wym glibc 2.38
This commit is contained in:
lda committed 2026-04-14 09:45:59 +07:00
1 parent 54d3f1e93d
commit ac824235ab
5 files changed
+139

No files matched your search

+100
View File
@@ -0,0 +1,100 @@
# Wakey Checkpoint (2026-04-11)
## Snapshot
This checkpoint captures the current state after control-plane migration, logging hardening, config ergonomics, and state storage upgrades.
## What Is Done
- Legacy router-hosted HTTP/static layer removed from `wakey` crate.
- New `wakey-control-plane` crate is active for:
- enroll-token issuance
- agent enrollment
- connected-agent websocket registry
- command relay to agent
- `wakey-agent` is active for outbound websocket execution and local command dispatch.
## Logging + Telemetry
- High-signal logs were added across:
- control-plane API relay path
- control-plane websocket lifecycle
- control-plane state lifecycle
- agent command lifecycle, session lifecycle, and enrollment
- Correlated relay spans include command context (`agent_id`, `request_id`, `command`).
- Control-plane telemetry is config-driven:
- optional OTLP endpoint
- optional JSON logs
- fallback local logs when OTLP endpoint is not set
## Config Ergonomics
### Control-plane
- Config file support is wired (`/etc/wakey-control-plane/config.toml` by default).
- `serve` can read defaults from config file and CLI can override.
- New `init-config` command scaffolds a control-plane config file.
### Agent
- Existing `init-config` command scaffolds agent config.
- Enrollment can optionally signal reload of running daemon.
## State Storage Upgrade
- Control-plane state backend moved from JSON snapshot to embedded `sled` DB.
- Default state path changed to `/var/lib/wakey-control-plane/state.db`.
- Legacy JSON migration support has been removed; sled is now the only supported
state format.
- Enroll tokens include persisted expiry timestamps and are validated on enroll.
- Periodic and explicit garbage collection remove expired tokens.
## Operator Commands
### Control-plane bootstrap
```sh
wakey-control-plane init-config
wakey-control-plane serve --config-file /etc/wakey-control-plane/config.toml
```
### Issue enroll token (live daemon path)
```sh
wakey-control-plane issue-enroll-token --public-url https://cp.example.com
```
### Agent bootstrap
```sh
wakey-agent enroll --server-url https://cp.example.com --enroll-token <token>
wakey-agent serve --config /etc/wakey-agent/config.toml
```
## Build Health
- Last verified passing:
- `cargo check --workspace`
- `cargo clippy --workspace`
## Test Coverage Added In This Pass
- Unit tests in `wakey-control-plane/src/state/store.rs` now cover:
- expired-token garbage collection removes persisted stale tokens
- enroll rejects expired tokens and consumes stale entries
- state stats counters for agents and expired token totals
## Known Tradeoffs / Follow-ups
- Reload semantics with `sled` are now mostly no-op for in-memory state (data is durable in DB).
- No dedicated state-inspection CLI command yet (suggestion: add `state-stats` command).
- OTLP configuration is currently control-plane focused; agent parity can be added if needed.
## Suggested Next Steps
1. Add a control-plane `state-stats` command to print DB path, agent count, token count.
2. Add symmetric telemetry config support in `wakey-agent` config file.
3. Add integration tests for:
- enroll + relay over websocket
- legacy JSON-to-sled migration path
- init-config command behavior and overrides
+82
View File
@@ -0,0 +1,82 @@
# Wakey Checkpoint (2026-04-12)
## Snapshot
This checkpoint captures progress after adding audit persistence, alert evaluation and transitions, control-plane alert APIs, and websocket timing diagnostics across agent and control-plane.
## Major Changes Landed
- Control-plane audit system implemented with persistent sled-backed events.
- Audit emission wired into:
- enroll accept/reject
- token issue/list/revoke
- command dispatch/result/error/timeout
- websocket auth accept/reject and disconnect
- Audit query API added:
- `GET /api/v1/control/audit/events`
- Active alert engine added with deterministic rules over audit + live session state.
- Alert APIs added:
- `GET /api/v1/control/alerts`
- `GET /api/v1/control/alerts/history`
- `GET /api/v1/control/alerts/ws`
- Alert transition persistence added (open/resolve transitions tracked across evaluations).
- Route classes split explicitly in runtime:
- public routes (enroll/ws/health)
- control routes (`/api/v1/control/*`)
- Caddy template added for edge policy and Cloudflare Access boundary:
- `deploy/Caddyfile.control-plane.example`
## Reliability and Diagnostics Improvements
- Agent websocket connect diagnostics now include:
- DNS resolution timing (`dns_resolve_ms`)
- websocket connect timing (`ws_connect_ms`)
- Control-plane websocket lifecycle logs include:
- connect-to-hello timing
- connect-to-auth timing
- hello-to-auth timing
- Slow connect warnings are now emitted when timing thresholds are exceeded.
## Root-Cause Findings Captured
- Long agent websocket connect delays were reproduced and traced to hostname resolution path.
- Switching agent `server_url` hostname to direct IP made connect immediate.
- This confirms app-level relay logic was not the source of the startup delay.
## Verification Status
- `cargo check --workspace` passing after all changes.
- Added and passing tests include:
- audit event append/filter in state store
- alert transition open/resolve persistence
- alert evaluator rule checks (offline + timeout, auth/enroll rejection spikes)
## Current API Surface for UI Start
- Agents and command execution:
- `GET /api/v1/control/agents`
- `POST /api/v1/control/agents/{agent_id}/command`
- Audits:
- `GET /api/v1/control/audit/events`
- Alerts:
- `GET /api/v1/control/alerts`
- `GET /api/v1/control/alerts/history`
- websocket subscribe: `GET /api/v1/control/alerts/ws`
## Remaining Plan Items (Most Significant)
1. UI implementation (`/ui` app shell and pages) is still open.
2. Alert dedupe/cooldown persistence and tuning are still basic and need hardening.
3. Audit retention pruning policy and long-run storage controls are not finalized.
4. Edge auth enforcement tests and deployment rehearsals remain to be added.
5. Multi-day soak drills and failure-injection validation remain open.
## Suggested Next Actions
1. Build minimal UI shell with three views:
- agents/commands
- audit timeline
- alerts panel (active + history + websocket stream)
2. Add periodic retention task for audit and alert transition trees.
3. Add proxy-level integration tests that assert private endpoints are blocked without Access headers.
4. Run a 48-72h soak with hostname vs IP connect-path metrics collected.
@@ -0,0 +1,14 @@
# Control Plane Review
## Findings
1. High: the admin/control API is effectively unauthenticated, so anyone who can reach the server can issue enroll tokens, list/revoke tokens, inspect audit/alerts, and send live commands to agents. In [`wakey-control-plane/src/runtime/mod.rs:61`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/runtime/mod.rs#L61) through [`wakey-control-plane/src/runtime/mod.rs:109`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/runtime/mod.rs#L109), `control_api_routes()` is merged straight into the app with no auth middleware despite the comment saying these routes are “admin-only.” The UI also calls those endpoints directly with plain `fetch` and no auth material in [`ui/src/api.ts:60`](c:/Users/Admin/Documents/realshit/wakey/ui/src/api.ts#L60) through [`ui/src/api.ts:123`](c:/Users/Admin/Documents/realshit/wakey/ui/src/api.ts#L123). This is a full remote-takeover issue for the control plane, not just a missing polish item.
2. High: a second websocket connection for the same `agent_id` silently replaces the current session, but the old authenticated socket is left alive and can still submit `result`/`error` frames against pending requests. In [`wakey-control-plane/src/ws.rs:195`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/ws.rs#L195) through [`wakey-control-plane/src/ws.rs:200`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/ws.rs#L200), a successful auth simply overwrites `sessions[agent_id] = tx.clone()`. The previous socket is not closed or demoted. Later, any authenticated socket can satisfy pending requests in [`wakey-control-plane/src/ws.rs:229`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/ws.rs#L229) through [`wakey-control-plane/src/ws.rs:247`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/ws.rs#L247), while requests are correlated only by `request_id` created in [`wakey-control-plane/src/api/commands.rs:91`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/api/commands.rs#L91) through [`wakey-control-plane/src/api/commands.rs:97`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/api/commands.rs#L97). That creates a split-brain/race condition where a stale or malicious prior session for the same agent can inject or win replies.
3. Medium: configured seed enroll tokens are reinserted into the database on every startup, so “one-time” tokens become reusable after a restart if they remain in config. In [`wakey-control-plane/src/state/store.rs:52`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/state/store.rs#L52) through [`wakey-control-plane/src/state/store.rs:68`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/state/store.rs#L68), `load_or_init()` blindly seeds `daemon.enroll_tokens` into sled every time the process starts. Enrollment consumes tokens in [`wakey-control-plane/src/state/store.rs:91`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/state/store.rs#L91) through [`wakey-control-plane/src/state/store.rs:109`](c:/Users/Admin/Documents/realshit/wakey/wakey-control-plane/src/state/store.rs#L109), but that consumption is undone on the next boot if the token still exists in config. That breaks the “short-lived/one-time” assumption and makes operational mistakes much more likely.
## Residual Risks
- The control-plane/UI surface is growing quickly and currently assumes a trusted environment in multiple places. Even after adding admin auth, I would expect more authz/session-boundary issues to surface.
- The command relay path is conceptually good, but it needs stronger session ownership rules before it is trustworthy under reconnect races or duplicated agents.