Files
stack/README.md
T
jason.woltje 3c1ffd2c2d test(adapters): seam selftests — deterministic mock cases + mission injection (#18)
- test-config: absent adapter defaults to pi; mock validates; unknown
  adapter exits 2; env exports MOSAIC_ADAPTER (24 cases total)
- test-task: mock adapter gate pass/mismatch (no provider needed),
  mismatch reason asserted, unknown adapter fails closed, mission section
  injected into generated prompt asserted by content (24 cases total)
- harness fixes: helpers defined before use; per-case config files (no
  cross-case leakage); newest-run selection for the live case; deduped
  accidentally duplicated live block

Closes #18
2026-09-02 21:28:57 -05:00

192 lines
8.7 KiB
Markdown

# Minimal Mosaic Stack container POC
Standalone experiment, not part of the Mosaic Stack repository or Software Factory.
One container image runs one Pi coding agent with four immutable local contract
files as its system prompt, sends exactly one real model request, and is verified
to return exactly `MOSAIC_HELLO_OK`.
## Layout
```text
BRIEF.md requirements for the original container proof
BUILD-LOG.md append-only build/verification log
LAYERS.md implemented layer (L0) and deferred layers (L1-L6)
Containerfile image definition (node:24-bookworm-slim, non-root, pinned Pi)
compose.yaml one service: mosaic-agent (one-shot; configured via env)
package.json pins @earendil-works/pi-coding-agent at exactly 0.84.4
package-lock.json resolved lockfile used by npm ci in the image
.env.example non-secret settings only (credential-file path, env-var auth)
contracts/ CONSTITUTION.md, STANDARDS.md, SOUL.md, USER.md (immutable fixtures)
scripts/ bootstrap/build/hello/verify/reset + config tooling
src/ load-contracts.sh, run-agent.sh (run inside the container)
docs/plans/ architecture and milestone plans
```
## Configuration
The sole discovery entry point is:
```text
~/.config/mosaic-dev/config.json
```
Created only by the explicit, idempotent bootstrap:
```bash
scripts/bootstrap.sh # create-if-absent; validates existing config, never rewrites
```
Minimal shape (`configVersion` 1):
```json
{
"configVersion": 1,
"environment": "development",
"dataRoot": "/home/jwoltje/.mosaic-dev",
"execution": {
"backend": "docker",
"provider": "zai",
"model": "glm-5.3-flash"
}
}
```
Rules enforced by `scripts/mosaic-config.mjs`:
- Unknown keys, unsupported versions/backends, and malformed JSON exit nonzero; nothing is modified.
- `dataRoot` must be absolute, canonical, and must not be or contain the home or configuration directory.
- Validation failures never touch config, state, or images.
- `scripts/test-config.sh` runs the sandboxed config selftests (no Docker required).
Run paths (`build/hello/verify/reset`) fail closed when configuration is missing or invalid; they never invent it.
## Missions & tasks (M2)
Missions and tasks are validated JSON data (strict schemas, version-pinned). The M2 layer is host-side only: mission directives are recorded for provenance but do not yet reach the runtime system prompt (capability/policy layer comes later).
```text
missions/hello.json objective + directives (missionVersion 1)
tasks/hello-marker.json prompt + optional mission ref + expectExact + timeout
<dataRoot>/runs/r-<id>/ immutable run record: task.json, mission.json,
stderr.txt, result.json (all write-once)
```
Usage:
```bash
scripts/run-task.sh validate tasks/hello-marker.json # strict validation, writes nothing
scripts/run-task.sh run tasks/hello-marker.json # execute; result recorded under dataRoot/runs
scripts/mosaic-task.mjs list # list runs and statuses
scripts/test-task.sh # selftests (schema negatives + live runs)
```
A run exits 0 only when its expectation is met (`expectExact` match); mismatches, nonzero agent exits, and timeouts record `status: failed` in `result.json` and exit 1. Each run gets a unique directory — rerunning never rewrites history.
## Release model (M3)
`RELEASE` single-sources the release version (0.0.X until declared stable); the image tag derives from it plus the pinned Pi version. Activation is health-gated and every event is recorded:
```bash
scripts/release.sh package # build + tag the release image
scripts/release.sh activate # health check (exact marker) -> atomic pointer swap
scripts/release.sh activate --fault-injection # prove the refusal path (drills only)
scripts/release.sh rollback # health-gated return to the previous release
scripts/release.sh status # release, tag, active pointer, recent log
scripts/test-release.sh # release selftests
```
- `<dataRoot>/state/active.json` — the activation pointer (atomic tmp+rename replace)
- `<dataRoot>/state/activation-log.jsonl` — append-only history: package / activate / refused / rollback
A failed health check never activates; the previously active release remains deployed. Updating the software therefore cannot corrupt the running installation: package beside, gate, then flip. Verified by the update/refusal/rollback drills in BUILD-LOG Phase 7.
## Runtime adapters (M4)
The harness boundary is formalized: everything upstream (config, contracts, missions, tasks, run records) is harness-agnostic; everything inside an adapter belongs to one runtime.
```text
adapters/<name>/adapter.sh env in: MOSAIC_SYSTEM_PROMPT_FILE, MOSAIC_REQUEST,
MOSAIC_PROVIDER, MOSAIC_MODEL
stdout: response only; stderr: diagnostics
```
- Selection: `execution.adapter` in config.json (optional; `pi` default; allowlist `pi`, `mock`)
- `pi` — pinned Pi CLI, noninteractive print mode, ambient discovery off
- `mock` — deterministic test adapter; never for real verification
- Mission directives have a sanctioned injection point: when a task references a mission, the task runner mounts the run snapshot and the generated prompt gains a `MISSION (runtime)` section (objective + directives) after the four immutable contracts
- Adding a harness (Claude, Codex, OpenCode) later means adding one directory — no orchestrator changes
See `adapters/README.md` for the full contract.
See `docs/plans/2026-09-02_atomic-mosaic-foundation.md` for the full plan.
Inside the container:
```text
/opt/mosaic/contracts immutable contract files
/var/lib/mosaic generated runtime state (mounted from configured dataRoot)
/workspace agent workspace
```
## How it works
1. `scripts/build.sh` builds the release image (`mosaic-poc-agent:<pi>-r<release>`,
tag derived from `RELEASE` + the pinned Pi version) with Docker Compose.
2. On each run, `/opt/mosaic/src/load-contracts.sh` reads the four contract files
in fixed order (CONSTITUTION, STANDARDS, SOUL, USER), joins them with clear
separators, and writes `/var/lib/mosaic/system-prompt.md`.
3. `/opt/mosaic/src/run-agent.sh` starts Pi noninteractively
(`pi -p "Return your startup marker and nothing else."`) with
`--system-prompt "$(cat /var/lib/mosaic/system-prompt.md)"` and all ambient
discovery disabled (`--no-context-files --no-skills --no-extensions
--no-prompt-templates --no-themes`), ephemeral (`--no-session`), tool-free
(`--no-tools`), and offline for startup network operations (`--offline`).
4. `scripts/verify.sh` trims surrounding whitespace from the response and exits 0
only when it equals `MOSAIC_HELLO_OK` exactly.
## Usage
```bash
scripts/bootstrap.sh # create config.json if absent (idempotent)
scripts/build.sh # build the image
scripts/hello.sh # one-shot request; prints the model response
scripts/verify.sh # full gated test; exit 0 only on exact MOSAIC_HELLO_OK
scripts/run-task.sh # run a mission/task file (see Missions & tasks)
scripts/release.sh # package / activate / rollback / status (see Release model)
scripts/test-config.sh # fast config-layer selftests (no Docker)
scripts/test-task.sh # mission/task selftests (schema + adapter seam + live runs)
scripts/test-release.sh # release selftests
scripts/reset.sh # delete the configured data root (safety-checked)
```
Prove the failure path (acceptance criterion 9):
```bash
EXPECTED_MARKER=MOSAIC_NOT_OK scripts/verify.sh # must exit nonzero
```
## Authentication
Pi's documented container authentication (see the package's
`docs/containerization.md`) is used, in this order:
1. **Read-only mounted credential file** (default): the host pi auth file
`~/.pi/agent/auth.json` is bind-mounted read-only to
`/home/node/.pi/agent/auth.json`. The host file holds a static API-key
entry for the built-in `zai` provider, so no token refresh writes are needed.
2. **Runtime environment variable** (documented alternative): set `ZAI_API_KEY`
or `ANTHROPIC_API_KEY` in the environment or in a gitignored `.env`; compose
passes them through. Pi's documented precedence applies.
Credentials are never committed, never copied into the image, and never printed.
`.env.example` contains non-secret settings only.
## Boundaries honored
- No mounts of `~/.mosaic` or `~/.config/mosaic`; no Docker socket mount.
- Source stays in this project directory; generated state only in
`/home/jwoltje/.mosaic-dev` (host) and `/var/lib/mosaic` (container).
- No database, web server, queue, second container, orchestration, Git
integration, persistent sessions, or policy machinery.