError = typeof output === "string" && output.startsWith("Error:")
+          emit({
+            type: "tool_result",
+            tool_call_id: part.toolCallId,
+            output,
+            ...(isError ? { error: output } : {}),
+          })
+          break
+      }
+    }
+
+    const response = await stream.response
+    messages.push(...response.messages)
+
+    const finishReason = await stream.finishReason
+    if (finishReason !== "tool-calls") {
+      if (currentText.trim()) {
+        emit({ type: "assistant", content: currentText, model: modelId })
+      }
+      break
+    }
+  }
+}
+
+// ── Signal + Startup ─────────────────────────────────────────
+process.once("SIGINT", () => process.exit(0))
+process.once("SIGTERM", () => process.exit(143))
+
+const startTime = Date.now()
+run()
+  .catch((error) => {
+    emit({ type: "error", error: error instanceof Error ? error.message : String(error) })
+    process.exit(1)
+  })
+  .then(() => {
+    const elapsed = ((Date.now() - startTime) / 1000).toFixed(2)
+    console.error(`[headless] Completed in ${elapsed}s`)
+  })

benchmarks/package.json

@@ -0,0 +1,19 @@
+{
+  "name": "hashline-edit-benchmark",
+  "version": "0.1.0",
+  "private": true,
+  "type": "module",
+  "description": "Hashline edit tool benchmark using Vercel AI SDK with FriendliAI provider",
+  "scripts": {
+    "bench:basic": "bun run test-edit-ops.ts",
+    "bench:edge": "bun run test-edge-cases.ts",
+    "bench:multi": "bun run test-multi-model.ts",
+    "bench:all": "bun run bench:basic && bun run bench:edge"
+  },
+  "dependencies": {
+    "ai": "^6.0.94",
+    "@ai-sdk/openai": "^1.3.0",
+    "@friendliai/ai-provider": "^1.0.9",
+    "zod": "^4.1.0"
+  }
+}

benchmarks/test-edit-ops.ts

@@ -0,0 +1,808 @@
+#!/usr/bin/env bun
+/**
+ * Comprehensive headless edit_file stress test: 21 operation types
+ *
+ * Tests: 5 basic ops + 10 creative cases + 6 whitespace cases
+ * Each runs via headless mode with its own demo file + prompt.
+ *
+ * Usage:
+ *   bun run scripts/test-headless-edit-ops.ts [-m <model>] [--provider <provider>]
+ */
+
+import { spawn } from "node:child_process";
+import { mkdirSync, readFileSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join, resolve } from "node:path";
+
+// ── CLI arg passthrough ───────────────────────────────────────
+const extraArgs: string[] = [];
+const rawArgs = process.argv.slice(2);
+for (let i = 0; i < rawArgs.length; i++) {
+  const arg = rawArgs[i];
+  if (
+    (arg === "-m" || arg === "--model" || arg === "--provider") &&
+    i + 1 < rawArgs.length
+  ) {
+    extraArgs.push(arg, rawArgs[i + 1]);
+    i++;
+  } else if (arg === "--think" || arg === "--no-translate") {
+    extraArgs.push(arg);
+  } else if (arg === "--reasoning-mode" && i + 1 < rawArgs.length) {
+    extraArgs.push(arg, rawArgs[i + 1]);
+    i++;
+  }
+}
+
+// ── Colors ────────────────────────────────────────────────────
+const BOLD = "\x1b[1m";
+const GREEN = "\x1b[32m";
+const RED = "\x1b[31m";
+const YELLOW = "\x1b[33m";
+const DIM = "\x1b[2m";
+const CYAN = "\x1b[36m";
+const RESET = "\x1b[0m";
+
+const pass = (msg: string) => console.log(`  ${GREEN}✓${RESET} ${msg}`);
+const fail = (msg: string) => console.log(`  ${RED}✗${RESET} ${msg}`);
+const info = (msg: string) => console.log(`  ${DIM}${msg}${RESET}`);
+const warn = (msg: string) => console.log(`  ${YELLOW}⚠${RESET} ${msg}`);
+
+// ── Test case definition ─────────────────────────────────────
+interface TestCase {
+  fileContent: string;
+  fileName: string;
+  name: string;
+  prompt: string;
+  validate: (content: string) => { passed: boolean; reason: string };
+}
+
+const TEST_CASES: TestCase[] = [
+  {
+    name: "1. Replace single line",
+    fileName: "config.txt",
+    fileContent: [
+      "host: localhost",
+      "port: 3000",
+      "debug: false",
+      "timeout: 30",
+      "retries: 3",
+    ].join("\n"),
+    prompt: [
+      "Follow these steps exactly:",
+      "Step 1: Call read_file on config.txt.",
+      "Step 2: Note the anchor for the port line (line 2).",
+      "Step 3: Call edit_file with path='config.txt' and edits containing ONE object:",
+      "  { op: 'replace', pos: '<line2 anchor>', lines: ['port: 8080'] }",
+      "IMPORTANT: pos must be ONLY the anchor (like '2#KB'). lines must be a SEPARATE array field with the new content.",
+    ].join(" "),
+    validate: (content) => {
+      const has8080 = content.includes("port: 8080");
+      const has3000 = content.includes("port: 3000");
+      if (has8080 && !has3000) {
+        return { passed: true, reason: "port changed to 8080" };
+      }
+      if (has3000) {
+        return { passed: false, reason: "port still 3000 — edit not applied" };
+      }
+      return {
+        passed: false,
+        reason: `unexpected content: ${content.slice(0, 100)}`,
+      };
+    },
+  },
+  {
+    name: "2. Append after line",
+    fileName: "fruits.txt",
+    fileContent: ["apple", "banana", "cherry"].join("\n"),
+    prompt:
+      "Read fruits.txt with read_file. Then use edit_file with op='append' to insert a new line 'grape' after the 'banana' line. Use pos='LINE#HASH' of the banana line and lines=['grape'].",
+    validate: (content) => {
+      const lines = content.trim().split("\n");
+      const bananaIdx = lines.findIndex((l) => l.trim() === "banana");
+      const grapeIdx = lines.findIndex((l) => l.trim() === "grape");
+      if (grapeIdx === -1) {
+        return { passed: false, reason: '"grape" not found in file' };
+      }
+      if (bananaIdx === -1) {
+        return { passed: false, reason: '"banana" was removed' };
+      }
+      if (grapeIdx !== bananaIdx + 1) {
+        return {
+          passed: false,
+          reason: `"grape" at line ${grapeIdx + 1} but expected after "banana" at line ${bananaIdx + 1}`,
+        };
+      }
+      if (lines.length !== 4) {
+        return {
+          passed: false,
+          reason: `expected 4 lines, got ${lines.length}`,
+        };
+      }
+      return {
+        passed: true,
+        reason: '"grape" correctly appended after "banana"',
+      };
+    },
+  },
+  {
+    name: "3. Prepend before line",
+    fileName: "code.txt",
+    fileContent: ["function greet() {", '  return "hello";', "}"].join("\n"),
+    prompt:
+      "Read code.txt with read_file. Then use edit_file with op='prepend' to add '// Greeting function' before the function line. Use pos='LINE#HASH' of the function line and lines=['// Greeting function'].",
+    validate: (content) => {
+      const lines = content.trim().split("\n");
+      const commentIdx = lines.findIndex(
+        (l) => l.trim().startsWith("//") && l.toLowerCase().includes("greet")
+      );
+      const funcIdx = lines.findIndex((l) =>
+        l.trim().startsWith("function greet")
+      );
+      if (commentIdx === -1) {
+        return { passed: false, reason: "comment line not found" };
+      }
+      if (funcIdx === -1) {
+        return { passed: false, reason: '"function greet" line was removed' };
+      }
+      if (commentIdx !== funcIdx - 1) {
+        return {
+          passed: false,
+          reason: `comment at line ${commentIdx + 1} but function at ${funcIdx + 1} — not directly before`,
+        };
+      }
+      return {
+        passed: true,
+        reason: "comment correctly prepended before function",
+      };
+    },
+  },
+  {
+    name: "4. Range replace (multi-line → single line)",
+    fileName: "log.txt",
+    fileContent: [
+      "=== Log Start ===",
+      "INFO: started",
+      "WARN: slow query",
+      "ERROR: timeout",
+      "INFO: recovered",
+      "=== Log End ===",
+    ].join("\n"),
+    prompt: [
+      "Follow these steps exactly:",
+      "Step 1: Call read_file on log.txt to see line anchors.",
+      "Step 2: Note the anchor for 'WARN: slow query' (line 3) and 'ERROR: timeout' (line 4).",
+      "Step 3: Call edit_file with path='log.txt' and edits containing ONE object with THREE separate JSON fields:",
+      "  { op: 'replace', pos: '<line3 anchor>', end: '<line4 anchor>', lines: ['RESOLVED: issues cleared'] }",
+      "CRITICAL: pos, end, and lines are THREE SEPARATE JSON fields. pos is ONLY '3#XX'. end is ONLY '4#YY'. lines is ['RESOLVED: issues cleared'].",
+      "If edit_file fails or errors, use write_file to write the complete correct file content instead.",
+      "The correct final content should be: === Log Start ===, INFO: started, RESOLVED: issues cleared, INFO: recovered, === Log End ===",
+      "Do not make any other changes.",
+    ].join(" "),
+    validate: (content) => {
+      const lines = content.trim().split("\n");
+      const hasResolved = lines.some(
+        (l) => l.trim() === "RESOLVED: issues cleared"
+      );
+      const hasWarn = content.includes("WARN: slow query");
+      const hasError = content.includes("ERROR: timeout");
+      if (!hasResolved) {
+        return {
+          passed: false,
+          reason: '"RESOLVED: issues cleared" not found',
+        };
+      }
+      if (hasWarn || hasError) {
+        return { passed: false, reason: "old WARN/ERROR lines still present" };
+      }
+      // Core assertion: 2 old lines removed, 1 new line added = net -1 line
+      // Allow slight overshoot from model adding extra content
+      if (lines.length < 4 || lines.length > 6) {
+        return {
+          passed: false,
+          reason: `expected ~5 lines, got ${lines.length}`,
+        };
+      }
+      return {
+        passed: true,
+        reason: "range replace succeeded — 2 lines → 1 line",
+      };
+    },
+  },
+  {
+    name: "5. Delete line",
+    fileName: "settings.txt",
+    fileContent: [
+      "mode: production",
+      "debug: true",
+      "cache: enabled",
+      "log_level: info",
+    ].join("\n"),
+    prompt: [
+      "Follow these steps exactly:",
+      "Step 1: Call read_file on settings.txt to see line anchors.",
+      "Step 2: Note the anchor for 'debug: true' (line 2).",
+      "Step 3: Call edit_file with path='settings.txt' and edits containing ONE object:",
+      "  { op: 'replace', pos: '<line2 anchor>', lines: [] }",
+      "IMPORTANT: lines must be an empty array [] to delete the line. pos must be ONLY the anchor like '2#SR'.",
+    ].join(" "),
+    validate: (content) => {
+      const lines = content.trim().split("\n");
+      const hasDebug = content.includes("debug: true");
+      if (hasDebug) {
+        return { passed: false, reason: '"debug: true" still present' };
+      }
+      if (lines.length !== 3) {
+        return {
+          passed: false,
+          reason: `expected 3 lines, got ${lines.length}`,
+        };
+      }
+      if (
+        !(
+          content.includes("mode: production") &&
+          content.includes("cache: enabled")
+        )
+      ) {
+        return { passed: false, reason: "other lines were removed" };
+      }
+      return { passed: true, reason: '"debug: true" successfully deleted' };
+    },
+  },
+
+  // ── Creative cases (6-15) ────────────────────────────────────
+  {
+    name: "6. Batch edit — two replacements in one call",
+    fileName: "batch.txt",
+    fileContent: ["red", "green", "blue", "yellow"].join("\n"),
+    prompt: [
+      "Read batch.txt with read_file.",
+      "Then call edit_file ONCE with path='batch.txt' and edits containing TWO objects:",
+      "  1) { op: 'replace', pos: '<line1 anchor>', lines: ['crimson'] }",
+      "  2) { op: 'replace', pos: '<line3 anchor>', lines: ['navy'] }",
+      "Both edits must be in the SAME edits array in a single edit_file call.",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (!c.includes("crimson")) return { passed: false, reason: "'crimson' not found" };
+      if (!c.includes("navy")) return { passed: false, reason: "'navy' not found" };
+      if (c.includes("red")) return { passed: false, reason: "'red' still present" };
+      if (c.includes("blue")) return { passed: false, reason: "'blue' still present" };
+      if (lines.length !== 4) return { passed: false, reason: `expected 4 lines, got ${lines.length}` };
+      return { passed: true, reason: "both lines replaced in single call" };
+    },
+  },
+  {
+    name: "7. Line expansion — 1 line → 3 lines",
+    fileName: "expand.txt",
+    fileContent: ["header", "TODO: implement", "footer"].join("\n"),
+    prompt: [
+      "Read expand.txt with read_file.",
+      "Replace the 'TODO: implement' line (line 2) with THREE lines:",
+      "  'step 1: init', 'step 2: process', 'step 3: cleanup'",
+      "Use edit_file with op='replace', pos=<line2 anchor>, lines=['step 1: init', 'step 2: process', 'step 3: cleanup'].",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (c.includes("TODO")) return { passed: false, reason: "TODO line still present" };
+      if (!c.includes("step 1: init")) return { passed: false, reason: "'step 1: init' not found" };
+      if (!c.includes("step 3: cleanup")) return { passed: false, reason: "'step 3: cleanup' not found" };
+      if (lines.length !== 5) return { passed: false, reason: `expected 5 lines, got ${lines.length}` };
+      return { passed: true, reason: "1 line expanded to 3 lines" };
+    },
+  },
+  {
+    name: "8. Append at EOF",
+    fileName: "eof.txt",
+    fileContent: ["line one", "line two"].join("\n"),
+    prompt: [
+      "Read eof.txt with read_file.",
+      "Use edit_file to append 'line three' after the LAST line of the file.",
+      "Use op='append', pos=<last line anchor>, lines=['line three'].",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (!c.includes("line three")) return { passed: false, reason: "'line three' not found" };
+      if (lines[lines.length - 1].trim() !== "line three")
+        return { passed: false, reason: "'line three' not at end" };
+      if (lines.length !== 3) return { passed: false, reason: `expected 3 lines, got ${lines.length}` };
+      return { passed: true, reason: "appended at EOF" };
+    },
+  },
+  {
+    name: "9. Special characters in content",
+    fileName: "special.json",
+    fileContent: [
+      '{',
+      '  "name": "old-value",',
+      '  "count": 42',
+      '}',
+    ].join("\n"),
+    prompt: [
+      "Read special.json with read_file.",
+      'Replace the line containing \"name\": \"old-value\" with \"name\": \"new-value\".',
+      "Use edit_file with op='replace', pos=<that line's anchor>, lines=['  \"name\": \"new-value\",'].",
+    ].join(" "),
+    validate: (c) => {
+      if (c.includes("old-value")) return { passed: false, reason: "'old-value' still present" };
+      if (!c.includes('"new-value"')) return { passed: false, reason: "'new-value' not found" };
+      if (!c.includes('"count": 42')) return { passed: false, reason: "other content was modified" };
+      return { passed: true, reason: "JSON value replaced with special chars intact" };
+    },
+  },
+  {
+    name: "10. Replace first line",
+    fileName: "first.txt",
+    fileContent: ["OLD HEADER", "body content", "footer"].join("\n"),
+    prompt: [
+      "Read first.txt with read_file.",
+      "Replace the very first line 'OLD HEADER' with 'NEW HEADER'.",
+      "Use edit_file with op='replace', pos=<line1 anchor>, lines=['NEW HEADER'].",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (c.includes("OLD HEADER")) return { passed: false, reason: "'OLD HEADER' still present" };
+      if (lines[0].trim() !== "NEW HEADER") return { passed: false, reason: "first line is not 'NEW HEADER'" };
+      if (!c.includes("body content")) return { passed: false, reason: "body was modified" };
+      return { passed: true, reason: "first line replaced" };
+    },
+  },
+  {
+    name: "11. Replace last line",
+    fileName: "last.txt",
+    fileContent: ["alpha", "bravo", "OLD_FOOTER"].join("\n"),
+    prompt: [
+      "Read last.txt with read_file.",
+      "Replace the last line 'OLD_FOOTER' with 'NEW_FOOTER'.",
+      "Use edit_file with op='replace', pos=<last line anchor>, lines=['NEW_FOOTER'].",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (c.includes("OLD_FOOTER")) return { passed: false, reason: "'OLD_FOOTER' still present" };
+      if (lines[lines.length - 1].trim() !== "NEW_FOOTER")
+        return { passed: false, reason: "last line is not 'NEW_FOOTER'" };
+      return { passed: true, reason: "last line replaced" };
+    },
+  },
+  {
+    name: "12. Adjacent line edits",
+    fileName: "adjacent.txt",
+    fileContent: ["aaa", "bbb", "ccc", "ddd"].join("\n"),
+    prompt: [
+      "Read adjacent.txt with read_file.",
+      "Replace line 2 ('bbb') with 'BBB' and line 3 ('ccc') with 'CCC'.",
+      "Use edit_file with TWO edits in the same call:",
+      "  { op: 'replace', pos: <line2 anchor>, lines: ['BBB'] }",
+      "  { op: 'replace', pos: <line3 anchor>, lines: ['CCC'] }",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (c.includes("bbb")) return { passed: false, reason: "'bbb' still present" };
+      if (c.includes("ccc")) return { passed: false, reason: "'ccc' still present" };
+      if (!c.includes("BBB")) return { passed: false, reason: "'BBB' not found" };
+      if (!c.includes("CCC")) return { passed: false, reason: "'CCC' not found" };
+      if (lines.length !== 4) return { passed: false, reason: `expected 4 lines, got ${lines.length}` };
+      return { passed: true, reason: "two adjacent lines replaced" };
+    },
+  },
+  {
+    name: "13. Prepend multi-line block",
+    fileName: "block.py",
+    fileContent: ["def main():", "    print('hello')", "", "main()"].join("\n"),
+    prompt: [
+      "Read block.py with read_file.",
+      "Prepend a 2-line comment block before 'def main():' (line 1).",
+      "The two lines are: '# Author: test' and '# Date: 2025-01-01'.",
+      "Use edit_file with op='prepend', pos=<line1 anchor>, lines=['# Author: test', '# Date: 2025-01-01'].",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (!c.includes("# Author: test")) return { passed: false, reason: "author comment not found" };
+      if (!c.includes("# Date: 2025-01-01")) return { passed: false, reason: "date comment not found" };
+      const defIdx = lines.findIndex((l) => l.startsWith("def main"));
+      const authorIdx = lines.findIndex((l) => l.includes("Author"));
+      if (authorIdx >= defIdx) return { passed: false, reason: "comments not before def" };
+      return { passed: true, reason: "2-line block prepended before function" };
+    },
+  },
+  {
+    name: "14. Delete range — 3 consecutive lines",
+    fileName: "cleanup.txt",
+    fileContent: ["keep1", "remove-a", "remove-b", "remove-c", "keep2"].join("\n"),
+    prompt: [
+      "Read cleanup.txt with read_file.",
+      "Delete lines 2-4 ('remove-a', 'remove-b', 'remove-c') using a single range replace.",
+      "Use edit_file with op='replace', pos=<line2 anchor>, end=<line4 anchor>, lines=[].",
+      "An empty lines array deletes the range.",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (c.includes("remove")) return { passed: false, reason: "'remove' lines still present" };
+      if (!c.includes("keep1")) return { passed: false, reason: "'keep1' was deleted" };
+      if (!c.includes("keep2")) return { passed: false, reason: "'keep2' was deleted" };
+      if (lines.length !== 2) return { passed: false, reason: `expected 2 lines, got ${lines.length}` };
+      return { passed: true, reason: "3 consecutive lines deleted via range" };
+    },
+  },
+  {
+    name: "15. Replace with duplicate-content line",
+    fileName: "dupes.txt",
+    fileContent: ["item", "item", "item", "item"].join("\n"),
+    prompt: [
+      "Read dupes.txt with read_file. All 4 lines have the same text 'item'.",
+      "Replace ONLY line 3 with 'CHANGED'. Do NOT modify any other line.",
+      "Use edit_file with op='replace', pos=<line3 anchor>, lines=['CHANGED'].",
+      "The anchor hash uniquely identifies line 3 even though the content is identical.",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (!c.includes("CHANGED")) return { passed: false, reason: "'CHANGED' not found" };
+      const changedCount = lines.filter((l) => l.trim() === "CHANGED").length;
+      const itemCount = lines.filter((l) => l.trim() === "item").length;
+      if (changedCount !== 1) return { passed: false, reason: `expected 1 CHANGED, got ${changedCount}` };
+      if (itemCount !== 3) return { passed: false, reason: `expected 3 item lines, got ${itemCount}` };
+      if (lines.length !== 4) return { passed: false, reason: `expected 4 lines, got ${lines.length}` };
+      return { passed: true, reason: "only line 3 changed among duplicates" };
+    },
+  },
+
+  // ── Whitespace cases (16-21) ──────────────────────────────────
+  {
+    name: "16. Fix indentation — 2 spaces → 4 spaces",
+    fileName: "indent.js",
+    fileContent: ["function foo() {", "  const x = 1;", "  return x;", "}"].join("\n"),
+    prompt: [
+      "Read indent.js with read_file.",
+      "Replace line 2 '  const x = 1;' (2-space indent) with '    const x = 1;' (4-space indent).",
+      "Use edit_file with op='replace', pos=<line2 anchor>, lines=['    const x = 1;'].",
+      "The ONLY change is the indentation: 2 spaces → 4 spaces. Content stays the same.",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.split("\n");
+      const line2 = lines[1];
+      if (!line2) return { passed: false, reason: "line 2 missing" };
+      if (line2 === "    const x = 1;") return { passed: true, reason: "indentation fixed to 4 spaces" };
+      if (line2 === "  const x = 1;") return { passed: false, reason: "still 2-space indent" };
+      return { passed: false, reason: `unexpected line 2: '${line2}'` };
+    },
+  },
+  {
+    name: "17. Replace preserving leading whitespace",
+    fileName: "preserve.py",
+    fileContent: [
+      "class Foo:",
+      "    def old_method(self):",
+      "        pass",
+    ].join("\n"),
+    prompt: [
+      "Read preserve.py with read_file.",
+      "Replace line 2 '    def old_method(self):' with '    def new_method(self):'.",
+      "Keep the 4-space indentation. Only change the method name.",
+      "Use edit_file with op='replace', pos=<line2 anchor>, lines=['    def new_method(self):'].",
+    ].join(" "),
+    validate: (c) => {
+      if (c.includes("old_method")) return { passed: false, reason: "'old_method' still present" };
+      const lines = c.split("\n");
+      const methodLine = lines.find((l) => l.includes("new_method"));
+      if (!methodLine) return { passed: false, reason: "'new_method' not found" };
+      if (!methodLine.startsWith("    ")) return { passed: false, reason: "indentation lost" };
+      return { passed: true, reason: "method renamed with indentation preserved" };
+    },
+  },
+  {
+    name: "18. Insert blank line between sections",
+    fileName: "sections.txt",
+    fileContent: ["[section-a]", "value-a=1", "[section-b]", "value-b=2"].join("\n"),
+    prompt: [
+      "Read sections.txt with read_file.",
+      "Insert a blank empty line between 'value-a=1' (line 2) and '[section-b]' (line 3).",
+      "Use edit_file with op='append', pos=<line2 anchor>, lines=[''].",
+      "lines=[''] inserts one empty line.",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.split("\n");
+      const valAIdx = lines.findIndex((l) => l.includes("value-a=1"));
+      const secBIdx = lines.findIndex((l) => l.includes("[section-b]"));
+      if (valAIdx === -1) return { passed: false, reason: "'value-a=1' missing" };
+      if (secBIdx === -1) return { passed: false, reason: "'[section-b]' missing" };
+      if (secBIdx - valAIdx < 2) return { passed: false, reason: "no blank line between sections" };
+      const between = lines[valAIdx + 1];
+      if (between.trim() !== "") return { passed: false, reason: `line between is '${between}', not blank` };
+      return { passed: true, reason: "blank line inserted between sections" };
+    },
+  },
+  {
+    name: "19. Delete blank line",
+    fileName: "noblank.txt",
+    fileContent: ["first", "", "second", "third"].join("\n"),
+    prompt: [
+      "Read noblank.txt with read_file.",
+      "Delete the empty blank line (line 2). Use edit_file with op='replace', pos=<line2 anchor>, lines=[].",
+    ].join(" "),
+    validate: (c) => {
+      const lines = c.trim().split("\n");
+      if (lines.length !== 3) return { passed: false, reason: `expected 3 lines, got ${lines.length}` };
+      if (lines[0].trim() !== "first") return { passed: false, reason: "'first' not on line 1" };
+      if (lines[1].trim() !== "second") return { passed: false, reason: "'second' not on line 2" };
+      return { passed: true, reason: "blank line deleted" };
+    },
+  },
+  {
+    name: "20. Tab → spaces conversion",
+    fileName: "tabs.txt",
+    fileContent: ["start", "\tindented-with-tab", "end"].join("\n"),
+    prompt: [
+      "Read tabs.txt with read_file.",
+      "Replace the tab-indented line 2 using edit_file with edits: [{ op: 'replace', pos: '<line2 anchor>', lines: ['    indented-with-spaces'] }].",
+      "Expected final line 2 to be 4 spaces followed by indented-with-spaces.",
+    ].join(" "),
+    validate: (c) => {
+      if (c.includes("\t")) return { passed: false, reason: "tab still present" };
+      if (!c.includes("    indented-with-spaces"))
+        return { passed: false, reason: "'    indented-with-spaces' not found" };
+      if (!c.includes("start")) return { passed: false, reason: "'start' was modified" };
+      return { passed: true, reason: "tab converted to 4 spaces" };
+    },
+  },
+  {
+    name: "21. Deeply nested indent replacement",
+    fileName: "nested.ts",
+    fileContent: [
+      "if (a) {",
+      "  if (b) {",
+      "    if (c) {",
+      "      old_call();",
+      "    }",
+      "  }",
+      "}",
+    ].join("\n"),
+    prompt: [
+      "Read nested.ts with read_file.",
+      "Replace line 4 '      old_call();' with '      new_call();'.",
+      "Preserve the exact 6-space indentation. Only change the function name.",
+      "Use edit_file with op='replace', pos=<line4 anchor>, lines=['      new_call();'].",
+    ].join(" "),
+    validate: (c) => {
+      if (c.includes("old_call")) return { passed: false, reason: "'old_call' still present" };
+      const lines = c.split("\n");
+      const callLine = lines.find((l) => l.includes("new_call"));
+      if (!callLine) return { passed: false, reason: "'new_call' not found" };
+      const leadingSpaces = callLine.match(/^ */)?.[0].length ?? 0;
+      if (leadingSpaces !== 6) return { passed: false, reason: `expected 6-space indent, got ${leadingSpaces}` };
+      return { passed: true, reason: "deeply nested line replaced with indent preserved" };
+    },
+  },
+];
+
+// ── JSONL event types ─────────────────────────────────────────
+interface ToolCallEvent {
+  tool_call_id: string;
+  tool_input: Record<string, unknown>;
+  tool_name: string;
+  type: "tool_call";
+}
+
+interface ToolResultEvent {
+  error?: string;
+  output: string;
+  tool_call_id: string;
+  type: "tool_result";
+}
+
+interface AnyEvent {
+  type: string;
+  [key: string]: unknown;
+}
+
+// ── Run single test case ─────────────────────────────────────
+async function runTestCase(
+  tc: TestCase,
+  testDir: string
+): Promise<{
+  passed: boolean;
+  editCalls: number;
+  editSuccesses: number;
+  duration: number;
+}> {
+  const testFile = join(testDir, tc.fileName);
+  writeFileSync(testFile, tc.fileContent, "utf-8");
+
+  const headlessScript = resolve(import.meta.dir, "headless.ts");
+  const headlessArgs = [
+    "run",
+    headlessScript,
+    "-p",
+    tc.prompt,
+    "--no-translate",
+    ...extraArgs,
+  ];
+
+  const startTime = Date.now();
+
+  const output = await new Promise<string>((res, reject) => {
+    const proc = spawn("bun", headlessArgs, {
+      cwd: testDir,
+      env: { ...process.env, BUN_INSTALL: process.env.BUN_INSTALL },
+      stdio: ["ignore", "pipe", "pipe"],
+    });
+
+    let stdout = "";
+    let stderr = "";
+
+    proc.stdout.on("data", (chunk: Buffer) => {
+      stdout += chunk.toString();
+    });
+    proc.stderr.on("data", (chunk: Buffer) => {
+      stderr += chunk.toString();
+    });
+
+    const timeout = setTimeout(
+      () => {
+        proc.kill("SIGTERM");
+        reject(new Error("Timed out after 4 minutes"));
+      },
+      4 * 60 * 1000
+    );
+
+    proc.on("close", (code) => {
+      clearTimeout(timeout);
+      if (code !== 0) {
+        reject(new Error(`Exit code ${code}\n${stderr.slice(-500)}`));
+      } else {
+        res(stdout);
+      }
+    });
+    proc.on("error", (err) => {
+      clearTimeout(timeout);
+      reject(err);
+    });
+  });
+
+  const duration = Date.now() - startTime;
+
+  // Parse events
+  const events: AnyEvent[] = [];
+  for (const line of output.split("\n").filter((l) => l.trim())) {
+    try {
+      events.push(JSON.parse(line) as AnyEvent);
+    } catch {
+      // skip non-JSON
+    }
+  }
+
+  const toolCalls = events.filter(
+    (e) => e.type === "tool_call"
+  ) as unknown as ToolCallEvent[];
+  const toolResults = events.filter(
+    (e) => e.type === "tool_result"
+  ) as unknown as ToolResultEvent[];
+
+  const editCalls = toolCalls.filter((e) => e.tool_name === "edit_file");
+  const editCallIds = new Set(editCalls.map((e) => e.tool_call_id));
+  const editResults = toolResults.filter((e) =>
+    editCallIds.has(e.tool_call_id)
+  );
+  const editSuccesses = editResults.filter((e) => !e.error);
+
+  // Show blocked calls
+  const editErrors = editResults.filter((e) => e.error);
+  for (const err of editErrors) {
+    const matchingCall = editCalls.find(
+      (c) => c.tool_call_id === err.tool_call_id
+    );
+    info(`  blocked: ${err.error?.slice(0, 120)}`);
+    if (matchingCall) {
+      info(`  input: ${JSON.stringify(matchingCall.tool_input).slice(0, 200)}`);
+    }
+  }
+
+  // Validate file content
+  let finalContent: string;
+  try {
+    finalContent = readFileSync(testFile, "utf-8");
+  } catch {
+    return {
+      passed: false,
+      editCalls: editCalls.length,
+      editSuccesses: editSuccesses.length,
+      duration,
+    };
+  }
+
+  const validation = tc.validate(finalContent);
+
+  return {
+    passed: validation.passed,
+    editCalls: editCalls.length,
+    editSuccesses: editSuccesses.length,
+    duration,
+  };
+}
+
+// ── Main ──────────────────────────────────────────────────────
+const main = async () => {
+  console.log(`\n${BOLD}Headless Edit Operations Test — ${TEST_CASES.length} Types${RESET}\n`);
+
+  const testDir = join(tmpdir(), `edit-ops-${Date.now()}`);
+  mkdirSync(testDir, { recursive: true });
+  info(`Test dir: ${testDir}`);
+  console.log();
+
+  let totalPassed = 0;
+  const results: { name: string; passed: boolean; detail: string }[] = [];
+
+  for (const tc of TEST_CASES) {
+    console.log(`${CYAN}${BOLD}${tc.name}${RESET}`);
+    info(`File: ${tc.fileName}`);
+    info(`Prompt: "${tc.prompt.slice(0, 80)}..."`);
+
+    try {
+      const result = await runTestCase(tc, testDir);
+      const status = result.passed
+        ? `${GREEN}PASS${RESET}`
+        : `${RED}FAIL${RESET}`;
+      const detail = `edit_file: ${result.editSuccesses}/${result.editCalls} succeeded, ${(result.duration / 1000).toFixed(1)}s`;
+
+      console.log(`  ${status} — ${detail}`);
+
+      if (result.passed) {
+        totalPassed++;
+        // Validate the file to show reason
+        const content = readFileSync(join(testDir, tc.fileName), "utf-8");
+        const v = tc.validate(content);
+        pass(v.reason);
+      } else {
+        const content = readFileSync(join(testDir, tc.fileName), "utf-8");
+        const v = tc.validate(content);
+        fail(v.reason);
+        info(
+          `Final content:\n${content
+            .split("\n")
+            .map((l, i) => `    ${i + 1}: ${l}`)
+            .join("\n")}`
+        );
+      }
+
+      results.push({ name: tc.name, passed: result.passed, detail });
+    } catch (error) {
+      const msg = error instanceof Error ? error.message : String(error);
+      console.log(`  ${RED}ERROR${RESET} — ${msg.slice(0, 200)}`);
+      fail(msg.slice(0, 200));
+      results.push({ name: tc.name, passed: false, detail: msg.slice(0, 100) });
+    }
+
+    // Reset file for next test (in case of side effects)
+    try {
+      rmSync(join(testDir, tc.fileName), { force: true });
+    } catch {}
+
+    console.log();
+  }
+
+  // Summary
+  console.log(`${BOLD}━━━ Summary ━━━${RESET}`);
+  for (const r of results) {
+    const icon = r.passed ? `${GREEN}✓${RESET}` : `${RED}✗${RESET}`;
+    console.log(`  ${icon} ${r.name} — ${r.detail}`);
+  }
+  console.log();
+  console.log(
+    `${BOLD}Result: ${totalPassed}/${TEST_CASES.length} passed (${Math.round((totalPassed / TEST_CASES.length) * 100)}%)${RESET}`
+  );
+
+  // Cleanup
+  try {
+    rmSync(testDir, { recursive: true, force: true });
+  } catch {}
+
+  if (totalPassed === TEST_CASES.length) {
+    console.log(
+      `\n${BOLD}${GREEN}🎉 ALL TESTS PASSED — 100% success rate!${RESET}\n`
+    );
+    process.exit(0);
+  } else {
+    console.log(`\n${BOLD}${RED}Some tests failed.${RESET}\n`);
+    process.exit(1);
+  }
+};
+
+main();

benchmarks/test-multi-model.ts

@@ -0,0 +1,280 @@
+#!/usr/bin/env bun
+/**
+ * Multi-model edit_file test runner
+ *
+ * Runs test-headless-edit-ops.ts against every available model
+ * and produces a summary table.
+ *
+ * Usage:
+ *   bun run scripts/test-multi-model-edit.ts [--timeout <seconds>]
+ */
+
+import { spawn } from "node:child_process";
+import { resolve } from "node:path";
+
+// ── Models ────────────────────────────────────────────────────
+const MODELS = [
+  { id: "MiniMaxAI/MiniMax-M2.5", short: "M2.5" },
+  // { id: "MiniMaxAI/MiniMax-M2.1", short: "M2.1" },  // masked: slow + timeout-prone
+  // { id: "zai-org/GLM-5", short: "GLM-5" },            // masked: API 503
+  { id: "zai-org/GLM-4.7", short: "GLM-4.7" },
+];
+
+// ── CLI args ──────────────────────────────────────────────────
+let perModelTimeoutSec = 900; // 15 min default per model (5 tests)
+const rawArgs = process.argv.slice(2);
+for (let i = 0; i < rawArgs.length; i++) {
+  if (rawArgs[i] === "--timeout" && i + 1 < rawArgs.length) {
+    const parsed = Number.parseInt(rawArgs[i + 1], 10);
+    if (Number.isNaN(parsed) || parsed <= 0) {
+      console.error(`Invalid --timeout value: ${rawArgs[i + 1]}`);
+      process.exit(1);
+    }
+    perModelTimeoutSec = parsed;
+    i++;
+}
+
+// ── Colors ────────────────────────────────────────────────────
+const BOLD = "\x1b[1m";
+const GREEN = "\x1b[32m";
+const RED = "\x1b[31m";
+const YELLOW = "\x1b[33m";
+const DIM = "\x1b[2m";
+const CYAN = "\x1b[36m";
+const RESET = "\x1b[0m";
+
+// ── Types ─────────────────────────────────────────────────────
+interface TestResult {
+  detail: string;
+  name: string;
+  passed: boolean;
+}
+
+interface ModelResult {
+  durationMs: number;
+  error?: string;
+  modelId: string;
+  modelShort: string;
+  tests: TestResult[];
+  totalPassed: number;
+  totalTests: number;
+}
+
+// ── Parse test-headless-edit-ops stdout ───────────────────────
+function parseOpsOutput(stdout: string): TestResult[] {
+  const results: TestResult[] = [];
+
+  // Match lines like: "  PASS — edit_file: 1/1 succeeded, 32.5s"
+  // or "  FAIL — edit_file: 0/3 succeeded, 15.2s"
+  // or "  ERROR — Timed out after 10 minutes"
+  // Following a line like: "1. Replace single line"
+  const lines = stdout.split("\n");
+
+  let currentTestName = "";
+  for (const line of lines) {
+    // Detect test name: starts with ANSI-colored bold cyan + "N. Name"
+    // Strip ANSI codes for matching
+    const stripped = line.replace(/\x1b\[[0-9;]*m/g, "");
+
+    // Test name pattern: "N. <name>"
+    const testNameMatch = stripped.match(/^\s*(\d+\.\s+.+)$/);
+    if (
+      testNameMatch &&
+      !stripped.includes("—") &&
+      !stripped.includes("✓") &&
+      !stripped.includes("✗")
+    ) {
+      currentTestName = testNameMatch[1].trim();
+      continue;
+    }
+
+    // Result line: PASS/FAIL/ERROR
+    if (currentTestName && stripped.includes("PASS")) {
+      const detail = stripped.replace(/^\s*PASS\s*—?\s*/, "").trim();
+      results.push({
+        name: currentTestName,
+        passed: true,
+        detail: detail || "passed",
+      });
+      currentTestName = "";
+    } else if (currentTestName && stripped.includes("FAIL")) {
+      const detail = stripped.replace(/^\s*FAIL\s*—?\s*/, "").trim();
+      results.push({
+        name: currentTestName,
+        passed: false,
+        detail: detail || "failed",
+      });
+      currentTestName = "";
+    } else if (currentTestName && stripped.includes("ERROR")) {
+      const detail = stripped.replace(/^\s*ERROR\s*—?\s*/, "").trim();
+      results.push({
+        name: currentTestName,
+        passed: false,
+        detail: detail || "error",
+      });
+      currentTestName = "";
+    }
+  }
+
+  return results;
+}
+
+// ── Run one model ────────────────────────────────────────────
+async function runModel(model: {
+  id: string;
+  short: string;
+}): Promise<ModelResult> {
+  const opsScript = resolve(import.meta.dir, "test-edit-ops.ts");
+  const startTime = Date.now();
+
+  return new Promise<ModelResult>((resolvePromise) => {
+    const proc = spawn(
+      "bun",
+      ["run", opsScript, "-m", model.id, "--no-translate"],
+      {
+        cwd: resolve(import.meta.dir),
+        env: { ...process.env, BUN_INSTALL: process.env.BUN_INSTALL },
+        stdio: ["ignore", "pipe", "pipe"],
+      }
+    );
+
+    let stdout = "";
+    let stderr = "";
+
+    proc.stdout.on("data", (chunk: Buffer) => {
+      stdout += chunk.toString();
+    });
+    proc.stderr.on("data", (chunk: Buffer) => {
+      stderr += chunk.toString();
+    });
+
+    const timeout = setTimeout(() => {
+      proc.kill("SIGTERM");
+      resolvePromise({
+        modelId: model.id,
+        modelShort: model.short,
+        tests: [],
+        totalPassed: 0,
+        totalTests: 0,
+        durationMs: Date.now() - startTime,
+        error: `Timed out after ${perModelTimeoutSec}s`,
+      });
+    }, perModelTimeoutSec * 1000);
+
+    proc.on("close", () => {
+      clearTimeout(timeout);
+      const tests = parseOpsOutput(stdout);
+      const totalPassed = tests.filter((t) => t.passed).length;
+
+      resolvePromise({
+        modelId: model.id,
+        modelShort: model.short,
+        tests,
+        totalPassed,
+        totalTests: Math.max(tests.length, 5),
+        durationMs: Date.now() - startTime,
+      });
+    });
+
+    proc.on("error", (err) => {
+      clearTimeout(timeout);
+      resolvePromise({
+        modelId: model.id,
+        modelShort: model.short,
+        tests: [],
+        totalPassed: 0,
+        totalTests: 0,
+        durationMs: Date.now() - startTime,
+        error: err.message,
+      });
+    });
+  });
+}
+
+// ── Main ──────────────────────────────────────────────────────
+const main = async () => {
+  console.log(`\n${BOLD}═══ Multi-Model edit_file Test Runner ═══${RESET}\n`);
+  console.log(`${DIM}Models: ${MODELS.map((m) => m.short).join(", ")}${RESET}`);
+  console.log(`${DIM}Timeout: ${perModelTimeoutSec}s per model${RESET}`);
+  console.log();
+
+  const allResults: ModelResult[] = [];
+
+  for (const model of MODELS) {
+    console.log(`${CYAN}${BOLD}▶ Testing ${model.short} (${model.id})${RESET}`);
+    const result = await runModel(model);
+    allResults.push(result);
+
+    const timeStr = `${(result.durationMs / 1000).toFixed(1)}s`;
+    if (result.error) {
+      console.log(`  ${RED}ERROR${RESET}: ${result.error} (${timeStr})`);
+    } else {
+      const color =
+        result.totalPassed === result.totalTests
+          ? GREEN
+          : result.totalPassed > 0
+            ? YELLOW
+            : RED;
+      console.log(
+        `  ${color}${result.totalPassed}/${result.totalTests} passed${RESET} (${timeStr})`
+      );
+      for (const t of result.tests) {
+        const icon = t.passed ? `${GREEN}✓${RESET}` : `${RED}✗${RESET}`;
+        console.log(`    ${icon} ${t.name}`);
+      }
+    }
+    console.log();
+  }
+
+  // ── Summary Table ──────────────────────────────────────────
+  console.log(`${BOLD}═══ Summary ═══${RESET}\n`);
+
+  // Per-model results
+  for (const r of allResults) {
+    const timeStr = `${(r.durationMs / 1000).toFixed(0)}s`;
+    const color = r.error ? RED : r.totalPassed === r.totalTests ? GREEN : r.totalPassed > 0 ? YELLOW : RED;
+    const label = r.error ? `ERROR: ${r.error}` : `${r.totalPassed}/${r.totalTests}`;
+    console.log(`  ${r.modelShort.padEnd(8)} ${color}${label}${RESET} (${timeStr})`);
+    for (const t of r.tests) {
+      const icon = t.passed ? `${GREEN}✓${RESET}` : `${RED}✗${RESET}`;
+      console.log(`    ${icon} ${t.name}`);
+    }
+  }
+
+  console.log();
+
+  // Overall
+  const totalModels = allResults.length;
+  const erroredModels = allResults.filter((r) => r.error).length;
+  const perfectModels = allResults.filter(
+    (r) => !r.error && r.totalPassed === r.totalTests && r.totalTests > 0
+  ).length;
+  console.log(
+    `${BOLD}Models with 100%: ${perfectModels}/${totalModels}${RESET}`
+  );
+
+  const overallPassed = allResults.reduce((sum, r) => sum + r.totalPassed, 0);
+  const overallTotal = allResults.reduce((sum, r) => sum + r.totalTests, 0);
+  console.log(
+    `${BOLD}Overall: ${overallPassed}/${overallTotal} (${Math.round((overallPassed / overallTotal) * 100)}%)${RESET}`
+  );
+
+  console.log();
+
+  if (erroredModels > 0) {
+    console.log(
+      `${BOLD}${RED}${erroredModels} model(s) errored. See details above.${RESET}\n`
+    );
+    process.exit(1);
+  } else if (perfectModels === totalModels) {
+    console.log(`${BOLD}${GREEN}🎉 ALL MODELS PASSED ALL TESTS!${RESET}\n`);
+    process.exit(0);
+  } else {
+    console.log(
+      `${BOLD}${YELLOW}Some models have failures. See details above.${RESET}\n`
+    );
+    process.exit(1);
+  }
+};
+
+main();

package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode",
-  "version": "3.8.5",
+  "version": "3.9.0",
   "description": "The Best AI Agent Harness - Batteries-Included OpenCode Plugin with Multi-Model Orchestration, Parallel Background Agents, and Crafted LSP/AST Tools",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -75,17 +75,17 @@
     "typescript": "^5.7.3"
   },
   "optionalDependencies": {
-    "oh-my-opencode-darwin-arm64": "3.8.5",
-    "oh-my-opencode-darwin-x64": "3.8.5",
-    "oh-my-opencode-darwin-x64-baseline": "3.8.5",
-    "oh-my-opencode-linux-arm64": "3.8.5",
-    "oh-my-opencode-linux-arm64-musl": "3.8.5",
-    "oh-my-opencode-linux-x64": "3.8.5",
-    "oh-my-opencode-linux-x64-baseline": "3.8.5",
-    "oh-my-opencode-linux-x64-musl": "3.8.5",
-    "oh-my-opencode-linux-x64-musl-baseline": "3.8.5",
-    "oh-my-opencode-windows-x64": "3.8.5",
-    "oh-my-opencode-windows-x64-baseline": "3.8.5"
+    "oh-my-opencode-darwin-arm64": "3.9.0",
+    "oh-my-opencode-darwin-x64": "3.9.0",
+    "oh-my-opencode-darwin-x64-baseline": "3.9.0",
+    "oh-my-opencode-linux-arm64": "3.9.0",
+    "oh-my-opencode-linux-arm64-musl": "3.9.0",
+    "oh-my-opencode-linux-x64": "3.9.0",
+    "oh-my-opencode-linux-x64-baseline": "3.9.0",
+    "oh-my-opencode-linux-x64-musl": "3.9.0",
+    "oh-my-opencode-linux-x64-musl-baseline": "3.9.0",
+    "oh-my-opencode-windows-x64": "3.9.0",
+    "oh-my-opencode-windows-x64-baseline": "3.9.0"
   },
   "trustedDependencies": [
     "@ast-grep/cli",

packages/darwin-arm64/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-darwin-arm64",
-  "version": "3.8.5",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (darwin-arm64)",
   "license": "MIT",
   "repository": {

packages/darwin-x64-baseline/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-darwin-x64-baseline",
-  "version": "3.1.1",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (darwin-x64-baseline, no AVX2)",
   "license": "MIT",
   "repository": {

packages/darwin-x64/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-darwin-x64",
-  "version": "3.8.5",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (darwin-x64)",
   "license": "MIT",
   "repository": {

packages/linux-arm64-musl/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-linux-arm64-musl",
-  "version": "3.8.5",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (linux-arm64-musl)",
   "license": "MIT",
   "repository": {

packages/linux-arm64/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-linux-arm64",
-  "version": "3.8.5",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (linux-arm64)",
   "license": "MIT",
   "repository": {

packages/linux-x64-baseline/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-linux-x64-baseline",
-  "version": "3.1.1",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (linux-x64-baseline, no AVX2)",
   "license": "MIT",
   "repository": {

packages/linux-x64-musl-baseline/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-linux-x64-musl-baseline",
-  "version": "3.1.1",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (linux-x64-musl-baseline, no AVX2)",
   "license": "MIT",
   "repository": {

packages/linux-x64-musl/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-linux-x64-musl",
-  "version": "3.8.5",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (linux-x64-musl)",
   "license": "MIT",
   "repository": {

packages/linux-x64/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-linux-x64",
-  "version": "3.8.5",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (linux-x64)",
   "license": "MIT",
   "repository": {

packages/windows-x64-baseline/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-windows-x64-baseline",
-  "version": "3.1.1",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (windows-x64-baseline, no AVX2)",
   "license": "MIT",
   "repository": {

packages/windows-x64/package.json

@@ -1,6 +1,6 @@
 {
   "name": "oh-my-opencode-windows-x64",
-  "version": "3.8.5",
+  "version": "3.9.0",
   "description": "Platform-specific binary for oh-my-opencode (windows-x64)",
   "license": "MIT",
   "repository": {

src/agents/hephaestus.ts

@@ -448,21 +448,6 @@ ${oracleSection}
 4. **Run build** if applicable — exit code 0 required
 5. **Tell user** what you verified and the results — keep it clear and helpful
 
-### Auto-Commit Policy (MANDATORY for implementation/fix work)
-
-1. **Auto-commit after implementation is complete** when the task includes feature/fix code changes
-2. **Commit ONLY after verification gates pass**:
-   - \`lsp_diagnostics\` clean on all modified files
-   - Related tests pass
-   - Typecheck/build pass when applicable
-3. **If any gate fails, DO NOT commit** — fix issues first, re-run verification, then commit
-4. **Use Conventional Commits format** with meaningful intent-focused messages:
-   - \`feat(scope): add ...\` for new functionality
-   - \`fix(scope): resolve ...\` for bug fixes
-   - \`refactor(scope): simplify ...\` for internal restructuring
-5. **Do not make placeholder commits** (\`wip\`, \`temp\`, \`update\`) or commit unverified code
-6. **If user explicitly says not to commit**, skip commit and report that changes are left uncommitted
-
 - **File edit** — \`lsp_diagnostics\` clean
 - **Build** — Exit code 0
 - **Tests** — Pass (or pre-existing failures noted)

src/features/background-agent/manager.test.ts

@@ -2987,6 +2987,28 @@ describe("BackgroundManager.handleEvent - session.deleted cascade", () => {
     manager.shutdown()
     resetToastManager()
   })
+
+  test("should clean pending notifications for deleted sessions", () => {
+    //#given
+    const manager = createBackgroundManager()
+    const sessionID = "session-pending-notifications"
+
+    manager.queuePendingNotification(sessionID, "<system-reminder>queued</system-reminder>")
+    expect(getPendingNotifications(manager).get(sessionID)).toEqual([
+      "<system-reminder>queued</system-reminder>",
+    ])
+
+    //#when
+    manager.handleEvent({
+      type: "session.deleted",
+      properties: { info: { id: sessionID } },
+    })
+
+    //#then
+    expect(getPendingNotifications(manager).has(sessionID)).toBe(false)
+
+    manager.shutdown()
+  })
 })
 
 describe("BackgroundManager.handleEvent - session.error", () => {

src/features/background-agent/manager.ts

@@ -830,6 +830,8 @@ export class BackgroundManager {
         tasksToCancel.set(descendant.id, descendant)
       }
 
+      this.pendingNotifications.delete(sessionID)
+
       if (tasksToCancel.size === 0) return
 
       for (const task of tasksToCancel.values()) {
@@ -866,6 +868,13 @@ export class BackgroundManager {
           subagentSessions.delete(task.sessionID)
         }
       }
+
+      for (const task of tasksToCancel.values()) {
+        if (task.parentSessionID) {
+          this.pendingNotifications.delete(task.parentSessionID)
+        }
+      }
+
       SessionCategoryRegistry.remove(sessionID)
     }
 

src/hooks/ralph-loop/completion-promise-detector.test.ts

@@ -0,0 +1,111 @@
+/// <reference types="bun-types" />
+import { describe, expect, test } from "bun:test"
+import type { PluginInput } from "@opencode-ai/plugin"
+import { detectCompletionInSessionMessages } from "./completion-promise-detector"
+
+type SessionMessage = {
+  info?: { role?: string }
+  parts?: Array<{ type: string; text?: string }>
+}
+
+function createPluginInput(messages: SessionMessage[]): PluginInput {
+  const pluginInput = {
+    client: { session: {} } as PluginInput["client"],
+    project: {} as PluginInput["project"],
+    directory: "/tmp",
+    worktree: "/tmp",
+    serverUrl: new URL("http://localhost"),
+    $: {} as PluginInput["$"],
+  } as PluginInput
+
+  pluginInput.client.session.messages =
+    (async () => ({ data: messages })) as unknown as PluginInput["client"]["session"]["messages"]
+
+  return pluginInput
+}
+
+describe("detectCompletionInSessionMessages", () => {
+  describe("#given session with prior DONE and new messages", () => {
+    test("#when sinceMessageIndex excludes prior DONE #then should NOT detect completion", async () => {
+      // #given
+      const messages: SessionMessage[] = [
+        {
+          info: { role: "assistant" },
+          parts: [{ type: "text", text: "Old completion <promise>DONE</promise>" }],
+        },
+        {
+          info: { role: "assistant" },
+          parts: [{ type: "text", text: "Working on the new task" }],
+        },
+      ]
+      const ctx = createPluginInput(messages)
+
+      // #when
+      const detected = await detectCompletionInSessionMessages(ctx, {
+        sessionID: "session-123",
+        promise: "DONE",
+        apiTimeoutMs: 1000,
+        directory: "/tmp",
+        sinceMessageIndex: 1,
+      })
+
+      // #then
+      expect(detected).toBe(false)
+    })
+
+    test("#when sinceMessageIndex includes current DONE #then should detect completion", async () => {
+      // #given
+      const messages: SessionMessage[] = [
+        {
+          info: { role: "assistant" },
+          parts: [{ type: "text", text: "Old completion <promise>DONE</promise>" }],
+        },
+        {
+          info: { role: "assistant" },
+          parts: [{ type: "text", text: "Current completion <promise>DONE</promise>" }],
+        },
+      ]
+      const ctx = createPluginInput(messages)
+
+      // #when
+      const detected = await detectCompletionInSessionMessages(ctx, {
+        sessionID: "session-123",
+        promise: "DONE",
+        apiTimeoutMs: 1000,
+        directory: "/tmp",
+        sinceMessageIndex: 1,
+      })
+
+      // #then
+      expect(detected).toBe(true)
+    })
+  })
+
+  describe("#given no sinceMessageIndex (backward compat)", () => {
+    test("#then should scan all messages", async () => {
+      // #given
+      const messages: SessionMessage[] = [
+        {
+          info: { role: "assistant" },
+          parts: [{ type: "text", text: "Old completion <promise>DONE</promise>" }],
+        },
+        {
+          info: { role: "assistant" },
+          parts: [{ type: "text", text: "No completion in latest message" }],
+        },
+      ]
+      const ctx = createPluginInput(messages)
+
+      // #when
+      const detected = await detectCompletionInSessionMessages(ctx, {
+        sessionID: "session-123",
+        promise: "DONE",
+        apiTimeoutMs: 1000,
+        directory: "/tmp",
+      })
+
+      // #then
+      expect(detected).toBe(true)
+    })
+  })
+})

src/hooks/ralph-loop/completion-promise-detector.ts

@@ -52,6 +52,7 @@ export async function detectCompletionInSessionMessages(
 		promise: string
 		apiTimeoutMs: number
 		directory: string
+		sinceMessageIndex?: number
 	},
 ): Promise<boolean> {
 	try {
@@ -75,7 +76,12 @@ export async function detectCompletionInSessionMessages(
 				? responseData
 				: []
 
-		const assistantMessages = (messageArray as OpenCodeSessionMessage[]).filter((msg) => msg.info?.role === "assistant")
+		const scopedMessages =
+			typeof options.sinceMessageIndex === "number" && options.sinceMessageIndex >= 0 && options.sinceMessageIndex < messageArray.length
+				? messageArray.slice(options.sinceMessageIndex)
+				: messageArray
+
+		const assistantMessages = (scopedMessages as OpenCodeSessionMessage[]).filter((msg) => msg.info?.role === "assistant")
 		if (assistantMessages.length === 0) return false
 
 		const pattern = buildPromisePattern(options.promise)

src/hooks/ralph-loop/index.test.ts

@@ -603,7 +603,7 @@ describe("ralph-loop", () => {
       expect(hook.getState()).toBeNull()
 
       // then - messages API was called with correct session ID
-      expect(messagesCalls.length).toBe(1)
+      expect(messagesCalls.length).toBe(2)
       expect(messagesCalls[0].sessionID).toBe("session-123")
     })
 
@@ -633,7 +633,7 @@ describe("ralph-loop", () => {
       expect(hook.getState()).toBeNull()
 
       // then - messages API was called with correct session ID
-      expect(messagesCalls.length).toBe(1)
+      expect(messagesCalls.length).toBe(2)
       expect(messagesCalls[0].sessionID).toBe("session-123")
     })
 
@@ -1075,7 +1075,7 @@ Original task: Build something`
       expect(promptCalls.length).toBe(0)
       expect(hook.getState()).toBeNull()
       // API should NOT be called since transcript found completion
-      expect(messagesCalls.length).toBe(0)
+      expect(messagesCalls.length).toBe(1)
     })
 
     test("should show ultrawork completion toast", async () => {

src/hooks/ralph-loop/loop-state-controller.ts

@@ -23,6 +23,7 @@ export function createLoopStateController(options: {
 			loopOptions?: {
 				maxIterations?: number
 				completionPromise?: string
+				messageCountAtStart?: number
 				ultrawork?: boolean
 				strategy?: "reset" | "continue"
 			},
@@ -34,6 +35,7 @@ export function createLoopStateController(options: {
 					loopOptions?.maxIterations ??
 					config?.default_max_iterations ??
 					DEFAULT_MAX_ITERATIONS,
+				message_count_at_start: loopOptions?.messageCountAtStart,
 				completion_promise:
 					loopOptions?.completionPromise ??
 					DEFAULT_COMPLETION_PROMISE,
@@ -93,5 +95,19 @@ export function createLoopStateController(options: {
 
 			return state
 		},
+
+		setMessageCountAtStart(sessionID: string, messageCountAtStart: number): RalphLoopState | null {
+			const state = readState(directory, stateDir)
+			if (!state || state.session_id !== sessionID) {
+				return null
+			}
+
+			state.message_count_at_start = messageCountAtStart
+			if (!writeState(directory, state, stateDir)) {
+				return null
+			}
+
+			return state
+		},
 	}
 }

src/hooks/ralph-loop/ralph-loop-event-handler.ts

@@ -84,6 +84,7 @@ export function createRalphLoopEventHandler(
 						promise: state.completion_promise,
 						apiTimeoutMs: options.apiTimeoutMs,
 						directory: options.directory,
+						sinceMessageIndex: state.message_count_at_start,
 					})
 
 				if (completionViaTranscript || completionViaApi) {

src/hooks/ralph-loop/ralph-loop-hook.ts

@@ -13,6 +13,7 @@ export interface RalphLoopHook {
     options?: {
       maxIterations?: number
       completionPromise?: string
+      messageCountAtStart?: number
       ultrawork?: boolean
       strategy?: "reset" | "continue"
     }
@@ -23,6 +24,19 @@ export interface RalphLoopHook {
 
 const DEFAULT_API_TIMEOUT = 5000 as const
 
+function getMessageCountFromResponse(messagesResponse: unknown): number {
+  if (Array.isArray(messagesResponse)) {
+    return messagesResponse.length
+  }
+
+  if (typeof messagesResponse === "object" && messagesResponse !== null && "data" in messagesResponse) {
+    const data = (messagesResponse as { data?: unknown }).data
+    return Array.isArray(data) ? data.length : 0
+  }
+
+  return 0
+}
+
 export function createRalphLoopHook(
   ctx: PluginInput,
   options?: RalphLoopOptions
@@ -51,7 +65,25 @@ export function createRalphLoopHook(
 
 	return {
 		event,
-		startLoop: loopState.startLoop,
+		startLoop: (sessionID, prompt, loopOptions): boolean => {
+			const startSuccess = loopState.startLoop(sessionID, prompt, loopOptions)
+			if (!startSuccess || typeof loopOptions?.messageCountAtStart === "number") {
+				return startSuccess
+			}
+
+			ctx.client.session
+				.messages({
+					path: { id: sessionID },
+					query: { directory: ctx.directory },
+				})
+				.then((messagesResponse: unknown) => {
+					const messageCountAtStart = getMessageCountFromResponse(messagesResponse)
+					loopState.setMessageCountAtStart(sessionID, messageCountAtStart)
+				})
+				.catch(() => {})
+
+			return startSuccess
+		},
 		cancelLoop: loopState.cancelLoop,
 		getState: loopState.getState as () => RalphLoopState | null,
 	}

src/hooks/ralph-loop/storage.ts

@@ -44,6 +44,12 @@ export function readState(directory: string, customPath?: string): RalphLoopStat
       active: isActive,
       iteration: iterationNum,
       max_iterations: Number(data.max_iterations) || DEFAULT_MAX_ITERATIONS,
+      message_count_at_start:
+        typeof data.message_count_at_start === "number"
+          ? data.message_count_at_start
+          : typeof data.message_count_at_start === "string" && data.message_count_at_start.trim() !== ""
+            ? Number(data.message_count_at_start)
+            : undefined,
       completion_promise: stripQuotes(data.completion_promise) || DEFAULT_COMPLETION_PROMISE,
       started_at: stripQuotes(data.started_at) || new Date().toISOString(),
       prompt: body.trim(),
@@ -72,13 +78,17 @@ export function writeState(
     const sessionIdLine = state.session_id ? `session_id: "${state.session_id}"\n` : ""
     const ultraworkLine = state.ultrawork !== undefined ? `ultrawork: ${state.ultrawork}\n` : ""
     const strategyLine = state.strategy ? `strategy: "${state.strategy}"\n` : ""
+    const messageCountAtStartLine =
+      typeof state.message_count_at_start === "number"
+        ? `message_count_at_start: ${state.message_count_at_start}\n`
+        : ""
     const content = `---
 active: ${state.active}
 iteration: ${state.iteration}
 max_iterations: ${state.max_iterations}
 completion_promise: "${state.completion_promise}"
 started_at: "${state.started_at}"
-${sessionIdLine}${ultraworkLine}${strategyLine}---
+${sessionIdLine}${ultraworkLine}${strategyLine}${messageCountAtStartLine}---
 ${state.prompt}
 `
 

src/hooks/ralph-loop/types.ts

@@ -4,6 +4,7 @@ export interface RalphLoopState {
   active: boolean
   iteration: number
   max_iterations: number
+  message_count_at_start?: number
   completion_promise: string
   started_at: string
   prompt: string

src/plugin/ultrawork-db-model-override.ts

@@ -23,8 +23,8 @@ function tryUpdateMessageModel(
   if (result.changes === 0) return false
   if (variant) {
     db.prepare(
-      `UPDATE message SET data = json_set(data, '$.variant', ?) WHERE id = ?`,
-    ).run(variant, messageId)
+      `UPDATE message SET data = json_set(data, '$.variant', ?, '$.thinking', ?) WHERE id = ?`,
+    ).run(variant, variant, messageId)
   }
   return true
 }

src/tools/delegate-task/token-limiter.test.ts

@@ -19,6 +19,8 @@ import {
   truncateToTokenBudget,
 } from "./token-limiter"
 
+const TRUNCATION_MARKER_TOKEN_OVERHEAD = estimateTokenCount("\n[TRUNCATED]")
+
 describe("token-limiter", () => {
   test("estimateTokenCount uses 1 token per 4 chars approximation", () => {
     // given
@@ -40,7 +42,62 @@ describe("token-limiter", () => {
     const result = truncateToTokenBudget(content, maxTokens)
 
     // then
-    expect(estimateTokenCount(result)).toBeLessThanOrEqual(maxTokens)
+    expect(estimateTokenCount(result)).toBeLessThanOrEqual(maxTokens + TRUNCATION_MARKER_TOKEN_OVERHEAD)
+  })
+
+  describe("truncateToTokenBudget", () => {
+    describe("#given content that exceeds budget", () => {
+      describe("#when content has newlines", () => {
+        test("#then should truncate at last newline boundary", () => {
+          // #given
+          const content = "line-1\nline-2\nline-3"
+
+          // #when
+          const result = truncateToTokenBudget(content, 2)
+
+          // #then
+          expect(result).toBe("line-1\n[TRUNCATED]")
+        })
+
+        test("#then should append [TRUNCATED] marker", () => {
+          // #given
+          const content = "line-1\nline-2\nline-3"
+
+          // #when
+          const result = truncateToTokenBudget(content, 2)
+
+          // #then
+          expect(result).toContain("[TRUNCATED]")
+        })
+      })
+
+      describe("#when content is single long line with no newlines", () => {
+        test("#then should slice and append [TRUNCATED] marker", () => {
+          // #given
+          const content = "A".repeat(30)
+
+          // #when
+          const result = truncateToTokenBudget(content, 2)
+
+          // #then
+          expect(result).toBe("AAAAAAAA\n[TRUNCATED]")
+        })
+      })
+    })
+
+    describe("#given content within budget", () => {
+      test("#then should return content unchanged without marker", () => {
+        // #given
+        const content = "line-1\nline-2"
+
+        // #when
+        const result = truncateToTokenBudget(content, 20)
+
+        // #then
+        expect(result).toBe(content)
+        expect(result).not.toContain("[TRUNCATED]")
+      })
+    })
   })
 
   test("buildSystemContentWithTokenLimit returns undefined when there is no content", () => {
@@ -76,10 +133,10 @@ describe("token-limiter", () => {
     const result = buildSystemContentWithTokenLimit(input, 80)
 
     // then
-    expect(result).toContain("AGENTS_CONTEXT:keep")
-    expect(result).toContain("CATEGORY_APPEND:keep")
+    expect(result).toContain("AGENTS_C")
+    expect(result).toContain("CATE")
     expect(result).toContain("SKILL_ALPHA:")
-    expect(estimateTokenCount(result as string)).toBeLessThanOrEqual(80)
+    expect(estimateTokenCount(result as string)).toBeLessThanOrEqual(80 + TRUNCATION_MARKER_TOKEN_OVERHEAD)
   })
 
   test("buildSystemContentWithTokenLimit truncates category after skills are exhausted", () => {
@@ -95,9 +152,9 @@ describe("token-limiter", () => {
     const result = buildSystemContentWithTokenLimit(input, 30)
 
     // then
-    expect(result).toContain("AGENTS_CONTEXT:keep")
+    expect(result).toContain("AGENTS_C")
     expect(result).not.toContain("SKILL_ALPHA:" + "a".repeat(80))
-    expect(estimateTokenCount(result as string)).toBeLessThanOrEqual(30)
+    expect(estimateTokenCount(result as string)).toBeLessThanOrEqual(30 + TRUNCATION_MARKER_TOKEN_OVERHEAD)
   })
 
   test("buildSystemContentWithTokenLimit truncates agents context last", () => {
@@ -116,6 +173,6 @@ describe("token-limiter", () => {
     expect(result).toContain("AGENTS_CONTEXT:")
     expect(result).not.toContain("SKILL_ALPHA:")
     expect(result).not.toContain("CATEGORY_APPEND:")
-    expect(estimateTokenCount(result as string)).toBeLessThanOrEqual(10)
+    expect(estimateTokenCount(result as string)).toBeLessThanOrEqual(10 + TRUNCATION_MARKER_TOKEN_OVERHEAD)
   })
 })

src/tools/delegate-task/token-limiter.ts

@@ -20,7 +20,13 @@ export function truncateToTokenBudget(content: string, maxTokens: number): strin
     return content
   }
 
-  return content.slice(0, maxCharacters)
+  const sliced = content.slice(0, maxCharacters)
+  const lastNewline = sliced.lastIndexOf("\n")
+  if (lastNewline > 0) {
+    return `${sliced.slice(0, lastNewline)}\n[TRUNCATED]`
+  }
+
+  return `${sliced}\n[TRUNCATED]`
 }
 
 function joinSystemParts(parts: string[]): string | undefined {

src/tools/hashline-edit/diff-utils.test.ts

@@ -1,5 +1,6 @@
 /// <reference types="bun-types" />
 import { describe, expect, it } from "bun:test"
+import { parsePatch } from "diff"
 import { generateUnifiedDiff } from "./diff-utils"
 
 function createNumberedLines(totalLineCount: number): string {
@@ -7,6 +8,66 @@ function createNumberedLines(totalLineCount: number): string {
 }
 
 describe("generateUnifiedDiff", () => {
+  describe("#given OpenCode compatibility format", () => {
+    it("#then includes the Index header emitted by diff library", () => {
+      //#given
+      const oldContent = "a\n"
+      const newContent = "b\n"
+
+      //#when
+      const diff = generateUnifiedDiff(oldContent, newContent, "test.ts")
+
+      //#then
+      expect(diff).toContain("Index: test.ts")
+    })
+
+    it("#then includes unified --- and +++ file headers", () => {
+      //#given
+      const oldContent = "a\n"
+      const newContent = "b\n"
+
+      //#when
+      const diff = generateUnifiedDiff(oldContent, newContent, "test.ts")
+
+      //#then
+      expect(diff).toContain("--- test.ts")
+      expect(diff).toContain("+++ test.ts")
+    })
+
+    it("#then remains parseable by OpenCode parsePatch flow", () => {
+      //#given
+      const oldContent = "line1\nline2\n"
+      const newContent = "line1\nline2-updated\n"
+
+      //#when
+      const diff = generateUnifiedDiff(oldContent, newContent, "test.ts")
+      const patches = parsePatch(diff)
+
+      //#then
+      expect(patches).toHaveLength(1)
+      expect(patches[0]?.oldFileName).toBe("test.ts")
+      expect(patches[0]?.newFileName).toBe("test.ts")
+      expect(patches[0]?.hunks).toHaveLength(1)
+    })
+  })
+
+  describe("#given content without trailing newline", () => {
+    it("#then keeps no-newline markers parseable", () => {
+      //#given
+      const oldContent = "a"
+      const newContent = "b"
+
+      //#when
+      const diff = generateUnifiedDiff(oldContent, newContent, "test.ts")
+      const patches = parsePatch(diff)
+      const hunkLines = patches[0]?.hunks[0]?.lines ?? []
+
+      //#then
+      expect(diff).toContain("\\ No newline at end of file")
+      expect(hunkLines).toEqual(["-a", "\\ No newline at end of file", "+b", "\\ No newline at end of file"])
+    })
+  })
+
   it("creates separate hunks for distant changes", () => {
     //#given
     const oldContent = createNumberedLines(60)

src/tools/hashline-edit/edit-deduplication.ts

@@ -1,18 +1,24 @@
 import type { HashlineEdit } from "./types"
 import { toNewLines } from "./edit-text-normalization"
+import { normalizeLineRef } from "./validation"
 
 function normalizeEditPayload(payload: string | string[]): string {
   return toNewLines(payload).join("\n")
 }
 
+function canonicalAnchor(anchor: string | undefined): string {
+  if (!anchor) return ""
+  return normalizeLineRef(anchor)
+}
+
 function buildDedupeKey(edit: HashlineEdit): string {
   switch (edit.op) {
     case "replace":
-      return `replace|${edit.pos}|${edit.end ?? ""}|${normalizeEditPayload(edit.lines)}`
+      return `replace|${canonicalAnchor(edit.pos)}|${edit.end ? canonicalAnchor(edit.end) : ""}|${normalizeEditPayload(edit.lines)}`
     case "append":
-      return `append|${edit.pos ?? ""}|${normalizeEditPayload(edit.lines)}`
+      return `append|${canonicalAnchor(edit.pos)}|${normalizeEditPayload(edit.lines)}`
     case "prepend":
-      return `prepend|${edit.pos ?? ""}|${normalizeEditPayload(edit.lines)}`
+      return `prepend|${canonicalAnchor(edit.pos)}|${normalizeEditPayload(edit.lines)}`
     default:
       return JSON.stringify(edit)
   }

src/tools/hashline-edit/edit-operations.test.ts

@@ -1,5 +1,5 @@
 import { describe, expect, it } from "bun:test"
-import { applyHashlineEdits } from "./edit-operations"
+import { applyHashlineEdits, applyHashlineEditsWithReport } from "./edit-operations"
 import { applyAppend, applyInsertAfter, applyPrepend, applyReplaceLines, applySetLine } from "./edit-operation-primitives"
 import { computeLineHash } from "./hash-computation"
 import type { HashlineEdit } from "./types"
@@ -389,3 +389,23 @@ describe("hashline edit operations", () => {
     expect(result).toEqual("replaced A\nline 3\nreplaced B")
   })
 })
+
+describe("dedupe anchor canonicalization", () => {
+  it("deduplicates edits with whitespace-variant anchors", () => {
+    //#given
+    const content = "line 1\nline 2"
+    const lines = content.split("\n")
+    const canonical = `1#${computeLineHash(1, lines[0])}`
+    const spaced = ` 1 # ${computeLineHash(1, lines[0])} `
+
+    //#when
+    const report = applyHashlineEditsWithReport(content, [
+      { op: "append", pos: canonical, lines: ["inserted"] },
+      { op: "append", pos: spaced, lines: ["inserted"] },
+    ])
+
+    //#then
+    expect(report.deduplicatedEdits).toBe(1)
+    expect(report.content).toBe("line 1\ninserted\nline 2")
+  })
+})

src/tools/hashline-edit/hashline-edit-executor.ts

@@ -33,7 +33,7 @@ function resolveToolCallID(ctx: ToolContextWithCallID): string | undefined {
 
 function canCreateFromMissingFile(edits: HashlineEdit[]): boolean {
   if (edits.length === 0) return false
-  return edits.every((edit) => edit.op === "append" || edit.op === "prepend")
+  return edits.every((edit) => (edit.op === "append" || edit.op === "prepend") && !edit.pos)
 }
 
 function buildSuccessMeta(
@@ -86,19 +86,19 @@ export async function executeHashlineEditTool(args: HashlineEditArgs, context: T
     const filePath = args.filePath
     const { delete: deleteMode, rename } = args
 
+    if (deleteMode && rename) {
+      return "Error: delete and rename cannot be used together"
+    }
+    if (deleteMode && args.edits.length > 0) {
+      return "Error: delete mode requires edits to be an empty array"
+    }
+
     if (!deleteMode && (!args.edits || !Array.isArray(args.edits) || args.edits.length === 0)) {
       return "Error: edits parameter must be a non-empty array"
     }
 
     const edits = deleteMode ? [] : normalizeHashlineEdits(args.edits)
 
-    if (deleteMode && rename) {
-      return "Error: delete and rename cannot be used together"
-    }
-    if (deleteMode && edits.length > 0) {
-      return "Error: delete mode requires edits to be an empty array"
-    }
-
     const file = Bun.file(filePath)
     const exists = await file.exists()
     if (!exists && !deleteMode && !canCreateFromMissingFile(edits)) {

src/tools/hashline-edit/tool-description.ts

@@ -10,7 +10,7 @@ WORKFLOW:
 VALIDATION:
   Payload shape: { "filePath": string, "edits": [...], "delete"?: boolean, "rename"?: string }
   Each edit must be one of: replace, append, prepend
-  Edit shape: { "op": "replace"|"append"|"prepend", "pos"?: "LINE#ID", "end"?: "LINE#ID", "lines"?: string|string[]|null }
+  Edit shape: { "op": "replace"|"append"|"prepend", "pos"?: "LINE#ID", "end"?: "LINE#ID", "lines": string|string[]|null }
   lines must contain plain replacement text only (no LINE#ID prefixes, no diff + markers)
   CRITICAL: all operations validate against the same pre-edit file snapshot and apply bottom-up. Refs/tags are interpreted against the last-read version of the file.
 

src/tools/hashline-edit/tools.test.ts

@@ -341,4 +341,81 @@ describe("createHashlineEditTool", () => {
     //#then
     expect(envelope.lineEnding).toBe("\r\n")
   })
+
+  it("rejects delete=true with non-empty edits before normalization", async () => {
+    //#given
+    const filePath = path.join(tempDir, "delete-reject.txt")
+    fs.writeFileSync(filePath, "line1")
+
+    //#when
+    const result = await tool.execute(
+      {
+        filePath,
+        delete: true,
+        edits: [{ op: "replace", pos: "1#ZZ", lines: "bad" }],
+      },
+      createMockContext(),
+    )
+
+    //#then
+    expect(result).toContain("delete mode requires edits to be an empty array")
+    expect(fs.existsSync(filePath)).toBe(true)
+  })
+
+  it("rejects delete=true combined with rename", async () => {
+    //#given
+    const filePath = path.join(tempDir, "delete-rename.txt")
+    fs.writeFileSync(filePath, "line1")
+
+    //#when
+    const result = await tool.execute(
+      {
+        filePath,
+        delete: true,
+        rename: path.join(tempDir, "new-name.txt"),
+        edits: [],
+      },
+      createMockContext(),
+    )
+
+    //#then
+    expect(result).toContain("delete and rename cannot be used together")
+    expect(fs.existsSync(filePath)).toBe(true)
+  })
+
+  it("rejects missing file creation with anchored append", async () => {
+    //#given
+    const filePath = path.join(tempDir, "nonexistent.txt")
+
+    //#when
+    const result = await tool.execute(
+      {
+        filePath,
+        edits: [{ op: "append", pos: "1#ZZ", lines: ["bad"] }],
+      },
+      createMockContext(),
+    )
+
+    //#then
+    expect(result).toContain("File not found")
+  })
+
+  it("allows missing file creation with unanchored append", async () => {
+    //#given
+    const filePath = path.join(tempDir, "newfile.txt")
+
+    //#when
+    const result = await tool.execute(
+      {
+        filePath,
+        edits: [{ op: "append", lines: ["created"] }],
+      },
+      createMockContext(),
+    )
+
+    //#then
+    expect(fs.existsSync(filePath)).toBe(true)
+    expect(fs.readFileSync(filePath, "utf-8")).toBe("created")
+    expect(result).toBe(`Updated ${filePath}`)
+  })
 })

src/tools/hashline-edit/tools.ts

@@ -31,7 +31,6 @@ export function createHashlineEditTool(): ToolDefinition {
             end: tool.schema.string().optional().describe("Range end anchor in LINE#ID format"),
             lines: tool.schema
               .union([tool.schema.string(), tool.schema.array(tool.schema.string()), tool.schema.null()])
-              .optional()
               .describe("Replacement or inserted lines. null/[] deletes with replace"),
           })
         )

src/tools/hashline-edit/validation.ts

@@ -15,7 +15,7 @@ const MISMATCH_CONTEXT = 2
 
 const LINE_REF_EXTRACT_PATTERN = /([0-9]+#[ZPMQVRWSNKTXJBYH]{2})/
 
-function normalizeLineRef(ref: string): string {
+export function normalizeLineRef(ref: string): string {
   const originalTrimmed = ref.trim()
   let trimmed = originalTrimmed
   trimmed = trimmed.replace(/^(?:>>>|[+-])\s*/, "")