balladeer 1.0.2 → 1.0.4

package/README.md

@@ -24,7 +24,8 @@ npx -y balladeer@latest setup
 
 Use `@latest` so a new session picks up the current command. Balladeer tells an older command when
 an update is available. Run `npx -y balladeer@latest setup --refresh` to update an existing
-repository connection and its instructions without replacing other tools' settings.
+repository connection and its instructions without replacing other tools' settings. For Codex, keep
+`--client codex` on the refresh command: `npx -y balladeer@latest setup --client codex --refresh`.
 
 Run it inside the repository you want to protect. It prints the boundary explanation below, then one
 step at a time, and it exits within seconds rather than blocking on anything.
@@ -40,10 +41,19 @@ already enrolled it connects this machine and writes its MCP configuration; when
 leaves the checkout unchanged and Balladeer shows that enforcement tracking is unavailable until an
 administrator connects a repository.
 
+## If setup cannot save its local connection
+
+Setup needs access to its private credential folder, normally `~/.config/balladeer`. If your
+terminal or coding agent blocks that access, allow the command to access the private folder named in
+the error, then repeat the same setup command with the same flags. Keep your existing saved
+credentials; do not delete the store or move it into your repository. Setup checks local write
+access before requesting a new browser pairing and reports a credential-store error separately from
+a connection failure.
+
 ## Requirements
 
-- Node 22 or newer. The command has no runtime dependencies at all: it uses Node builtins and
-  nothing else, so installing it pulls nothing else down.
+- Node 22 or newer. The command uses Node builtins and the pinned `@iarna/toml` parser to preserve
+  Codex project configuration safely.
 - `git`, and the GitHub CLI (`gh`) signed in to an account with push access to the repository. Your
   `gh` is how the command reads the repository's ids and opens the pull request that carries the
   workflow. Balladeer itself never holds a GitHub credential.
@@ -100,10 +110,10 @@ with your checkout and no Balladeer credential.
 It writes, locally and visibly: your credentials to `$XDG_CONFIG_HOME/balladeer/credentials.json`,
 or `~/.config/balladeer/credentials.json`, with mode 600 inside a directory with mode 700, and it
 refuses to write them anywhere a commit could pick them up; a `balladeer` entry in the repository's
-`.mcp.json`, added beside whatever is already there and never over the top of somebody else's entry;
-and a marker-fenced block in the repository's `CLAUDE.md` or `AGENTS.md`, replaced between the
-markers on each run and never outside them. A file whose markers are damaged is left alone and
-reported rather than appended to.
+`.mcp.json` for Claude, or `.codex/config.toml` with `--client codex`, preserving unrelated entries;
+and a marker-fenced block in `AGENTS.md` for Codex (the existing `CLAUDE.md` or `AGENTS.md` for
+Claude), replaced between the markers on each run and never outside them. A file whose markers are
+damaged is left alone and reported rather than appended to.
 
 It changes, on your GitHub repository and through your own `gh`: the repository variables the
 workflow reads, so the workflow file itself names no host by hand.
@@ -175,10 +185,35 @@ keep both.
 
 Both credentials are yours to end. Revoke the setup session or the coding agent's connection at any
 time in Balladeer; the setup session also ends on its own. Removing the `balladeer` entry from
-`.mcp.json` disconnects the agent locally, and deleting the credential store leaves this machine
-with nothing of yours on it.
+`.mcp.json` (Claude) or the managed table in `.codex/config.toml` (Codex) disconnects that project
+host locally after its next reload. Archiving the credential store removes this machine's saved
+Balladeer access without changing the server's records.
 
 ## Licence
 
 Apache License 2.0. The full text ships in this package as `LICENSE`, so you can read the rights you
 have from the copy on your own machine rather than taking a field in a manifest on trust.
+
+## Choose Codex or reconnect a workspace
+
+Run `npx -y balladeer@latest setup --client codex` in the repository to write its
+`.codex/config.toml` and the managed Balladeer block in `AGENTS.md`. This leaves Claude files,
+Claude Desktop, global Codex settings and unrelated MCP entries alone. The default client remains
+Claude; choose it explicitly with `--client claude`.
+
+Open and trust the project in Codex if appropriate, then start a fresh session. Project settings are
+ignored in untrusted projects; the setup connection probe does not prove the host loaded its tools.
+Check the actual Balladeer tools in that new session. Refresh only local Codex configuration with
+`setup --client codex --refresh`; no pairing or network request is made.
+
+To reconnect a different existing workspace, run:
+
+```sh
+npx -y balladeer@latest setup --client codex --existing --choose-workspace --wait
+```
+
+Choose the workspace and approve pairing in the browser. The current control plane's setup session
+is replaced only after a new pairing starts; all stored agent connections and other control planes
+remain. `--create-workspace "Team name"` likewise opens fresh pairing instead of silently reusing
+the previous workspace. It pre-fills the name; a human still creates the workspace. An interrupted
+command can resume the same pending choice by running it again with the same flags.

package/dist/cli.d.ts

@@ -6,6 +6,8 @@ type Parsed = Readonly<{
     wait: boolean;
     /** Repair the files a previous setup wrote, and do nothing else. */
     refresh: boolean;
+    client: "codex" | "claude" | undefined;
+    chooseWorkspace: boolean;
     /** `setup --force`: set up even though an earlier Balladeer is still installed. */
     force: boolean;
     /** `setup --existing`: connect only when this checkout is already enrolled. */

package/dist/cli.js

@@ -22,13 +22,24 @@ import { CLI_INVOCATION, CLI_VERSION, DEFAULT_CONTROL_PLANE } from "./wire.js";
 const USAGE = `balladeer ${CLI_VERSION}
 
   ${CLI_INVOCATION} setup [--repository owner/name]... [--json] [--wait] [--control-plane <url>]
-                  [--create-workspace <name>] [--refresh] [--force]
+                  [--create-workspace <name>] [--choose-workspace] [--refresh] [--force]
+                  [--client codex|claude]
                   [--existing]
                   [--claude-desktop | --no-claude-desktop]
       Pair this session, add this repository, connect this coding agent and CI,
       and report what a person still has to do. Name --repository more than once
       to add other repositories to the same workspace; each of those is added and
       nothing more, because connecting an agent and CI happens in a checkout.
+      --client codex writes project .codex/config.toml and AGENTS.md.
+      Trust the project in Codex and start a fresh session to load its tools.
+      --client claude (the default) preserves the Claude setup behavior.
+      For --refresh only, a lone managed Codex entry selects Codex automatically.
+      If both clients have Balladeer entries, choose --client explicitly.
+      --choose-workspace starts fresh browser pairing for this control plane,
+      preserving other saved agent connections. Use it with --existing to
+      reconnect an existing workspace. --create-workspace also starts fresh
+      pairing instead of reusing the previously selected workspace.
+
       --create-workspace carries a name to the approval page, where a person
       signs in and creates the workspace themselves; this command never creates
       one.
@@ -36,8 +47,8 @@ const USAGE = `balladeer ${CLI_VERSION}
       If this repository is already connected, connect this machine; otherwise
       continue in the browser while an administrator connects it.
       --refresh does one thing and talks to nobody: it rewrites this
-      repository's balladeer entry in .mcp.json, its Balladeer instructions
-      block and its Claude desktop chat entry to the current form, leaves every
+      selected client's project MCP entry and Balladeer instructions block
+      (and Claude desktop entry for Claude) to the current form, leaves every
       other entry in those files alone, and prints what it changed. Run it when
       Balladeer says a newer version is available.
       --force sets up even though an earlier Balladeer is still installed on
@@ -49,6 +60,10 @@ const USAGE = `balladeer ${CLI_VERSION}
       skipping quietly; --no-claude-desktop leaves that file alone entirely.
       Quit and reopen the app afterwards: it reads its configuration at startup.
 
+      If setup cannot access its private credential folder, allow this command
+      to access the folder named in the error, then repeat the same setup command.
+      Keep existing saved credentials; do not delete the store to retry.
+
   ${CLI_INVOCATION} repositories [--json] [--control-plane <url>]
       List the repositories this machine's GitHub account can see, marking the
       one you are standing in and the ones Balladeer already has.
@@ -151,6 +166,8 @@ export function parseArguments(argv) {
     let json = false;
     let wait = false;
     let refresh = false;
+    let client;
+    let chooseWorkspace = false;
     let force = false;
     let existingOnly = false;
     let claudeDesktop;
@@ -175,6 +192,7 @@ export function parseArguments(argv) {
         "--owner": "a membership id",
         "--create-workspace": "a workspace name",
         "--role": "contributor, viewer, or administrator",
+        "--client": "codex or claude",
         "--runner": "a path to the pinned runner's cli.js",
     };
     const value = (flag, inline) => {
@@ -207,6 +225,14 @@ export function parseArguments(argv) {
             wait = true;
         else if (name === "--refresh")
             refresh = true;
+        else if (name === "--choose-workspace")
+            chooseWorkspace = true;
+        else if (name === "--client") {
+            const requested = value("--client", inline);
+            if (requested !== "codex" && requested !== "claude")
+                throw new StoreError("usage", "--client needs codex or claude.");
+            client = requested;
+        }
         else if (name === "--force")
             force = true;
         else if (name === "--existing")
@@ -264,6 +290,16 @@ export function parseArguments(argv) {
     // `--refresh` repairs the files setup writes, so it belongs to setup and to
     // nothing else. Accepting it silently elsewhere would let somebody run
     // `status --refresh`, see no error, and believe their install was repaired.
+    if ((client !== undefined || chooseWorkspace || createWorkspace !== undefined) &&
+        command !== "setup") {
+        throw new StoreError("usage", "--client, --choose-workspace and --create-workspace belong to setup.");
+    }
+    if (refresh && (chooseWorkspace || createWorkspace !== undefined)) {
+        throw new StoreError("usage", "--refresh only repairs local files; omit it to choose or create a workspace.");
+    }
+    if (client === "codex" && claudeDesktop === true) {
+        throw new StoreError("usage", "--client codex leaves Claude configuration alone; run a separate --client claude setup to connect Claude Desktop.");
+    }
     if (refresh && command !== "setup") {
         throw new StoreError("usage", `--refresh belongs to setup: run \`${CLI_INVOCATION} setup --refresh\`.`);
     }
@@ -298,6 +334,8 @@ export function parseArguments(argv) {
         json,
         wait,
         refresh,
+        client,
+        chooseWorkspace,
         force,
         existingOnly,
         claudeDesktop,
@@ -347,6 +385,8 @@ async function dispatch(parsed, write) {
                 json: parsed.json,
                 wait: parsed.wait,
                 refresh: parsed.refresh,
+                chooseWorkspace: parsed.chooseWorkspace,
+                ...(parsed.client === undefined ? {} : { client: parsed.client }),
                 force: parsed.force,
                 existingOnly: parsed.existingOnly,
                 ...(parsed.claudeDesktop === undefined ? {} : { claudeDesktop: parsed.claudeDesktop }),

package/dist/codex-config.d.ts

@@ -0,0 +1,7 @@
+import { type MergeResult, type StdioMcpEntry } from "./mcp-config.js";
+export declare const CODEX_CONFIG_FILE = ".codex/config.toml";
+export declare function readCodexEntry(root: string): unknown;
+/** A client hint only when the complete file and our exact fenced table agree. */
+export declare function hasManagedCodexEntry(root: string, controlPlane: string): boolean;
+/** Only our fenced table is replaced; every unrelated byte and parsed setting survives. */
+export declare function mergeCodexConfig(root: string, entry: StdioMcpEntry, controlPlane: string): MergeResult;

package/dist/codex-config.js

@@ -0,0 +1,159 @@
+import { randomBytes } from "node:crypto";
+import { closeSync, existsSync, fsyncSync, lstatSync, mkdirSync, openSync, readFileSync, renameSync, unlinkSync, writeSync, } from "node:fs";
+import { join } from "node:path";
+import { parse } from "@iarna/toml";
+import { isOurEntry } from "./mcp-config.js";
+export const CODEX_CONFIG_FILE = ".codex/config.toml";
+const START = "# balladeer:mcp:start";
+const END = "# balladeer:mcp:end";
+function render(entry) {
+    return [
+        START,
+        "[mcp_servers.balladeer]",
+        `command = ${JSON.stringify(entry.command)}`,
+        `args = ${JSON.stringify(entry.args)}`,
+        ...(entry.env
+            ? [
+                "[mcp_servers.balladeer.env]",
+                ...Object.entries(entry.env).map(([key, value]) => `${JSON.stringify(key)} = ${JSON.stringify(value)}`),
+            ]
+            : []),
+        END,
+        "",
+    ].join("\n");
+}
+function withoutOurs(value) {
+    const copy = structuredClone(value);
+    const servers = copy.mcp_servers;
+    if (servers &&
+        typeof servers === "object" &&
+        !Array.isArray(servers) &&
+        !(servers instanceof Date)) {
+        delete servers.balladeer;
+        if (Object.keys(servers).length === 0)
+            delete copy.mcp_servers;
+    }
+    return JSON.stringify(copy);
+}
+export function readCodexEntry(root) {
+    try {
+        const config = parse(readFileSync(join(root, CODEX_CONFIG_FILE), "utf8"));
+        const servers = config.mcp_servers;
+        return servers &&
+            typeof servers === "object" &&
+            !Array.isArray(servers) &&
+            !(servers instanceof Date)
+            ? servers.balladeer
+            : undefined;
+    }
+    catch {
+        return undefined;
+    }
+}
+/** A client hint only when the complete file and our exact fenced table agree. */
+export function hasManagedCodexEntry(root, controlPlane) {
+    try {
+        const text = readFileSync(join(root, CODEX_CONFIG_FILE), "utf8");
+        const starts = [...text.matchAll(/^# balladeer:mcp:start\r?$/gm)];
+        const ends = [...text.matchAll(/^# balladeer:mcp:end\r?$/gm)];
+        if (starts.length !== 1 ||
+            ends.length !== 1 ||
+            !starts[0] ||
+            !ends[0] ||
+            starts[0].index >= ends[0].index)
+            return false;
+        const block = parse(text.slice(starts[0].index + starts[0][0].length, ends[0].index));
+        const servers = block.mcp_servers;
+        if (Object.keys(block).length !== 1 ||
+            !servers ||
+            typeof servers !== "object" ||
+            Array.isArray(servers) ||
+            servers instanceof Date ||
+            Object.keys(servers).length !== 1)
+            return false;
+        const own = readCodexEntry(root);
+        return (isOurEntry(own, controlPlane) && JSON.stringify(own) === JSON.stringify(servers.balladeer));
+    }
+    catch {
+        return false;
+    }
+}
+/** Only our fenced table is replaced; every unrelated byte and parsed setting survives. */
+export function mergeCodexConfig(root, entry, controlPlane) {
+    const block = render(entry);
+    const refused = (reason) => ({
+        kind: "refused",
+        reason: `${CODEX_CONFIG_FILE}: ${reason} No configuration was changed.`,
+        block,
+    });
+    const directory = join(root, ".codex");
+    const path = join(root, CODEX_CONFIG_FILE);
+    if (existsSync(directory) &&
+        (!lstatSync(directory).isDirectory() || lstatSync(directory).isSymbolicLink()))
+        return refused("the directory is not an ordinary project directory.");
+    if (existsSync(path) && (!lstatSync(path).isFile() || lstatSync(path).isSymbolicLink()))
+        return refused("the file is not an ordinary project file.");
+    let previous = "";
+    let parsed;
+    try {
+        if (existsSync(path))
+            previous = readFileSync(path, "utf8");
+        parsed = parse(previous);
+    }
+    catch {
+        return refused("could not read valid TOML; repair the file before running setup again.");
+    }
+    const starts = [...previous.matchAll(/^# balladeer:mcp:start\r?$/gm)];
+    const ends = [...previous.matchAll(/^# balladeer:mcp:end\r?$/gm)];
+    if (starts.length !== ends.length ||
+        starts.length > 1 ||
+        (starts[0] && ends[0] && starts[0].index >= ends[0].index))
+        return refused("the Balladeer markers are damaged; repair them before running setup again.");
+    const servers = parsed.mcp_servers;
+    if (servers !== undefined &&
+        (servers === null ||
+            typeof servers !== "object" ||
+            Array.isArray(servers) ||
+            servers instanceof Date))
+        return refused("mcp_servers must be a table.");
+    const own = servers && typeof servers === "object" && !Array.isArray(servers) && !(servers instanceof Date)
+        ? servers.balladeer
+        : undefined;
+    if (own !== undefined && (!starts[0] || !isOurEntry(own, controlPlane)))
+        return refused("an unmanaged or unrecognized balladeer table already exists. Review it and merge the supplied block yourself; setup will not replace it.");
+    if (starts[0] && own === undefined)
+        return refused("the marked block does not contain a Balladeer server; review the markers first.");
+    const next = starts[0] && ends[0]
+        ? previous.slice(0, starts[0].index) +
+            block.trimEnd() +
+            previous.slice(ends[0].index + ends[0][0].length)
+        : previous + (previous.endsWith("\n") || previous === "" ? "" : "\n") + "\n" + block;
+    try {
+        const result = parse(next);
+        if (withoutOurs(parsed) !== withoutOurs(result))
+            return refused("the marked replacement would change another setting; review the block manually.");
+    }
+    catch {
+        return refused("the proposed block conflicts with existing TOML tables; merge the supplied block manually.");
+    }
+    if (next === previous)
+        return { kind: "written", changed: false };
+    mkdirSync(directory, { recursive: true, mode: 0o700 });
+    const temporary = join(directory, `.balladeer.${randomBytes(8).toString("hex")}.tmp`);
+    let descriptor;
+    try {
+        descriptor = openSync(temporary, "wx", existsSync(path) ? lstatSync(path).mode & 0o777 : 0o600);
+        writeSync(descriptor, next);
+        fsyncSync(descriptor);
+        closeSync(descriptor);
+        descriptor = undefined;
+        renameSync(temporary, path);
+    }
+    finally {
+        if (descriptor !== undefined)
+            closeSync(descriptor);
+        if (existsSync(temporary))
+            unlinkSync(temporary);
+    }
+    return { kind: "written", changed: true };
+}

package/dist/commands/setup.d.ts

@@ -4,6 +4,8 @@ export type SetupOptions = Readonly<{
     wait: boolean;
     repo?: string;
     createWorkspace?: string;
+    client?: "codex" | "claude";
+    chooseWorkspace?: boolean;
     /**
      * Every repository this run was told to set up, in the order they were named.
      *

package/dist/commands/setup.js

@@ -3,18 +3,19 @@ import { existsSync, mkdtempSync, writeFileSync } from "node:fs";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { ClientTooOldError, RefusalError, TransportError, proposalReviewLink, reviewLink, request, } from "../client.js";
+import { CODEX_CONFIG_FILE, hasManagedCodexEntry, mergeCodexConfig, readCodexEntry, } from "../codex-config.js";
 import { writeConventions } from "../conventions.js";
 import { APPROVE_IN_THE_RIGHT_WORKSPACE, DISCOVERY_PLAYBOOK, JOIN_OR_CREATE } from "../copy.js";
 import { ghLogin, readPublicRepositoryFacts, readRepositoryFacts, runCommand, setRepositoryVariable, } from "../gh.js";
 import { CI_BRANCH, commitWorkflowOnBranch, repositoryRoot, workflowOnDefaultBranch, } from "../git.js";
-import { MCP_CONFIG_FILE, currentEntry, entryRepositoryId, mergeMcpConfig, readMcpConfig, stdioEntry, } from "../mcp-config.js";
+import { MCP_CONFIG_FILE, currentEntry, entryRepositoryId, isOurEntry, mergeMcpConfig, readMcpConfig, stdioEntry, } from "../mcp-config.js";
 import { DESKTOP_CONFIG_FILE, desktopConfigLocation, desktopServerKey, desktopStdioEntry, mergeDesktopConfig, } from "../desktop-config.js";
 import { selectAgent } from "../agent.js";
 import { findLegacyInstall, legacyMessage } from "../legacy.js";
 import { formatInstant } from "../local-time.js";
 import { checkoutEntryPath, commandLine, runningFromRegistryInstall } from "../release.js";
 import { hostHint, repositoryHint } from "../repository.js";
-import { StoreError, credentialsPath, dropPendingPairing, findPendingPairing, findSession, putAgent, putPendingPairing, putSession, readCredentials, writeCredentials, } from "../store.js";
+import { StoreError, assertStoreWritable, credentialsPath, dropPendingPairing, findPendingPairing, findSession, putAgent, putPendingPairing, putSession, readCredentials, writeCredentials, } from "../store.js";
 import { CLI_VERSION, DELEGATED_SCOPES, } from "../wire.js";
 import { explainText } from "./explain.js";
 import { probeAgentConnection } from "./mcp.js";
@@ -167,10 +168,11 @@ function pendingLines(pending, repository, host, waiting, approvalUri) {
         ...JOIN_OR_CREATE.split("\n").map((line) => (line.length === 0 ? "" : `  ${line}`)),
     ];
 }
-function storeSession(credentials, controlPlane, token, session) {
+function storeSession(credentials, controlPlane, token, session, workspaceChoice) {
     return putSession(dropPendingPairing(credentials, controlPlane), {
         controlPlane,
         sessionId: session.sessionId,
+        ...(workspaceChoice?.startsWith("create:") ? { workspaceChoice } : {}),
         token,
         workspaceId: session.workspaceId,
         workspaceName: session.workspaceName,
@@ -267,6 +269,7 @@ export async function runSetup(options) {
     const { controlPlane } = options;
     const now = options.now ?? (() => new Date());
     say(options, `Balladeer setup, version ${CLI_VERSION}.\n`);
+    say(options, `Agent client: ${options.client === "codex" ? "Codex" : "Claude"}. Use --client codex or --client claude to choose.`);
     emit(options, { step: "explain", version: CLI_VERSION });
     say(options, explainText(controlPlane));
     const chosen = chooseRepositories(namedRepositories(options), options.cwd);
@@ -288,8 +291,29 @@ export async function runSetup(options) {
     catch (error) {
         return storeFailure(options, error);
     }
+    // A completed create intent resumes only its own authenticated session. The
+    // name is local intent, never a lookup or grant of workspace authority. An
+    // explicit chooser always starts a switch. Keep the old session until the
+    // new pairing is claimed, so a failed start or claim loses no credentials.
+    const workspaceChoice = options.createWorkspace !== undefined
+        ? `create:${carriedName}`
+        : options.chooseWorkspace === true
+            ? "choose"
+            : undefined;
+    const previousSession = findSession(credentials, controlPlane);
+    const freshChoice = workspaceChoice !== undefined &&
+        (options.chooseWorkspace === true || previousSession?.workspaceChoice !== workspaceChoice);
+    if (workspaceChoice !== undefined) {
+        credentials = {
+            ...credentials,
+            pendingPairings: credentials.pendingPairings.filter((entry) => entry.controlPlane !== controlPlane || entry.workspaceChoice === workspaceChoice),
+        };
+    }
+    if (freshChoice) {
+        say(options, "Choose the workspace in the new browser pairing. Existing agent connections remain saved.");
+    }
     // A still-valid session carries straight on to the steps that need it.
-    const stored = findSession(credentials, controlPlane);
+    const stored = freshChoice ? undefined : previousSession;
     if (stored && Date.parse(stored.expiresAt) > now().getTime()) {
         try {
             const answer = await request(controlPlane, {
@@ -361,6 +385,9 @@ export async function runSetup(options) {
     }
     let started;
     try {
+        // An unsaved claim secret cannot be recovered by retrying the browser link.
+        // Check local access before creating an unapproved remote pairing.
+        assertStoreWritable(options.environment);
         const { verifier, digest } = newVerifier();
         started = await request(controlPlane, {
             method: "POST",
@@ -375,6 +402,7 @@ export async function runSetup(options) {
             verificationUri: started.verificationUriComplete,
             expiresAt: started.expiresAt,
             startedAt: now().toISOString(),
+            ...(workspaceChoice === undefined ? {} : { workspaceChoice }),
         };
         credentials = putPendingPairing(credentials, pending);
         writeCredentials(credentials, options.environment);
@@ -432,7 +460,7 @@ async function claimOnce(options, credentials, pending) {
         if (answer.status === "pending") {
             return { kind: "pending", pollIntervalSeconds: answer.pollIntervalSeconds };
         }
-        const updated = storeSession(credentials, options.controlPlane, answer.token, answer.session);
+        const updated = storeSession(credentials, options.controlPlane, answer.token, answer.session, pending.workspaceChoice);
         writeCredentials(updated, options.environment);
         reportPaired(options, answer.session);
         const session = findSession(updated, options.controlPlane);
@@ -839,7 +867,7 @@ function blockedRepository(options, role, name, reason, next, publicFacts) {
 }
 const NO_DESKTOP = { lines: [], json: undefined };
 function connectClaudeDesktop(options, repositoryId, repositoryName) {
-    if (options.claudeDesktop === false)
+    if (options.claudeDesktop === false || options.client === "codex")
         return NO_DESKTOP;
     // Asked for by name, or taken by default. The difference is only whether a
     // machine with no Claude desktop on it hears about it: a person who typed the
@@ -898,11 +926,28 @@ function connectClaudeDesktop(options, repositoryId, repositoryName) {
         },
     };
 }
-function hostNextSteps() {
+function hostConfigFile(options) {
+    return options.client === "codex" ? CODEX_CONFIG_FILE : MCP_CONFIG_FILE;
+}
+function mergeHostConfig(options, root, repositoryId, published) {
+    const entry = stdioEntry(repositoryId, published);
+    return options.client === "codex"
+        ? mergeCodexConfig(root, { ...entry, args: [...entry.args, "--control-plane", options.controlPlane] }, options.controlPlane)
+        : mergeMcpConfig(root, entry, options.controlPlane);
+}
+function writeHostConventions(options, root) {
+    return writeConventions(root, options.client === "codex" ? { file: "AGENTS.md" } : {});
+}
+function hostNextSteps(options) {
+    if (options.client === "codex")
+        return [
+            "  Codex project configuration is ready. Open this repository in Codex and trust the project if appropriate; Codex ignores project configuration in untrusted projects.",
+            "  Start a fresh Codex session to load .codex/config.toml and AGENTS.md, then check that Balladeer's tools are available. The connection probe does not prove your Codex host loaded them.",
+        ];
     return [
-        `  Your agent host reads ${MCP_CONFIG_FILE} when it starts, so the tools appear in its next session rather than in one already running. Nothing here needs a restart: finish setup first.`,
+        `  Your agent host reads ${hostConfigFile(options)} when it starts, so the tools appear in its next session rather than in one already running. Nothing here needs a restart: finish setup first.`,
         "  Claude Code asks you to approve a project MCP server the first time it sees one, so approve balladeer when it asks. In a session already running, /mcp lists the servers and reconnects them.",
-        "  Another host reads its own configuration rather than this file: add the same entry there, and start it again.",
+        "  For Codex, run setup --client codex --refresh to configure this project's Codex MCP and AGENTS.md, then review project trust and start a fresh Codex session.",
     ];
 }
 /**
@@ -942,11 +987,11 @@ async function configureAgentHost(options, repositoryId, name, published) {
     let merged;
     let conventions;
     if (root !== undefined) {
-        merged = mergeMcpConfig(root, stdioEntry(repositoryId, published), options.controlPlane);
+        merged = mergeHostConfig(options, root, repositoryId, published);
         if (merged.kind === "written") {
             if (merged.changed)
-                files.push(MCP_CONFIG_FILE);
-            conventions = writeConventions(root);
+                files.push(hostConfigFile(options));
+            conventions = writeHostConventions(options, root);
             if (conventions.kind === "written" && conventions.changed)
                 files.push(conventions.file);
         }
@@ -961,7 +1006,7 @@ async function configureAgentHost(options, repositoryId, name, published) {
 }
 function reportAgentHost(options, configured, published) {
     if (configured.root === undefined) {
-        say(options, `  This directory is not a git work tree, so I wrote no ${MCP_CONFIG_FILE} and no instructions block.`);
+        say(options, `  This directory is not a git work tree, so I wrote no ${hostConfigFile(options)} and no instructions block.`);
     }
     else {
         if (configured.merged?.kind === "refused") {
@@ -975,7 +1020,7 @@ function reportAgentHost(options, configured, published) {
             say(options, `  Repair the markers in ${configured.conventions.file}, or delete the block between them, then run this command again.`);
         }
         if (configured.merged?.kind === "written") {
-            say(options, `  The ${MCP_CONFIG_FILE} entry runs this command's own MCP forwarder for this repository and reads the credential from that store, so there is nothing else to set. This command never prints the credential.`);
+            say(options, `  The ${hostConfigFile(options)} entry runs this command's own MCP forwarder for this repository and reads the credential from that store, so there is nothing else to set. This command never prints the credential.`);
             if (published === null && !existsSync(checkoutEntryPath())) {
                 say(options, `  That entry runs ${checkoutEntryPath()}, which this checkout has not built yet. Run \`pnpm --filter balladeer build\` once so your agent host can start it.`);
             }
@@ -983,11 +1028,11 @@ function reportAgentHost(options, configured, published) {
     }
     for (const line of configured.desktop.lines)
         say(options, line);
-    for (const line of hostNextSteps())
+    for (const line of hostNextSteps(options))
         say(options, line);
     say(options, published === null
         ? `  Setup does not install a global balladeer executable. This checkout's CLI runs as \`${commandLine(null, "<command>")}\` after it is built.`
-        : `  Setup does not install a global balladeer executable. Run the CLI as \`${commandLine(published, "<command>")}\`; ${MCP_CONFIG_FILE} invokes that same package automatically.`);
+        : `  Setup does not install a global balladeer executable. Run the CLI as \`${commandLine(published, "<command>")}\`; ${hostConfigFile(options)} invokes that same package automatically.`);
     if (configured.files.length > 0) {
         say(options, `  ${configured.files.join(" and ")} ${configured.files.length === 1 ? "is an uncommitted change" : "are uncommitted changes"} in your working tree. Commit ${configured.files.length === 1 ? "it" : "them"} when you are ready; they carry no secret.`);
     }
@@ -1090,7 +1135,7 @@ async function stepAgent(options, credentials, session, view, name, published) {
     // sends are.
     const proof = await proveAgent(agent);
     const wrote = configured.merged?.kind === "written" && configured.conventions?.kind === "written"
-        ? `, ${MCP_CONFIG_FILE} and ${configured.conventions.file} updated`
+        ? `, ${hostConfigFile(options)} and ${configured.conventions.file} updated`
         : "";
     say(options, proof.kind === "answered"
         ? `Step 3 of 5  Connected this coding agent on this machine  credential stored in ${credentialsPath(options.environment)}${wrote}`
@@ -1589,13 +1634,21 @@ export function refreshPublishedForm(existing) {
  * changed. Run twice it changes nothing the second time, which is what makes it
  * safe to put in a nag an agent will act on without asking.
  */
-async function runRefresh(options) {
+async function runRefresh(input) {
+    let options = input;
     say(options, `Balladeer setup --refresh, version ${CLI_VERSION}.\n`);
     const root = await repositoryRoot(options.cwd);
     if (root === undefined) {
         return failRefresh(options, "not_a_repository", "This directory is not a git work tree, so there is no repository configuration to refresh.");
     }
-    const existing = currentEntry(readMcpConfig(root));
+    if (options.client === undefined && hasManagedCodexEntry(root, options.controlPlane)) {
+        if (isOurEntry(currentEntry(readMcpConfig(root)), options.controlPlane)) {
+            return failRefresh(options, "client_ambiguous", "Both Codex and Claude have Balladeer project entries. Run setup --client codex --refresh or setup --client claude --refresh. No files were changed.");
+        }
+        options = { ...options, client: "codex" };
+        say(options, "Using Codex for refresh because this repository has only its managed Balladeer entry.");
+    }
+    const existing = options.client === "codex" ? readCodexEntry(root) : currentEntry(readMcpConfig(root));
     let stored;
     let storedName;
     try {
@@ -1616,9 +1669,9 @@ async function runRefresh(options) {
     // that never paired can still repair a checkout somebody else set up.
     const repositoryId = stored ?? entryRepositoryId(existing);
     if (repositoryId === undefined) {
-        return failRefresh(options, "not_connected", `Nothing here names a repository to refresh: this machine holds no Balladeer credential for it and ${MCP_CONFIG_FILE} has no balladeer entry. Run \`${commandLine(null, "setup")}\` to connect it.`);
+        return failRefresh(options, "not_connected", `Nothing here names a repository to refresh: this machine holds no Balladeer credential for it and ${hostConfigFile(options)} has no balladeer entry. Run \`${commandLine(null, "setup")}\` to connect it.`);
     }
-    const merged = mergeMcpConfig(root, stdioEntry(repositoryId, refreshPublishedForm(existing)), options.controlPlane);
+    const merged = mergeHostConfig(options, root, repositoryId, refreshPublishedForm(existing));
     if (merged.kind === "refused") {
         say(options, merged.reason);
         say(options, "Merge this block into it yourself:");
@@ -1626,7 +1679,7 @@ async function runRefresh(options) {
             say(options, `  ${line}`);
         return failRefresh(options, "mcp_config_refused", merged.reason);
     }
-    const conventions = writeConventions(root);
+    const conventions = writeHostConventions(options, root);
     if (conventions.kind === "refused") {
         say(options, conventions.reason);
         return failRefresh(options, "conventions_refused", conventions.reason);
@@ -1639,12 +1692,12 @@ async function runRefresh(options) {
     const desktop = connectClaudeDesktop(options, repositoryId, storedName ?? repositoryHint(options.cwd));
     const files = [];
     if (merged.changed)
-        files.push(MCP_CONFIG_FILE);
+        files.push(hostConfigFile(options));
     if (conventions.changed)
         files.push(conventions.file);
     say(options, merged.changed
-        ? `Rewrote the balladeer entry in ${MCP_CONFIG_FILE}. Every other entry in that file is untouched.`
-        : `${MCP_CONFIG_FILE} already names the current command; I left it alone.`);
+        ? `Rewrote the balladeer entry in ${hostConfigFile(options)}. Every other entry in that file is untouched.`
+        : `${hostConfigFile(options)} already names the current command; I left it alone.`);
     say(options, conventions.changed
         ? `Rewrote the Balladeer block in ${conventions.file} to conventions v${conventions.version}${conventions.previousVersion === undefined
             ? ""
@@ -1655,6 +1708,9 @@ async function runRefresh(options) {
     // never a change anybody commits: it lives in this person's home directory.
     for (const line of desktop.lines)
         say(options, line.replace(/^ {2}/, ""));
+    if (options.client === "codex")
+        for (const line of hostNextSteps(options))
+            say(options, line);
     if (files.length === 0 && desktop.json?.status !== "connected") {
         say(options, "Nothing to change: this repository is already on the current form.");
     }

package/dist/conventions.d.ts

@@ -27,7 +27,7 @@ export declare const CONVENTIONS_END = "<!-- balladeer:conventions:end -->";
  * instructions marker instead. A bump here with no change to the block would
  * rewrite twelve repositories to say exactly what they already said.
  */
-export declare const CONVENTIONS_VERSION = 16;
+export declare const CONVENTIONS_VERSION = 17;
 export declare function managedByLine(version: number): string;
 /**
  * Which version of the block a file already carries, or nothing when it carries
@@ -74,4 +74,5 @@ export type ConventionsResult = Readonly<{
 export declare function writeConventions(repositoryRoot: string, options?: Readonly<{
     version?: number;
     body?: string;
+    file?: "CLAUDE.md" | "AGENTS.md";
 }>): ConventionsResult;

package/dist/conventions.js

@@ -31,7 +31,7 @@ export const CONVENTIONS_END = "<!-- balladeer:conventions:end -->";
  * instructions marker instead. A bump here with no change to the block would
  * rewrite twelve repositories to say exactly what they already said.
  */
-export const CONVENTIONS_VERSION = 16;
+export const CONVENTIONS_VERSION = 17;
 /** Both spellings: the stable marker, and the versioned one version 1 wrote. */
 const START_MARKER = /<!-- balladeer:conventions:start(?: v(\d{1,4}))? -->/g;
 const END_MARKER = /<!-- balladeer:conventions:end -->/g;
@@ -114,7 +114,7 @@ function writeAtomically(path, contents) {
 export function writeConventions(repositoryRoot, options = {}) {
     const version = options.version ?? CONVENTIONS_VERSION;
     const existingName = CANDIDATE_FILES.find((name) => existsSync(join(repositoryRoot, name)));
-    const name = existingName ?? "CLAUDE.md";
+    const name = options.file ?? existingName ?? "CLAUDE.md";
     const path = join(repositoryRoot, name);
     const block = fenced(version, options.body ?? CONVENTIONS_BLOCK);
     let existing = "";

package/dist/copy.d.ts

@@ -115,7 +115,7 @@ export declare const CAPTURE_SEAM = "### When to say nothing, and what a yes is
  * initialize, and the page `/agent` serves. A packaging test holds them
  * together.
  */
-export declare const SETTLING_QUESTIONS = "### When somebody says \"tell me about\" one\n\nAn id is how a person points at something here, and every promise page and every proposal page\ncarries one. A promise id starts with prom_ and a proposal id starts with cand_: expand a promise\nwith get_promise and a proposal with get_proposal, and read neither of them out as a list of fields.\nSay in four or five sentences what it is for, who it is for, when it applies and what must then be\ntrue. Then say where it stands: a promise is agreed, and either protected or not yet checked by\nanything; a proposal is agreed by nobody and waiting on the person it names.\n\nA proposal that still carries open questions is not finished, and settling them is usually why\nsomebody asked. Offer to work through them, then take them one at a time in the order they come\nback. For each one, say what it decides in their words rather than in the question's; give your\nrecommendation and the reason you hold it, drawn from this repository and from the promise itself;\nand stop there. When they answer, record what they said with resolve_question, in their own words\nwhere they gave you any, and where their answer changes the promise, follow it with update_proposal,\nadd_case or remove_case and tell them what you changed. None of that agrees to anything: the named\nowner agrees, in their own browser or through a sign-off you carry, and the questions they answered\nstay on the proposal in their name.\n\nThe same rule governs every question you leave open in the first place. A question that names a gap\nand stops is a note, and nobody can answer a note. Say what answering it decides, in the words a\ncustomer would use, and carry your own best guess with the reason behind it, so that the shortest\ntrue answer is yes. For example: \"Decides: whether a worker that is running but reconciling nothing\ncounts as an outage this promise covers. Best guess: yes, because the promise is about somebody\nhearing before a customer does, and a wedged worker is invisible to every check this repository has.\nSay yes, or tell me otherwise.\" Balladeer refuses a question filed without both halves and says\nwhich one is missing.";
+export declare const SETTLING_QUESTIONS = "### When somebody says \"tell me about\" one\n\nAn id is how a person points at something here, and every promise page and every proposal page\ncarries one. A promise id starts with prom_ and a proposal id starts with cand_: expand a promise\nwith get_promise and a proposal with get_proposal, and read neither of them out as a list of fields.\nSay in four or five sentences what it is for, who it is for, when it applies and what must then be\ntrue. Then say where it stands: a promise is agreed, and either protected or not yet checked by\nanything; a proposal is agreed by nobody and waiting on the person it names.\n\nClarify ambiguity that materially changes the behavior during capture, within the existing\nquestion budget; never invent an answer.\n\nOrdinary open questions record uncertainty; they do not prevent the named owner from agreeing to\nthe behavior as written. Agreement does not answer them or add an unstated guarantee. Separate\ncatalog-conflict questions can hold agreement until the owner rules on the stated conflict. Do not\ncall a proposal unready just because it has ordinary open questions. Offer to discuss them if\nuseful; if the person wants to, take them one at a time in the order they come back. For each ordinary question,\nsay what it decides in their words rather than in the question's; give your\nrecommendation and the reason you hold it, drawn from this repository and from the promise itself; and stop there. When\nthey answer, record what they said with resolve_question, in their own words where they gave you\nany, and where their answer changes the promise, follow it with update_proposal, add_case or\nremove_case and tell them what you changed. Never invent an answer or clear uncertainty because\nthey chose to agree. None of that agrees to anything: the named owner agrees, in their own browser\nor through a sign-off you carry, and the questions they answered stay on the proposal in their name.\n\nThe same rule governs every question you leave open in the first place. A question that names a gap\nand stops is a note, and nobody can answer a note. Say what answering it decides, in the words a\ncustomer would use, and carry your own best guess with the reason behind it, so that the shortest\ntrue answer is yes. For example: \"Decides: whether a worker that is running but reconciling nothing\ncounts as an outage this promise covers. Best guess: yes, because the promise is about somebody\nhearing before a customer does, and a wedged worker is invisible to every check this repository has.\nSay yes, or tell me otherwise.\" Balladeer refuses a question filed without both halves and says\nwhich one is missing.";
 /**
  * What a session may spend on reading, in three sentences.
  *
@@ -193,7 +193,7 @@ export declare const SESSION_STAMP = "Run `npx -y balladeer@latest session` when
  * block that drifts per run would rewrite a customer's committed file on every
  * setup and the diff would say nothing.
  */
-export declare const CONVENTIONS_BLOCK = "## Balladeer promises\n\nBefore planning work in this repository, read the promises this team has already approved through\nthe Balladeer MCP server. They are the behaviors a named person has agreed the software keeps, so\nyour plan has to hold them, not just read them.\n\n### Reading what is already agreed\n\nRetrieve by promise, and only when this session has a reason to. One repository here can hold a\nthousand agreed promises, and a plan built from whatever survived a truncated catalog read is worse\nthan a plan built from none of it, because nothing tells you which half went missing.\n\nWhen a person gives you a promise id, expand exactly that one with get_promise and stop there. Every\npromise page carries a control that copies its id, so an id is what a person hands you when they\nmean a particular promise. Ask for one rather than searching for what they meant.\n\nConsult list_promises in two situations and no others. The person asks what this repository has\npromised, in which case page the index they asked for. Or the change you are about to make touches\npaths that carry promises, in which case give those paths to list_promises: it answers with the\npromises whose scope overlaps them, closest first, and with the few that name no path and so cover\nthe whole repository. That answer is a selection rather than a page and does not continue with a\ncursor, so when the total beside it is larger than what you were handed, narrow the paths rather\nthan asking for more.\n\nWhen you do not yet know which paths you are about to touch, do not call the index at all. Work from\nids until you do, because a page you did not ask a question of is not about your change, and reading\none as though it were is how a plan quietly misses the promise it breaks.\n\nTo learn which paths those are, read `paths` on a list_promises answer you asked for\nwithout paths of your own. It is the set of\nrepository paths this repository's promises are scoped to, deduplicated and bounded, with\n`pathsTruncated` saying whether there were more than the answer carries. Compare the files you are\nabout to change against it. Nothing matching means there is nothing here to read, and saying so is a\nbetter answer than a page of promises about somewhere else.\n\nNever call get_promise_context at the start of a session. It answers the markers you give it, and\nbefore you know what you are changing there are no markers to give: what comes back is a slice of\nthe catalog chosen by nothing. Call it once the work is in front of you, with that work's markers.\n\nEvery one of these reads is bounded and none of them returns the whole catalog. Read the total\nbeside the rows and the sentence in `scope` that says what the total is a total of, and when the\ntotal is larger than what you were handed, page or narrow rather than planning as though you had\nseen everything.\n\nEach index row carries what it takes to rule that promise out without fetching it: the repository,\nthe paths it covers, its one-sentence claim, what protects it and why, and when anything last\nchecked it.\n\nPrefer the measurement over the markers. Where this repository has a touch map, `npx -y balladeer@latest affected\n<paths...>` answers which promises ran the files in front of you, out of what each verifier actually\nexecuted the last time `npx -y balladeer@latest touch-map` measured it. Both commands run on this machine and\nsend Balladeer nothing, so you may name any path in the change.\n\nAn answer marked stale was measured against a verifier that has changed since, so read it as the\nlast thing anybody measured rather than as fact, and offer to run `npx -y balladeer@latest touch-map` again.\nWhere there is no map, where it does not name your paths, or where it reports a promise it could not\nmeasure, fall back to the paths on the index rows. The map narrows which promises are worth\nfetching. It never widens what you may read: the catalog is still never enumerated, and the fetch\nbudget is unchanged.\n\nWhen a retrieval receipt asks why you read a row, the answer for one the map named is `touch`.\n\nFetch by id, and only for an id the person gave you or an index row whose paths match the change in\nfront of you. Never enumerate the catalog. Never chain one fetch into the next to see the whole of\nsomething: when the rows do not settle it, narrow the filter rather than expanding another promise.\n\n### When to say nothing, and what a yes is worth\n\nMost rules are said in passing, in the middle of something else. Somebody saying one has not asked\nyou to record anything, so propose nothing and start no interview. When you do ask, it is one line\nappended to the end of the reply you were already going to give, never a message of its own, never\nasked twice about the same rule, and nothing is proposed until they say yes. A no ends it, and that\nrule is not raised again for the rest of the conversation.\n\nSilence is per conversation rather than per message. Once a conversation is one of these, you ask\nnothing for the rest of it, however good the rule sounds:\n\n- A question, or working out how something already behaves. Nothing is filed and nothing is offered.\n- A refactor. Nobody predicts a promise from a rewrite. Offer the promises this area already carries\n  that nothing is checking yet, once, and then wait. Ask nothing about new ones.\n- A change to wording alone. Wording somebody may change again tomorrow is not a rule.\n- An incident, while it is still being fixed. Ask nothing at all until the fix is merged or they say\n  it is done, however good tonight's failing case would be.\n- Somebody still weighing options. A decision nobody has made yet is not a rule.\n- An exploration or spike. It ends in nothing or in a plan, and its sentences sound like rules and\n  are not.\n- Reading somebody else's change, while you are still reading it. Nothing is offered until they\n  give a verdict.\n\nThe words to ask in, the moment to offer, and the shape a proposal takes are not in this file. The\nBalladeer server sends them at the start of every session, and its copy is the current one: read\nwhat it sent this session rather than what this file remembers.\n\n### Two promises Balladeer cannot keep\n\nA promise about speed needs three things before it is a promise at all: a number, a percentile, and\nwhere it is measured. \"The quote page answers in under one second at p95, measured in production at\npeak load\" is one Balladeer can keep. \"The quote page has to be fast\" is not. Say this, and ask for\nthe part that is missing rather than filing it:\n\n\"I can keep that once it has a number, a percentile, and a place it is measured. Without those it is\na wish, not a promise.\"\n\nBalladeer refuses one without all three and names which of them is missing. When all three are\nthere, write the measurement method into the promise, and tell them plainly that it reads agreed and\nunprotected until a test that actually measures it exists.\n\nA promise about how the team works is not something the software does, so no check can ever catch\nit. \"We must provide a low-friction capture experience\" is one of these. File nothing and say:\n\n\"That is a promise about how we work, not something the software does that a check can fail.\nBalladeer only keeps promises a check can catch. If a customer would notice something when this\nslips, say that and I will keep that instead.\"\n\nThen take the customer-visible half if they give you one, and file that instead.\n\nRun `npx -y balladeer@latest session` when you start work here and pass the id it prints to every promise read\nyou make, as `session`. It is the same id for the whole of one piece of work, and it costs nothing:\na read without it is answered exactly the same way.\n\nIf local file permissions refuse that session write, no session stamp was saved. Your saved\nconnection is unchanged; this does not prove its current server authorization. Continue\nalready-authorized MCP reads, coding and explicitly requested capture without the optional\n`session` field. Do not invent an ID, add a session trailer, or record an unsaved stamp. Do not\nbroaden filesystem access or move credentials to retry this write. A missing credential or server\nauthorization refusal is different: follow that refusal. Capture still needs the person's request\nor accepted offer, and human meaning approval is unchanged.\n\nIf a session ID was saved, near the end of the work call `mark_promise_used` with the ids you actually cited\nor acted on. Naming none is a fine answer. It is the only way anybody finds out which rows were\nworth sending you, because reading a promise and then ignoring it is not the same as using it.\n\nIf a session ID was saved, write `Balladeer-Session: <the id>` into the commit message or pull-request body you produce,\nand run `npx -y balladeer@latest session --record` once the commit exists. That is what lets a check that goes red\nnext week be read back against what Balladeer told you before you started. Your commit message never\nleaves your machine: only the session id and the commit SHA are sent.\n\n### When somebody asks you to protect a behavior\n\nSomebody has asked you to protect a behavior when they say what the software must do, or must never\ndo again, and mean it as a rule rather than as this one bug. That is one promise, for the behavior\nthey named, and nothing else: if you notice others worth protecting, say so in a sentence and let\nthem choose, and file none of them. Never propose from a conversation that did not ask you to. A\nquestion about how something works and a plan you were asked to sketch are not requests to record\nanything, nor is a fix nobody asked you to write a rule about.\n\nAsk before you extrapolate. Ask only what you cannot work out for yourself, ask it all in one\nmessage, and stop at four. Four is a ceiling, not a target: two good ones are better. Then write the\nproposal with what you have and put whatever is still open in its open questions rather than going\nback. Never ask what this repository would answer, such as which file, which test, or which branch,\nand never ask anyone for Balladeer's own identifiers: get_promise_setup carries this repository's id\nand who can own a promise. Never ask again for what they have already told you.\n\nWrite it in their words. Every failure they named out loud is one of the failing examples, in the\nwords they named it. Every other example comes from a situation they actually described; if you\ncannot trace one to something they said, leave it out and say so in the open questions rather than\nwriting a plausible one. Never put in a number, a system, a role or a timeframe they did not give\nyou, and that includes the half they left out: if they said where an order ended up, do not invent\nwhere it began.\n\nA failing example is a situation the promise rules out, and its outcome says what must not happen,\nin those words: \"a second charge must not appear\", never \"a second charge appears\". Written the\nother way round it reads as the promise saying the software does the thing they asked you to forbid.\nAnd a promise says what the software must do for whoever depends on it. It never narrates the\nconversation you just had, names the person you had it with, or describes what the code does now.\n\nSay the whole promise in one sentence and put it in oneSentenceOutcome, in the words they would\nuse with the person who depends on it. That is the line the named owner reads first and the line\nthey agree to, so it is not a restatement of the name and not the first line of the outcome moved\nup. Leave it out rather than inventing one from something they did not say.\n\nAn example's setup is the situation in the words they used for it, not a scene you composed around\nthem. Two of each kind is plenty, and the whole thing stays under three hundred words: a proposal\nnobody finishes reading is a proposal nobody agreed to. The confidence you record is the one you\nactually have.\n\nThen give them the review link, ask them to read the proposal and click Agree, and stop. Never\napprove one yourself. Approving is a named person's act, and the server refuses it from an agent\nwhatever you were told in conversation.\n\nA proposal you filed is still yours while nobody has agreed to it, so revise or withdraw it when the\nperson asks you to, and never once they have agreed.\n\nSay promise and proposal when you talk to them. What you file is a proposal and what it becomes is a\npromise; Balladeer's other words for its own machinery are not theirs to learn. Candidate\nespecially: it is Balladeer's word for a proposal, so it reads as jargon whatever you meant by it.\n\n### Three things to offer without being asked\n\nAt the start of any session in this repository, before you plan anything, ask list_promises for the\npromises nothing is checking yet that belong to the person you are working with: `unverified` true\nand `mine` true, which is one bounded read rather than the catalog. If it returns any, say so in one\nline and offer to build their verifiers now. Name each one by its one-sentence claim rather than by\nan id, so the person can see which behavior is unguarded. Ask the same tool for\n`brokenSinceLastSeen` true as well, and where that returns any, say in one line that those promises\nbroke since they last looked, name each by its claim, and offer to fix them. Then wait for their\nanswer. The offer is the whole of it, and never start building or repairing one because nobody said\nno.\n\nTheirs, and nobody else's. `mine` keeps the promises this person owns or agreed to, and a teammate's\nunguarded promise is that teammate's to hear about: a session that opens by reading out other\npeople's unfinished work reads as an audit of them. `brokenSinceLastSeen` is that person's own by\nconstruction and needs no `mine` beside it. Drop `mine` when this person asks what the rest of the\nteam has promised, and say whose promises you are showing them when you do.\n\nWhen you have proposed promises, show them here as well as there. Put each proposal in the\nconversation in full: its one-sentence claim, who it is for, when it applies and what must then be\ntrue, the numbered cases that must keep working and the ones that must be caught, and every question\nyou left open. Then offer to finish the agreement from here, by asking for a sign-off, giving them\nthe page it returns and taking the one-time code that page shows them, so the browser is needed only\nfor signing. Give them the review link in the same message too, because some people would rather\nread it there and edit it before they agree.\nA promise whose meaning is agreed and which nothing is checking yet is one you can finish. When\nsomebody gives you a promise id, run `npx -y balladeer@latest status <promise id>`, or expand it with get_promise.\nIf it comes back agreed with nothing checking it, say so in one line and offer to prepare and build\nits verifier now. Only if they say yes: `prepare_qualification` mints the one-time setup, or\n`npx -y balladeer@latest prepare <promise id>` mints it and writes it where the sealed run reads it. You need no\nsign-off for that and there is no code to ask anybody for, because agreeing the meaning was their\nact and building the check that proves it is yours. Then build the verifier, seal it, push, and tell\nthem protection starts by itself when that run qualifies. The offer is the whole of it: never\nprepare one because nobody said no.\n\n### When somebody says \"tell me about\" one\n\nAn id is how a person points at something here, and every promise page and every proposal page\ncarries one. A promise id starts with prom_ and a proposal id starts with cand_: expand a promise\nwith get_promise and a proposal with get_proposal, and read neither of them out as a list of fields.\nSay in four or five sentences what it is for, who it is for, when it applies and what must then be\ntrue. Then say where it stands: a promise is agreed, and either protected or not yet checked by\nanything; a proposal is agreed by nobody and waiting on the person it names.\n\nA proposal that still carries open questions is not finished, and settling them is usually why\nsomebody asked. Offer to work through them, then take them one at a time in the order they come\nback. For each one, say what it decides in their words rather than in the question's; give your\nrecommendation and the reason you hold it, drawn from this repository and from the promise itself;\nand stop there. When they answer, record what they said with resolve_question, in their own words\nwhere they gave you any, and where their answer changes the promise, follow it with update_proposal,\nadd_case or remove_case and tell them what you changed. None of that agrees to anything: the named\nowner agrees, in their own browser or through a sign-off you carry, and the questions they answered\nstay on the proposal in their name.\n\nThe same rule governs every question you leave open in the first place. A question that names a gap\nand stops is a note, and nobody can answer a note. Say what answering it decides, in the words a\ncustomer would use, and carry your own best guess with the reason behind it, so that the shortest\ntrue answer is yes. For example: \"Decides: whether a worker that is running but reconciling nothing\ncounts as an outage this promise covers. Best guess: yes, because the promise is about somebody\nhearing before a customer does, and a wedged worker is invisible to every check this repository has.\nSay yes, or tell me otherwise.\" Balladeer refuses a question filed without both halves and says\nwhich one is missing.\n\n### Files that are sealed, and the one reason to edit one\n\nEvery file under .continuity/promises/ is sealed. The promise that owns that directory records the\nexact bytes of each file in it, so editing one, adding one there, renaming one or deleting one\nbreaks the seal. A promise whose seal is broken stops being checked, and it stays that way until a\nnamed person qualifies it again, which is their afternoon rather than your commit. Nothing in there\nis ordinary source: keep it out of refactors, formatting runs and dependency upgrades.\n\nFind out which directories are sealed before you plan an edit, not after. npx -y balladeer@latest status lists\nthem, and list_promises names a promise's sealed directory on its row once a verifier is bound to\nit. Then, before you push, run npx -y balladeer@latest check-seals. It prints nothing and exits zero when your\nchange touches no seal, and names the promise, its owner and its page when your change would break\none.\n\nThe one reason to edit a sealed file is to repair a verifier that can no longer run: something it\nimports moved, or the language it is written in changed under it. Never edit one to make a failing\ncheck pass. A check going red is the promise doing its job, and the repair for that belongs in the\nbehavior it protects. A repair is not finished until the promise is sealed again with the runner\nthis repository is pinned to. get_promise_setup carries that exact seal command, and\nnpx -y balladeer@latest check-seals prints it beside any promise it names.\n\n### The rest of a promise's life\nWhen a promise is obsolete, finished, deliberately off for a while, or owned by the wrong person,\npropose the change and hand them the promise page. Deciding is theirs.\n\nYou can finish one of those acts here, and only one way. Ask for a sign-off with\nrequest_owner_signoff, give them the page it returns, and ask for the one-time code that page shows\nthem. Then call the act's own tool with that code and their own words. Never call one on your own\ninitiative, never on a general approval of some earlier act, and never ask for a code you were not\ngiven: a refusal is the person's to resolve, not yours to retry. Balladeer records them as the\nperson who acted and you as the messenger.\n\nReport the promise's state exactly as Balladeer reported it: proposed, agreed, or protected, never\none in place of another.\n\n### Where your team watches this\n\nBalladeer is a web app as well as these tools, at the address setup printed. Its catalog lists every\npromise with who owns it and whether anything is checking it, and each promise has a page of its own\nshowing what this team agreed the software must do and then every run that has checked it since,\nnewest first, with the commit each one checked. Whenever there is a link to give, give the link\nrather than a summary of it: the page says what you would have said, and it stays true after this\nconversation has ended.\n\nBalladeer also has a Slack app, which a workspace administrator installs from workspace settings.\nOnce it is installed, whoever owns a promise gets a direct message when theirs goes live and when a\nrun on the default branch breaks it. Until somebody installs it, nothing is sent anywhere, so say it\nis available rather than saying they will be told.\n\n### Questions people ask\n\nAnswer these when they come up. Where you do not know, say so and point at the address setup\nprinted: a confident wrong answer about what a vendor can see is worse than no answer.\n\nWhat it does: it holds the behaviors this team has agreed the software must keep, and reports\nwhether each one is still being kept, from this repository's own tests running in its own CI.\n\nWhat it sees: the text of each promise somebody approves, this repository's numeric ids and the name\nof its default branch, and from CI the pass or fail outcome, the commit checked, and content hashes.\nNever the code, the tests, the fixtures, the logs, the prompts, or the transcripts.\n\nWhat stopping costs: nothing that matters to their tests. The verifier package, its fixtures and the\nworkflow file are theirs, in their repository, running in their CI, and disconnecting changes none\nof them. An administrator can download everything Balladeer holds at any time from workspace\nsettings, and disconnecting hands them that same download in the response that ends access.\n\nWho can approve one: the named person who owns it, in their own browser. Not an administrator on\ntheir behalf, and never you.\n\nWhether it blocks a merge: no. The check is advisory on Balladeer's side, and their own branch\nprotection is what decides whether a failing check stops anything.\n\nWho can invite people and change setup: a workspace administrator. A contributor can read the\nworkspace and propose promises, and a viewer can read it. If somebody asks you to add a teammate,\ninvite_teammate returns the page an administrator sends the invitation from, with the address filled\nin for them to read. The tool sends nothing itself: an administrator presses Send.";
+export declare const CONVENTIONS_BLOCK = "## Balladeer promises\n\nBefore planning work in this repository, read the promises this team has already approved through\nthe Balladeer MCP server. They are the behaviors a named person has agreed the software keeps, so\nyour plan has to hold them, not just read them.\n\n### Reading what is already agreed\n\nRetrieve by promise, and only when this session has a reason to. One repository here can hold a\nthousand agreed promises, and a plan built from whatever survived a truncated catalog read is worse\nthan a plan built from none of it, because nothing tells you which half went missing.\n\nWhen a person gives you a promise id, expand exactly that one with get_promise and stop there. Every\npromise page carries a control that copies its id, so an id is what a person hands you when they\nmean a particular promise. Ask for one rather than searching for what they meant.\n\nConsult list_promises in two situations and no others. The person asks what this repository has\npromised, in which case page the index they asked for. Or the change you are about to make touches\npaths that carry promises, in which case give those paths to list_promises: it answers with the\npromises whose scope overlaps them, closest first, and with the few that name no path and so cover\nthe whole repository. That answer is a selection rather than a page and does not continue with a\ncursor, so when the total beside it is larger than what you were handed, narrow the paths rather\nthan asking for more.\n\nWhen you do not yet know which paths you are about to touch, do not call the index at all. Work from\nids until you do, because a page you did not ask a question of is not about your change, and reading\none as though it were is how a plan quietly misses the promise it breaks.\n\nTo learn which paths those are, read `paths` on a list_promises answer you asked for\nwithout paths of your own. It is the set of\nrepository paths this repository's promises are scoped to, deduplicated and bounded, with\n`pathsTruncated` saying whether there were more than the answer carries. Compare the files you are\nabout to change against it. Nothing matching means there is nothing here to read, and saying so is a\nbetter answer than a page of promises about somewhere else.\n\nNever call get_promise_context at the start of a session. It answers the markers you give it, and\nbefore you know what you are changing there are no markers to give: what comes back is a slice of\nthe catalog chosen by nothing. Call it once the work is in front of you, with that work's markers.\n\nEvery one of these reads is bounded and none of them returns the whole catalog. Read the total\nbeside the rows and the sentence in `scope` that says what the total is a total of, and when the\ntotal is larger than what you were handed, page or narrow rather than planning as though you had\nseen everything.\n\nEach index row carries what it takes to rule that promise out without fetching it: the repository,\nthe paths it covers, its one-sentence claim, what protects it and why, and when anything last\nchecked it.\n\nPrefer the measurement over the markers. Where this repository has a touch map, `npx -y balladeer@latest affected\n<paths...>` answers which promises ran the files in front of you, out of what each verifier actually\nexecuted the last time `npx -y balladeer@latest touch-map` measured it. Both commands run on this machine and\nsend Balladeer nothing, so you may name any path in the change.\n\nAn answer marked stale was measured against a verifier that has changed since, so read it as the\nlast thing anybody measured rather than as fact, and offer to run `npx -y balladeer@latest touch-map` again.\nWhere there is no map, where it does not name your paths, or where it reports a promise it could not\nmeasure, fall back to the paths on the index rows. The map narrows which promises are worth\nfetching. It never widens what you may read: the catalog is still never enumerated, and the fetch\nbudget is unchanged.\n\nWhen a retrieval receipt asks why you read a row, the answer for one the map named is `touch`.\n\nFetch by id, and only for an id the person gave you or an index row whose paths match the change in\nfront of you. Never enumerate the catalog. Never chain one fetch into the next to see the whole of\nsomething: when the rows do not settle it, narrow the filter rather than expanding another promise.\n\n### When to say nothing, and what a yes is worth\n\nMost rules are said in passing, in the middle of something else. Somebody saying one has not asked\nyou to record anything, so propose nothing and start no interview. When you do ask, it is one line\nappended to the end of the reply you were already going to give, never a message of its own, never\nasked twice about the same rule, and nothing is proposed until they say yes. A no ends it, and that\nrule is not raised again for the rest of the conversation.\n\nSilence is per conversation rather than per message. Once a conversation is one of these, you ask\nnothing for the rest of it, however good the rule sounds:\n\n- A question, or working out how something already behaves. Nothing is filed and nothing is offered.\n- A refactor. Nobody predicts a promise from a rewrite. Offer the promises this area already carries\n  that nothing is checking yet, once, and then wait. Ask nothing about new ones.\n- A change to wording alone. Wording somebody may change again tomorrow is not a rule.\n- An incident, while it is still being fixed. Ask nothing at all until the fix is merged or they say\n  it is done, however good tonight's failing case would be.\n- Somebody still weighing options. A decision nobody has made yet is not a rule.\n- An exploration or spike. It ends in nothing or in a plan, and its sentences sound like rules and\n  are not.\n- Reading somebody else's change, while you are still reading it. Nothing is offered until they\n  give a verdict.\n\nThe words to ask in, the moment to offer, and the shape a proposal takes are not in this file. The\nBalladeer server sends them at the start of every session, and its copy is the current one: read\nwhat it sent this session rather than what this file remembers.\n\n### Two promises Balladeer cannot keep\n\nA promise about speed needs three things before it is a promise at all: a number, a percentile, and\nwhere it is measured. \"The quote page answers in under one second at p95, measured in production at\npeak load\" is one Balladeer can keep. \"The quote page has to be fast\" is not. Say this, and ask for\nthe part that is missing rather than filing it:\n\n\"I can keep that once it has a number, a percentile, and a place it is measured. Without those it is\na wish, not a promise.\"\n\nBalladeer refuses one without all three and names which of them is missing. When all three are\nthere, write the measurement method into the promise, and tell them plainly that it reads agreed and\nunprotected until a test that actually measures it exists.\n\nA promise about how the team works is not something the software does, so no check can ever catch\nit. \"We must provide a low-friction capture experience\" is one of these. File nothing and say:\n\n\"That is a promise about how we work, not something the software does that a check can fail.\nBalladeer only keeps promises a check can catch. If a customer would notice something when this\nslips, say that and I will keep that instead.\"\n\nThen take the customer-visible half if they give you one, and file that instead.\n\nRun `npx -y balladeer@latest session` when you start work here and pass the id it prints to every promise read\nyou make, as `session`. It is the same id for the whole of one piece of work, and it costs nothing:\na read without it is answered exactly the same way.\n\nIf local file permissions refuse that session write, no session stamp was saved. Your saved\nconnection is unchanged; this does not prove its current server authorization. Continue\nalready-authorized MCP reads, coding and explicitly requested capture without the optional\n`session` field. Do not invent an ID, add a session trailer, or record an unsaved stamp. Do not\nbroaden filesystem access or move credentials to retry this write. A missing credential or server\nauthorization refusal is different: follow that refusal. Capture still needs the person's request\nor accepted offer, and human meaning approval is unchanged.\n\nIf a session ID was saved, near the end of the work call `mark_promise_used` with the ids you actually cited\nor acted on. Naming none is a fine answer. It is the only way anybody finds out which rows were\nworth sending you, because reading a promise and then ignoring it is not the same as using it.\n\nIf a session ID was saved, write `Balladeer-Session: <the id>` into the commit message or pull-request body you produce,\nand run `npx -y balladeer@latest session --record` once the commit exists. That is what lets a check that goes red\nnext week be read back against what Balladeer told you before you started. Your commit message never\nleaves your machine: only the session id and the commit SHA are sent.\n\n### When somebody asks you to protect a behavior\n\nSomebody has asked you to protect a behavior when they say what the software must do, or must never\ndo again, and mean it as a rule rather than as this one bug. That is one promise, for the behavior\nthey named, and nothing else: if you notice others worth protecting, say so in a sentence and let\nthem choose, and file none of them. Never propose from a conversation that did not ask you to. A\nquestion about how something works and a plan you were asked to sketch are not requests to record\nanything, nor is a fix nobody asked you to write a rule about.\n\nAsk before you extrapolate. Ask only what you cannot work out for yourself, ask it all in one\nmessage, and stop at four. Four is a ceiling, not a target: two good ones are better. Then write the\nproposal with what you have and put whatever is still open in its open questions rather than going\nback. Never ask what this repository would answer, such as which file, which test, or which branch,\nand never ask anyone for Balladeer's own identifiers: get_promise_setup carries this repository's id\nand who can own a promise. Never ask again for what they have already told you.\n\nWrite it in their words. Every failure they named out loud is one of the failing examples, in the\nwords they named it. Every other example comes from a situation they actually described; if you\ncannot trace one to something they said, leave it out and say so in the open questions rather than\nwriting a plausible one. Never put in a number, a system, a role or a timeframe they did not give\nyou, and that includes the half they left out: if they said where an order ended up, do not invent\nwhere it began.\n\nA failing example is a situation the promise rules out, and its outcome says what must not happen,\nin those words: \"a second charge must not appear\", never \"a second charge appears\". Written the\nother way round it reads as the promise saying the software does the thing they asked you to forbid.\nAnd a promise says what the software must do for whoever depends on it. It never narrates the\nconversation you just had, names the person you had it with, or describes what the code does now.\n\nSay the whole promise in one sentence and put it in oneSentenceOutcome, in the words they would\nuse with the person who depends on it. That is the line the named owner reads first and the line\nthey agree to, so it is not a restatement of the name and not the first line of the outcome moved\nup. Leave it out rather than inventing one from something they did not say.\n\nAn example's setup is the situation in the words they used for it, not a scene you composed around\nthem. Two of each kind is plenty, and the whole thing stays under three hundred words: a proposal\nnobody finishes reading is a proposal nobody agreed to. The confidence you record is the one you\nactually have.\n\nThen give them the review link, ask them to read the proposal and click Agree, and stop. Never\napprove one yourself. Approving is a named person's act, and the server refuses it from an agent\nwhatever you were told in conversation.\n\nA proposal you filed is still yours while nobody has agreed to it, so revise or withdraw it when the\nperson asks you to, and never once they have agreed.\n\nSay promise and proposal when you talk to them. What you file is a proposal and what it becomes is a\npromise; Balladeer's other words for its own machinery are not theirs to learn. Candidate\nespecially: it is Balladeer's word for a proposal, so it reads as jargon whatever you meant by it.\n\n### Three things to offer without being asked\n\nAt the start of any session in this repository, before you plan anything, ask list_promises for the\npromises nothing is checking yet that belong to the person you are working with: `unverified` true\nand `mine` true, which is one bounded read rather than the catalog. If it returns any, say so in one\nline and offer to build their verifiers now. Name each one by its one-sentence claim rather than by\nan id, so the person can see which behavior is unguarded. Ask the same tool for\n`brokenSinceLastSeen` true as well, and where that returns any, say in one line that those promises\nbroke since they last looked, name each by its claim, and offer to fix them. Then wait for their\nanswer. The offer is the whole of it, and never start building or repairing one because nobody said\nno.\n\nTheirs, and nobody else's. `mine` keeps the promises this person owns or agreed to, and a teammate's\nunguarded promise is that teammate's to hear about: a session that opens by reading out other\npeople's unfinished work reads as an audit of them. `brokenSinceLastSeen` is that person's own by\nconstruction and needs no `mine` beside it. Drop `mine` when this person asks what the rest of the\nteam has promised, and say whose promises you are showing them when you do.\n\nWhen you have proposed promises, show them here as well as there. Put each proposal in the\nconversation in full: its one-sentence claim, who it is for, when it applies and what must then be\ntrue, the numbered cases that must keep working and the ones that must be caught, and every question\nyou left open. Then offer to finish the agreement from here, by asking for a sign-off, giving them\nthe page it returns and taking the one-time code that page shows them, so the browser is needed only\nfor signing. Give them the review link in the same message too, because some people would rather\nread it there and edit it before they agree.\nA promise whose meaning is agreed and which nothing is checking yet is one you can finish. When\nsomebody gives you a promise id, run `npx -y balladeer@latest status <promise id>`, or expand it with get_promise.\nIf it comes back agreed with nothing checking it, say so in one line and offer to prepare and build\nits verifier now. Only if they say yes: `prepare_qualification` mints the one-time setup, or\n`npx -y balladeer@latest prepare <promise id>` mints it and writes it where the sealed run reads it. You need no\nsign-off for that and there is no code to ask anybody for, because agreeing the meaning was their\nact and building the check that proves it is yours. Then build the verifier, seal it, push, and tell\nthem protection starts by itself when that run qualifies. The offer is the whole of it: never\nprepare one because nobody said no.\n\n### When somebody says \"tell me about\" one\n\nAn id is how a person points at something here, and every promise page and every proposal page\ncarries one. A promise id starts with prom_ and a proposal id starts with cand_: expand a promise\nwith get_promise and a proposal with get_proposal, and read neither of them out as a list of fields.\nSay in four or five sentences what it is for, who it is for, when it applies and what must then be\ntrue. Then say where it stands: a promise is agreed, and either protected or not yet checked by\nanything; a proposal is agreed by nobody and waiting on the person it names.\n\nClarify ambiguity that materially changes the behavior during capture, within the existing\nquestion budget; never invent an answer.\n\nOrdinary open questions record uncertainty; they do not prevent the named owner from agreeing to\nthe behavior as written. Agreement does not answer them or add an unstated guarantee. Separate\ncatalog-conflict questions can hold agreement until the owner rules on the stated conflict. Do not\ncall a proposal unready just because it has ordinary open questions. Offer to discuss them if\nuseful; if the person wants to, take them one at a time in the order they come back. For each ordinary question,\nsay what it decides in their words rather than in the question's; give your\nrecommendation and the reason you hold it, drawn from this repository and from the promise itself; and stop there. When\nthey answer, record what they said with resolve_question, in their own words where they gave you\nany, and where their answer changes the promise, follow it with update_proposal, add_case or\nremove_case and tell them what you changed. Never invent an answer or clear uncertainty because\nthey chose to agree. None of that agrees to anything: the named owner agrees, in their own browser\nor through a sign-off you carry, and the questions they answered stay on the proposal in their name.\n\nThe same rule governs every question you leave open in the first place. A question that names a gap\nand stops is a note, and nobody can answer a note. Say what answering it decides, in the words a\ncustomer would use, and carry your own best guess with the reason behind it, so that the shortest\ntrue answer is yes. For example: \"Decides: whether a worker that is running but reconciling nothing\ncounts as an outage this promise covers. Best guess: yes, because the promise is about somebody\nhearing before a customer does, and a wedged worker is invisible to every check this repository has.\nSay yes, or tell me otherwise.\" Balladeer refuses a question filed without both halves and says\nwhich one is missing.\n\n### Files that are sealed, and the one reason to edit one\n\nEvery file under .continuity/promises/ is sealed. The promise that owns that directory records the\nexact bytes of each file in it, so editing one, adding one there, renaming one or deleting one\nbreaks the seal. A promise whose seal is broken stops being checked, and it stays that way until a\nnamed person qualifies it again, which is their afternoon rather than your commit. Nothing in there\nis ordinary source: keep it out of refactors, formatting runs and dependency upgrades.\n\nFind out which directories are sealed before you plan an edit, not after. npx -y balladeer@latest status lists\nthem, and list_promises names a promise's sealed directory on its row once a verifier is bound to\nit. Then, before you push, run npx -y balladeer@latest check-seals. It prints nothing and exits zero when your\nchange touches no seal, and names the promise, its owner and its page when your change would break\none.\n\nThe one reason to edit a sealed file is to repair a verifier that can no longer run: something it\nimports moved, or the language it is written in changed under it. Never edit one to make a failing\ncheck pass. A check going red is the promise doing its job, and the repair for that belongs in the\nbehavior it protects. A repair is not finished until the promise is sealed again with the runner\nthis repository is pinned to. get_promise_setup carries that exact seal command, and\nnpx -y balladeer@latest check-seals prints it beside any promise it names.\n\n### The rest of a promise's life\nWhen a promise is obsolete, finished, deliberately off for a while, or owned by the wrong person,\npropose the change and hand them the promise page. Deciding is theirs.\n\nYou can finish one of those acts here, and only one way. Ask for a sign-off with\nrequest_owner_signoff, give them the page it returns, and ask for the one-time code that page shows\nthem. Then call the act's own tool with that code and their own words. Never call one on your own\ninitiative, never on a general approval of some earlier act, and never ask for a code you were not\ngiven: a refusal is the person's to resolve, not yours to retry. Balladeer records them as the\nperson who acted and you as the messenger.\n\nReport the promise's state exactly as Balladeer reported it: proposed, agreed, or protected, never\none in place of another.\n\n### Where your team watches this\n\nBalladeer is a web app as well as these tools, at the address setup printed. Its catalog lists every\npromise with who owns it and whether anything is checking it, and each promise has a page of its own\nshowing what this team agreed the software must do and then every run that has checked it since,\nnewest first, with the commit each one checked. Whenever there is a link to give, give the link\nrather than a summary of it: the page says what you would have said, and it stays true after this\nconversation has ended.\n\nBalladeer also has a Slack app, which a workspace administrator installs from workspace settings.\nOnce it is installed, whoever owns a promise gets a direct message when theirs goes live and when a\nrun on the default branch breaks it. Until somebody installs it, nothing is sent anywhere, so say it\nis available rather than saying they will be told.\n\n### Questions people ask\n\nAnswer these when they come up. Where you do not know, say so and point at the address setup\nprinted: a confident wrong answer about what a vendor can see is worse than no answer.\n\nWhat it does: it holds the behaviors this team has agreed the software must keep, and reports\nwhether each one is still being kept, from this repository's own tests running in its own CI.\n\nWhat it sees: the text of each promise somebody approves, this repository's numeric ids and the name\nof its default branch, and from CI the pass or fail outcome, the commit checked, and content hashes.\nNever the code, the tests, the fixtures, the logs, the prompts, or the transcripts.\n\nWhat stopping costs: nothing that matters to their tests. The verifier package, its fixtures and the\nworkflow file are theirs, in their repository, running in their CI, and disconnecting changes none\nof them. An administrator can download everything Balladeer holds at any time from workspace\nsettings, and disconnecting hands them that same download in the response that ends access.\n\nWho can approve one: the named person who owns it, in their own browser. Not an administrator on\ntheir behalf, and never you.\n\nWhether it blocks a merge: no. The check is advisory on Balladeer's side, and their own branch\nprotection is what decides whether a failing check stops anything.\n\nWho can invite people and change setup: a workspace administrator. A contributor can read the\nworkspace and propose promises, and a viewer can read it. If somebody asks you to add a teammate,\ninvite_teammate returns the page an administrator sends the invitation from, with the address filled\nin for them to read. The tool sends nothing itself: an administrator presses Send.";
 /**
  * How an agent turns a repository that has just been connected into a catalog a
  * person can read. Setup's last step prints it, and the control plane's `/agent`

package/dist/copy.js

@@ -289,15 +289,21 @@ Say in four or five sentences what it is for, who it is for, when it applies and
 true. Then say where it stands: a promise is agreed, and either protected or not yet checked by
 anything; a proposal is agreed by nobody and waiting on the person it names.
 
-A proposal that still carries open questions is not finished, and settling them is usually why
-somebody asked. Offer to work through them, then take them one at a time in the order they come
-back. For each one, say what it decides in their words rather than in the question's; give your
-recommendation and the reason you hold it, drawn from this repository and from the promise itself;
-and stop there. When they answer, record what they said with resolve_question, in their own words
-where they gave you any, and where their answer changes the promise, follow it with update_proposal,
-add_case or remove_case and tell them what you changed. None of that agrees to anything: the named
-owner agrees, in their own browser or through a sign-off you carry, and the questions they answered
-stay on the proposal in their name.
+Clarify ambiguity that materially changes the behavior during capture, within the existing
+question budget; never invent an answer.
+
+Ordinary open questions record uncertainty; they do not prevent the named owner from agreeing to
+the behavior as written. Agreement does not answer them or add an unstated guarantee. Separate
+catalog-conflict questions can hold agreement until the owner rules on the stated conflict. Do not
+call a proposal unready just because it has ordinary open questions. Offer to discuss them if
+useful; if the person wants to, take them one at a time in the order they come back. For each ordinary question,
+say what it decides in their words rather than in the question's; give your
+recommendation and the reason you hold it, drawn from this repository and from the promise itself; and stop there. When
+they answer, record what they said with resolve_question, in their own words where they gave you
+any, and where their answer changes the promise, follow it with update_proposal, add_case or
+remove_case and tell them what you changed. Never invent an answer or clear uncertainty because
+they chose to agree. None of that agrees to anything: the named owner agrees, in their own browser
+or through a sign-off you carry, and the questions they answered stay on the proposal in their name.
 
 The same rule governs every question you leave open in the first place. A question that names a gap
 and stops is a note, and nobody can answer a note. Say what answering it decides, in the words a

package/dist/store.d.ts

@@ -11,8 +11,12 @@ export type PendingPairing = Readonly<{
     verificationUri: string;
     expiresAt: string;
     startedAt: string;
+    /** Local pairing intent; lets an explicit workspace choice resume its own code. */
+    workspaceChoice?: string;
 }>;
 export type StoredSession = Readonly<{
+    /** Claimed local create intent, not workspace identity or server authority. */
+    workspaceChoice?: string;
     controlPlane: string;
     sessionId: string;
     token: string;
@@ -74,6 +78,8 @@ export declare function credentialsPath(environment?: NodeJS.ProcessEnv): string
  */
 export declare function assertSafeStoreLocation(directory: string, environment?: NodeJS.ProcessEnv, cwd?: string): void;
 export declare function ensureStoreDirectory(environment?: NodeJS.ProcessEnv): string;
+/** Check actual write permission before requesting a pairing whose secret must be saved. */
+export declare function assertStoreWritable(environment?: NodeJS.ProcessEnv): void;
 export declare function readCredentials(environment?: NodeJS.ProcessEnv): Credentials;
 /**
  * Atomic, and never through a path an attacker could have pre-planted: the

package/dist/store.js

@@ -86,12 +86,38 @@ function assertPrivate(path, kind) {
         throw new StoreError("credential_store_unsafe", `${path} is readable by other users. Run: chmod ${kind === "directory" ? "700" : "600"} ${path}`);
     }
 }
+function storeAccessError(error, path, operation) {
+    if (error instanceof StoreError)
+        return error;
+    const code = error?.code;
+    const permissionDenied = code === "EACCES" || code === "EPERM";
+    return new StoreError(operation === "read" ? "credential_store_unreadable" : "credential_store_unwritable", permissionDenied
+        ? `Balladeer cannot ${operation} its local credential store at ${path}. Allow this command to access that private folder in your terminal or coding agent, then rerun the same command. Existing saved credentials were not replaced.`
+        : `Balladeer could not ${operation} its local credential store at ${path} (${code ?? "filesystem error"}). Check that this location is a private, accessible directory, then retry. Existing saved credentials were not replaced.`, { cause: error });
+}
 export function ensureStoreDirectory(environment = process.env) {
     const directory = configHome(environment);
-    assertSafeStoreLocation(directory, environment);
-    mkdirSync(directory, { recursive: true, mode: 0o700 });
-    assertPrivate(directory, "directory");
-    return directory;
+    try {
+        assertSafeStoreLocation(directory, environment);
+        mkdirSync(directory, { recursive: true, mode: 0o700 });
+        assertPrivate(directory, "directory");
+        return directory;
+    }
+    catch (error) {
+        throw storeAccessError(error, directory, "write");
+    }
+}
+/** Check actual write permission before requesting a pairing whose secret must be saved. */
+export function assertStoreWritable(environment = process.env) {
+    const probe = `.write-check-${randomBytes(8).toString("hex")}`;
+    const path = join(configHome(environment), probe);
+    try {
+        writeStoreFile(probe, "", environment);
+        unlinkSync(path);
+    }
+    catch (error) {
+        throw storeAccessError(error, configHome(environment), "write");
+    }
 }
 export function readCredentials(environment = process.env) {
     const path = credentialsPath(environment);
@@ -101,9 +127,10 @@ export function readCredentials(environment = process.env) {
         raw = readFileSync(path, "utf8");
     }
     catch (error) {
-        if (error instanceof StoreError)
-            throw error;
-        return { ...EMPTY, pendingPairings: [], sessions: [], agents: [] };
+        if (error?.code === "ENOENT") {
+            return { ...EMPTY, pendingPairings: [], sessions: [], agents: [] };
+        }
+        throw storeAccessError(error, path, "read");
     }
     try {
         const parsed = JSON.parse(raw);
@@ -164,9 +191,7 @@ export function writeStoreFile(fileName, contents, environment = process.env) {
         catch {
             // The temporary file may never have been created.
         }
-        if (error instanceof StoreError)
-            throw error;
-        throw new StoreError("credential_store_unwritable", `Balladeer could not write ${path}: ${error instanceof Error ? error.message : "unknown reason"}`, { cause: error });
+        throw storeAccessError(error, path, "write");
     }
 }
 /** Origins, not URL strings, so `https://host/../` cannot alias a stored entry. */

package/dist/wire.d.ts

@@ -5,10 +5,10 @@
  * runtime dependency at all: Node 22 builtins and global fetch, nothing else.
  * A contract test compares the scope list below against the server's.
  */
-export declare const CLI_VERSION = "1.0.2";
+export declare const CLI_VERSION = "1.0.4";
 export declare const CLI_INVOCATION = "npx -y balladeer@latest";
 export declare const CLIENT_HEADER = "x-balladeer-client";
-export declare const CLIENT_HEADER_VALUE = "balladeer/1.0.2";
+export declare const CLIENT_HEADER_VALUE = "balladeer/1.0.4";
 export declare const DEFAULT_CONTROL_PLANE = "https://envelopes.balladeer.ai";
 export type DelegatedScope = "repository:enroll" | "agent:issue" | "ci:connect" | "workspace:invite" | "candidate:propose";
 export declare const DELEGATED_SCOPES: readonly DelegatedScope[];

package/dist/wire.js

@@ -5,7 +5,7 @@
  * runtime dependency at all: Node 22 builtins and global fetch, nothing else.
  * A contract test compares the scope list below against the server's.
  */
-export const CLI_VERSION = "1.0.2";
+export const CLI_VERSION = "1.0.4";
 export const CLI_INVOCATION = "npx -y balladeer@latest";
 export const CLIENT_HEADER = "x-balladeer-client";
 export const CLIENT_HEADER_VALUE = `balladeer/${CLI_VERSION}`;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "balladeer",
-  "version": "1.0.2",
+  "version": "1.0.4",
   "description": "Set up Balladeer from your terminal, or from a coding agent's.",
   "license": "Apache-2.0",
   "private": false,
@@ -26,5 +26,8 @@
   },
   "devDependencies": {
     "typescript": "5.9.3"
+  },
+  "dependencies": {
+    "@iarna/toml": "2.2.5"
   }
 }