gaslamp 2.2.2 → 3.0.0

package/src/consult.mjs

@@ -1,67 +1,43 @@
-// consult.mjs — the heart of gaslamp 2.0: one blocking consult, spawned as a
-// shell child of whichever agent called it.
+// consult.mjs — one blocking consult, spawned as a shell child of whichever
+// agent called it.
 //
 //   gaslamp claude [opts] <prompt|->    spawns `claude -p`
 //   gaslamp codex  [opts] <prompt|->    spawns `codex exec`
 //   gaslamp gemini [opts] <prompt|->    spawns `agy` (Antigravity CLI) in print mode
-//   gaslamp grok   [opts] <prompt|->    spawns `grok` (Grok Build) in headless mode
+//   gaslamp grok   [opts] <prompt|->    spawns `grok` (Grok Build) headless
 //
 // Block-until-reply, deliberately: asynchrony is the HARNESS's job. Each agent
-// backgrounds the shell call with its own facility (Claude Code:
-// run_in_background; Codex: its background terminal) and is told when it
-// finishes — so consults run in parallel and replies trickle in, with no
-// daemon, no job spool, and no second async layer inside gaslamp.
+// backgrounds the shell call with its own facility and is told when it
+// finishes, so consults run in parallel with no daemon and no job spool.
 //
-// What survives of a spool is passive write-through durability (see jobs.mjs):
-// every consult records prompt/events/reply/stderr as it runs, and the session
-// id is captured the moment the backend reports it. If the harness kills the
-// wrapper we kill the whole child process group — deliberate orphans would
-// invisibly double-burn API quota — and recovery is `--resume <session>`, not
-// a surviving child.
+// What survives of a spool is passive write-through durability (jobs.mjs):
+// the session id is recorded the moment the backend reports it. If the
+// harness kills the wrapper, the whole child process group dies with it (an
+// orphan would invisibly double-burn quota), and recovery is continuing the
+// conversation, not a surviving child.
 //
-// The recursion guard is ONE mechanism now (1.0 needed two): children are
-// spawned with GASLAMP_NESTED=1 and the consult verbs refuse under it. A
-// consult is one hop; GASLAMP_ALLOW_RECURSION=1 opts out. For Codex the
-// sentinel is *also* forced into its shell env via
+// The recursion guard: children run with GASLAMP_NESTED=1 and the consult
+// verbs refuse under it. Codex also gets it forced into its shell env via
 // `-c shell_environment_policy.set.GASLAMP_NESTED="1"`, because a
-// `shell_environment_policy` with `inherit = "core"` would otherwise strip it
-// before any nested `gaslamp` call could see it. (The dotted leaf MERGES with
-// the user's [shell_environment_policy.set] table — verified on codex 0.139 —
-// it does not replace it.) It is loop prevention, not a security boundary: a
-// shell-capable agent could unset the env var.
+// `shell_environment_policy` with `inherit = "core"` would strip it (the
+// dotted leaf merges with the user's own [shell_environment_policy.set]).
+// Loop prevention, not a security boundary.
 //
 // Exit codes (also in --help):
 //   0 reply delivered · 1 consult failed or killed · 2 usage · 3 nested
-//   (one-hop) refusal · 4 network-disabled sandbox · 5 session busy
+//   (one-hop) refusal · 4 this environment can't run it (no network, or no OS
+//   sandbox for the mode) · 5 conversation busy
 
 import { execFileSync, spawn } from "node:child_process";
-import { createHash } from "node:crypto";
-import { createWriteStream, existsSync, mkdirSync, readdirSync, readFileSync, realpathSync, unlinkSync, writeFileSync, writeSync } from "node:fs";
-import { homedir } from "node:os";
-import { dirname, join } from "node:path";
-import { fileURLToPath } from "node:url";
-import { BACKENDS } from "./backends.mjs";
+import { appendFileSync, createWriteStream, existsSync, mkdirSync, readFileSync, realpathSync, unlinkSync, writeFileSync, writeSync } from "node:fs";
+import { homedir, tmpdir } from "node:os";
+import { dirname, isAbsolute, join, resolve } from "node:path";
+import { BACKENDS, SANDBOXES } from "./backends.mjs";
 import { which } from "./which.mjs";
-import {
-  alive, createJob, jobDir, jobsDir, locksDir, looksLikeJobId,
-  newJobId, readMeta, renderTrailer, writeMeta,
-} from "./jobs.mjs";
+import { alive, createJob, envelope, jobDir, jobsDir, locksDir, looksLikeJobId, newJobId, readMeta, renderTrailer, shellQuote, writeMeta } from "./jobs.mjs";
 import { readThread, validThreadName, writeThread } from "./threads.mjs";
 import { readStdin } from "./stdin.mjs";
 
-// Tools permitted under the claude `read-only` override (advisory reviewer).
-// The git reads matter: a reviewer that can't run `git diff`/`git log` is
-// blind to exactly the evidence it's usually asked about (codex's read-only
-// sandbox can run them; this keeps the two read-onlys comparable). Note the
-// override is "read-capable", not purely tool-read-only.
-const DEFAULT_ALLOWED_TOOLS = "Read,Grep,Glob,WebFetch,WebSearch," +
-  "Bash(git diff:*),Bash(git log:*),Bash(git show:*),Bash(git status:*)";
-// Comma-split when a comma is present (entries like `Bash(git diff:*)` carry
-// spaces — a naive /[\s,]+/ shreds them); legacy whitespace lists still work.
-const ALLOWED_TOOLS_RAW = process.env.GASLAMP_ALLOWED_TOOLS || DEFAULT_ALLOWED_TOOLS;
-const ALLOWED_TOOLS = ALLOWED_TOOLS_RAW
-  .split(ALLOWED_TOOLS_RAW.includes(",") ? "," : /\s+/)
-  .map((s) => s.trim()).filter(Boolean).join(",");
 const ALLOW_RECURSION = !!process.env.GASLAMP_ALLOW_RECURSION;
 const DEBUG = !!process.env.GASLAMP_DEBUG;
 const log = (...a) => { if (DEBUG) writeSync(2, "[gaslamp] " + a.join(" ") + "\n"); };
@@ -70,51 +46,49 @@ const log = (...a) => { if (DEBUG) writeSync(2, "[gaslamp] " + a.join(" ") + "\n
 // pipe is async in Node; a plain write could be truncated).
 const die = (code, msg) => { writeSync(2, `gaslamp: ${msg}\n`); process.exit(code); };
 
+const realHome = () => { try { return realpathSync(homedir()); } catch { return homedir(); } };
+
 // The default Gemini tier when --model is omitted. Antigravity also serves
 // Claude and GPT-OSS models under the same account, and an account default
 // pointing at one of them would turn a "gemini" consult into a second opinion
 // from a model the caller already has — so the target always pins a Gemini.
 const GEMINI_DEFAULT_TIER = "argon";
 
-// Tool names that reach the live web (or X) per backend, for the provenance
-// record: what the consultee could retrieve, and what it actually did.
-const RETRIEVAL_TOOLS = {
-  claude: /^(WebSearch|WebFetch)$/,
-  codex: /^web_search/,
-  gemini: /^(search_web|read_url_content|open_browser_url|read_browser_page|browser_)/,
-  grok: /^(web_search|web_fetch|x_)/,
-};
-
-// Grok's X search is a backend tool: init never lists it, and a call arrives
-// as a tool_use named "X search:" ({"variant":"XSearch","backend":true}) whose
-// tool_result names the tool that ran (x_keyword_search, x_semantic_search).
-const toolKey = (name) => String(name).toLowerCase().replace(/[^a-z0-9]+/g, "_").replace(/^_+|_+$/g, "");
+// The consult preamble: the one piece of framing a consulted agent needs.
+// Subagents are told their final text IS the return value; a consult deserves
+// the same, or it replies conversationally and promises work instead of
+// delivering it. claude takes it as --append-system-prompt and grok as
+// --rules; codex and agy have no such flag and get it prepended to the prompt
+// as a <gaslamp_consult> block. prompt.md records the caller's content only.
+const PREAMBLE =
+  "You are being consulted by another coding agent via gaslamp; this is one hop, and you cannot consult back. " +
+  "Your final message is returned verbatim to the calling agent as the consult's result — make it the deliverable " +
+  "itself (findings, verdict, diff, answer), not a conversational reply or a promise of future work. " +
+  "Distinguish what you verified from what you infer.";
 
 // Variables a third-party harness has no business inheriting from the calling
 // agent: Claude Code's session id, messaging socket token, account and OAuth
 // details; Codex's and the API keys' equivalents. They reach the consultee's
-// shell and its MCP servers — the end-conversation server keyed its
-// destructive path on CLAUDE_CODE_SESSION_ID before it learned to stop at the
-// nearest harness — and nothing in a Gemini or Grok consult needs them.
+// shell and its MCP servers — the end-conversation server once resolved its
+// harness to the CALLING Claude session through them.
 const FOREIGN_ENV = /^(CLAUDE|ANTHROPIC_|CODEX_|OPENAI_)/;
 
 // Grok Build reads other harnesses' configuration by default — ~/.claude's
 // CLAUDE.md as its own instructions, Claude's skills, MCP servers and hooks,
 // Cursor's the same. A consult must not hand the consultee the caller's
-// identity file or run the caller's hooks. Machines bootstrapped from a9lim's
-// ledger turn these off in ~/.grok/config.toml; this covers every other one.
-// (Grok also reads ~/.claude/settings.json permissions, with no switch; the
-// explicit --permission-mode below overrides its defaultMode.)
+// identity file or run the caller's hooks. (Grok also reads
+// ~/.claude/settings.json permissions, with no switch; the explicit
+// --permission-mode below overrides its defaultMode.)
 const GROK_COMPAT_OFF = Object.fromEntries(
   ["CLAUDE", "CURSOR"].flatMap((v) => ["AGENTS", "SKILLS", "RULES", "MCPS", "HOOKS"]
     .map((c) => [`GROK_${v}_${c}_ENABLED`, "0"])));
 
 // The environment every process gaslamp starts for a backend gets — the
-// consult itself and its probes (`--version`, `agy models`) alike.
+// consult itself and the `agy models` probe alike.
 function backendEnv(backend) {
   const env = { ...process.env, GASLAMP_NESTED: "1" };
   // Keychain OAuth is authoritative for the consulted Claude: a stale env key
-  // 401s every call (1.0's hardest auth lesson). No Codex analog needed.
+  // 401s every call.
   if (backend === "claude") delete env.ANTHROPIC_API_KEY;
   if (backend === "gemini" || backend === "grok")
     for (const k of Object.keys(env)) if (FOREIGN_ENV.test(k)) delete env[k];
@@ -123,84 +97,98 @@ function backendEnv(backend) {
   return env;
 }
 
-// Read-only for a gemini consult, on macOS: agy runs under Seatbelt, writing
-// nothing but its own state directory and temp — the guarantee no longer
-// depends on Antigravity's permission settings, which a user may keep lax
-// for interactive work (toolPermission always-proceed, agentMode auto) and
-// which no flag overrides per run. Inside that state directory the files
-// that configure agy stay write-denied (later deny rules win in SBPL).
-// (bin/ stays writable: agy extracts a helper, bin/agentapi, before running
-// any shell command, and denying it breaks read-only commands like git diff.)
-// Elsewhere — Linux, or inside a sandbox of Codex's own, where Seatbelt does
-// not nest — the preset check below is the guarantee.
+// ---- sandboxes ----------------------------------------------------------------
+// Every fenced mode is enforced by the OS, and a fence that can't start
+// refuses the consult rather than running it unfenced:
+//   claude  Claude Code's own sandbox (Seatbelt / bubblewrap) around Bash,
+//           switched on per run with --settings; it does not cover the file
+//           tools, so permissions fence those
+//   codex   its sandbox_mode
+//   gemini  gaslamp runs agy under Seatbelt, so macOS only
+//   grok    its --sandbox (Seatbelt / Landlock)
+// read-only means no writes outside the harness's own state and temp dirs.
+// It does not cut the network: a read-only consultee can still search.
+
+// The file tools run outside Claude Code's sandbox.
+const CLAUDE_FILE_TOOLS = "Edit,Write,NotebookEdit";
+
+// The user's own settings can puncture the fence and --settings can't take
+// it back: lists merge across scopes (excludedCommands run unfenced,
+// allowWrite adds directories), and sandbox.filesystem.disabled is honored
+// from user settings only. Probed: either one let a read-only consult write.
+// So a fenced consult loads no user, project or local settings at all
+// (--setting-sources ""): managed settings and gaslamp's --settings only.
+function claudeSandbox(sandbox, cwd) {
+  // allowUnsandboxedCommands:false drops the unsandboxed retry; without
+  // failIfUnavailable a sandbox that can't start (no socat on Linux, nested
+  // under another sandbox) silently runs commands unfenced.
+  const s = { enabled: true, failIfUnavailable: true, allowUnsandboxedCommands: false, autoAllowBashIfSandboxed: true };
+  // The write set is the cwd and temp; read-only denies home and the cwd on
+  // top (deny wins). Not "/": temp stays writable, which Claude Code's
+  // per-command bookkeeping needs.
+  if (sandbox === "read-only") s.filesystem = { denyWrite: [realHome(), cwd] };
+  return JSON.stringify({ sandbox: s });
+}
+
 const SANDBOX_EXEC = "/usr/bin/sandbox-exec";
-const useSeatbelt = () => process.platform === "darwin" && existsSync(SANDBOX_EXEC)
-  && !process.env.CODEX_SANDBOX && process.env.GASLAMP_GEMINI_SEATBELT !== "0";
-function agySeatbeltProfile() {
+// Seatbelt does not nest, so inside Codex's own sandbox it isn't available.
+const seatbelt = () => process.platform === "darwin" && existsSync(SANDBOX_EXEC) && !process.env.CODEX_SANDBOX;
+
+// agy writes nothing but its own state directory and temp, plus the cwd under
+// workspace-write. Inside the state directory the files that configure agy
+// stay write-denied (in SBPL the later rule wins). bin/ stays writable: agy
+// extracts a helper there (bin/agentapi) before every shell command.
+function agyProfile(sandbox, cwd) {
   const q = (p) => '"' + p.replace(/[\\"]/g, (c) => "\\" + c) + '"';
-  const state = join(homedir(), ".gemini", "antigravity-cli");
+  const state = join(realHome(), ".gemini", "antigravity-cli");
+  const writable = [state, "/private/tmp", "/private/var/folders", "/dev", ...(sandbox === "workspace-write" ? [cwd] : [])];
   return [
     "(version 1)",
     "(allow default)",
     "(deny file-write*)",
-    `(allow file-write* (subpath ${q(state)}) (subpath "/private/tmp") (subpath "/private/var/folders") (subpath "/dev"))`,
+    `(allow file-write* ${writable.map((p) => `(subpath ${q(p)})`).join(" ")})`,
     `(deny file-write* ${["settings.json", "keybindings.json"].map((f) => `(literal ${q(join(state, f))})`).join(" ")} ` +
       `${["rules", "plugins", "skills", "updater"].map((d) => `(subpath ${q(join(state, d))})`).join(" ")})`,
   ].join("\n");
 }
 
-// What a read-only gemini consult may find pre-allowed in Antigravity's
-// permission files: reads, read-only git (read-capable like the claude read
-// set — git's own --output flags aside), and the end_session off-ramp. Any
-// other grant (writes, other commands, unsandboxed, browser actuation, other
-// MCP tools) would run unprompted under the request-review preset, so its
-// presence refuses the consult rather than letting "read-only" mean less.
-const AGY_READ_GRANT = /^(read_file|read_url)\(|^command\(git (diff|log|show|status)\)$|^mcp\(end-conversation\/(\*|end_session)\)$/;
-const AGY_LAX_PRESETS = new Set(["always-proceed", "proceed-in-sandbox"]);
-// agentMode "auto" approves file edits and commands itself: under it a print
-// run wrote a file and ran `touch` while init still reported request-review
-// (1.4.3), so the init check cannot see it. accept-edits is refused likewise.
-const AGY_LAX_MODES = new Set(["auto", "accept-edits"]);
-
-// Allow rules and the tool-permission preset from every file the CLI merges:
-// its settings.json, the settings shared with the Antigravity app, and the
-// per-project files (all of them — which one applies is not knowable here).
-function agyPermissionState() {
-  const root = join(homedir(), ".gemini");
-  const files = [join(root, "antigravity-cli", "settings.json"), join(root, "config", "config.json")];
+// ---- the end_session off-ramp ---------------------------------------------------
+// The end-conversation server (a9lim/end-conversation-mcp) logs to
+// ~/.claude/conversation-ends.jsonl, which no sandbox here lets a consultee's
+// MCP server write. Every consult names a spool in the per-user temp dir,
+// which every fence allows; the server falls back to it, and gaslamp,
+// unfenced, files the records into the log when the consult ends. Only
+// records addressed to that log are filed — the spool is consultee-writable,
+// so its own say-so cannot choose where gaslamp appends.
+const OFF_RAMP_LOG = () => join(homedir(), ".claude", "conversation-ends.jsonl");
+
+// Returns whether the consultee ended its own harness (scope "main").
+function fileOffRamp(spool, id) {
+  if (!existsSync(spool)) return false;
+  let ended = false;
   try {
-    for (const f of readdirSync(join(root, "config", "projects")))
-      if (f.endsWith(".json")) files.push(join(root, "config", "projects", f));
-  } catch {}
-  const grants = [], presets = [], modes = [];
-  const walk = (node, file) => {
-    if (Array.isArray(node)) return node.forEach((n) => walk(n, file));
-    if (!node || typeof node !== "object") return;
-    for (const [k, v] of Object.entries(node)) {
-      if (k === "allow" && Array.isArray(v)) for (const r of v) if (typeof r === "string") grants.push({ rule: r, file });
-      if (k === "toolPermission" && typeof v === "string") presets.push({ preset: v, file });
-      if (k === "agentMode" && typeof v === "string") modes.push({ mode: v, file });
-      walk(v, file);
+    for (const line of readFileSync(spool, "utf8").split("\n")) {
+      let r; try { r = JSON.parse(line); } catch { continue; }
+      if (!r || typeof r !== "object" || r.log !== OFF_RAMP_LOG() || !["main", "instance"].includes(r.scope)) {
+        if (line.trim()) writeSync(2, "[gaslamp] ignored an off-ramp spool line not addressed to the log\n");
+        continue;
+      }
+      const { log: dest, ...record } = r;
+      record.via = `gaslamp ${id}`;
+      mkdirSync(dirname(dest), { recursive: true });
+      appendFileSync(dest, JSON.stringify(record) + "\n");
+      if (r.scope === "main") {
+        ended = true;
+        writeSync(2, `[gaslamp] the consultee ended its session (end_session scope=main${r.reason ? `: ${r.reason}` : ""})\n`);
+      }
     }
-  };
-  for (const f of files) {
-    try { walk(JSON.parse(readFileSync(f, "utf8")), f); } catch {}
+    unlinkSync(spool);
+  } catch (e) {
+    writeSync(2, `[gaslamp] could not file off-ramp records from ${spool}: ${e.message}\n`);
   }
-  return { grants, presets, modes };
+  return ended;
 }
 
-// The consult preamble: the one piece of framing a consulted agent needs.
-// Workflow subagents are told their final text IS the return value; a consult
-// deserves the same, or it replies conversationally and promises work instead
-// of delivering it. Fixed and tiny by design — not a context-injection hook.
-// claude carries it as --append-system-prompt; codex (no such flag) gets it
-// prepended to the prompt as a <gaslamp_consult> block. --raw drops it.
-const PREAMBLE =
-  "You are being consulted by another coding agent via gaslamp; this is one hop, and you cannot consult back. " +
-  "Your final message is returned verbatim to the calling agent as the consult's result — make it the deliverable " +
-  "itself (findings, verdict, diff, answer), not a conversational reply or a promise of future work. " +
-  "Distinguish what you verified from what you infer.";
-
 function resolveBin(backend) {
   if (backend === "gemini") {
     for (const c of [process.env.AGY_BIN, join(homedir(), ".local/bin/agy")].filter(Boolean))
@@ -241,12 +229,11 @@ function parseArgs(argv) {
     else if (a === "--model" || a === "-m") o.model = val();
     else if (a === "--sandbox" || a === "-s") o.sandbox = val();
     else if (a === "--effort" || a === "-e") o.effort = val();
-    else if (a === "--label" || a === "-l") o.label = val();
     else if (a === "--cwd" || a === "-C") o.cwd = val();
     else if (a === "--schema") o.schema = val();
-    else if (a === "--raw") o.raw = true;
     else if (a === "--json") o.json = true;
     else if (a === "-") o.prompt = "-";
+    else if (a === "--label" || a === "-l") die(2, "--label is gone in gaslamp 3: --thread <name> names a consult, and reusing the name continues it");
     else if (a.startsWith("-") && a.length > 1) die(2, `unknown flag ${a} (see gaslamp --help)`);
     else if (o.prompt == null) o.prompt = a;
     else die(2, "more than one prompt argument — quote the prompt, or pipe it on stdin");
@@ -254,6 +241,18 @@ function parseArgs(argv) {
   return o;
 }
 
+// Canonical, so thread and job directories compare equal however they were
+// spelled (/tmp vs /private/tmp), and Seatbelt matches the real path.
+function resolveCwd(raw) {
+  try { return realpathSync(resolve(raw ?? process.cwd())); }
+  catch { die(2, `--cwd ${raw}: no such directory`); }
+}
+
+const canonicalDir = (p) => {
+  if (typeof p !== "string" || !isAbsolute(p)) return null;
+  try { return realpathSync(p); } catch { return p; }
+};
+
 // --schema takes inline JSON (anything starting with "{") or a file path.
 // Returns the schema text, checked to parse — a bad schema should die here as
 // a usage error, not forty seconds into a consult as a backend error.
@@ -270,9 +269,9 @@ function resolveSchema(raw) {
 // OpenAI's structured-output endpoint only accepts STRICT schemas: every
 // object node must carry additionalProperties:false and a `required` listing
 // EVERY key in properties (both verified live — it 400s otherwise). Normalize
-// on the way to codex; claude takes the schema as written. The one semantic
-// this costs is optional-by-omission — spell optional fields nullable
-// ({"type":["string","null"]}) if a codex consult may leave them out.
+// on the way to codex; the others take the schema as written. The one
+// semantic this costs is optional-by-omission — spell optional fields
+// nullable ({"type":["string","null"]}) if a codex consult may leave them out.
 export function strictify(node) {
   if (Array.isArray(node)) return node.map(strictify);
   if (node === null || typeof node !== "object") return node;
@@ -290,25 +289,29 @@ export function strictify(node) {
   return out;
 }
 
-// Map --resume to a concrete session id. Accepts a prior JOB id too (the
-// handle the caller usually has at hand) and resolves it via the job record.
-function resolveResume(backend, resume) {
-  if (!looksLikeJobId(resume)) return resume;
-  const m = readMeta(resume);
-  if (!m) die(2, `no job record ${resume} under ${jobsDir()}`);
+// --resume takes a prior job id: the handle every trailer prints.
+function resolvePrior(backend, id) {
+  if (!looksLikeJobId(id)) die(2, `--resume takes a job id (from a trailer or \`gaslamp jobs\`); a named conversation continues with --thread`);
+  const m = readMeta(id);
+  if (!m) die(2, `no job record ${id} under ${jobsDir()}`);
   if (m.backend !== backend)
-    die(2, `job ${resume} was a ${m.backend} consult — use \`gaslamp ${m.backend} --resume ${resume}\``);
+    die(2, `job ${id} was a ${m.backend} consult — use \`gaslamp ${m.backend} --resume ${id}\``);
+  // A thread's conversation continues through its name: resuming one of its
+  // jobs would fork it past the pointer, under a different lock.
+  if (m.thread) die(2, `job ${id} belongs to thread ${m.thread} — continue it with --thread ${m.thread}`);
   if (!m.sessionId)
-    die(2, `job ${resume} recorded no session id (it ${m.status === "running" ? "is still starting" : "died before the backend reported one"})`);
-  return m.sessionId;
+    die(2, `job ${id} recorded no session id (it ${m.status === "running" ? "is still starting" : "died before the backend reported one"})`);
+  return m;
 }
 
-// Two concurrent consults resuming the SAME session is undefined behavior on
-// both backends, so it's refused (exit 5). Lockfiles are atomic-create
-// (flag "wx"); a lock whose holder pid is dead is stale — reap and retry.
-function acquireLock(backend, sid, jobId) {
+// One conversation, one consult at a time: two consults on the same session
+// is undefined behavior on every backend, and two fresh consults racing to
+// bind the same thread name would each start a conversation. Lockfiles are
+// atomic-create (flag "wx"); a lock whose holder pid is dead is stale — reap
+// and retry.
+function acquireLock(backend, key, jobId) {
   mkdirSync(locksDir(), { recursive: true });
-  const path = join(locksDir(), `${backend}--${sid}`);
+  const path = join(locksDir(), `${backend}--${key}`);
   const payload = JSON.stringify({ pid: process.pid, job: jobId, at: new Date().toISOString() }) + "\n";
   for (let attempt = 0; attempt < 3; attempt++) {
     try { writeFileSync(path, payload, { flag: "wx" }); return path; }
@@ -317,11 +320,11 @@ function acquireLock(backend, sid, jobId) {
       let holder = null;
       try { holder = JSON.parse(readFileSync(path, "utf8")); } catch { /* malformed = stale */ }
       if (holder?.pid && alive(holder.pid))
-        die(5, `session ${sid} is busy — job ${holder.job ?? "?"} (pid ${holder.pid}) is mid-consult on it. \`gaslamp poll ${holder.job ?? "--last"}\` to watch, or wait for it.`);
+        die(5, `that conversation is busy — job ${holder.job ?? "?"} (pid ${holder.pid}) is mid-consult on it. \`gaslamp poll ${holder.job ?? "--last"}\` to watch, or wait for it.`);
       try { unlinkSync(path); } catch { /* raced another reaper */ }
     }
   }
-  die(5, `could not acquire the lock for session ${sid} under ${locksDir()}`);
+  die(5, `could not acquire ${path}`);
 }
 
 // ---- model tiers --------------------------------------------------------------
@@ -378,147 +381,102 @@ function resolveModel(backend, model, bin, env) {
   return model;
 }
 
-// ---- provenance -----------------------------------------------------------------
-// Vendors can change the model or the system prompt behind a stable product
-// name, so every job records what it can about who actually answered.
-const PKG_ROOT = join(dirname(fileURLToPath(import.meta.url)), "..");
-const PKG_VERSION = JSON.parse(readFileSync(join(PKG_ROOT, "package.json"), "utf8")).version;
-let revisionMemo;
-function gaslampRevision() {
-  if (revisionMemo !== undefined) return revisionMemo;
-  try {
-    const git = (...a) => execFileSync("git", ["-C", PKG_ROOT, ...a], { encoding: "utf8", stdio: ["ignore", "pipe", "ignore"], timeout: 5000 }).trim();
-    const rev = git("rev-parse", "--short=12", "HEAD");
-    revisionMemo = git("status", "--porcelain", "--untracked-files=no") ? `${rev}+dirty` : rev;
-  } catch { revisionMemo = "unknown"; }
-  return revisionMemo;
-}
+// ---- backend argv ---------------------------------------------------------------
 
-function harnessVersion(bin, env) {
-  try {
-    const out = execFileSync(bin, ["--version"], { encoding: "utf8", timeout: 10000, stdio: ["ignore", "pipe", "ignore"], env });
-    return out.split("\n").find((l) => l.trim())?.trim() || "unknown";
-  } catch { return "unknown"; }
-}
-
-// Grok writes each session's full system prompt next to its transcript —
-// the one backend whose vendor prompt is knowable, so it is hashed into the
-// record (a change between jobs is visible without storing the text twice).
-function grokSystemPrompt(sessionId, cwd) {
-  const root = join(process.env.GROK_HOME || join(homedir(), ".grok"), "sessions");
-  const candidates = [join(root, encodeURIComponent(cwd), sessionId, "system_prompt.txt")];
-  try { for (const d of readdirSync(root)) candidates.push(join(root, d, sessionId, "system_prompt.txt")); } catch {}
-  for (const path of candidates) {
-    try {
-      const text = readFileSync(path, "utf8");
-      return { known: true, sha256: createHash("sha256").update(text).digest("hex"), chars: text.length, path };
-    } catch {}
-  }
-  return { known: false };
-}
-
-function geminiArgs(o, resumeSid, schemaText, model, sandbox) {
-  // Print mode's -p takes the prompt as an argument; streaming input reads it
-  // from stdin as one {"event":"user"} line and ends the session at EOF, which
-  // keeps the prompt-on-stdin contract (no argv-size or quoting traps).
-  const args = ["--input-format", "stream-json", "--output-format", "stream-json", "--model", model];
-  if (o.effort) args.push("--effort", o.effort);
-  if (resumeSid) args.push("--conversation", resumeSid);
-  if (schemaText) args.push("--json-schema", schemaText);
-  // read-only adds no flag: the default request-review preset already denies
-  // writes, shell and web reads in print mode. Before spawn, the permission
-  // files are checked for grants that would let anything else through; after,
-  // the init event's permission_mode must confirm the preset (fail closed).
-  if (sandbox === "danger-full-access") args.push("--dangerously-skip-permissions");
-  return args;
-}
-
-function grokArgs(o, resumeSid, schemaText, sandbox) {
-  // streaming-messages-json is Anthropic's stream-json shape (system init with
-  // session_id and model first, result last) — the claude parser reads it.
-  const args = ["--prompt-file", "/dev/stdin", "--output-format", "streaming-messages-json", "--no-auto-update"];
-  // --rules appends to the system prompt: claude's --append-system-prompt.
-  if (!o.raw) args.push("--rules", PREAMBLE);
-  if (schemaText) args.push("--json-schema", schemaText);
-  // The kernel sandbox (Seatbelt / Landlock) is the read-only guarantee. The
-  // permission layer is not: under --permission-mode default a headless run
-  // denied the write tool but ran an arbitrary `touch` through the shell
-  // (probed on 1.0.62). read-only still writes ~/.grok (sessions) and temp dirs.
-  if (sandbox === "read-only") args.push("--sandbox", "read-only", "--permission-mode", "default");
-  else if (sandbox === "workspace-write") args.push("--sandbox", "workspace", "--permission-mode", "acceptEdits");
-  else if (sandbox === "danger-full-access") args.push("--sandbox", "off", "--always-approve");
-  if (resumeSid) args.push("--resume", resumeSid);
-  if (o.model) args.push("-m", o.model);
-  if (o.effort) args.push("--effort", o.effort);
-  return args;
-}
-
-function claudeArgs(o, resumeSid, schemaText) {
+function claudeArgs(o, model, resumeSid, schemaText, sandbox, cwd) {
   // stream-json (not plain json) so events land incrementally: the session id
   // arrives in the init event within seconds, and the write-through record
   // survives a killed wrapper. Plain json would buffer everything to the end.
-  const args = ["-p", "--output-format", "stream-json", "--verbose"];
-  if (!o.raw) args.push("--append-system-prompt", PREAMBLE);
+  const args = ["-p", "--output-format", "stream-json", "--verbose", "--append-system-prompt", PREAMBLE];
   // Native structured output: the result event carries the parsed object in
   // `structured_output` alongside the JSON text in `result`.
   if (schemaText) args.push("--json-schema", schemaText);
-  if (o.sandbox === "read-only") {
-    // --permission-mode default overrides the user's defaultMode (even
-    // bypass); the allowlist restricts to read tools. A real read-only.
-    args.push("--permission-mode", "default", "--allowedTools", ALLOWED_TOOLS);
-  } else if (o.sandbox === "danger-full-access") {
+  // --permission-mode overrides the user's defaultMode. In -p a call that
+  // would need approval is denied, never left waiting.
+  if (sandbox === "read-only")
+    args.push("--permission-mode", "manual", "--disallowedTools", CLAUDE_FILE_TOOLS,
+      "--setting-sources", "", "--settings", claudeSandbox(sandbox, cwd));
+  else if (sandbox === "workspace-write")
+    args.push("--permission-mode", "acceptEdits", "--setting-sources", "", "--settings", claudeSandbox(sandbox, cwd));
+  else if (sandbox === "danger-full-access")
     args.push("--dangerously-skip-permissions");
-  }
-  // else: no permission flag — the consulted Claude uses the user's own
-  // ~/.claude config, mirroring Codex deferring to config.toml.
-
   if (resumeSid) args.push("--resume", resumeSid);
-  if (o.model) args.push("--model", o.model);
+  if (model) args.push("--model", model);
   if (o.effort) args.push("--effort", o.effort);
   return args; // prompt goes on stdin
 }
 
-function codexArgs(o, resumeSid, replyPath, schemaPath) {
+function codexArgs(o, model, resumeSid, replyPath, schemaPath, sandbox) {
   const args = ["exec"];
   if (resumeSid) args.push("resume", resumeSid);
   // -o is the authoritative reply text (the JSONL item schema is less stable);
   // --json is the session-id channel (thread.started arrives up front).
   args.push("--json", "-o", replyPath, "--skip-git-repo-check");
   // Native structured output; codex wants a FILE, so the schema is materialized
-  // in the job dir (works on `exec resume` too — verified on codex 0.142).
+  // in the job dir (works on `exec resume` too).
   if (schemaPath) args.push("--output-schema", schemaPath);
   // Force the sentinel into the consulted Codex's shell env (see header).
   args.push("-c", 'shell_environment_policy.set.GASLAMP_NESTED="1"');
   // `codex exec resume` has no -s flag; -c sandbox_mode works on both forms.
-  if (o.sandbox) args.push("-c", `sandbox_mode="${o.sandbox}"`);
-  if (o.model) args.push("-m", o.model);
+  if (sandbox) args.push("-c", `sandbox_mode="${sandbox}"`);
+  if (model) args.push("-m", model);
   if (o.effort) args.push("-c", `model_reasoning_effort="${o.effort}"`);
   args.push("-"); // prompt on stdin, always — no argv-size or quoting traps
   return args;
 }
 
+function geminiArgs(o, model, resumeSid, schemaText) {
+  // Print mode's -p takes the prompt as an argument; streaming input reads it
+  // from stdin as one {"event":"user"} line and ends the session at EOF, which
+  // keeps the prompt-on-stdin contract.
+  //
+  // agy's own permission layer is always skipped: the fence is Seatbelt (or
+  // nothing, under danger-full-access). Under its default request-review
+  // preset print mode soft-denies writes and shell commands and ends the turn
+  // empty, and users' interactive settings vary; skipping it makes every mode
+  // behave the same whatever those settings say.
+  const args = ["--input-format", "stream-json", "--output-format", "stream-json", "--model", model, "--dangerously-skip-permissions"];
+  if (o.effort) args.push("--effort", o.effort);
+  if (resumeSid) args.push("--conversation", resumeSid);
+  if (schemaText) args.push("--json-schema", schemaText);
+  return args;
+}
+
+function grokArgs(o, model, resumeSid, schemaText, sandbox) {
+  // streaming-messages-json is Anthropic's stream-json shape (system init with
+  // session_id first, result last) — the claude parser reads it. --rules
+  // appends to the system prompt: claude's --append-system-prompt.
+  const args = ["--prompt-file", "/dev/stdin", "--output-format", "streaming-messages-json", "--no-auto-update", "--rules", PREAMBLE];
+  if (schemaText) args.push("--json-schema", schemaText);
+  // The kernel sandbox is the fence, and grok's permission layer is skipped,
+  // as agy's is: headless it denies whatever isn't pre-approved (under
+  // acceptEdits, write and run_terminal_command alike), and in another
+  // version it let `touch` through under default. It is no guarantee.
+  args.push("--sandbox", { "read-only": "read-only", "workspace-write": "workspace", "danger-full-access": "off" }[sandbox], "--always-approve");
+  if (resumeSid) args.push("--resume", resumeSid);
+  if (model) args.push("-m", model);
+  if (o.effort) args.push("--effort", o.effort);
+  return args;
+}
+
 export function runConsult(backend, argv) {
   const o = parseArgs(argv);
 
   // ---- preflights -----------------------------------------------------------
   if (process.env.GASLAMP_NESTED && !ALLOW_RECURSION)
     die(3, "this agent is itself a gaslamp consult — a consult is one hop. (GASLAMP_ALLOW_RECURSION=1 opts out.)");
-  // Codex sets this in sandboxed shell envs when network is off. Neither
-  // backend could phone home, so refuse pointedly instead of timing out.
+  // Codex sets this in sandboxed shell envs when network is off. No backend
+  // could phone home, so refuse pointedly instead of timing out.
   if (process.env.CODEX_SANDBOX_NETWORK_DISABLED === "1")
-    die(4, "network is disabled in this sandbox (CODEX_SANDBOX_NETWORK_DISABLED=1) — a consulted agent could reach neither Anthropic nor OpenAI. Re-run with network access (escalated permissions, or sandbox_mode/network_access in ~/.codex/config.toml).");
-
-  const facts = BACKENDS[backend];
-  if (o.sandbox && !facts.sandboxes.includes(o.sandbox)) {
-    die(2, backend === "claude" && o.sandbox === "workspace-write"
-      ? "claude has no filesystem sandbox, so there is no honest workspace-write — use read-only, danger-full-access, or omit --sandbox to defer to the user's own config"
-      : backend === "gemini" && o.sandbox === "workspace-write"
-        ? "Antigravity print mode denies file writes unless every permission is skipped, so there is no honest workspace-write — use read-only (the default) or danger-full-access"
-        : `bad --sandbox "${o.sandbox}" (${backend}: ${facts.sandboxes.join(" | ")})`);
-  }
-  // claude/codex: omitted = the consulted agent's own config. gemini/grok:
-  // omitted = read-only (BACKENDS.defaultSandbox).
-  const sandbox = o.sandbox ?? facts.defaultSandbox;
+    die(4, "network is disabled in this sandbox (CODEX_SANDBOX_NETWORK_DISABLED=1), so a consulted agent could not reach its provider. The consult was not started.");
+
+  const sandbox = o.sandbox ?? BACKENDS[backend].defaultSandbox;
+  if (sandbox && !SANDBOXES.includes(sandbox)) die(2, `bad --sandbox "${sandbox}" (${SANDBOXES.join(" | ")})`);
+  if (backend === "gemini" && sandbox !== "danger-full-access" && !seatbelt())
+    die(4, `a ${sandbox} gemini consult runs agy under macOS Seatbelt, which this environment doesn't have ` +
+      "(Linux, or inside another sandbox), so the consult was not started. " +
+      "--sandbox danger-full-access would run it with no filesystem guarantee.");
+  const cwd = resolveCwd(o.cwd);
 
   // ---- prompt (argv, or stdin via `-` / piped-with-no-arg) -------------------
   let prompt = o.prompt;
@@ -527,9 +485,9 @@ export function runConsult(backend, argv) {
     prompt = readStdin();
   } else if (!process.stdin.isTTY) {
     // Prompt on argv AND piped stdin: stdin is evidence — the diff, the log,
-    // the failing output — appended as a <stdin> block. This is codex exec's
-    // own convention for exactly this case, applied uniformly to both backends.
-    const evidence = readStdin();
+    // the failing output — appended as a <stdin> block (codex exec's own
+    // convention for exactly this case, applied to every backend).
+    const evidence = readStdin({ evidence: true });
     if (evidence.trim())
       prompt += "\n\n<stdin>\n" + (evidence.endsWith("\n") ? evidence : evidence + "\n") + "</stdin>";
   }
@@ -537,129 +495,91 @@ export function runConsult(backend, argv) {
 
   const schemaText = o.schema ? resolveSchema(o.schema) : null;
 
-  // ---- thread → resume + lock -------------------------------------------------
-  // A thread is resume-if-bound: an existing pointer supplies the session id
-  // (and the session lock below serializes concurrent consults on it); an
-  // unbound name runs fresh and binds when the backend reports its session.
-  // (Two concurrent consults racing to BIND the same new name is a caller bug;
-  // the loser's session stays reachable through its job record.)
-  if (o.thread) {
-    if (o.resume) die(2, "--thread and --resume are exclusive — a thread IS a resume handle");
-    if (!validThreadName(o.thread)) die(2, `bad thread name "${o.thread}" (letters/digits then letters, digits, . _ -; max 64)`);
-  }
-  const thread = o.thread ? readThread(backend, o.thread) : null;
-  const resumeSid = o.resume ? resolveResume(backend, o.resume) : thread?.sessionId ?? null;
-  // Everything that runs the harness synchronously (the tier lookup can take
-  // agy ~2s, the version probe a few ms) happens before the lock is taken and
-  // the record exists: no signal handler is installed yet, and a kill here
-  // must not strand either.
+  // ---- the conversation: a thread name, or a prior job ----------------------
+  if (o.thread && o.resume) die(2, "--thread and --resume are exclusive — a thread IS a continuation");
+  if (o.thread && !validThreadName(o.thread))
+    die(2, `bad thread name "${o.thread}" (letters/digits, then letters, digits, . _ -; max 64)`);
+  const prior = o.resume ? resolvePrior(backend, o.resume) : null;
+
+  // Everything that runs the harness synchronously (agy's tier lookup takes
+  // ~2s) happens before the lock is taken and the record exists: no signal
+  // handler is installed yet.
   const bin = resolveBin(backend);
   const env = backendEnv(backend);
-  let grants = null, agySettings = null, wrap = null;
-  let enforcement = sandbox === "read-only" ? {
-    claude: "permission-mode default + read allowlist", codex: "codex read-only sandbox",
-    grok: "grok --sandbox read-only (kernel)", gemini: null,
-  }[backend] : null;
-  if (backend === "gemini" && sandbox === "read-only") {
-    const state = agyPermissionState();
-    grants = state.grants.map((g) => g.rule);
-    agySettings = [...state.presets.map((p) => `toolPermission ${p.preset}`), ...state.modes.map((m) => `agentMode ${m.mode}`)];
-    const lax = state.presets.filter((p) => AGY_LAX_PRESETS.has(p.preset));
-    const modes = state.modes.filter((m) => AGY_LAX_MODES.has(m.mode));
-    const open = state.grants.filter((g) => !AGY_READ_GRANT.test(g.rule));
-    if (useSeatbelt()) {
-      wrap = [SANDBOX_EXEC, "-p", agySeatbeltProfile()];
-      enforcement = "seatbelt (sandbox-exec)";
-    } else if (lax.length || modes.length || open.length)
-      die(1, "Antigravity's settings would let a read-only consult act unprompted — " +
-        [...lax.map((p) => `toolPermission "${p.preset}" (${p.file})`),
-          ...modes.map((m) => `agentMode "${m.mode}" (${m.file})`),
-          ...open.map((g) => `${g.rule} (${g.file})`)].join(", ") +
-        ". Remove them, or pass --sandbox danger-full-access to consult without the guarantee.");
-    else enforcement = "agy request-review preset";
-  }
   const model = resolveModel(backend, o.model, bin, env);
-  const version = harnessVersion(bin, env);
   const id = newJobId(backend);
-  let lockPath = resumeSid ? acquireLock(backend, resumeSid, id) : null;
+
+  let lockPath = o.thread || prior
+    ? acquireLock(backend, o.thread ? `thread-${o.thread}` : `session-${prior.sessionId}`, id)
+    : null;
   const releaseLock = () => { if (lockPath) { try { unlinkSync(lockPath); } catch {} lockPath = null; } };
+  const refuse = (code, msg) => { releaseLock(); die(code, msg); };
+
+  // Read under the lock, so a consult that just finished on this thread has
+  // re-pointed it already. A conversation runs where it started: reusing a
+  // name from another directory is far more often a collision than a
+  // continuation, and Claude Code keeps its sessions per project directory.
+  const thread = o.thread ? readThread(backend, o.thread) : null;
+  const home = thread ?? prior;
+  if (home) {
+    const what = o.thread ? `thread ${o.thread}` : `job ${prior.id}`;
+    // 2.x thread pointers recorded no directory (their last job did), and 2.x
+    // jobs stored --cwd as typed. A stored path is canonicalized; one that
+    // can't be (relative, or missing) is adopted only when --cwd names it.
+    const stored = canonicalDir(home.cwd ?? (thread ? readMeta(thread.lastJobId)?.cwd : null));
+    if (stored == null && o.cwd == null)
+      refuse(2, `${what} predates gaslamp 3 and recorded no usable directory. Pass --cwd <dir> to continue it there.`);
+    if (stored != null && stored !== cwd)
+      refuse(2, `${what} lives in ${stored}. Continue it with --cwd ${shellQuote(stored)}` +
+        `${o.thread ? ", or give this conversation its own name" : ""}.`);
+  }
+  const resumeSid = prior?.sessionId ?? thread?.sessionId ?? null;
 
   // ---- job record ------------------------------------------------------------
   const dir = jobDir(id);
   const meta = {
-    id, backend, status: "running",
-    sessionId: resumeSid, resumedFrom: resumeSid, thread: o.thread ?? null,
-    label: o.label ?? null,
-    model, sandbox, effort: o.effort ?? null,
-    cwd: o.cwd || process.cwd(),
-    pid: process.pid, childPid: null, exitCode: null, signal: null,
-    startedAt: new Date().toISOString(), endedAt: null,
-    promptChars: prompt.length, schema: !!schemaText, raw: !!o.raw,
-    provenance: {
-      provider: facts.provider, harness: facts.harness, harnessVersion: version,
-      // `model` is what the backend reported answering with; unknown until it
-      // does (codex never reports one in its event stream).
-      model: "unknown", modelRequested: o.model ?? null, modelPassed: model,
-      permissionMode: null, sandbox, enforcement,
-      ...(grants ? { grants, agySettings } : {}),
-      systemPrompt: { known: false },
-      retrieval: { available: null, used: [] },
-      gaslamp: { version: PKG_VERSION, revision: gaslampRevision() },
-    },
+    id, backend, status: "running", error: null,
+    sessionId: resumeSid, thread: o.thread ?? null, cwd,
+    pid: process.pid, startedAt: new Date().toISOString(), endedAt: null,
   };
   createJob(meta, prompt);
 
   // ---- spawn -----------------------------------------------------------------
   const replyPath = join(dir, "reply.md");
-  // schema.json lands in the job dir for BOTH backends: codex needs the file,
-  // and the record should show what was actually sent — for codex that is the
-  // strictified form (see strictify above), for claude the schema as written.
+  // schema.json records what was actually sent: for codex the strictified
+  // form, for the others the schema as written. codex also needs it as a file.
   const schemaPath = schemaText ? join(dir, "schema.json") : null;
   if (schemaPath) {
-    const sent = backend === "codex"
-      ? JSON.stringify(strictify(JSON.parse(schemaText)), null, 2)
-      : schemaText;
+    const sent = backend === "codex" ? JSON.stringify(strictify(JSON.parse(schemaText)), null, 2) : schemaText;
     writeFileSync(schemaPath, sent.endsWith("\n") ? sent : sent + "\n");
   }
-  const args = backend === "claude" ? claudeArgs({ ...o, model }, resumeSid, schemaText)
-    : backend === "codex" ? codexArgs({ ...o, model }, resumeSid, replyPath, schemaPath)
-    : backend === "gemini" ? geminiArgs(o, resumeSid, schemaText, model, sandbox)
-    : grokArgs(o, resumeSid, schemaText, sandbox);
-
-  // sandbox-exec applies the profile and execs the harness in place (same
-  // pid), so the group, the signals and the process name are agy's own.
+  const args = backend === "claude" ? claudeArgs(o, model, resumeSid, schemaText, sandbox, cwd)
+    : backend === "codex" ? codexArgs(o, model, resumeSid, replyPath, schemaPath, sandbox)
+    : backend === "gemini" ? geminiArgs(o, model, resumeSid, schemaText)
+    : grokArgs(o, model, resumeSid, schemaText, sandbox);
+
+  const spool = join(tmpdir(), `gaslamp-${id}.end-conversation.jsonl`);
+  env.END_CONVERSATION_SPOOL = spool;
+  // sandbox-exec applies the profile and execs agy in place (same pid), so the
+  // group, the signals and the process name are agy's own.
+  const wrap = backend === "gemini" && sandbox !== "danger-full-access" ? [SANDBOX_EXEC, "-p", agyProfile(sandbox, cwd)] : null;
   const [cmd, argv0] = wrap ? [wrap[0], [...wrap.slice(1), bin, ...args]] : [bin, args];
   log("spawn", cmd, JSON.stringify(argv0));
   // detached → the child leads its own process group, so signal forwarding can
   // kill the whole tree (backend + whatever it spawned), never just the top.
-  const child = spawn(cmd, argv0, { cwd: meta.cwd, env, stdio: ["pipe", "pipe", "pipe"], detached: true });
-  meta.childPid = child.pid ?? null;
-  writeMeta(meta);
+  const child = spawn(cmd, argv0, { cwd, env, stdio: ["pipe", "pipe", "pipe"], detached: true });
   const killGroup = (sig) => {
     try { process.kill(-child.pid, sig); }
     catch { try { child.kill(sig); } catch {} }
   };
 
-  writeSync(2, `[gaslamp] → ${backend} · job ${id} · watch: tail -f ${join(dir, "events.jsonl")}\n`);
-
-  // Early machine receipt for the fleet runner (gated on GASLAMP_FLEET, which the
-  // fleet sets on its children): the parent records each child's job id the
-  // moment it spawns — durable even if the fleet is killed mid-run — without
-  // scraping the human stderr breadcrumb. The lone --json line is the envelope
-  // (it carries `backend`/`status`); this receipt carries `type:"started"`.
-  if (o.json && process.env.GASLAMP_FLEET === "1")
-    process.stdout.write(JSON.stringify({ type: "started", jobId: id, childPid: meta.childPid }) + "\n");
+  const conversation = o.thread ? ` · ${thread ? "continuing" : "new"} thread ${o.thread}` : prior ? ` · continuing ${prior.id}` : "";
+  writeSync(2, `[gaslamp] → ${backend} · job ${id}${conversation}\n`);
 
   child.stdin.on("error", () => {}); // EPIPE if the child dies before reading
-  // prompt.md records the CALLER's content; the preamble is mechanism, applied
-  // here at the spawn boundary (codex has no --append-system-prompt analog).
-  const preamble = o.raw ? "" : "<gaslamp_consult>\n" + PREAMBLE + "\n</gaslamp_consult>\n\n";
-  if (backend === "gemini") {
-    child.stdin.write(JSON.stringify({ event: "user", message: { content: preamble + prompt } }) + "\n");
-  } else {
-    if (backend === "codex") child.stdin.write(preamble);
-    child.stdin.write(prompt);
-  }
+  const preamble = "<gaslamp_consult>\n" + PREAMBLE + "\n</gaslamp_consult>\n\n";
+  if (backend === "gemini") child.stdin.write(JSON.stringify({ event: "user", message: { content: preamble + prompt } }) + "\n");
+  else child.stdin.write(backend === "codex" ? preamble + prompt : prompt);
   child.stdin.end();
 
   const events = createWriteStream(join(dir, "events.jsonl"), { flags: "a" });
@@ -669,40 +589,32 @@ export function runConsult(backend, argv) {
   child.stdout.setEncoding("utf8");
   child.stderr.setEncoding("utf8");
 
-  // A consult the harness can't start honestly is stopped, not left to run.
+  // A consult the harness can't run honestly is stopped, not left to run.
   let refusal = null;
-  const refuse = (why) => {
+  const stop = (why) => {
     if (refusal) return;
     refusal = why;
-    writeSync(2, `[gaslamp] ${why}\n`);
     killGroup("SIGTERM");
     setTimeout(() => killGroup("SIGKILL"), 2000).unref();
   };
   // `agy` without a cached sign-in prints a Google sign-in URL and waits for a
-  // pasted code forever in print mode (1.4.3; its docs say it exits). Only
-  // what the harness prints before its init event counts — stderr and any
-  // non-JSON stdout, accumulated across chunks — never protocol content,
-  // where a reply may well say "authentication required".
+  // pasted code forever in print mode. Only what it prints before its init
+  // event counts — stderr and any non-JSON stdout, accumulated across chunks —
+  // never protocol content, where a reply may well say "authentication
+  // required".
   let initSeen = false, preInit = "";
   const AUTH_NOTICE = /Authentication required/i;
-  const notSignedIn = () => refuse("agy is not signed in — run `agy` once in a terminal to sign in, then retry");
+  const notSignedIn = () => stop("agy is not signed in — run `agy` once in a terminal to sign in, then retry");
   const watchAuth = (text) => {
     if (backend !== "gemini" || initSeen) return;
     preInit = (preInit + text).slice(-4096);
     if (AUTH_NOTICE.test(preInit)) notSignedIn();
   };
-  child.stderr.on("data", (d) => { errlog.write(d); watchAuth(d); });
+  let errTail = "";
+  child.stderr.on("data", (d) => { errlog.write(d); errTail = (errTail + d).slice(-4096); watchAuth(d); });
 
-  const prov = meta.provenance;
-  const noteTool = (name) => {
-    if (name && RETRIEVAL_TOOLS[backend].test(name) && !prov.retrieval.used.includes(name))
-      prov.retrieval.used.push(name);
-  };
-
-  // ---- event stream: write through; fish out session id + reply --------------
-  let buf = "", resultText = null, resultErr = false, lastAgentText = null, structuredOut;
-  let initValidated = false; // gemini read-only: the harness confirmed its preset
-  const usageAcc = { input: 0, output: 0 }; // codex and agy report per turn/step; sum them
+  // ---- event stream: write through; fish out the session id and the reply -------
+  let buf = "", resultText = null, resultErr = false, backendError = null, lastAgentText = null, structuredOut;
   const claudeShaped = backend === "claude" || backend === "grok";
   const handleLine = (raw) => {
     const line = raw.trim();
@@ -714,93 +626,29 @@ export function runConsult(backend, argv) {
       : (j.event === "init" || j.event === "result" ? (j.conversation_id ?? j.result?.conversation_id) : null);
     if (sid) {
       if (meta.sessionId !== sid) { meta.sessionId = sid; writeMeta(meta); } // early: the recovery handle
-      // Bind/re-point the thread the moment the session is known — claude
+      // Bind or re-point the thread the moment the session is known — claude
       // resumes fork a fresh session id, and the pointer must chase it.
       if (o.thread) writeThread({
-        backend, name: o.thread, sessionId: sid, lastJobId: id,
+        backend, name: o.thread, sessionId: sid, cwd, lastJobId: id,
         createdAt: thread?.createdAt ?? meta.startedAt, updatedAt: new Date().toISOString(),
       });
     }
-    if (claudeShaped && j.type === "system" && j.subtype === "init") {
-      if (j.model) prov.model = j.model;
-      prov.permissionMode = j.permissionMode ?? null;
-      if (Array.isArray(j.tools)) prov.retrieval.available = j.tools.filter((t) => RETRIEVAL_TOOLS[backend].test(t));
-      // Grok's backend X search is never listed but is there (see toolKey).
-      if (backend === "grok" && prov.retrieval.available) prov.retrieval.available.push("x_search");
-      writeMeta(meta);
-    } else if (claudeShaped && j.type === "assistant") {
-      for (const c of j.message?.content ?? []) {
-        if (c.type === "tool_use" || c.type === "server_tool_use") noteTool(toolKey(c.name));
-        if (c.type === "tool_use" && c.name === "use_tool") noteTool(c.input?.tool_name);
-      }
-    } else if (backend === "grok" && j.type === "user") {
-      for (const c of j.message?.content ?? []) {
-        if (c.type !== "tool_result" || typeof c.content !== "string") continue;
-        let r; try { r = JSON.parse(c.content); } catch { continue; }
-        if (typeof r?.name === "string") noteTool(r.name);
-      }
+    if (claudeShaped && j.type === "result") {
+      resultText = j.result ?? j.content ?? null;
+      resultErr = !!j.is_error;
+      if (j.structured_output !== undefined) structuredOut = j.structured_output;
+    } else if (backend === "codex" && j.type === "item.completed") {
+      // Tolerant fallback if -o never lands (schema drift across versions).
+      const item = j.item ?? {};
+      if ((item.type ?? item.item_type) === "agent_message" && typeof item.text === "string") lastAgentText = item.text;
     } else if (backend === "gemini" && j.event === "init") {
       initSeen = true;
-      const init = j.init ?? {};
-      if (init.model) prov.model = init.model;
-      prov.permissionMode = init.permission_mode ?? null;
-      // The browser toolset is ~20 tools; one entry says the same thing.
-      if (Array.isArray(init.tools)) prov.retrieval.available = [...new Set(init.tools
-        .filter((t) => RETRIEVAL_TOOLS.gemini.test(t)).map((t) => t.startsWith("browser_") ? "browser_*" : t))];
-      writeMeta(meta);
-      // Without Seatbelt, read-only rests on Antigravity's request-review (or
-      // strict) preset; a laxer one would let the consult act unprompted.
-      if (sandbox === "read-only" && !wrap && !["request-review", "strict"].includes(prov.permissionMode))
-        refuse(`agy reports permission mode "${prov.permissionMode}", not request-review — refusing a read-only consult it would not enforce`);
-      else initValidated = true;
-    } else if (backend === "gemini" && j.event === "step_update") {
-      const st = j.step_update ?? {};
-      if (st.step_type === "tool" && st.state === "DONE") noteTool(st.tool_name);
-      if (st.step_type === "agent_response" && st.state === "DONE" && st.usage) {
-        // Per step; the result event's usage is cumulative over the whole
-        // conversation, resumes included. input_tokens EXCLUDES cache reads
-        // (a resumed turn: 2376 input + 5118 cache read ≈ the full prompt),
-        // so both count, as for claude.
-        usageAcc.input += (st.usage.input_tokens ?? 0) + (st.usage.cache_read_tokens ?? 0);
-        usageAcc.output += st.usage.output_tokens ?? 0;
-        meta.usage = { inputTokens: usageAcc.input, outputTokens: usageAcc.output, costUsd: null };
-      }
     } else if (backend === "gemini" && j.event === "result") {
       const r = j.result ?? {};
       resultText = r.response ?? null;
       resultErr = r.status !== "SUCCESS";
-      if (r.error) writeSync(2, `[gaslamp] agy: ${r.error}\n`);
+      if (r.error) backendError = String(r.error);
       if (r.structured_output !== undefined) structuredOut = r.structured_output;
-    } else if (claudeShaped && j.type === "result") {
-      resultText = j.result ?? j.content ?? null;
-      resultErr = !!j.is_error;
-      if (j.structured_output !== undefined) structuredOut = j.structured_output;
-      // Grok bills by the model it actually ran (grok-4.7-build behind a
-      // reported grok-4.7), which only shows up here.
-      if (j.modelUsage) prov.billedModels = Object.keys(j.modelUsage);
-      // Fleets are quota-shaped, so quota should be visible: inputTokens
-      // counts cache reads/writes too (that's what the meter meters).
-      if (j.usage) meta.usage = {
-        inputTokens: (j.usage.input_tokens ?? 0) + (j.usage.cache_read_input_tokens ?? 0)
-          + (j.usage.cache_creation_input_tokens ?? 0),
-        outputTokens: j.usage.output_tokens ?? 0,
-        costUsd: j.total_cost_usd ?? null,
-      };
-    } else if (backend === "codex" && j.type === "turn.completed" && j.usage) {
-      // OpenAI semantics differ from Anthropic's: input_tokens already
-      // INCLUDES the cached portion (cached_input_tokens is a subset, and
-      // codex's own total_tokens = input + output), so adding it again
-      // double-counts cache hits.
-      usageAcc.input += j.usage.input_tokens ?? 0;
-      usageAcc.output += j.usage.output_tokens ?? 0;
-      meta.usage = { inputTokens: usageAcc.input, outputTokens: usageAcc.output, costUsd: null };
-    } else if (backend === "codex" && (j.type === "item.started" || j.type === "item.completed")) {
-      const item = j.item ?? {};
-      const t = item.type ?? item.item_type;
-      if (t === "web_search") noteTool("web_search");
-      // Tolerant fallback if -o never lands (schema drift across versions).
-      if (j.type === "item.completed" && t === "agent_message" && typeof item.text === "string")
-        lastAgentText = item.text;
     }
   };
   child.stdout.on("data", (d) => {
@@ -824,33 +672,31 @@ export function runConsult(backend, argv) {
     process.on(sig, () => {
       if (killedBy) return killGroup("SIGKILL");
       killedBy = sig;
-      writeSync(2, `[gaslamp] ${sig} — stopping job ${id}; the session survives, recover with --resume\n`);
+      writeSync(2, `[gaslamp] ${sig} — stopping job ${id}; the conversation survives and can be continued\n`);
       killGroup(sig);
       setTimeout(() => killGroup("SIGKILL"), 2000).unref();
     });
   }
 
-  // One terminal state, published once: the reply is written and validated
-  // before meta.json leaves "running" and the session lock is released.
-  const finalize = (status, code, reply) => {
+  // One terminal state, published once: the reply and its data are written
+  // before meta.json leaves "running" and the lock is released.
+  const finalize = (status, error, reply, data) => {
     if (meta.status !== "running") return;
     if (reply != null) writeFileSync(replyPath, reply);
-    if (backend === "grok" && meta.sessionId) prov.systemPrompt = grokSystemPrompt(meta.sessionId, meta.cwd);
+    if (data != null) writeFileSync(join(dir, "data.json"), JSON.stringify(data) + "\n");
     meta.status = status;
-    meta.exitCode = code;
-    meta.signal = killedBy;
+    meta.error = error;
     meta.endedAt = new Date().toISOString();
-    if (refusal) meta.refusal = refusal;
     writeMeta(meta);
     releaseLock();
   };
 
   child.on("error", (e) => {
-    finalize("failed", null);
+    finalize("failed", `failed to launch ${bin}: ${e.message}`);
     die(1, `failed to launch ${bin}: ${e.message}`);
   });
 
-  child.on("close", (code) => {
+  child.on("close", (code, signal) => {
     if (buf.trim()) handleLine(buf); // a final event without its newline
     buf = "";
     events.end();
@@ -859,61 +705,51 @@ export function runConsult(backend, argv) {
     // its inherited pipes, so `close` could fire) may not be: on any kill
     // path, finish the group now rather than trusting the unref'd timer.
     if (killedBy || refusal) killGroup("SIGKILL");
-
-    // The deliverable. codex writes it via -o (agent_message as fallback);
-    // the others hand it over in the result event — as text, or for a
-    // --schema consult possibly only as structured_output.
-    let reply;
-    if (backend === "codex") reply = existsSync(replyPath) ? readFileSync(replyPath, "utf8") : lastAgentText;
-    else reply = resultText != null && String(resultText).trim() ? String(resultText)
+    const endedSession = fileOffRamp(spool, id);
+
+    // The deliverable. codex writes it via -o (agent_message as fallback); the
+    // others hand it over in the result event — as text, or for a --schema
+    // consult possibly only as structured_output.
+    const codexWrote = backend === "codex" && existsSync(replyPath);
+    const reply = backend === "codex" ? (codexWrote ? readFileSync(replyPath, "utf8") : lastAgentText)
+      : resultText != null && String(resultText).trim() ? String(resultText)
       : structuredOut !== undefined ? JSON.stringify(structuredOut) : resultText;
 
-    let status = killedBy ? "killed" : (refusal || code !== 0 || resultErr) ? "failed" : "done";
-    if (status === "done" && backend === "gemini" && sandbox === "read-only" && !wrap && !initValidated) {
+    const lastErrLine = errTail.trim().split("\n").at(-1)?.trim().slice(0, 300);
+    let status = "done", error = null;
+    if (killedBy) { status = "killed"; error = `killed by ${killedBy}`; }
+    else if (refusal) { status = "failed"; error = refusal; }
+    else if (code !== 0) {
       status = "failed";
-      refusal = "agy reported no init event, so its permission mode — the read-only guarantee — went unconfirmed";
-      writeSync(2, `[gaslamp] ${refusal}\n`);
-    }
-    // agy and grok report success for a turn that ended on a denied tool call
-    // with nothing to say (agy: "no output produced", on stderr). An empty
-    // reply is not a deliverable.
-    if (status === "done" && (backend === "gemini" || backend === "grok") && !reply?.trim()) {
+      error = endedSession ? "the consultee ended its session (end_session)"
+        : lastErrLine || `${backend} exited with ${code ?? signal}`;
+    } else if (resultErr) {
+      status = "failed";
+      error = backendError || String(resultText ?? "").trim().split("\n")[0].slice(0, 300) || `${backend} reported an error`;
+    } else if (!reply?.trim()) {
+      // A turn that ended on a denied tool call, or with nothing to say, is
+      // not a deliverable, whatever the harness reported.
       status = "failed";
-      writeSync(2, `[gaslamp] ${backend} returned an empty reply — marking the consult failed\n`);
+      error = "empty reply";
     }
-    // A schema consult's contract is a typed reply: parse it (claude hands the
-    // object over in the result event; codex's reply.md IS the JSON), and an
-    // unparseable reply is a FAILED consult, not a quiet string.
+    // A schema consult's contract is a typed reply: an unparseable one is a
+    // FAILED consult, not a quiet string.
     let data = null;
     if (schemaText) {
       data = structuredOut ?? null;
       if (data == null && reply != null) { try { data = JSON.parse(reply); } catch {} }
-      if (data == null && status === "done") {
-        status = "failed";
-        writeSync(2, "[gaslamp] --schema was set but the reply is not valid JSON — marking the consult failed\n");
-      }
+      if (data == null && status === "done") { status = "failed"; error = "--schema was set but the reply is not valid JSON"; }
     }
-    finalize(status, code, backend === "codex" && existsSync(replyPath) ? null : reply);
-
-    if (o.json) {
-      process.stdout.write(JSON.stringify({
-        backend, jobId: id, sessionId: meta.sessionId, status: meta.status,
-        exitCode: code, content: reply ?? "",
-        ...(meta.label ? { label: meta.label } : {}),
-        ...(meta.usage ? { usage: meta.usage } : {}),
-        ...(schemaText ? { data } : {}),
-      }) + "\n");
-    } else {
+    finalize(status, error, codexWrote ? null : reply, data);
+
+    if (o.json) process.stdout.write(JSON.stringify(envelope(meta)) + "\n");
+    else {
       if (reply) process.stdout.write(reply.endsWith("\n") ? reply : reply + "\n");
-      if (!reply && meta.status !== "done") {
-        try { // surface the stderr tail so a failure isn't a shrug
-          const tail = readFileSync(join(dir, "stderr.log"), "utf8").trim().split("\n").slice(-6).join("\n");
-          if (tail) writeSync(2, tail + "\n");
-        } catch {}
-      }
+      // surface the backend's stderr tail so a failure isn't a shrug
+      else if (status !== "done" && errTail.trim()) writeSync(2, errTail.trim().split("\n").slice(-6).join("\n") + "\n");
       process.stdout.write(renderTrailer(meta) + "\n");
     }
     // exitCode (not exit()) so stdout drains; nothing else holds the loop open.
-    process.exitCode = meta.status === "done" ? 0 : 1;
+    process.exitCode = status === "done" ? 0 : 1;
   });
 }