Files
stack/docs/DEVELOPER-GUIDE/architecture/fleet-first-start-identity.md
T
2026-08-16 19:01:26 -05:00

82 lines
4.3 KiB
Markdown

# Fleet First-Start Identity Boundary
> **Status:** Implemented by issue #1264. Requirements: `FCM-REQ-12`, `AC-FCM-10`.
## Problem
`launchRuntime()` called `checkSoul()` before runtime execution. A missing top-level `SOUL.md`
caused `checkSoul()` to spawn a child `mosaic wizard` with inherited stdio. Under a systemd-created
fleet pane with no TTY, that child blocked or failed before the runtime boundary even though generic
`defaults/SOUL.md` and `defaults/USER.md` already shipped in the same `MOSAIC_HOME`.
## Chosen boundary
The fix remains at `checkSoul()` and does not add flags to `yolo`, fleet commands, systemd units, or
`start-agent-session.sh`:
1. A present, nonblank, whitespace-exact `MOSAIC_AGENT_NAME` selects the fleet path.
2. `resolveFleetIdentity()` must resolve that exact member through the existing roster/helper
boundary, and any defined `MOSAIC_AGENT_CLASS` (including blank/whitespace) must canonicalize to
the roster class, before any identity seed. Only undefined means absent.
3. `lstatSync()` preflights every destination directory entry without following links, so a dangling
link is rejected before its counterpart can be published.
4. Safe bounded snapshots are read from only the missing contracts under `defaults/`.
5. Each snapshot is written to a random owner-private temporary file in `MOSAIC_HOME`.
6. `linkSync()` publishes the complete file without overwriting an existing path. `EEXIST` means a
concurrent seat or operator won; the existing path is preserved and revalidated.
7. Temporary files are removed, and both installed contracts are re-opened through the no-symlink
secure-file reader before launch continues.
8. `composeContract()` independently re-resolves the roster, securely reads fleet `USER.md` through
a Linux descriptor at the point of use, and injects exact member identity and communications data.
A standalone launch with no `MOSAIC_AGENT_NAME` retains the portable tolerant USER read and the
interactive wizard. Fleet-only no-follow enforcement must not make supported standalone macOS
composition depend on Linux `/proc` descriptor traversal.
## Identity and authority
The copied defaults deliberately say “Mosaic agent”; they are a generic behavioral base. They are
not the source of a fleet seat's identity. The canonical roster controls:
- exact agent/session name;
- canonical role/class and persona;
- peer rows and point of contact;
- tmux socket and helper target; and
- communications generation.
An unknown/padded ambient name, mismatched class, or explicitly blank/whitespace class fails before
any file is seeded. This avoids replacing the interactive wall with a fleet of indistinguishable or
ambiently invented identities.
## Concurrency and filesystem properties
- Sources and final destinations are bounded regular files beneath `MOSAIC_HOME`; target and dangling
symlinks are not followed.
- New files have mode `0600`.
- Hard-link publication is same-filesystem, atomic, and no-clobber.
- A temporary path is removed only when this process successfully created it.
- All required source snapshots are validated before the first destination is published, preventing
a missing second default from leaving a partial seed.
- Existing operator files are never chmodded or rewritten.
## Verification
`src/commands/launch-first-start.spec.ts` uses the production-kind boundary: the real built CLI in a
no-TTY subprocess, not a direct wizard test. The package `test:vitest` gate builds Mosaic before
Vitest, while the clean-checkout command builds its workspace dependencies first, so ignored
`dist/cli.js` cannot be absent or stale. A fake lease launcher records whether execution reached the
runtime boundary and captures the composed prompt.
Positive and negative cases prove the check can both proceed and refuse. Fleet composition coverage
replaces a previously validated `USER.md` with an external symlink and proves point-of-use refusal;
a standalone unreadable-optional-USER case proves the portable tolerant branch remains separate.
Real Pi authentication and provider task execution remain environment tests, not claims of this
fixture.
## Non-goals
- Runtime installation or pane-PATH resolution (#1256/#1258).
- The held `~/.mosaic` launch-composition layer in PR #1213.
- Personalizing the operator's standalone identity without a wizard.
- Changing fleet systemd or shell launcher code.