Files
platform/src/servers/data-path.ts
T
pastilhasandClaude Opus 5 b7112f6e5d a commander records what it is and what it spawned, so deleting one means it
Without this record a commander cannot be stopped or fully deleted, and the delete route's comment claimed
otherwise: "the running background process (if any) dies on its own". It does not. Nothing in that route
touched a process, so a deleted commander left its loop running — still polling, still able to fire, still
able to spawn a deploying agent, with no chip to stop it from and no row left to disarm.

THE RECORD, and why the loop has to write it. Three ways to recover the link were tried and none works: the
PROCESS holds nothing (no open transcript descriptor, nothing identifying in its environment — tested); the
TRANSCRIPT is indistinguishable, because queue-operation/ai-title/last-prompt look like platform markers and
are the harness's own, appearing in every session equally; CONTENT matching finds ordinary chats that quote a
procedure as readily as the agent following it. So the loop records, because it is the only party that knows.

Two files under DATA_PATH/<email>/commander-state/<session>/: `loop-pid` and `spawned`. The reader is
deliberately tolerant — a truncated write or a stray line is dropped, not fatal, because a record that failed
closed would make a commander undeletable over one bad line.

STOPPING needs no new mechanism, only the pid: descendantPids + SIGTERM from the loop takes the loop and any
agent under it, which is exactly what that helper was written for. A commander with no recorded pid stops
nothing and SAYS so — the honest answer for one armed before this existed.

DELETE now cascades in a deliberate order: stop the processes, then the children's transcripts, then the
commander. Stopping first because a live agent writing to a transcript being deleted is a race, and because
the loop could otherwise spawn a replacement between the last two steps. It returns counts rather than a bare
ok — this is the one irreversible action here.

The children's transcripts are deleted with the EXISTING deleteClaudeSession: its cwd argument is documented
as "a hint, not an authority", so it locates a child that ran in the platform checkout rather than the
commanders directory. No new function.

THE CHAIN'S WEAK LINK was that the loop cannot know which commander it belongs to. The paths are now handed
to it in the background-job prompt, fetched from a new endpoint because they are rooted in DATA_PATH — a
server fact. Paths rather than an upload endpoint on purpose: a shell redirect needs no credential, and a
background job has no business holding one.

The prompt rule is numbered properly rather than landing on 7.5, and the output rule shifts down one when it
is present. Verified both shapes produce consecutive numbers.

The manager now exposes claudeSessionId, because a commander is keyed on the TRANSCRIPT id everywhere and
officer's own session handle is not interchangeable — that mismatch is what once rendered a working commander
as an empty pane.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-22 17:18:30 +00:00

287 lines
16 KiB
TypeScript

import { join, resolve } from 'node:path';
import { chmodSync, existsSync, mkdirSync, readFileSync, readdirSync, statSync } from 'node:fs';
import { homedir } from 'node:os';
// ── The install root, and why it is derived rather than configured ──
//
// An install is one directory with everything under it:
//
// $OFFICER_ROOT/
// platform/ the repo — the working directory every process runs in
// data/ DATA_PATH
// capabilities/ OFFICER_ITEMS_DIR
// dockers/ what the app store provisions
//
// These were three environment variables until 2026-08-12, which meant three answers that had to agree
// with each other and with the layout on disk. They are one fact now: the root is the parent of the
// working directory, and everything else is a fixed name under it. Nothing to set and nothing to
// disagree.
//
// This depends on the working directory being the repo, which is why pm2 pins `cwd` in
// ecosystem.profile.cjs — read the comment there before changing either. `assertInstallLayout` below
// is the check that says so out loud instead of silently writing to the wrong place.
export const OFFICER_ROOT = resolve(process.cwd(), '..');
/**
* The repo itself — the working directory, named rather than re-derived at each call site.
*
* Plugins live under this rather than beside it (`platform/plugins/<app-name>/`), which is the whole
* developer story: bun links the workspace packages into the root `node_modules`, so anything inside the
* repo can `import { useClient } from 'hooks/useClient'` with no publishing and no version negotiation.
* A plugin one directory higher would resolve none of it.
*/
export const PLATFORM_DIR = process.cwd();
export const DATA_PATH = join(OFFICER_ROOT, 'data');
// Unified, file-based store for all agent items, living outside the repo. Every skill/tool/task/
// process/extension is a directory under one of these type subfolders — no scope tiers, no DB.
export const OFFICER_ITEMS_DIR = join(OFFICER_ROOT, 'capabilities');
/**
* Refuse to boot when the working directory is not the platform repo.
*
* Same posture as `assertPermissionTotality` and `assertSecretsClosed`: a prerequisite that silently
* not holding is worse than one that fails. Every path in this file hangs off `resolve(process.cwd(), '..')`,
* so a process started from the wrong directory does not error — it computes a plausible root somewhere
* else and writes managed homes, capabilities and agent runs into it. The install looks empty and the
* data looks lost, with nothing naming the cause.
*
* `ecosystem.profile.cjs` already pins `cwd` for exactly this reason. This is the check that the pin
* still works, which is the part that was missing when the ecosystem files moved directory.
*/
export function assertInstallLayout(): void {
const manifest = join(process.cwd(), 'package.json');
let name: string | undefined;
try {
name = JSON.parse(readFileSync(manifest, 'utf8')).name;
} catch {
// Absent or unreadable is the same answer as wrong: this is not the repo.
}
if (name === 'officer') return;
throw new Error(
[
`Officer must run from the platform repo, but the working directory is ${process.cwd()}`,
'',
` expected a directory containing the platform's package.json ("officer")`,
` found ${name ? `package.json for "${name}"` : 'no readable package.json'}`,
'',
'Every path is derived from this — the install root is its parent, and data/, capabilities/ and',
`dockers/ hang off that. Continuing would write to ${OFFICER_ROOT} instead of the real install.`,
'',
'Under pm2 this means the `cwd` pin in ecosystem.profile.cjs no longer points at the repo.',
].join('\n'),
);
}
export type ItemType = 'skills' | 'tools' | 'tasks' | 'processes' | 'extensions' | 'agents';
export const ITEM_TYPES: ItemType[] = ['skills', 'tools', 'tasks', 'processes', 'extensions', 'agents'];
export const itemsDir = (type: ItemType) => join(OFFICER_ITEMS_DIR, type);
export const ensureItemDirs = () => {
for (const type of ITEM_TYPES) mkdirSync(itemsDir(type), { recursive: true });
};
export const SERVER_CONFIG_DIR = join(DATA_PATH, 'server-settings');
// Every run of a given agent shares one working directory. That is deliberate and load-bearing: the
// `claude` CLI groups transcripts by cwd (see api/chat/claude-sessions.ts), so a shared cwd is what
// makes an agent's runs show up as their own project group in /chat, listed newest-first, with no
// database row anywhere. The accumulated CLAUDE.md / LEARNINGS.md for the agent live here too.
export const getAgentRunsDir = (dirName: string) => join(DATA_PATH, 'agentic_runs', dirName);
// The `pi`/opencode agent's own config dir (auth.json + models.json). External-tool path.
export const AGENT_CONFIG_DIR = join(homedir(), '.pi', 'agent');
export const SEED_PATH = resolve(import.meta.dir, '../../seed');
// The managed home under DATA_PATH. A remnant of the first architecture, where every user ran inside
// their own Docker container and this was that container's home — seeded by provisioning, described to
// the agent by a generated CLAUDE.md. Both of those are gone, and the OWNER's sessions never come here
// any more — terminals, chats and task runs all use getOwnerHomeDir below.
//
// It is not dead, though: user-home.ts returns it for a NON-owner, where it is deliberately the same
// path as osUserHome, and pipeline-executor still calls it.
//
// It was also getOwnerHomeDir's fallback until 2026-08-12, which is the only reason an unset HOME_DIR
// used to run the owner's terminals in a directory nobody meant.
export const getHomeDir = (email: string) => join(DATA_PATH, email, 'home');
// Where the owner's sessions actually run: their real login home, so platform terminals/chats/tasks
// share config and credentials with the shell they use outside Officer.
//
// `homedir()` is right here for one reason, and it is worth stating because everything below rests on
// it: the server process runs AS the owner. It is not a general "whose home is this" helper — a member
// never reaches this function, because their sessions go through os-user.ts and setpriv. It ignores the
// email it is passed, which it also did before; the callers that must not are already commented as such.
//
// ── Why this is captured once, and not read per call ──
//
// Measured on bun 1.3.10: BOTH `os.homedir()` and `os.userInfo().homedir` return $HOME when it is set,
// rather than reading the password file. And `sidecar/claude/user-instance.ts` assigns `process.env.HOME`
// on its way to spawning an agent. So a lazy read here would hand back whichever home was most recently
// spawned into — the owner's on the first call and something else afterwards.
//
// This module imports nothing but node builtins, so it is evaluated before any of that can run. The
// value is the owner's home, taken while $HOME still means what it says.
const OWNER_HOME = homedir();
// This was HOME_DIR in .env until 2026-08-12, whose absence fell back to getHomeDir — the managed home
// under DATA_PATH, not a login home at all. So forgetting to set it did not fail; it quietly ran every
// terminal somewhere else.
export const getOwnerHomeDir = (_email: string): string => OWNER_HOME;
// The directory skeleton a new account gets under DATA_PATH.
//
// Most of these are also created on demand by whichever feature owns them, so pre-creating them buys
// legibility more than function — the tree shows what an account has without it having to be used first.
// `home` is the exception and the reason this exists: nothing else creates it, and it is where a
// non-owner's sessions would run.
//
// Single-sourced here rather than in the script that used to own the list, because there are now two
// callers — `scripts/provision-user-dirs.ts` and the owner's create-account handler — and a skeleton
// that differs depending on how the account was made is a bug nobody would think to look for.
export const USER_DIRS = ['home', 'attachments', 'cache', 'dashboards', 'email_accounts', 'logs', 'sidecar'] as const;
/**
* Create an account's root and its skeleton, closed by default.
*
* Keyed on email because that is what the on-disk layout uses everywhere else (`DATA_PATH/<email>/…`).
* Renaming an account's email would orphan its directory; that is pre-existing and not this function's
* problem, but it is the reason nothing here derives a path from the id.
*
* ── Why the modes are set here and not only by os-user.ts ──
*
* `711` on the account directory, `700` on everything inside it. Measured while testing per-user Linux
* accounts: at the default umask these came out `755`, and a member with a shell could read ANOTHER
* member's home directory just by naming it — the parent being unlistable is not protection when the
* child itself is world-readable. "Locked unless something opens it" has to be the resting state, so it
* belongs at creation rather than in the confinement pass, which only ever runs for accounts that have an
* OS user.
*
* `chmod` explicitly rather than mkdir's `mode`, which is masked by the umask and does nothing at all for
* a directory that already exists.
*/
export const provisionUserDirs = (email: string): void => {
const accountDir = join(DATA_PATH, email);
for (const dir of USER_DIRS) mkdirSync(join(accountDir, dir), { recursive: true });
// Traversable, not listable: reaching `home` must not mean enumerating the platform's tree beside it.
chmodSync(accountDir, 0o711);
for (const dir of USER_DIRS) {
try {
chmodSync(join(accountDir, dir), 0o700);
} catch {
// A directory that is no longer OURS to chmod. `home` becomes the member's on the first successful
// provision, and `chmod` requires ownership — so re-running this threw EPERM and took every RETRY down
// before it began, which is how this was found. os-user.ts sets the home's mode through sudo and is the
// authority for it; here the mode is a default for directories we are creating, not an assertion about
// ones that already exist.
}
}
};
export const getTmpAttachmentsDir = (email: string) => join(DATA_PATH, email, 'attachments', 'tmp');
export const getAttachmentsDir = (email: string, sessionId: string) => join(DATA_PATH, email, 'attachments', sessionId);
// Per-account email storage: DATA_PATH/<owner>/email_accounts/<accountEmail>/emails.db, with a
// single attachment_cache/ shared across the owner's accounts.
export const getEmailAccountsDir = (ownerEmail: string) => join(DATA_PATH, ownerEmail, 'email_accounts');
export const getEmailDbPath = (ownerEmail: string, accountEmail: string) =>
join(getEmailAccountsDir(ownerEmail), accountEmail, 'emails.db');
export const getEmailAttachmentCacheDir = (ownerEmail: string) =>
join(getEmailAccountsDir(ownerEmail), 'attachment_cache');
/**
* Every commander a user has shares this ONE directory as its cwd — deliberately not one directory per
* commander. The first design was per-commander, and was dropped the same night it was proposed: a fresh
* directory would have to exist before or exactly when a commander's first turn fires, and racing that
* creation against the harness's own spawn is precisely the kind of timing hazard Claude Code's
* directory-per-project convention was never built to be driven by. One shared, pre-existing directory
* removes the race instead of handling it.
*
* `DATA_PATH`-rooted even for the OWNER, which is a deliberate departure from ordinary `/chat` — an owner's
* normal chat runs in `getOwnerHomeDir` (their real login home). Commanders use the managed directory
* instead, on both sides of the account split, which is what keeps them out of the owner's own project
* transcripts as surely as `command_hq_sessions` keeps them out of the `/chat` list. See `docs/command-hq.md`.
*
* **A sibling of `home`, but reachable like `home` — not confined like every other sibling.** Every other
* directory alongside a member's `home` (`attachments`, `email_accounts`, …) is 700 and owned by the
* SERVICE user, unreachable by the member "even by name" — correct for platform-written data, and exactly
* wrong for a directory a member's own `claude` must `cd` into via `setpriv`. `confineUserTree`
* (`os-user.ts`) chowns this one to the member the same way it chowns `home`. An account provisioned
* BEFORE this existed will not have that chown until it is re-provisioned once — see `docs/command-hq.md`.
*/
export const getCommandersDir = (email: string) => join(DATA_PATH, email, 'commanders');
/**
* Where a commander records what it spawned — one directory per commander, named by its session id.
*
* Beside the transcripts rather than inside them: a transcript is the harness's file and the platform does
* not write into it. `spawned` is a plain newline-delimited list of session ids, appended by the loop the
* commander started, because nothing else CAN know them — a spawned agent's id is not recoverable from its
* process (no open descriptor, nothing in its environment) and its transcript is indistinguishable from any
* other session's.
*
* That makes this record load-bearing rather than convenient. Without it a commander cannot be deleted
* completely, its children cannot be followed, and its cost cannot be totalled.
*/
export const getCommanderStateDir = (email: string, sessionId: string) =>
join(DATA_PATH, email, 'commander-state', sessionId);
/**
* Which mail account's store to use, found on DISK rather than by asking the database.
*
* Email is a plugin (`plugins/email/`) as of 2026-08-18, and the platform may not import a plugin's
* queries — that direction is the one the plugin system forbids, and here it is also the direction that has
* already broken this server: `getEmailAccounts` was awaited at the top level of the agent sidecar, and
* when the table went away on 2026-08-13 the rejection took `officer-agent` into a PM2 restart loop with
* every live chat session inside it (`fbb6d6c7`).
*
* Two callers still legitimately need an answer — the agent's `email_db` MCP tool, and a chat opened
* against a mailbox — and neither needs to know anything about mail to get one. The directory layout is
* already the platform's (see `USER_DIRS`), so reading it answers the question with no coupling: no import,
* no query, and nothing to throw if the plugin is not installed.
*
* Newest by mtime, because the most recently synced account is the one a fresh question is most likely
* about. Returns the DIRECTORY NAME (the account's address), which is what both callers compose with.
*/
export function findEmailAccountName(ownerEmail: string): string | null {
try {
const root = getEmailAccountsDir(ownerEmail);
const candidates = readdirSync(root, { withFileTypes: true })
.filter((entry) => entry.isDirectory() && entry.name !== 'attachment_cache')
.map((entry) => ({ name: entry.name, db: join(root, entry.name, 'emails.db') }))
.filter((candidate) => existsSync(candidate.db))
.map((candidate) => ({ ...candidate, at: statSync(candidate.db).mtimeMs }))
.sort((a, b) => b.at - a.at);
return candidates[0]?.name ?? null;
} catch {
// No directory, no permission, no plugin. All of them mean "no account", which is a normal state.
return null;
}
}
// Sanitises a display username or email into a bare, lowercase, shell-safe token. The name and the
// 32-char Linux limit are the last trace of the per-container architecture, where this really did name
// a Linux user inside the user's container. Nothing creates a Linux user now — the value is carried
// through the websocket/job payloads and ends up only as a claim inside the signed task token, so this
// is a sanitiser rather than an account name. Left in place because unpicking it means changing what
// goes into that token and into WSData, which is a wider change than a cleanup.
export const toShellUsername = (username: string, email: string): string => {
const raw = username || email.split('@')[0]!;
// Replace invalid chars, lowercase, truncate to 32 chars
return (
raw
.replace(/@.*$/, '')
.replace(/[^a-zA-Z0-9._-]/g, '_')
.toLowerCase()
.slice(0, 32) || 'officer'
);
};