fix(#1264): harden unattended identity bootstrap
ci/woodpecker/pr/ci Pipeline failed

This commit is contained in:
goals
2026-08-16 18:33:39 -05:00
parent 43fa047787
commit 9dc90be7e1
14 changed files with 448 additions and 288 deletions
@@ -1,185 +1,201 @@
# #1264 Unattended Fleet First-Start Verification
> Status: **IN PROGRESS** | Executor: goals | Date: 2026-08-16 | Target: isolated local fixtures only
> Status: **IN PROGRESS — review remediation complete locally; push/re-review pending** | Executor:
> goals | Date: 2026-08-16 | Target: isolated local fixtures only
## Objective
Verify that a named fleet seat launched through a systemd-equivalent, no-TTY environment on a clean
host reaches its runtime boundary without an interactive Mosaic identity wizard. Preserve standalone
wizard behavior and canonical roster ownership of exact seat identity.
wizard behavior and canonical-roster ownership of exact seat identity.
## Source evidence accepted for local verification
Daphne's canary investigation was read from jarvis-brain commit
`6c0b6fc70ae6a179a1b7ff9dedfc54e9adccd19a`, report
`docs/reports/2026-08-16_sbx-canary-greenfield-e2e.md`. It measured:
Daphne's canary Run-7 report is reachable from jarvis-brain `origin/main` at
`8bf94afeb8c7d5df96cdd4a4508e75a1d2999710`,
`docs/reports/2026-08-16_sbx-canary-greenfield-e2e.md`. The earlier local object
`6c0b6fc70ae6a179a1b7ff9dedfc54e9adccd19a` is not reachable from an origin ref and is not used as
shipping provenance. Run 7 measured:
```text
systemd -> start-agent-session.sh -> mosaic yolo pi (PID 3726)
-> child mosaic wizard (PID 3762)
```
The canary pane remains preserved and was not accessed. Product behavior is independently tested here
with temporary roots and fake runtime executables; no canary or installed-host inference is scored as
local PASS evidence.
The pane was preserved when Run 7 was captured. Formal review ID 168 records that an authorized
rollback occurred later. This task never accessed or altered the canary VM, pane, snapshot, or
rollback state. Product behavior is independently tested here with temporary roots and fake runtime
executables.
## Controls
- Worktree base: `origin/next@476db12b92971634b67fd2057b7577ee5894e449`.
- `DATABASE_URL` remains unset.
- No real credential, token, provider, VM, installed Mosaic tree, unit, timer, PATH profile, or live
tmux session is read or mutated.
- Tiny's concurrent runtime-preflight and `start-agent-session.sh` PATH work are out of scope.
- Original base: `origin/next@476db12b92971634b67fd2057b7577ee5894e449`.
- PR: #1268, first pushed head `43fa0477877e0d0f110da8d11c3033b40ddeb191`.
- `DATABASE_URL` remains unset for local tests.
- No runtime/provider credential or token value, VM, installed Mosaic tree, unit, timer, PATH profile,
or live tmux session is read or mutated. Standard Gitea/Woodpecker wrappers authenticate metadata
reads/writes without exposing credential values.
- Tiny's runtime-preflight and `start-agent-session.sh` PATH work remain out of scope.
- Held PR #1213 is not a dependency.
## Requirements-to-evidence map
| Acceptance criterion | Method | Evidence |
| ---------------------------------------------------------------------- | ----------------------------------------------------- | ------------------------------ |
| No-TTY fleet first start avoids wizard and reaches runtime | Real built-CLI subprocess with piped stdin | Focused GREEN, CLI test 1 |
| Missing top-level identity files are initialized from shipped defaults | Exact-byte and `0600` assertions | Focused GREEN, CLI tests 1/11 |
| Exact seat identity remains roster-owned | Captured argv plus unknown/padded/blank-name refusals | Focused GREEN, CLI tests 1/69 |
| Existing operator identity is never overwritten | Custom bytes/mode with defaults removed | Focused GREEN, CLI tests 2/3 |
| Concurrent/repeated first start is safe | Four parallel CLIs plus repeated launch | Focused GREEN, CLI tests 2/11 |
| Missing/unsafe defaults fail without prompting | Missing, symlink, oversized, and installed-link tests | Focused GREEN, unit/CLI tests |
| Standalone launch retains wizard | Same real CLI without fleet identity | Focused GREEN, CLI test 10 |
| Acceptance criterion | Method | Evidence |
| ---------------------------------------------------------------------- | --------------------------------------------------------------- | ----------------------------- |
| No-TTY fleet first start avoids wizard and reaches runtime | Exact-source built CLI with piped stdin | CLI GREEN |
| Missing top-level identity files are initialized from shipped defaults | Exact-byte and `0600` assertions | CLI + filesystem GREEN |
| Exact seat identity remains roster-owned | Captured argv; name/class mismatch no-side-effect refusals | CLI GREEN |
| Existing operator identity is never overwritten | Custom bytes/mode with defaults removed | CLI + filesystem GREEN |
| Concurrent/repeated first start is safe | Four parallel CLIs plus repeated launch | CLI GREEN |
| Missing/unsafe defaults and destinations fail before partial mutation | Missing, target/dangling symlink, oversized, invalid-root cases | Filesystem/CLI GREEN |
| Validated `USER.md` cannot be replaced by an external symlink | Seed, replace, compose at point of use | Composition GREEN |
| Standalone launch retains wizard | Same built CLI without fleet identity | CLI GREEN |
| Built-CLI evidence cannot use stale ignored `dist/` | Build-with-dependencies gate before Vitest | Package script + command gate |
## Command evidence
## Initial RED
### Worktree helper refusal and sanctioned fallback
Command:
```bash
~/bin/mosaic-worktree.sh new fix/1264-fleet-unattended-first-start --from origin/next
```
Exit: `1`. Stderr was retained; the helper refused because the derived worktree path was under
`/var/home/jason.woltje`, while `/src` does not exist on this host. Fred explicitly authorized the
plain-git fallback and path used for this task.
Command:
```bash
git -C /var/home/jason.woltje/src/stack worktree add \
/var/home/jason.woltje/agent-work/1264-unattended-first-start \
-b fix/1264-fleet-unattended-first-start origin/next
```
Exit: `0`; HEAD `476db12b92971634b67fd2057b7577ee5894e449`.
### RED
Production source remained unchanged after adding the reproducer. The built CLI represented
`origin/next@476db12` behavior.
Command:
Production source remained unchanged after adding the first reproducer. The CLI was built from
`origin/next@476db12` before the test.
```bash
env -u DATABASE_URL pnpm --filter @mosaicstack/mosaic exec vitest run \
src/commands/launch-first-start.spec.ts
```
Exit: `1`.
Exit `1`; one file and one test failed. Output included:
```text
Test Files 1 failed (1)
Tests 1 failed (1)
[mosaic] SOUL.md not found. Running setup wizard...
◆ What would you like to do?
[mosaic] Setup failed. Run: mosaic wizard
AssertionError: expected 1 to be +0
```
The fixture used piped stdin (not a TTY), a temporary `HOME`/`MOSAIC_HOME`, shipped default bytes,
a canonical one-seat roster, a fake `pi`, and a fake lease-runtime boundary. The wizard rendered and
the runtime-boundary capture was never created. Complete combined stdout/stderr was retained at
`/tmp/1264-red.out` during execution.
The fake runtime-boundary capture was not created. Complete stdout/stderr was retained at
`/tmp/1264-red.out` during that work session.
### GREEN and baseline
## Formal-review remediation RED
After the production change, the original one-test command exited `0` with `1/1` passing. The final
focused command was:
Daphne's exact-head review ID 168 requested changes at `43fa0477`. Before changing production code,
new regressions were run against an exact-source build. Four tests failed while the existing 1,568
passed:
1. valid roster name plus mismatched ambient class seeded both files before refusal;
2. dangling `SOUL.md` allowed `USER.md` to be published before refusal;
3. dangling `USER.md` allowed `SOUL.md` to be published before refusal; and
4. replacing a securely validated `USER.md` with an external symlink was followed by composition.
This establishes that all four reviewer findings were observable on the pushed implementation.
## Final GREEN
The production-kind command builds Mosaic and all workspace dependencies before invoking Vitest,
because `dist/` is ignored and may otherwise be absent or stale:
```bash
env -u DATABASE_URL pnpm --filter @mosaicstack/mosaic exec vitest run \
src/commands/fleet-first-start-identity.spec.ts \
src/commands/launch-first-start.spec.ts \
src/commands/launch.spec.ts \
src/commands/compose-contract.spec.ts \
src/config/file-adapter.test.ts \
src/cli-smoke.spec.ts
env -u DATABASE_URL sh -c '
pnpm --filter @mosaicstack/mosaic... build &&
pnpm --filter @mosaicstack/mosaic exec vitest run \
src/commands/fleet-first-start-identity.spec.ts \
src/commands/launch-first-start.spec.ts \
src/commands/launch.spec.ts \
src/commands/compose-contract.spec.ts \
src/config/file-adapter.test.ts \
src/cli-smoke.spec.ts
'
```
Exit: `0`; `6/6` files and `119/119` tests passed. The 11 production-kind CLI tests cover no-TTY
launch, captured exact roster name/class, byte-equal `0600` seeds, no-clobber/idempotence, partial
seed, missing/symlink defaults, unknown/padded/blank ambient members, standalone interactive control,
and four concurrent first starts with no temporary residue. Nine direct filesystem tests cover
successful/no-clobber hard-link publication, unexpected link errors, source prevalidation, existing
operator contracts, idempotence, symlink sources/destinations, and oversized input.
Exit `0`: `6/6` files, `124/124` tests.
Full package Vitest:
- 12 real-CLI/no-TTY tests cover exact roster name/class, byte-equal `0600` seeds, no-clobber,
partial seed, missing/symlink defaults, unknown/padded/blank name, mismatched class, standalone
wizard preservation, and four concurrent starts.
- 12 direct filesystem tests cover complete publication, existing operators, idempotence, source
prevalidation, target and dangling destination links, invalid roots, oversized input, and
unexpected link errors.
- Composition coverage deterministically replaces validated `USER.md` with an external symlink and
requires refusal at point of use.
Full package gate (which rebuilds Mosaic itself after the clean-checkout dependency build):
```bash
env -u DATABASE_URL pnpm --filter @mosaicstack/mosaic run test:vitest
```
Exit `0`: `88/88` files, `1,573/1,573` tests.
Focused helper + point-of-use coverage:
```text
Test Files 88 passed (88)
Tests 1568 passed (1568)
Exit 0
2 files, 52/52 tests
Statements 97.97% | Branches 88% | Functions 100% | Lines 97.97%
Exit 0
```
New helper coverage:
Final repository gates after remediation:
```text
Statements 100% | Branches 93.33% | Functions 100% | Lines 100%
9/9 tests passed; coverage command exit 0
pnpm preflight exit 0
pnpm typecheck 45/45 tasks, exit 0
pnpm lint 25/25 tasks, exit 0
pnpm build 25/25 tasks, exit 0
pnpm format:check exit 0
git diff --check exit 0
```
Baseline commands:
Pre-PR targeted shell runs on the unchanged shell surfaces also passed:
```text
pnpm preflight exit 0
pnpm typecheck 45/45 tasks, exit 0
pnpm lint 25/25 tasks, exit 0
pnpm build 25/25 tasks, exit 0
pnpm format:check exit 0
pnpm --filter @mosaicstack/mosaic build exit 0
bash framework/tools/fleet/test-start-agent-session.sh
exit 0; retained expected fixture LD_PRELOAD warning
bash framework/tools/quality/scripts/test-install-migration.sh
21 passed, 0 failed, exit 0
bash framework/tools/_scripts/test-mosaic-init-rce.sh
PASS, exit 0
git diff --check exit 0
bash framework/tools/fleet/test-start-agent-session.sh exit 0 locally
bash framework/tools/quality/scripts/test-install-migration.sh 21 passed, 0 failed
bash framework/tools/_scripts/test-mosaic-init-rce.sh PASS
```
The aggregate `test:framework-shell` command exited `1` at `invariant_r_unittest.py`: `5/6` tests
passed and the remaining test refused the operator-global Pi drift from measured `0.84.1` to
installed `0.84.2`. This is retained as an environment/version-coupling failure, not scored as a
#1264 code failure and not retried. The aggregate stopped there; later aggregate stages are
**UNTESTED** except for the three targeted shell suites listed above.
The aggregate local `test:framework-shell` run stopped at `invariant_r_unittest.py`: installed
operator-global Pi is `0.84.2`, while the invariant is measured for `0.84.1`. Later aggregate stages
remain unmeasured except the targeted suites above. Root `pnpm test` remains locally **UNTESTED**
because this checkout prohibits the PostgreSQL-dependent gateway isolation path.
Root `pnpm test` is **UNTESTED** because it can execute the prohibited local PostgreSQL-dependent
gateway isolation path. No PostgreSQL service, connection, migration, or initialization was used.
CI is **UNTESTED — pending PR**.
## Review and security evidence
- Initial Codex review found padded-name mutation-before-refusal; it was fixed with three
no-side-effect regressions.
- Codex review of the formal-review remediation: `approve`, confidence `0.88`, 6 files, no findings.
- Codex security review of the remediation: risk `none`, confidence `0.93`, 9 files, no findings.
Its sandbox could not execute Vitest because Vite attempted a write on a read-only mount; the
executor-owned results above are the test evidence.
- Daphne formal review ID 168 at exact head `43fa0477`: `REQUEST_CHANGES`, four blocking groups. All
four have red-first regressions and local green remediation. Re-review of the next pushed exact
head is necessarily pending until that head exists.
## CI evidence and external blocker
Pipeline 2445 ran against exact first head `43fa0477`:
- install, sanitization, upgrade guard, typecheck, lint, and format passed;
- Mosaic Vitest passed `88/88`, `1,568/1,568`; and
- the test step emitted exactly one `FAIL:` line:
```text
FAIL: host provides 'pi' in the system path; missing-binary cases are not measurable here
```
That line comes from the inherited `test-start-agent-session.sh` CI-fit guard, not #1264. Fred filed
the correction as PR #1270. Its pipeline 2448 is terminal green and proves the four formerly masked
suites execute, but #1270 is not merged, so `next` still carries the failing chain. A new #1268
pipeline is pending the remediation push. Terminal-green #1268 CI is not claimed.
PR #1268's envelope was read back as `user.login=mos-dt-0`; its commit is explicitly authored and
committed by `goals <[email protected]>`. No goals Gitea login exists on this host, and no
other principal was borrowed. The cross-wrapper principal defect is tracked in #1272.
## Explicitly untested
- Canary VM remediation or restart: **UNTESTED and prohibited for this task**.
- Canary VM remediation/restart: **UNTESTED and prohibited**.
- Real Pi authentication/provider prompt and task execution: **UNTESTED**.
- PR #1213 composition layer: **UNTESTED and not required**.
- Deployment/published npm package behavior: **UNTESTED until CI/release; deployment is not this PR's scope**.
- Deployment/published npm behavior: **UNTESTED until merge/release**.
- Local PostgreSQL execution/migration: **UNTESTED and prohibited**.
## Review and residual risks
Initial independent Codex code review returned `request-changes` with one should-fix: a padded
`MOSAIC_AGENT_NAME` could be trimmed for pre-seed validation and later rejected in composition,
leaving seeds behind. The implementation now rejects blank or padded values before roster lookup or
writes; three no-side-effect regression cases pass. Codex re-review approved the remediated delta
with no findings (`confidence=0.86`). Its read-only sandbox could not execute Vitest because Vite
needed a temporary config artifact; executor-owned focused/full results above are the test evidence.
Final Codex security review found no confirmed vulnerabilities (`risk=none`, confidence `0.91`).
Formal PR review by a reviewer other than goals/Fred and CI remain pending.
Real Pi authentication/provider prompt and task execution remain unmeasured. The local gate proves
that Mosaic crosses its identity boundary and reaches the fake lease-runtime boundary; it does not
claim provider readiness or deployment.
The local gate proves Mosaic crosses its identity boundary and reaches a fake lease-runtime boundary;
it does not claim provider readiness, deployment, or a currently running canary seat.