embedded-lab-control 0.2.0__tar.gz

embedded_lab_control-0.2.0/CHANGELOG.md

@@ -0,0 +1,1514 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/).
+
+## [Unreleased]
+
+## [0.2.0] - 2026-10-31
+
+First release on PyPI, as `embedded-lab-control` (the import package and
+the CLI are still `labctl`): `pipx install "embedded-lab-control[web,mcp]"`,
+adding `kasa`, `sdwire` and `actuators` as needed (see the README for what
+each extra enables). Python 3.10-3.13; the `sdwire` extra needs 3.12+, and
+Kasa KLAP v2 devices need 3.11+. Everything below the upgrade notes is
+the full history since development started; there was no earlier public
+release.
+
+**Highlights**
+
+- MCP server on the mcp 2.x SDK, over stdio or HTTP (API-key auth, TLS,
+  image uploads), with a two-step confirmation for destructive tools and
+  a deny-all allowlist for host file paths.
+- Hardware claims enforced on every surface (CLI, REST, web UI, MCP),
+  including boards that share a plug outlet, SDWire, console adapter or
+  relay channel; long calls keep the caller's claim alive.
+- Power control for Tasmota, Kasa (incl. KLAP v2 strips) and Shelly;
+  SDWire SD-card multiplexers; USB relay actuators with per-board
+  bindings (`pre_power` straps applied on every power-on path); ser2net
+  serial access with udev-stable names; health monitoring; activity
+  stream.
+- Session-based web UI and REST API with optional HTTPS.
+
+**Upgrading from a source install (`labctl` 0.1.0)**
+
+- The distribution was renamed `labctl` → `embedded-lab-control`.
+  `scripts/update.sh` and `scripts/install-services.sh` uninstall the old
+  distribution first; if you upgrade a venv by hand, run
+  `pip uninstall labctl` before installing (both own `labctl/`, and
+  removing the old one later would delete the new install's files).
+- The database migrates itself to schema v12 on first use.
+- `labctl mcp --http PORT` now really binds PORT (it always used 8000):
+  the shipped unit's server moves to `127.0.0.1:8080`. Clients wired to
+  `:8000` must move. A non-loopback `--host` requires `auth.enabled`;
+  HTTP clients then send `Authorization: Bearer <api_key>`.
+- MCP host file access is deny-all until `mcp.allowed_read_paths` /
+  `mcp.allowed_write_paths` are set (`flash_image`, `sdwire_update`
+  copies, `boot_test`); `update.sh` creates `/var/lib/labctl/images` and
+  `/var/lib/labctl/output` and prints the snippet to add. The CLI is
+  unaffected.
+- Destructive MCP tools return a plan and a `confirm_token` first
+  (`mcp.confirm_destructive`, `mcp.confirm_exempt`).
+- Commands that operate or reconfigure a board another user has claimed
+  are refused unless overridden (`--override-claim`, REST
+  `"override_claim": true`, operators only on REST/web), and the override
+  is logged.
+- Removed config keys (`proxy.enabled`, `proxy.log_retention_days`,
+  `proxy.idle_timeout`, `health.update_status_on_check`) are ignored
+  with a warning. Several commands that used to read a confirmation from
+  stdin now need `--yes` without a terminal.
+- After updating a host with generated udev rules, run
+  `labctl serial udev --install --reload` (rules are now
+  `root:dialout 0660`).
+
+**For packagers**: `mcp>=2.2,<3` pulls in `mcp-types` (pinned exactly
+by mcp), `httpx2` and `opentelemetry-api`. The sdist includes the tests,
+config samples and systemd units, docs and scripts; the wheel carries only
+the `labctl` package.
+
+### Added (P2 publish)
+- Release workflow (`.github/workflows/release.yml`): pushing a `vX.Y.Z`
+  tag runs the CI gates (`ci.yml`, now also callable as a reusable
+  workflow), checks the tag matches `labctl.__version__` and that this
+  file has the version's section, builds and `twine check`s the sdist
+  and wheel once, publishes them to TestPyPI and then, after approval of
+  the protected `pypi` environment, to PyPI with trusted publishing (no
+  API tokens), and creates the GitHub Release. Running it by hand only
+  builds and checks. `docs/RELEASING.md` has the one-time setup, the
+  release steps and what to do when a step fails.
+
+### Changed (P2 release prep)
+- A command whose optional dependency is missing now says which extra to
+  install, in one line, instead of a traceback or `pip install <package>`:
+  `labctl web` (and REST) and `labctl user hash-password`/`add`/`verify`
+  (werkzeug) → `[web]`, `labctl mcp` → `[mcp]`, Kasa plugs → `[kasa]`,
+  SDWire → `[sdwire]`, USB relays → `[actuators]`. Each message gives
+  `pip install "embedded-lab-control[<extra>]"` and the `pipx inject`
+  equivalent with the extra's version floors for the running Python;
+  the same text reaches the CLI, REST and MCP (`labctl/extras.py`).
+  Below Python 3.12 the SDWire message says it needs 3.12+ (the extra
+  installs nothing there) and gives no install command. Only the extra's
+  own package being absent is reported this way: an installed but broken
+  or incompatible one gets its own one-line error carrying the real
+  ImportError and the versions the extra wants (still an error line,
+  not a traceback). A missing Kasa library
+  no longer also logs a "Failed to load Kasa credentials" warning, and
+  isn't remembered for the life of the process.
+- README: install from PyPI with pipx or pip, a table of what each extra
+  enables, and Python version notes; the source install and the
+  systemd install from a git checkout (`scripts/install-services.sh`)
+  are unchanged.
+- `docs/HARDWARE_MAP.md` is now a generic guide to mapping USB serial
+  adapters (USB paths, `labctl serial discover/add/udev`, assigning
+  ports, troubleshooting) with placeholder paths and names.
+- `scripts/install-udev.sh` requires the rules file as its argument (it
+  defaulted to a repository file that no longer exists) and refuses a
+  file with no `SYMLINK` assignment or with a non-comment line that
+  isn't a udev rule, such as the message `labctl serial udev` prints
+  when no adapters are registered. The README and `HARDWARE_MAP.md` note
+  that installing a rules file by hand skips the claim check
+  `labctl serial udev --install` makes. `labctl ports -v` with no ports points
+  at `labctl serial discover` instead of `discover-usb-serial.sh`.
+- The source distribution now includes the tests, `config/` (sample
+  config, systemd units, udev mapping example), `docs/`, `scripts/`,
+  `skills/`, `CHANGELOG.md` and `SECURITY.md` (`MANIFEST.in`); the wheel
+  is unchanged. The test suite passes from an unpacked sdist.
+- CI runs on pushes to `main`, on pull requests (so a PR branch is
+  tested once, not twice) and by hand (`workflow_dispatch`).
+
+### Removed (P2 release prep)
+- This lab's `config/udev/99-lab-serial.rules`,
+  `config/udev/port-mapping.yaml` and `config/ser2net/labctl-ser2net.yaml`.
+  Rules are generated with `labctl serial udev` and the ser2net config
+  with `labctl ser2net generate`; `config/udev/port-mapping.yaml.example`
+  (now with placeholder paths and names) remains for the legacy
+  `generate-udev-rules.py`.
+
+### Added (#2: long calls keep the caller's claim alive)
+- A claim's deadline moved only when a call started, so one long call
+  (a 10-run `boot_test`, a long `serial_capture`, flashing a large
+  image) could outlast it, and another agent could then claim the board
+  and power-cycle it mid-run. MCP `boot_test`, `serial_capture`,
+  `sdwire_update` and `flash_image`, and CLI `boot-test`, now heartbeat a
+  claim the caller already holds on the board from a background thread
+  while they run (every quarter of the claim's current duration, 5-60 s,
+  so a renew to a shorter claim is followed). They never claim a board
+  for the caller. Each heartbeat is one conditional UPDATE
+  (`ResourceManager.heartbeat_held_claim`, which replaces the unused
+  `heartbeat_claim`) that never revives a released, expired or
+  taken-over claim; `touch_claim` now says whether a row changed. A
+  database error is a missed heartbeat (the call goes on) until the
+  claim's deadline passes without a successful one, or, for a caller
+  holding no claim, until the checks have failed for 60 s in a row; if
+  the claim can't be read when the call starts, it runs without
+  keepalive (logged). The keepalive follows the claim the call started
+  with (its heartbeat matches that claim's id): a claim released and
+  taken again, even by the same user, counts as lost.
+- Boot tests check again before each run's power cycle and stop once
+  someone else claims the board or a board sharing its hardware, the
+  caller's claim is lost (released, force-released or expired), or, over
+  MCP with `claims.require_claim`, nobody holds it. The check runs
+  before the heartbeat, so a refused check extends nothing. A stopped
+  test says `STOPPED after N of M runs: <reason>` and reports
+  `Result: INCOMPLETE, ...` (or `Result: STOPPED before the first run`,
+  noting that the board is left off if the deploy powered it off)
+  instead of a pass rate; MCP results start with `Stopped: <reason>` and
+  the CLI exits 1. A CLI test started with `--override-claim` isn't
+  stopped by the claims it overrode, but is by any claim taken after it
+  started.
+- `sdwire_update` and `flash_image` with `reboot=true` leave the board
+  off, and say why, when the check fails after the card is written:
+  `Not power cycled: <reason>` (`Not powered on:`) when the caller's own
+  claim was lost, `<board> is left off (not power cycled) because
+  '<board>' is now claimed by '<agent>'` when someone else claimed the
+  board or a board sharing its hardware.
+- `/deploy-and-test` skill: claim after a successful build, sizing the
+  claim for the run (for `boot_test`, at least runs × (timeout + cycle
+  delay + a few seconds) plus the deploy and a margin); renew it between
+  long steps; release it on every path (deploy failure, boot timeout, a
+  confirmation not given). Resolves the claims spec's open question on
+  the skill.
+
+### Changed (#41: options and config keys that did nothing)
+- `serial.base_tcp_port` (and `LABCTL_BASE_TCP_PORT`) now sets where TCP
+  port auto-assignment starts (`port assign` without `-t`, MCP, REST, web,
+  `import`): one above the highest port in use, and at least the base. It
+  must be 1-65535 (anything else keeps 4000, with a warning). The web
+  settings page showed it blank (it read `ser2net.base_tcp_port`).
+- `ser2net generate` no longer gives every port without a TCP port
+  4000. With `--install` each such port is assigned one and saved, as
+  `port assign` without `-t` would (one above the highest in use, at
+  least `serial.base_tcp_port`; refused past 65535), and the command
+  says which, with `port assign`'s claim checks (one shared helper). If
+  any such port is on a board another user has claimed, or one sharing
+  a claimed board's console device, `--install` refuses and saves and
+  writes nothing (leaving the port out would change that board's
+  console), listing them; `--override-claim` (new; recorded in the
+  activity log) assigns them anyway. Claimed boards whose ports already
+  have a TCP port are written as before. Nothing is saved unless the
+  config file can be written (checked first), and each save repeats
+  `port assign`'s check, so a claim taken meanwhile still refuses; any
+  failure after a save lists the TCP ports already saved. A port whose
+  save fails is left out and reported; the config is still written and
+  the command exits 1.
+  Without `--install` nothing is saved: those ports are left out of the
+  output and the number each would get is shown (stderr).
+- `ser2net generate --install` writes `ser2net.config_file` (default
+  `/etc/ser2net.yaml`, unchanged); it always wrote `/etc/ser2net.yaml`.
+  The write is atomic when labctl can create files in the directory (a
+  temporary file there, given the old file's owner and mode, fsynced and
+  renamed over it), and otherwise rewritten in place (e.g. a labctl group
+  member, who can write `/etc/ser2net.yaml` but not `/etc`); a symlink on
+  the path or a hard-linked file is refused, since it may run as root. The
+  permission-denied message names the configured path; when the file is
+  really absent from a directory labctl can reach, it says to create it
+  with `sudo labctl ser2net generate --install` (and, for
+  `/etc/ser2net.yaml`, then rerun `install-services.sh` or `update.sh`
+  to make it group-writable); when the path can't be reached, it says so.
+  `install-services.sh`, and now `update.sh` (for a file created after the
+  install), make only `/etc/ser2net.yaml` group-writable (`root:labctl
+  664`; a symlink there is left alone).
+- `health.status_retention_days` now prunes status history: the monitor
+  deletes older entries at startup and then hourly; `0` keeps all of it.
+  Each board's newest entry per status, and the one in effect at the
+  cutoff, are always kept, so a board stable for longer than the
+  retention keeps its uptime. Schema v12 indexes `status_log(logged_at)`.
+  An hourly run is cheap lookups (newest row, rows past the cutoff, all
+  rows) and prunes only when one changed since the monitor's last prune.
+  Pruning deletes in transactions of at most 5,000 rows, so the first
+  prune of a long history doesn't hold the write lock for long.
+- Uptime (`GET /api/sbcs/<name>/uptime`): the 24-hour figure starts from
+  the status in effect 24 hours ago, so a board up the whole day shows
+  100% (it showed 0% with no status change in the window), and times are
+  compared in the database's clock (UTC), not local time.
+- `claims.require_agent_name` is enforced: MCP `claim_sbc` and
+  `POST /api/claims/<sbc>` refuse (`agent_name_required`) a claim that
+  would be recorded under a placeholder name: no `agent_name` and no
+  authenticated user (or, over MCP stdio, no name given earlier in the
+  session). CLI claims are always named (`--name` or `$USER`). A REST
+  `agent_name` that isn't a string is a 400.
+- MCP over unauthenticated HTTP no longer remembers an `agent_name`
+  across calls: the server can't tell its clients apart, so one client's
+  name named every other client's claims and releases. It is used for
+  the call that gives it; stdio and per-user (authenticated) names are
+  still remembered.
+- MCP `claim_sbc` defaults `duration_minutes` to
+  `claims.default_duration_minutes`, like the CLI and REST (it was 30).
+- `post-power` bindings can no longer be created: nothing acts on them
+  yet. `bind --phase` no longer offers `post-power`, and labctl
+  (`ResourceManager`) refuses `post_power`; MCP `bind` refuses it before
+  its claim checks. Bindings
+  stored with it before are kept and listed as inactive (`bindings list`
+  marks them `*`, `bindings status` says so, MCP output carries
+  `sample_phase_note`).
+- `ports --all` lists the USB serial devices (`/dev/ttyUSB*`,
+  `/dev/ttyACM*`) no `/dev/lab` symlink points at; the flag did nothing.
+- `remove`, `sdwire remove`, `serial remove`, `actuator remove` and
+  `power-all` ask only on a terminal. Without one they no longer read the
+  answer from stdin (a piped `y` used to confirm) and exit 1 asking for
+  `--yes`. Declining `actuator remove` now exits 1, like the others (it
+  printed "Aborted." and exited 0).
+- `install-services.sh` and `update.sh` make the services' database
+  writable by the `labctl` group, so group members can write it from the
+  CLI without `sudo -u labctl`: a new config puts it in
+  `/var/lib/labctl/db` (setgid `2775`; the file, created empty if
+  missing, and any `-wal`/`-shm` are `664`; SQLite gives new `-wal`/`-shm`
+  files the database's mode). A database configured elsewhere is left
+  alone, with a note. `database_path` is compared with the directory
+  resolved (a symlinked `/var/lib/labctl` still matches; the database file
+  itself is never followed). The labctl home gets group search (`g+x`,
+  nothing else) when it belongs to the `labctl` group and lacks it; if
+  group members still can't traverse it, the scripts warn. Both scripts
+  run the same root-safe step,
+  `python -m labctl.core.admin_files shared-db` (new; also new: `create`),
+  and a problem there (a symlinked database file, a non-regular `-wal`, a
+  config that won't parse) is a warning: the scripts carry on, and
+  `update.sh` still restarts the services.
+  **Upgrading:** an existing install keeps its database where it is
+  (e.g. `/var/lib/labctl/labctl.db`). To share it with the group, as root
+  throughout (your account may not see into `/var/lib/labctl`): stop the
+  services, move `labctl.db` (and any `labctl.db-wal`/`-shm`) into
+  `/var/lib/labctl/db/`, check it arrived (`sudo ls -l
+  /var/lib/labctl/db/labctl.db`; if it isn't there, `update.sh` creates a
+  new, empty database), set `database_path: /var/lib/labctl/db/labctl.db`
+  in `/etc/labctl/config.yaml` and copy it to the service config, run
+  `update.sh` (or `install-services.sh`), and start the services. README
+  "Moving the database" has the commands.
+- REST `POST /api/sbcs/<name>/power` returns 400, not 500, for a
+  non-string `action` or a JSON body that isn't an object.
+
+### Removed
+- Config keys that were parsed but never acted on: `proxy.enabled`
+  (and `LABCTL_PROXY_ENABLED`), `proxy.log_retention_days`,
+  `proxy.idle_timeout` and `health.update_status_on_check` (use
+  `monitor --no-update`). A config that still sets one loads, with a
+  warning that it's ignored.
+
+### Security
+- MCP `sdwire_to_host` fails closed (#41): it switches the card only when
+  the board's plug reports it off. It used to go ahead when the board had
+  no plug or its state couldn't be read; it now refuses, like
+  `sdwire_ls`/`cat`/`info`, unless `force=true`. CLI `labctl sdwire host`
+  does the same: it refuses a board with no plug or an unreadable or
+  unknown plug state unless `--force` (which, as before, skips the power
+  check), and so do CLI `sdwire ls`/`cat`/`info` (they refused only a
+  plug reporting ON). These now also read a card already on the host in
+  place, as the MCP read tools do, and take `--force` to skip the check.
+  One check (`labctl.sdwire.power_guard`) serves all of them. The MCP
+  read tools' refusal points at `sdwire_to_host(force=true)` for a board
+  you've made sure is off.
+- A per-board cycle delay stored before `edit --cycle-delay` checked the
+  range (e.g. `1e9`, NaN, negative) is refused on every path before
+  anything is switched or written (#41): MCP/REST/web cycles, the reboot
+  after `sdwire flash`/`update`, `boot-test` and the recovery commands, CLI
+  and MCP. Only CLI `power cycle` checked it before; the others slept on
+  it. Clear it with `labctl edit <name> --cycle-delay -1`.
+- `labctl serial remove` and `labctl serial udev --install` check claims
+  (#41). `serial remove` refuses an adapter that a board someone else has
+  claimed has a port on, by any path that reaches it (including a port
+  assigned by `/dev/ttyUSB0` behind the adapter's link, which isn't linked
+  to the registration). `serial udev --install` refuses when the new rules
+  would remove or re-point a `/dev/lab/<name>` link such a board uses
+  (every link, if the installed rules file can't be read). Both take
+  `--override-claim`. `serial remove` asks for confirmation only on a
+  terminal; without one it refuses unless `--yes` (see the no-terminal
+  entry above). MCP
+  `remove_serial_device` refuses an adapter a board someone else has
+  claimed has a port on, the same way.
+- Writing a board's SD card fails closed when the board has no power plug
+  (#41): `sdwire flash`/`update` (CLI) and MCP `sdwire_update`/`flash_image`
+  used to switch the card silently; they now refuse unless `--no-plug-ok`
+  (MCP `no_plug_ok=true`), which says you've made sure the board is off.
+  The flag never skips the power-off of a board with a plug nor excuses a
+  failed one. The MCP plans show `will_do.power_off`.
+- Under `sudo`, the CLI acts for the user who ran sudo (`SUDO_USER`): its
+  claim identity (`cli-<user>@<host>`), agent name and activity-log actor,
+  so e.g. `sudo labctl serial udev --install` doesn't turn your own claim
+  into someone else's.
+- MCP confirmation tokens are bound to the plan they were issued with
+  (`will_do`): if it differs at confirmation (e.g. the board's stored cycle
+  delay changed), the call is refused and the caller asks for a new plan.
+  A claimed board's power-cycling tools return the claim refusal, not a
+  plan.
+- REST and web-UI power cycles don't overlap: the guard and the 5 s rate
+  limit are per plug outlet, so a cycle of an outlet (by any board on it)
+  requested while another is running (it blocks for the whole delay) is
+  refused with `429`, and `on`/`off` of that outlet meanwhile with `409`.
+  The guard covers REST and the web UI in one server process only: power
+  operations from the CLI, MCP, `sdwire` reboots or `boot-test` aren't
+  seen by it (a lab-wide outlet lock is tracked in #38).
+- `labctl edit --cycle-delay` (setting or clearing it) is claim-checked
+  against the boards sharing the board's hardware as well, since a stored
+  delay is the minimum of every board on the outlet. Other `edit` changes
+  still check the board alone.
+
+### Changed
+- One cycle-delay rule on every surface (#41): the board's stored delay
+  (`labctl edit --cycle-delay`), or 3 s if none is set, is the default
+  and the minimum of every power cycle, not only CLI `power cycle` and
+  recovery: MCP `power_cycle` (whose `delay` now defaults to it), REST
+  `cycle` (likewise; the response adds `delay_seconds`), the web UI's
+  power-cycle button, the reboot after `sdwire flash`/`update` and every
+  `boot-test` run (CLI and MCP). A smaller requested delay is raised to
+  it, so MCP and REST no longer allow less than 3 s on a board without a
+  stored delay. Boards sharing a plug outlet are cycled together, so the
+  largest of their minimums applies (each board's stored delay, or 3 s
+  without one; an out-of-range one on any of them refuses), and the CLI
+  warning, MCP results and REST `delay_source` name the board it came
+  from. A negative requested delay is an error everywhere (CLI
+  `power cycle --delay` used to raise it to the minimum). The `--dry-run` plans of
+  `sdwire flash`/`update` and the MCP confirmations of `power_cycle`,
+  `enter_/exit_recovery`, `boot_test` and `sdwire_update` (with
+  `reboot=true`) show the delay and where it came from
+  (`will_do.cycle_delay_seconds`, `cycle_delay_source`).
+- Generated udev rules (`labctl serial udev`, and the legacy
+  `scripts/generate-udev-rules.py`) make `/dev/lab/*` devices
+  `root:dialout 0660`. They were world-writable (`0666`), so any local
+  user could open a console or drive a relay. The labctl services are in
+  `dialout` and ser2net runs as root; a user who drives relays from the
+  CLI must now be in `dialout` (`labctl serial udev --install` says so
+  when the user isn't). Installed rules change on the next
+  `labctl serial udev --install --reload`.
+- `install-services.sh` and `update.sh` make the service copy of the config
+  (`/var/lib/labctl/.config/labctl/config.yaml`, which holds the same API
+  keys and credentials as `/etc/labctl/config.yaml`) mode `640`; it was
+  created world-readable (`644`), so any local user could read the keys
+  wherever the labctl home is traversable by others (its mode comes from
+  `useradd`'s `HOME_MODE`, `0750` on current Ubuntu/Debian, or from an
+  admin: it is `2775` on the reference lab host).
+- The install and update scripts (run as root) no longer follow symlinks
+  or hard links planted under `/var/lib/labctl` (writable by the `labctl`
+  service user) when they seed `/etc/labctl/config.yaml` from the service
+  config, create the service config, change its owner and mode, or create
+  and re-permission the `images`/`output` directories. These go through
+  `labctl.core.admin_files`: no symlink on any path component below the
+  (resolved) labctl home, single-link regular files only, descriptor-based
+  ownership changes, `O_EXCL` creation, byte-for-byte copies. Previously a
+  compromised service could have root hand it `/etc/labctl/config.yaml`,
+  or copy an arbitrary root-only file into a config its group can read.
+  Remaining by-path root operations there (`chown -R`, `mkdir -p`, the
+  drift `cmp` and the allowlist config read) are tracked in #19.
+- **MCP over HTTP now authenticates with the web users' API keys (breaking
+  for HTTP clients when `auth.enabled` is true).** With auth enabled, every
+  request to `labctl mcp --http` needs `Authorization: Bearer <api_key>`
+  for a user in `auth.users` (401 otherwise); keys are re-read per request,
+  so revoking one needs no restart. Sessions are bound to the user that
+  opened them, destructive-tool confirm tokens to the user they were issued
+  to, and each user is a separate claimant (one user's claim blocks
+  another's mutating calls); claims and audit entries record the user. The
+  server refuses to start without auth on a non-loopback `--host` or with
+  non-loopback names in `mcp.allowed_hosts` / `allowed_origins` (a proxy
+  that forwards `Host: 127.0.0.1` can't be detected: enable auth before
+  adding one), with auth enabled but no API key configured, or when a
+  config file it should read can't be loaded (a fallback to another file
+  or to defaults could drop the authentication the intended config
+  enables; this includes `/etc/labctl/config.yaml` being unreadable to a
+  user outside the `labctl` group, who must then pass `-c`). A running
+  MCP server (either transport) now keeps the configuration it started
+  with until restarted; only API keys are re-read, from the same file, per
+  request (a broken file refuses every key). Users
+  without a string username can't authenticate by API key. New `mcp.allowed_hosts` / `mcp.allowed_origins` accept extra
+  `Host` / `Origin` headers (e.g. a reverse proxy's public name) on top of
+  the loopback names. See `docs/MCP_SERVER.md` (D012).
+- API key lookups (web and MCP) no longer raise on a non-ASCII key in a
+  request header, and a non-string `api_key` in the config never matches.
+- **Destructive MCP tools now require a two-step confirmation (breaking for
+  MCP clients).** The 23 tools annotated destructive (power off/cycle,
+  flashing, SD updates, `serial_send`, deletions, actuator/binding verbs,
+  recovery, force release) return a plan and a single-use `confirm_token`
+  on the first call and act only when called again with the same arguments
+  plus that token (120 s expiry, bound to tool and arguments). Configure
+  with `mcp.confirm_destructive` (default `true`) and `mcp.confirm_exempt`.
+  See `docs/MCP_SERVER.md` (D012).
+- **MCP host file access is now allowlisted (breaking).** `flash_image`
+  (`image_path`, `post_flash_copies` sources), `sdwire_update` (`copies`
+  sources) and `boot_test` (`image`, `output_dir`) only accept absolute host
+  paths inside the new `mcp.allowed_read_paths` / `mcp.allowed_write_paths`
+  settings. Both default to empty, which denies all host file access over
+  MCP. Paths are resolved (symlinks, `..`) before the check and are refused
+  before any power or SD-mux change. Checked files and directories are
+  opened at check time without following symlinks (`labctl.core.safe_open`)
+  and all later I/O uses those handles, so a symlink swapped in after the
+  check can't redirect it; raw images are streamed into `sudo dd` instead
+  of being opened by root. Image format and copied file names follow the
+  path given. The CLI is unaffected. Fresh installs
+  get `/var/lib/labctl/images` (read) and `/var/lib/labctl/output` (write);
+  `scripts/update.sh` creates those directories on existing hosts and prints
+  the config to add (D012).
+- Web UI: without a configured `auth.secret_key`, the session-signing key is
+  now random per process. It used to fall back to the constant
+  `labctl-dev-key` when auth was disabled; because the signed session
+  carries the CSRF token, anyone could forge it. Browser sessions now reset
+  when the web service restarts unless a key is configured.
+- MCP `sdwire_ls` / `sdwire_cat` / `sdwire_info` no longer have side effects.
+  A card already switched to the host is read in place and left there
+  (host mode is confirmed with a one-sector `sudo dd` read, because the
+  reader's reported size lags a switch in either direction); a
+  card on the DUT side is switched to the host and back only when the
+  board's power plug reports OFF. With no power plug, an unreadable plug, or
+  an unknown state they now refuse (`power_unknown`) instead of switching
+  the card away from a possibly running board. They are annotated read-only.
+  Callers that relied on reading a plug-less SBC's card must switch it
+  explicitly with `sdwire_to_host` first.
+- Read-only SD card mounts (`sdwire_ls`/`sdwire_cat` and the CLI
+  equivalents) no longer replay a dirty filesystem journal, which wrote to
+  the card despite `-o ro`: ext2/3/4 mount with `noload`, XFS/F2FS with
+  `norecovery`, btrfs with `rescue=nologreplay`.
+- MCP `sdwire_to_host` now waits (up to 10 s) until the card is readable on
+  the host before returning, and warns if it isn't, so an immediate
+  follow-up read doesn't misjudge the mux position.
+
+### Changed (documentation audit)
+- README, CLI help, MCP tool descriptions and docs were audited against
+  the code and corrected. Users will notice:
+  - README examples that failed as written: `labctl power on <sbc>` (not
+    `power <sbc> on`), `port assign` with its port type, `monitor` (no
+    `--foreground`), `-r` on `claim`/`force-release`/`request-release`.
+    The REST table lists the real routes, with the power endpoint's
+    `delay`, `override_claim` and status codes (400/409 `switched:
+    false`/429/500).
+  - Install instructions use the extras from a checkout (`pip install -e
+    ".[web,mcp,kasa,sdwire,actuators]"`): the package isn't on PyPI yet.
+    The README gives the Python requirements, the config search order
+    (a personal `~/.config/labctl/config.yaml` wins over `/etc`), and the
+    `labctl serial udev` flow in place of `install-udev.sh`. The SDWire
+    sudoers rule adds `parted` and `blkid` and gets its own file.
+  - New README sections on relays, bindings and recovery (`pre_power`)
+    and on the common flags (`--yes`, `--override-claim`, `--dry-run`).
+  - `docs/MCP_SERVER.md` documents the actuator, binding and recovery
+    tools; the prompt names are `debug_sbc` and `lab_report`; 24 tools
+    need confirmation; every image-upload error is listed.
+  - CLI help and MCP tool descriptions say what commands actually do,
+    e.g. `sdwire flash`/`update` power the board off and leave it off
+    without a reboot; `power on`/`cycle` set the `pre_power` relays of
+    boards sharing the outlet too; MCP/REST cycles don't use the board's
+    stored cycle delay; `actuator set` points at `labctl bindings
+    actuate`; `bind --phase post-power` is stored but not acted on yet.
+  - `SECURITY.md` lists the controls and states the web UI's default
+    exposure: the shipped unit listens on `0.0.0.0` over HTTP, auth is
+    off by default, and CSRF applies only with auth.
+  - `config/labctl.yaml.example` lists the environment variables that
+    actually override settings, and the `claims`, `proxy`, `health` and
+    `database` keys and `mcp.allow_admin_actuator_ops` with their
+    defaults.
+  - `docs/OPERATIONS.md`: `run_health_check` is `shared-resource` (it
+    writes nothing); claim-unchecked commands and aliases are complete.
+  - `docs/HARDWARE_MAP.md` is a generic example; `SPEC_actuators.md`,
+    `STATUS.md`, `CLAUDE.md` (schema v11) and `IMPLEMENTATION.md` match
+    what was built; the `deploy-and-test` skill covers confirmation,
+    claims and image upload.
+
+### Fixed
+- The SD card is no longer switched to the host while the board may be
+  running (bus contention can corrupt the card). `boot-test` with an image
+  (CLI and MCP) didn't power the board off at all; `sdwire flash` and
+  `sdwire update` (CLI and MCP) powered it off on a best-effort basis and
+  carried on when that failed. All of them now refuse, writing nothing,
+  when the board can't be switched off. A `boot-test` deploy that fails
+  once the board is off hands the card back to the board and says the
+  board was left off (it used to be a traceback for errors other than
+  RuntimeError).
+- `mcp.allow_admin_actuator_ops` turns on only for a YAML boolean `true`.
+  It was read with `bool()`, so a quoted `"false"` (or any other string)
+  enabled `actuator_add`/`actuator_remove`/`actuator_set` over MCP; any
+  non-boolean value now leaves them off, with a warning.
+- `labctl edit --cycle-delay` refuses a delay over 86400 seconds, NaN or
+  infinity (nothing is written; `-inf` no longer counts as "clear"). The
+  stored delay is the default and the floor of `power cycle`, so such a
+  value made every later cycle of the board fail, with no `--delay` able
+  to get round it; `power cycle` now says how to clear one stored
+  earlier.
+
+- `pre_power` relay bindings (e.g. a `recovery_mode` strap) are now set
+  before power-on on every path, not just CLI `power on`/`power cycle` and
+  the recovery composites (#36). The MCP `power_on`/`power_cycle`/
+  `boot_test` tools and the reboot after `sdwire_update`/`flash_image`,
+  the web UI and REST power actions, and CLI `power-all on`, `boot-test`
+  and the reboot after `sdwire flash`/`update` used to switch power and
+  leave the strap however it was last set, so a board could boot in or out
+  of recovery depending on which interface turned it on. Boards that share
+  the plug outlet have their straps set too, since the outlet powers them
+  all (this also applies to `enter-recovery`/`exit-recovery`). A relay
+  that can't be opened, or whose channel is busy with another operation,
+  now stops these before power is switched (cycles check before powering
+  off; a flash or file update with a reboot checks before it powers the
+  board off and writes the card): MCP and the CLI say so in one line, the
+  REST API answers 409 with `"switched": false` when nothing was switched
+  (a failure after power-off stays a 500), the web UI flashes it, after a
+  flash or file update the tools report that the power cycle failed and
+  why, and `boot-test` records the run as failed. A strap that fails after
+  power-off leaves the board off, and the error says so. A power-cycle
+  delay over 86400 seconds, or not a number, is now refused before
+  power-off: by CLI `power cycle` (also with `--dry-run`, and when the
+  board's stored delay is out of range; a delay below the board's minimum
+  is still raised to it), and, with negative delays too, by MCP
+  `power_cycle` and by the REST API with a 400. It used to switch the
+  board off, then fail.
+- A relay that labctl can't open no longer crashes `power on`/`power
+  cycle` (and the binding verbs and recovery composites) with a
+  traceback; nor does an actuator whose driver isn't implemented. The
+  error is one `Error:` line, and the board is never powered on without
+  its strap. `power cycle`, `enter-recovery` and `exit-recovery` now
+  check every relay their pre-power bindings need before powering off,
+  so a relay they can't open no longer leaves the board off. A
+  permission error says what to do: join the group that owns the device
+  (only when that group grants read/write and isn't root's), then start
+  a new login and restart any service or MCP client running labctl. `labctl actuator set` and the MCP
+  `actuator_set` tool report the same error instead of raising.
+- Over stdio, the MCP `list_images` tool said only that uploads need the
+  HTTP transport, which read as a dead end. It now lists what a stdio
+  session (which runs on the lab host) can do instead, with this server's
+  paths: pass an absolute path inside `mcp.allowed_read_paths`, copy the
+  image into `mcp.upload_dir` (under a unique temp name, then renamed into place;
+  it says whether the session can write there) and use its bare name, or
+  upload to the host's HTTP service. The last two are offered only when
+  bare names resolve (`upload_dir` inside `allowed_read_paths`). With
+  uploads off or `upload_dir` unreadable, a stdio session is still pointed
+  at host paths, and the server instructions no longer say files must be
+  uploaded over HTTP.
+- Kasa strips that only answer KLAP discovery (e.g. HS300 firmware 1.1.2,
+  which closed the legacy port 9999) are controlled per outlet again.
+  python-kasa types such a device by its family (`IOT.SMARTPLUGSWITCH`)
+  as a single plug with no outlets: state reads failed with
+  `KeyError: 'relay_state'`, and an on/off for any outlet was sent to the
+  strip as a whole, not to that outlet. labctl now reads the sysinfo and
+  lets python-kasa pick the class from it, as python-kasa's own `connect`
+  does, rebuilding the device on the same connection when discovery
+  picked another. A Kasa device that reports no outlets now refuses an
+  outlet index above 1 instead of switching the whole device.
+- Kasa power control works again with IOT devices whose firmware moved to
+  KLAP login version 2 (e.g. HS300 hardware 2.0, firmware 1.1.2): every
+  command failed with `AuthenticationError: Device response did not match
+  our challenge` even with correct credentials, because python-kasa before
+  0.12 always used KLAP v1 for IOT devices. The `kasa` extra now requires
+  `python-kasa>=0.12` on Python 3.11+. python-kasa 0.8 and later need
+  Python 3.11, so on 3.10 it pins 0.7.7 (it allowed 0.5, which lacks
+  APIs labctl uses) and those devices still can't be controlled there.
+  When an authentication error persists with credentials configured and
+  python-kasa < 0.12, the error message now says so and names the upgrade.
+- `labctl mcp --http` waits up to 15 s for a port that is in use before
+  starting (seen once on a deploy: the port was still in use ~1 s after
+  the old server had exited, cause unknown; the new server's start failed
+  and systemd restarted it 10 s later). It probes the addresses uvicorn
+  will bind, then lets uvicorn bind as before. `update.sh` also waits,
+  for every service, while a unit is `deactivating`, `activating` (as
+  while systemd restarts a failed start) or `reloading`, within
+  `MCP_START_TIMEOUT` (now measured in seconds), and says when systemd
+  had to restart one; it had reported `labctl-mcp` as failed.
+- The test suite no longer loads the machine's real config. About 115
+  tests loaded config without naming a file, so on a lab host whose user
+  can read `/etc/labctl` they used the production config and database; a
+  run during the image-upload work migrated a production database to a
+  pre-release schema v10 (an `images` table without `file_id`, so uploads
+  failed with 500). Schema v11 adds the missing column to such a
+  database, so a backup taken in between can be restored safely. Every test, and every process it starts, now gets its own home
+  and no system config. New `LABCTL_SYSTEM_CONFIG` environment variable:
+  replaces `/etc/labctl/config.yaml` in the config search order.
+- REST `DELETE /api/sbcs/<name>` on a claimed board answered `500`; it is
+  now `409 "sbc_claimed"` (#21). A non-numeric `duration_minutes` on
+  claim or renew answered `500`, now `400`; renew durations are now
+  bounded like claims. The web plug form refuses an outlet index below 1
+  (index 0 switches every outlet on a Tasmota device).
+- REST and web claims used a session ID the next request often didn't
+  share (a per-request ID without auth, an API-key prefix with it), so a
+  claim made over REST could not always be renewed or released; they now
+  use `web:<user>` (`web:anonymous` without auth) and record the user as
+  owner. Claims made before the upgrade keep their old ID: release them
+  with force-release if needed.
+- A command given a config file by name (`-c FILE` or `LABCTL_CONFIG`)
+  fails if it doesn't load (YAML error or invalid value) instead of
+  running on built-in defaults (commands that don't read the config,
+  e.g. `user generate-key` / `hash-password` / `add`, `completion` and
+  `--help`, still run). `labctl mcp` handles this itself: over
+  HTTP it refuses to start, over stdio it starts with a warning.
+  `labctl web` refuses to start if any config file it reads doesn't load
+  (e.g. `/etc/labctl/config.yaml` unreadable to a user outside the
+  `labctl` group: pass `-c`), since it would otherwise serve without the
+  login the intended config enables.
+- A config value that is valid YAML but not a valid setting (e.g.
+  `timeout_seconds: 10s`, or a list at the top level) is a load error
+  (warning; the next file in the search order, or built-in defaults)
+  instead of a crash.
+- An MCP `boot_test` `output_dir` that is an existing file is reported as
+  "not a directory" instead of a possible race.
+- A config section header with nothing under it (e.g. `mcp:` followed only
+  by commented-out entries) no longer crashes every labctl command; it
+  counts as an empty section. A section that isn't a mapping is ignored
+  with a warning.
+- Boot-test per-run output files (`run_NN.txt`) are written with
+  `O_NOFOLLOW` and refused if hard-linked, so a link planted in a shared
+  output directory can't redirect the (DUT-controlled) output into another
+  file. The default `/var/lib/labctl/output` is now also sticky (`3775`).
+  A run file that can't be written no longer fails the whole boot test;
+  the summary (also the MCP result) lists every file that wasn't saved.
+- `labctl -c FILE mcp` now makes the MCP server use FILE. The server loaded
+  its config itself and ignored `-c`, so it read
+  `~/.config/labctl/config.yaml` first (for the systemd unit: the service
+  user's config, not the `/etc/labctl/config.yaml` the unit names). This
+  matters now that the MCP host-file allowlist lives in config. FILE is
+  exclusive, as with any explicit path: never another config. The same
+  exclusive mode is available as `LABCTL_CONFIG_EXCLUSIVE=1` alongside
+  `LABCTL_CONFIG`. (How a running server treats later changes to FILE:
+  see Security, MCP over HTTP.)
+- MCP claim sweeper (issue #3): a crashed `mcp-stdio` session whose PID was
+  reused by another process is now recognised as gone (the process start
+  time is compared with the session's epoch), so its claim is released as
+  `session-lost` rather than later as `expired`.
+
+### Changed
+- Kasa: driver errors (bad outlet, unreachable plug) inside the MCP server
+  are reported as such, not as an unrelated asyncio error.
+- Power plugs: `labctl plug assign --index` and MCP `assign_power_plug`
+  refuse an outlet index below 1 (Tasmota's `Power0` switches every
+  relay). Rows saved with such an index (earlier, or by `labctl import`)
+  count as sharing every outlet on that address for claim checks.
+- MCP `serial_capture` is claim-checked like `serial_send`: it holds the
+  board's ser2net session.
+- HTTP MCP sessions are labelled `mcp-http` (claims' `session_kind` and
+  activity-log actors), not `mcp-stdio`; the dead-session sweep covers both.
+- **Package renamed for PyPI**: the distribution is now
+  `embedded-lab-control` (`pip install embedded-lab-control`); the import
+  package and CLI remain `labctl`. The version is single-sourced from
+  `labctl.__version__` (also used by `/api/health` and the web footer, which
+  previously hard-coded `0.1.0`). License metadata is now the SPDX
+  expression `MIT`, and the LICENSE file ships in the wheel and sdist.
+- **MCP server migrated to the mcp 2.x SDK** (`mcp>=2.2,<3`, `MCPServer`
+  replaces `FastMCP`). `mcp>=1.0` had started resolving to 2.2, where
+  `mcp.server.fastmcp` no longer exists and `labctl mcp` crashed on import.
+  Tools, resources and prompts are unchanged. mcp 2.x runs sync tools and
+  resources on worker threads; labctl keeps 1.x semantics by serializing tool
+  calls and the hardware-touching resources (`lab://power/{sbc}`,
+  `lab://health/{sbc}`, `lab://status`) under one lock (D011). The server
+  now reports labctl's version in `serverInfo`.
+- `scripts/update.sh` and `scripts/install-services.sh` uninstall the old
+  `labctl` distribution before installing `embedded-lab-control`, so the
+  first upgrade doesn't leave two distributions owning `labctl/` (a later
+  `pip uninstall labctl` would otherwise delete the new install).
+- The `sdwire` extra only installs on Python ≥ 3.12 (PyPI `sdwire` requires
+  it). On 3.10/3.11 SDWire commands report that the package is missing.
+- Development: `src/` and `tests/` are now clean under `black`, `isort` and
+  `flake8`, which are pinned exactly in the `dev` extra
+  (`black==26.5.1`, `isort==9.0.1`, `flake8==7.4.1`). Formatting-only apart
+  from removing dead code flagged by flake8 (unused imports/variables).
+
+### Added
+- MCP `serial_send` takes `keys` (named keys: ESC, Enter, Tab, Space,
+  Backspace (0x08), arrows, Home/End/Insert/Delete/PgUp/PgDn, F1-F12,
+  Ctrl-A..Ctrl-Z, in xterm encodings; each key is written 50 ms after
+  the previous one, for consoles that time out a lone ESC quickly) and
+  `data_hex` (raw bytes, e.g. for firmware whose terminal type wants
+  other function-key codes, or 0x7F to erase at a Linux prompt), so
+  agents can drive UEFI/U-Boot menus: a raw ESC in `data` is dropped
+  before it reaches the server (#33). `newline` defaults to off when
+  either is given. Its confirmation now shows the exact bytes it will
+  send (`will_do`), and an unknown key or bad hex is refused before a
+  token is issued. F11 has been checked on a Jetson's UEFI; the other
+  encodings have not been checked on real firmware yet.
+- **Image uploads over MCP HTTP** (D014): `PUT /images/<name>` on the MCP
+  HTTP port, with the same API key, stores a file in the new
+  `mcp.upload_dir` (default empty: off); `flash_image`, `sdwire_update`
+  copies and `boot_test` take an uploaded image's bare name. New tools
+  `list_images` (with upload instructions) and `delete_image`. Limits:
+  bare names, `mcp.max_upload_mb` (8192), 1 GiB kept free, optional
+  `?sha256=` check, atomic replace with `?overwrite=1`. Only the uploader
+  or an operator may replace or delete an image; uploads and deletions are
+  in the activity log. Schema v10 (`images` table).
+- **TLS for the MCP HTTP server**: `labctl mcp --http PORT --cert FILE
+  --key FILE`, or `mcp.cert_file`/`mcp.key_file`. A non-loopback bind
+  without TLS logs a warning. `labctl mcp --check` prints the `tls:` and
+  `uploads:` settings.
+- Claim enforcement, part 3 (D013): the web UI and REST API refuse to
+  operate or reconfigure a board another user has claimed (REST: `409`
+  `"sbc_claimed"`; web UI: a page naming the holder), including boards
+  sharing its hardware and boards already wired to a port or plug being
+  assigned. Operators may override (`"override_claim": true`, or the
+  button on the refusal page), recorded as `claim_override`; agents
+  (`role: agent`) get `403 "not_operator"`, also for force-release (which
+  any REST or web user could do before). `claims.require_claim` applies
+  to the REST API. REST claims record the user as owner, so one user is
+  one claimant across REST, the web UI and MCP over HTTP. See
+  docs/OPERATIONS.md.
+- Claim enforcement, part 2 (D013): CLI commands that operate or
+  reconfigure a board refuse when another user has claimed it (or a board
+  sharing its plug outlet, SDWire, console adapter or relay channel), naming
+  the holder, before any confirmation question. `--override-claim` acts
+  anyway and is recorded in the activity log as `claim_override`;
+  `power-all` skips claimed boards; `import` checks every board before
+  changing any; `claim` refuses a board whose shared hardware someone else
+  holds. `remove` releases the user's own CLI claim before deleting, and
+  `remove --force` is now an alias for `--override-claim` (also logged).
+  See docs/OPERATIONS.md for the list. Assigning a serial port a TCP port
+  another board's port already uses is now refused (CLI, MCP, web): the
+  two would share one ser2net console, reaching a claimed board's console
+  through another board.
+- Claim enforcement, part 1 (D013): MCP tools that change a board's
+  configuration (`update_sbc`, `assign_*`/`remove_*` for ports, plugs and
+  addresses, `sdwire_assign`/`unassign`, `bind`/`unbind`) now refuse under
+  another user's claim, as do `actuator_set`/`actuator_remove` on channels
+  bound to a claimed board. A claim also covers the board's hardware:
+  operating a board that shares a plug outlet, SDWire, serial device or
+  relay channel with a board someone else claimed is refused, as are
+  claiming it and assigning such hardware to another board.
+- `auth.users[].role`: `operator` (default) or `agent`. Only operators may
+  `force_release_sbc`; agents get `not_operator`.
+- `claims.require_claim` (default `false`): MCP calls that change a board
+  refuse with `claim_required` unless the caller has claimed it.
+- Claims record their authenticated owner (schema v9): the same user's
+  other MCP sessions, and the same user after a server restart, count as
+  the claimant (use, renew, release). Claiming a board you already hold
+  returns `already_claimed` and renews it for the requested duration
+  (boards sharing its hardware aren't re-checked).
+- `docs/CLAUDE_CODE_AGENTS.md`: connecting Claude Code agents to the HTTP
+  MCP server with one labctl user and API key each, so claims keep them
+  apart. Covers `claude mcp add` (manual) and a shared `--mcp-config` file
+  with the key taken from `LABCTL_API_KEY` at launch (streamlined), agent
+  instructions, identity checks and key revocation.
+- `SECURITY.md`: how to report vulnerabilities privately, and what's in
+  scope. `docs/MCP_SERVER.md` opens its security section with an overview
+  of every control (defaults, settings), known limits and a deployment
+  checklist; the README links them.
+- `labctl mcp --http PORT --check` runs the HTTP startup checks (the same
+  code the server runs) without serving: whether clients need API keys
+  (`auth: required` / `none`), the config file loaded, and startup
+  warnings; exit 1 if the server would refuse to start or a config file it
+  would read is unparseable, unreadable, or (when named by
+  `LABCTL_CONFIG`) missing. A `-c FILE` that is missing or unreadable is
+  rejected by the CLI itself (exit status 2).
+- `scripts/update.sh`, after restarting, shows `labctl-mcp`'s journal from
+  this restart (auth mode, startup warnings) and, whenever auth is
+  enabled, reminds that HTTP clients must send their API key; for any
+  service that failed to start it prints that service's log since the
+  restart. Before installing anything it refuses a config that doesn't
+  load (YAML error or an invalid value). The MCP server prints its auth
+  mode on every HTTP start and
+  logs startup warnings (short keys, unknown confirm exemptions, unchecked
+  Host headers; over stdio also config files it couldn't load) whatever
+  `log_level` says; over HTTP an unloadable config makes it refuse to
+  start instead (see Security).
+- Destructive CLI commands (19: `power off`/`cycle`, `sdwire flash`/`update`/
+  `host`/`unassign`, `serial send`, `actuator set`, binding verbs, recovery,
+  `force-release`, `boot-test`, record removals) take `--yes/-y` and ask for
+  confirmation, but only when stdin is a terminal: cron, CI, systemd, pipes
+  and agents' tool calls are not asked. Scripts run from an interactive
+  shell inherit the terminal and will be asked: add `--yes` to them. `power cycle`, `sdwire flash` and `sdwire update` take
+  `--dry-run` (validate and print the plan; change nothing).
+- `docs/OPERATIONS.md`: every CLI command, MCP tool and MCP resource
+  classified as read / db-write / shared-resource / system-write /
+  hardware / destructive (issue #7). A test keeps it in sync with the code.
+- MCP tool annotations (`title`, `readOnlyHint`, `destructiveHint`,
+  `idempotentHint`, `openWorldHint`) on all 52 tools, defined in one table in
+  `mcp_server.py`. A unit test fails if a tool is added without an entry.
+- `labctl mcp --host ADDR` (default `127.0.0.1`) for the HTTP transport.
+- `labctl mcp` without the `mcp` extra now prints an install hint instead of
+  a traceback.
+- GitHub Actions CI (`.github/workflows/ci.yml`) on push and pull request:
+  Python 3.10–3.13 matrix running flake8, black, isort and pytest with
+  coverage (XML uploaded as an artifact), plus a gitleaks scan of the full
+  git history. Reviewed gitleaks false positives are listed in
+  `.gitleaksignore`.
+- `labctl serial send` send pacing to avoid receiver UART RX-FIFO overrun
+  on sustained sends (issue #8). A gapless burst can leave a receiver no
+  window to drain its RX FIFO, dropping a contiguous mid-payload byte run.
+  Two new mutually-exclusive knobs feed the payload to ser2net in delayed
+  slices within a single TCP connection:
+  - `--interbyte-delay-ms N` — one byte at a time, N ms between bytes.
+  - `--chunk-size B --chunk-delay-ms N` — B bytes per slice, N ms between.
+  Pacing applies to both send-only and `--capture` paths. Hardware/software
+  flow control (RTS/CTS, XON/XOFF) is tracked separately in issue #9.
+
+### Fixed
+- `labctl mcp --http PORT` ignored PORT and always bound `127.0.0.1:8000`.
+  It now binds `127.0.0.1:PORT` (or `--host`). **Deployment note:** the
+  shipped systemd unit uses `--http 8080`, so a deployed server moves from
+  `:8000` to `:8080`. With a loopback bind, the SDK's DNS-rebinding guard
+  rejects non-localhost `Host` headers (HTTP 421).
+- `labctl.sdwire` failed to import on Python 3.10 (`datetime.UTC` is
+  3.11+), despite `requires-python >= 3.10`.
+- Hotfix: per-connection SQLite pragmas (`busy_timeout`, `synchronous=NORMAL`,
+  `foreign_keys=ON`) are now best-effort. After the WAL switch, callers
+  without OS-level write permission on the DB file (e.g. users outside
+  the `labctl` group running `labctl serial send`) hit
+  `sqlite3.OperationalError: attempt to write a readonly database`
+  at connection open, because SQLite auto-downgrades the connection to
+  read-only and then rejects write-side pragmas. We now wrap each
+  pragma in `try/except sqlite3.OperationalError` and log at debug.
+  Writable connections still get full tuning; read-only callers keep
+  working. Regression test added in `tests/unit/test_database.py`.
+- Transient `sqlite3.OperationalError: database is locked` errors under
+  multi-process load (monitor daemon + CLI + MCP + batch jobs hitting
+  the same SQLite file). Root cause was the combination of default
+  rollback-journal mode (readers and writers exclude each other across
+  the whole DB), the default 5s `sqlite3.connect()` busy timeout, and
+  `Database.initialize()` re-running its `sqlite_master` probe on every
+  `get_database()` call. Fix:
+  - `PRAGMA journal_mode=WAL` applied on first `initialize()` (persists
+    in the DB file header — concurrent processes inherit it). Readers
+    no longer block writers and vice versa; only writer-vs-writer is
+    serialized, and labctl writes commit sub-millisecond.
+  - `PRAGMA synchronous=NORMAL` (WAL-safe pairing, faster commits, a
+    few-second worst-case write loss on power failure — fine for a lab
+    DB).
+  - Per-connection `busy_timeout` mirrors `Config.database.timeout_seconds`
+    (new; default **10s**), and the same value is passed as
+    `sqlite3.connect(timeout=…)`. Backstops the narrow windows WAL
+    still serializes (schema migrations, WAL checkpoints).
+  - `initialize()` is now cached per-process: subsequent calls on the
+    same `Database` instance return immediately without re-opening a
+    connection. Long-lived processes (daemon, MCP, web) pay the
+    sqlite_master probe exactly once; short-lived CLI invocations still
+    pay one initialize() per run.
+  - New `database` config section (`database.timeout_seconds`, default
+    10.0) exposed via `Config.database`. Wired through `get_database` /
+    `get_manager` and the four call sites (CLI, web app, MCP server).
+  - Note: WAL mode creates `-wal` and `-shm` sidecar files next to the
+    DB file. Backup scripts that previously copied just `labctl.db` now
+    need to include those sidecars (or quiesce the daemon and run
+    `PRAGMA wal_checkpoint(TRUNCATE)` before the backup).
+  - `LABCTL_DATABASE_TIMEOUT` env var overrides
+    `database.timeout_seconds` at load time, parallel to the other
+    `LABCTL_*` overrides. Useful for ad-hoc tuning under contention
+    without editing config.yaml.
+  - `INSERT OR IGNORE` on the `schema_version` bootstrap inserts
+    eliminates a latent cross-process race where two fresh-DB
+    initializers could both pass the `sqlite_master` probe and the
+    losing one would fail with a PRIMARY KEY conflict. Independent
+    of the lock fix but the WAL change made it slightly more likely
+    by shrinking the serialization window.
+  - `Database._init_lock` is a `threading.RLock` (not `Lock`) so a
+    callback into `initialize()` while the lock is held does not
+    self-deadlock. Defensive — nothing in the current codebase
+    triggers re-entry.
+
+### Added (actuators feature, P1–P7, PR #4)
+
+First-class support for USB-controlled relays as provisioned lab
+resources, with per-target purpose **bindings** (recovery_mode,
+boot_select, power_button, …) so verbs are semantic
+(`labctl enter-recovery jetson-nano-2`) rather than device-specific.
+Built per `docs/SPEC_actuators.md`. Since #37, `pre_power` straps are
+set on every power-on path (see Fixed above).
+
+- **Schema v8** — `actuators`, `actuator_channels`, `bindings` tables
+  with CHECK constraints on enum columns, UNIQUE on
+  `actuator_channel_id` (one binding per channel), UNIQUE on
+  `(sbc_id, purpose)` (one purpose per SBC). Migration v7→v8.
+- **Driver layer** (`labctl.actuators`):
+  - `RelayDriver` ABC with structured `WriteOutcome` /
+    `ProbeResult` so composites can abort safely on USB drop mid-sequence.
+  - `lcus1_serial` — CH340-based 4-byte ASCII protocol at 9600/8N1,
+    write-only, software-tracked state, side-effect-free probe.
+  - `MockRelayDriver` for tests with programmable failure modes.
+  - `get_driver()` registry; future drivers raise NotImplementedError.
+- **Runtime layer** (`labctl.actuators.runtime`):
+  - Per-channel non-blocking locks ⇒ ChannelBusyError for concurrent
+    verbs; verb/shape congruence enforced (BindingShapeError).
+  - `actuate_binding` / `release_binding` / `press_binding` —
+    desired_state persisted only on WriteOutcome.OK.
+  - `apply_pre_power_bindings` — re-applies pre_power straps before
+    `power_on` / `power_cycle` so recovery_mode survives power cycles.
+  - `enter_recovery` / `exit_recovery` — pre-flight probe + power_off
+    + sleep + (de)assert + power_on, with explicit abort if the
+    actuator is unreachable BEFORE any power transition.
+  - `apply_safe_drive_on_startup` — daemon-start reconciliation that
+    leaves `desired_state=asserted` channels alone (recovery
+    mid-flash survives a daemon restart) and drives everything else
+    to `default_state`.
+- **CLI**:
+  - `labctl actuator add/list/probe/remove/set`
+  - `labctl bind`, `labctl unbind`, `labctl bindings list`
+  - `labctl bindings actuate / release / press / status`
+  - `labctl enter-recovery / exit-recovery`
+  - `power on` / `power cycle` now consult pre_power bindings.
+- **MCP**: complete tool set — provisioning, binding management,
+  operation verbs (claim-gated), and recovery composites.
+- **SBC claim composition**: `claim_sbc` snapshot-checks per-channel
+  locks; refuses with ChannelBusyError if a verb is mid-flight.
+- **Wiring-polarity validation** at bind time: refuses
+  `default_state == active`.
+- **`pyproject.toml`**: new optional `actuators` extra pulling
+  `pyserial>=3.5`.
+
+Tests: **897/897 pass.** ~140 new tests across schema migration,
+manager CRUD, driver ABC contract, LCUS-1 frame builder + I/O,
+runtime verbs, composite power integration, daemon-start safe-drive,
+claim composition, CLI surface, and MCP tools.
+
+#### Post-review hardening (PR #4)
+
+- Hardware writes from automatic power-flow paths now leave audit
+  trails alongside operator-initiated verbs: `apply_pre_power_bindings`
+  emits `pre_power_apply` events (one per strap driven); daemon-start
+  `apply_safe_drive_on_startup` emits one synthetic `safe_drive`
+  summary event with held/drove/failed counts and details.
+- `_validate_actuator_device_path` now rejects path traversal:
+  refuses any input containing `..` outright, then normalizes
+  remaining paths with `os.path.normpath` before the allow-list
+  check. Previously `/dev/lab/../etc/passwd` slipped through.
+- `_audit_actuator_event` / `_audit_raw_actuator_set` / new
+  `_audit_safe_drive` wrap their emit call in try/except so a
+  bookkeeping failure can't fail a verb that physically succeeded.
+- `_admin_actuator_ops_allowed` now logs at debug on config-load
+  errors so silent gate-disables are diagnosable.
+- `pre_power_apply` audit details now include `cycle_changed` so a
+  paired `actuate` → `pre_power_apply` row pair (the normal
+  `enter_recovery` shape) reads as "operator drove the strap, then
+  the power-flow rechecked it" rather than as a double-toggle.
+
+### Fixed
+- Kasa multi-outlet strips no longer hit themselves with parallel KLAP
+  handshakes during monitor cycles. Previously each outlet of an N-outlet
+  strip opened its own KLAP session in parallel, so the strip
+  rate-limited bursts of N concurrent auth attempts as
+  `AuthenticationError`. Symptom in production logs: every monitor cycle
+  produced a wave of "challenge mismatch" errors against `192.168.4.140`
+  (a 6-outlet HS300) followed by an exhausted-retry ERROR per outlet,
+  then a fresh wave 60 s later. Retries never helped because the cause
+  was contention, not transience. Fix: a process-wide per-host state
+  cache (5 s TTL) with a per-host serialization lock. The first caller
+  in a cycle pays one KLAP handshake and populates is-on for every
+  outlet; concurrent callers on sibling outlets read from the cache.
+  Writes invalidate so the next read reflects the change.
+
+### Added
+- `labctl status` now appends the last 5 audit-log events under the SBC
+  table for at-a-glance recent-action context (oldest first so the newest
+  sits closest to the prompt).
+- Per-SBC `power_cycle_delay_seconds` (schema v7). Acts as both the
+  default when `labctl power cycle --delay` is omitted *and* the floor
+  when a smaller value is passed (the CLI raises it with a warning).
+  Set/clear with `labctl edit <name> --cycle-delay <seconds>`
+  (negative value clears the override). Global fallback default is 3 s.
+
+### Changed
+- Kasa retry path now logs at WARNING for each retry attempt (was DEBUG,
+  invisible at default log levels), logs at INFO when a retry succeeds,
+  and logs at ERROR on final failure. Default retry count bumped from 1
+  to 2 (3 total attempts) to absorb intermittent KLAP auth flakes.
+- `labctl power cycle --delay` is now optional. Default and floor are
+  resolved per-SBC: SBC value if set, otherwise 3 s. Help text corrected
+  (was "default: 2s", code default was 2.0 — now 3.0).
+- Monitor daemon now runs two cadences (2026-04-30):
+  - **Fast track (`health.check_interval`, default 10 s)** — ping +
+    serial probes, parallelized across SBCs via a thread pool. Sets
+    `sbcs.status` more responsively.
+  - **Slow track (`health.power_check_interval`, default 60 s)** —
+    power probes, also parallelized. Stays at 60 s by default to limit
+    pressure on plug controllers (Kasa KLAP especially).
+  - **`health.min_sleep_seconds` (default 1.0)** — floor on the
+    between-cycle sleep so a runaway cycle (elapsed ≥ interval) can't
+    pin a CPU spinning.
+  - **Default `check_interval` 60 s → 10 s** for fresh installs;
+    existing configs keep their explicit value. Set
+    `power_check_interval: 10` in config to also tighten power
+    polling once Phase 2 lands (Kasa broadcast discovery).
+- `HealthChecker.check_all` is now parallel (ThreadPoolExecutor,
+  max 16 workers). Wall time is the *max* of per-SBC probe latencies
+  rather than the sum. Per-SBC failures isolate; one bad probe
+  returns an empty summary instead of dropping the row.
+- `labctl status --fast` staleness threshold derives from
+  `2 × power_check_interval` (was hardcoded 120 s) so it adapts when
+  you change the cadence.
+
+### Added
+- `labctl status --fast` (alias `-f`) renders from the daemon's cached
+  power observation instead of probing every plug live. Boot times go
+  from "stuck on the slowest plug" to instant: 6.6s → 0.15s on a 6-SBC
+  lab. Stale observations (older than 2× the monitor interval, default
+  120s) render dimmed so it's obvious when the daemon is behind. The
+  default mode still does live probing — `--fast` is opt-in. Schema v6
+  adds `last_power_state` and `last_power_at` columns to `sbcs`; the
+  monitor daemon UPDATEs them every cycle. 11 new tests cover the
+  schema migration, manager observation API, and `--fast` rendering
+  including staleness dimming and the no-observation-yet case
+  (2026-04-27)
+
+### Fixed
+- `labctl connect` now puts the local TTY in raw mode for the duration
+  of the session, so each keystroke is forwarded immediately. Previously
+  it shelled out to `nc`/`telnet`, which left the local terminal in
+  cooked/line-buffered mode — boot menus with ~3s timeouts (Jetson
+  U-Boot, Raspberry Pi, etc.) silently dropped keystrokes because they
+  sat in the kernel until the user hit Enter. Native loop with select()
+  + tty.setraw + TCP_NODELAY; non-TTY stdin/stdout still falls back to
+  `nc`. New escape sequence: Ctrl+] then 'q' (or Ctrl+\) to disconnect,
+  Ctrl+] then Ctrl+] to send a literal Ctrl+]. 12 regression tests
+  cover the escape state machine and TTY/non-TTY dispatch (2026-04-26)
+
+- `labctl serial list` showed registered adapters as "unassigned" when
+  their assignment rows had `serial_device_id = NULL`. Root cause:
+  `ResourceManager.assign_serial_port()` only populated the FK when
+  the caller explicitly supplied `serial_device_id`. Callers that
+  passed only a `/dev/lab/<name>` path (MCP `assign_serial_port`,
+  REST API, and CLI without `--serial-device`) left the FK NULL, so
+  the two-tier adapter→port link was broken while the port itself
+  kept working at the device layer.
+  Fix: `assign_serial_port()` now auto-resolves the FK from
+  `device_path` (expects `/dev/lab/<name>` or a bare name). New
+  `labctl serial repair [--apply]` command backfills the FK for
+  existing orphan rows (dry-run by default). 6 new regression tests
+  cover the auto-resolve and repair paths (2026-04-19)
+
+### Added
+- Activity stream Phase D — polish (2026-04-20)
+  - MCP resources `lab://activity/recent` and
+    `lab://activity/{sbc_name}` now expose the most recent activity
+    events as JSON
+  - `labctl activity export --format ndjson` exports activity events
+    for downstream processing
+  - The existing claim-sweep worker now also prunes activity events
+    older than 30 days
+  - Regression coverage added for MCP resources, CLI export, and
+    activity retention pruning
+- Activity stream Phase C — source fidelity (2026-04-20)
+  - MCP mutating tools now run inside `activity_context()` so audit
+    events record `source=mcp` and the active MCP session ID as `actor`
+    instead of falling back to `internal`
+  - Flask request lifecycle now sets per-request audit attribution for
+    authenticated REST API and web dashboard mutations (`api:<user>` /
+    `web:<user>`)
+  - Web claim endpoints now derive the claimant name from the correct
+    session key (`session["user"]`)
+  - Regression tests added for MCP, API-key, and logged-in web
+    mutation attribution
+- Activity stream Phase B — live (2026-04-19)
+  - `ActivityBroadcaster` background thread polls `audit_log` every
+    500ms and fans new rows to per-subscriber queues. Cross-process:
+    events from CLI/MCP/daemon reach browsers because everyone writes
+    to the same DB.
+  - `GET /activity` — web page with live feed, source/result filter
+    chips, pause/resume, and server-rendered hydration of the last
+    200 events.
+  - `GET /activity/stream` — Server-Sent Events endpoint, session-
+    authed. 15s keepalive heartbeats.
+  - `GET /api/activity` — JSON query endpoint (API-key auth), same
+    filters as the CLI.
+  - `labctl activity tail --follow` — streams live events by polling
+    the DB directly (works when the web service is stopped).
+  - Activity link added to the nav bar.
+  - 9 broadcaster tests (no-replay semantics, fan-out, unsubscribe,
+    SSE frame format, slow-consumer drain, hydration + live).
+- `labctl services status` — reports on systemd units labctl depends
+  on: `labctl-monitor`, `labctl-mcp`, `labctl-web`, and `ser2net`.
+  Shows ActiveState/SubState, uptime, restart counter, and recent
+  error-level journal lines for any unit that is unhealthy or has
+  restarted. Exits non-zero if anything is not running. Extend the
+  list with `--unit`. Non-privileged (uses `systemctl show` and
+  `journalctl -u`) (2026-04-19)
+- Activity stream Phase A — foundation (2026-04-19)
+  - `labctl.core.audit` module: `emit()`, `activity_context()`,
+    `set_context()`, redaction, millisecond timestamps
+  - Schema v5: extends `audit_log` with `actor`, `source`, `result`,
+    `claim_id` columns + indexes
+  - CLI boundary sets actor/source context in `main()`
+  - `labctl activity tail` queries the stream with
+    `--sbc/--actor/--source/--result/--since/-n` filters
+  - Power CLI commands emit activity events on success/failure
+  - Existing manager `_audit_log` calls route through `audit.emit()`
+  - 18 tests: schema migration, emit, contextvars, redaction, truncation
+  - Spec: `docs/RESOLVED_SPEC_activity_stream.md`
+
+
+- Hardware claims polish, Phase E (2026-04-16)
+  - `ClaimsConfig.validate()` clamps invalid bounds at load time
+  - `prune_released_claims()` auto-deletes old released claims (wired
+    into CLI expire + MCP sweep)
+  - `get_claim_metrics()` aggregate stats — exposed via `labctl claims stats`
+    and `lab://claims/metrics` MCP resource
+  - `AGENT_RULES.md` section 11: claim workflow, naming, duration guide
+- Hardware claims operator tooling, Phase D (2026-04-16)
+  - Web REST API: full claim lifecycle at `/api/claims/*`
+  - Dashboard: claim badges on SBC cards (agent, time remaining, request warning)
+  - SBC detail page: claim section with holder info, pending requests,
+    and force-release button
+  - MCP tool responses: `_claim_advisory()` appends pending release request
+    notices to all 10 gated tools on success
+- Hardware claims expiry and dead-session handling, Phase C (2026-04-16)
+  - MCP `atexit` handler: releases this session's claims on clean exit
+  - `release_dead_sessions()`: checks `kill -0 <pid>` for mcp-stdio claims;
+    dead process + expired grace → auto-release as `session-lost`
+  - Background daemon thread in MCP server sweeps every 30s
+  - `labctl claims expire` CLI command for cron-driven sweeps
+  - Grace period respected: dead sessions within grace window are left alone
+- Hardware claims MCP integration, Phase B (2026-04-16)
+  - MCP tools: `claim_sbc`, `release_sbc`, `renew_sbc_claim`, `list_claims`,
+    `get_claim`, `request_sbc_release`, `force_release_sbc`
+  - MCP resources: `lab://claims`, `lab://claims/{sbc_name}`,
+    `lab://claims/history/{sbc_name}`
+  - Claim enforcement on 10 mutating MCP tools — other-agent claims
+    return structured `sbc_claimed` JSON errors with hints
+  - Session identity via `mcp-stdio:<pid>-<start_epoch>` for stdio transport
+  - Heartbeat fires on every claimant tool call (reads and writes)
+- Hardware claims (exclusive access coordination), Phase A (2026-04-16)
+  - SQLite schema v4 adds `claims` and `claim_requests` tables
+  - Partial unique index enforces at most one active claim per SBC
+  - CLI: `labctl claim`, `release`, `renew`, `force-release`, `request-release`
+  - CLI: `labctl claims list | show <sbc> | history <sbc>`
+  - `labctl status` surfaces the current claim holder and pending release requests
+  - `labctl remove <sbc>` refuses while a claim is active (use `--force` to override)
+  - Manager ops: `claim_sbc`, `release_claim`, `renew_claim`, `heartbeat_claim`,
+    `get_active_claim`, `list_active_claims`, `list_claim_history`,
+    `force_release_claim`, `expire_stale_claims`, `record_release_request`
+  - Structured exceptions: `ClaimConflict`, `ClaimNotFoundError`, `NotClaimantError`,
+    `UnknownSBCError`
+  - Acquisition runs expire-stale sweep inside the transaction so past-deadline
+    claims don't block new acquisitions even when no sweeper has run
+  - Spec: `docs/RESOLVED_SPEC_claims.md`
+- SDWire power safety interlock (2026-04-09)
+  - `sdwire_to_host` rejects if SBC is powered on (prevents SD card bus contention)
+  - `--force` flag (CLI) / `force` parameter (MCP) to override when SBC is halted but power on
+  - `sdwire_update` and `flash_image` auto power-off SBC before switching to host
+  - Best-effort: allows operation when power state is unknown
+- `flash_image` MCP tool and enhanced `labctl sdwire flash` CLI (2026-04-06)
+  - Supports .img, .img.xz, .img.gz compressed formats
+  - Safety checks: block device size validation (max 256 GB), mount check, image format validation
+  - Post-flash file copies to boot partition (e.g., custom config.txt)
+  - On flash failure, SD card left on host for inspection
+  - Automatic power-off before flash, optional reboot after
+- MCP CRUD tools for full device lifecycle management (2026-04-04)
+  - Remove tools: `remove_serial_port`, `remove_network_address`, `remove_power_plug`
+  - Serial device tools: `add_serial_device`, `remove_serial_device`, `serial_discover`
+  - SDWire tools: `sdwire_add`, `sdwire_remove`, `sdwire_assign`, `sdwire_unassign`, `sdwire_discover`
+  - AI assistants can now fully manage lab resources without CLI
+- `/deploy-and-test` Claude Code skill (2026-04-02)
+  - Orchestrates build, deploy, boot, capture, and analysis via MCP tools
+  - Auto-discovers config from `.claude/deploy-config.json`, `CLAUDE.md`, or user prompts
+  - Boot reliability testing mode with `--runs N`
+  - Located at `skills/deploy-and-test.md`, opt in via project settings.json
+- Serial I/O and boot test tools (2026-04-02)
+  - `labctl serial capture <port> --timeout N --until <pattern>` — capture serial output with timeout or pattern match
+  - `labctl serial send <port> <data>` — send data to serial port, optional `--capture` for response
+  - `labctl boot-test <sbc>` — automated boot reliability testing: deploy, reboot N times, report success rate
+  - MCP tools: `serial_capture`, `serial_send`, `boot_test`
+  - Port name resolution: accepts port alias or SBC name (defaults to console port)
+  - 33 new tests for serial capture, send, and boot test
+- SDWire SD card multiplexer support (2026-03-31)
+  - Supports SDWire, SDWireC, and SDWire3 devices via `sdwire` Python library
+  - `labctl sdwire` CLI command group: discover, add, remove, list, assign, unassign, dut, host, flash, update
+  - `labctl sdwire flash <sbc> <image>` — automated workflow: switch to host, dd image, switch to DUT, power cycle
+  - `labctl sdwire update <sbc> -p N -c src:dest -r old:new -d file` — atomic partition operations: copy, rename, delete files, optional reboot
+  - Destination paths relative to partition root (e.g., `kernel.img` not `/boot/firmware/kernel.img`)
+  - Database schema v3: `sdwire_devices` and `sdwire_assignments` tables
+  - SDWire shown in `labctl info` output
+  - MCP resource `lab://sdwire-devices` and tools `sdwire_to_dut`, `sdwire_to_host`, `sdwire_update`
+  - Export/import includes SDWire device assignments
+  - Sudoers-based privilege escalation for SD card mount/umount/dd/sync operations (enables MCP server to manage SD cards without running as root)
+  - 55 new tests (322 total)
+- Global `--delay` / `-d` CLI option (2026-03-29)
+  - Adds a delay in seconds before any command executes
+  - Useful for chaining commands after power cycles (e.g., `labctl -d 30 health-check`)
+  - Suppressed with `--quiet`, prints "Waiting Xs..." otherwise
+- MCP (Model Context Protocol) server for AI integration (2026-03-28)
+  - `labctl mcp` command — stdio (default) and HTTP transports
+  - 7 resources: lab://sbcs, lab://sbcs/{name}, lab://power/{name},
+    lab://serial-devices, lab://ports, lab://health/{name}, lab://status
+  - 10 tools: power_on/off/cycle, run_health_check, add/remove/update_sbc,
+    assign_serial_port, assign_power_plug, set_network_address
+  - 2 prompts: debug-sbc (guided debugging), lab-report (status overview)
+  - Uses official `mcp` Python SDK with FastMCP
+  - 28 comprehensive tests for tools and resources
+- Two-tier serial device management (2026-03-28)
+  - `serial_devices` database table for registering physical USB-serial adapters
+  - `labctl serial` CLI command group: discover, add, remove, list, rename, udev
+  - Port assignments now support `--alias` for human-friendly names and `--serial-device` to link to registered adapters
+  - `labctl connect` resolves by alias, SBC name, or device path
+  - `labctl serial udev --install --reload` generates udev rules from database (replaces manual YAML editing)
+  - Pure-Python USB discovery via `/sys` and `udevadm` (no external script dependency)
+  - Database schema v2 migration (adds serial_devices table, alias/serial_device_id columns to serial_ports)
+  - Export/import includes serial devices and port aliases
+  - Web UI updated: alias field in port assignment form, alias column in port table
+  - REST API updated: serial device endpoints, alias in port endpoints
+  - Udev rules file group-writable by labctl group (no sudo needed for install)
+  - Sudoers entries for passwordless udevadm and ser2net restart
+- Kasa power retry logic (2026-03-28)
+  - Retries up to 2 times on transient KLAP authentication failures
+  - Handles intermittent HS300 firmware handshake errors transparently
+- Native HTTPS support for web server (2026-03-28)
+  - `--cert` and `--key` CLI flags on `labctl web` command
+  - `web:` config section for persistent SSL configuration (cert_file, key_file)
+  - Documentation for self-signed certificate generation and setup
+  - Systemd service updated with SSL cert read access and port 443 capability
+- Kasa Smart Power Strip support (2026-03-28)
+  - Auto-detects plugs vs strips via `kasa.Discover.discover_single()`
+  - Multi-outlet control using `plug_index` (1-based) for individual outlets
+  - TP-Link cloud credential support via `kasa:` config section (required for KLAP devices)
+  - Proper async cleanup (`device.disconnect()`) to prevent unclosed session warnings
+  - Detailed logging at all levels (DEBUG/INFO/ERROR) for troubleshooting
+- SBC renaming support (2026-03-28)
+  - `labctl edit <name> --rename <new-name>` CLI command
+  - Rename via REST API (`PUT /api/sbcs/<name>` with `{"name": "new-name"}`)
+  - Rename via web UI edit form
+- CLI logging initialization (2026-03-28)
+  - `logging.basicConfig()` wired to config `log_level` in CLI entrypoint
+  - `-v` flag sets DEBUG, `-q` flag sets WARNING
+  - Log output goes to stderr to avoid interfering with command output
+- `scripts/kasa-debug.py` — standalone debug script for Kasa device troubleshooting
+  - Device discovery, status, and power control
+  - `test-auth` action with credential hash diagnostics
+  - Verbose mode includes python-kasa internal debug output
+
+### Changed
+- Default log level changed from INFO to WARNING (2026-03-28)
+- Kasa controller logging lowered from INFO/ERROR to DEBUG (2026-03-28)
+  - Prevents log noise in `labctl status` output
+- `labctl ser2net generate` uses port alias as connection name when available (2026-03-28)
+- `labctl ser2net reload` auto-escalates to sudo when needed (2026-03-28)
+
+### Fixed
+- Critical: `get_status_history()` and `get_uptime()` called nonexistent `execute_query()` (2026-04-03)
+  - Replaced with `execute()`, fixed tuple indexing to dict-style access for `sqlite3.Row`
+- Critical: Open redirect in login — `next` parameter now rejects absolute and protocol-relative URLs (2026-04-03)
+- Path traversal protection in SDWire file operations (2026-04-03)
+  - Copy, rename, and delete paths validated via `realpath` to stay under mount point
+- Null checks after `execute_one()` on 5 insert-then-retrieve sites in manager (2026-04-03)
+- Power controller error handling improvements (2026-04-03)
+  - Added warning-level logging to Tasmota/Shelly bare exception handlers
+  - Added timeout to Kasa `future.result()` to prevent indefinite blocking
+  - Fixed missing space in Kasa error message formatting
+- SerialProxy thread safety fixes (2026-04-03)
+  - `asyncio.Lock` for write lock acquisition to prevent TOCTOU race
+  - Snapshot `clients` dict before broadcast iteration
+  - Guard orphaned `create_task(self.stop())` with state check
+- ser2net generate: add /dev/lab/ prefix for bare device names (2026-04-09)
+- ser2net format string: revert to `<baud><parity><databits><stopbits>` (e.g. `115200n81`) — ser2net rejects the comma-separated form with "unknown parameter 8n1" (2026-04-16)
+- Monitor daemon: `AlertManager.trigger()` call on power state change used kwargs instead of a single `Alert` object, raising `TypeError` each cycle. The uncaught exception aborted `run_once()` mid-iteration, preventing DB status updates for any SBCs processed after the crashing one. Added regression tests for `MonitorDaemon` (2026-04-17)
+- Consolidated SBC.to_dict() serialization, eliminating duplication between MCP and API (2026-04-03)
+- Atomic upsert (INSERT...ON CONFLICT) for serial ports, network addresses, power plugs, SDWire assignments (2026-04-03)
+- Config load exception now logs warning instead of silently continuing (2026-04-03)
+- SessionLogger rotation handles disk errors without leaking file handles (2026-04-03)
+- ser2net format string: extracted PARITY_CHARS constant (2026-04-03)
+- Extracted `_load_serial_device()` helper, eliminating 3x code duplication in manager (2026-04-03)
+- Removed unused `_connection` field from Database class (2026-04-03)
+- Simplified redundant `is True`/`is False` checks in Shelly controller (2026-04-03)
+- SessionLogger timestamps now include milliseconds to prevent collision (2026-04-03)
+- MCP parameter validation: `power_cycle` rejects negative delay, `boot_test` requires runs >= 1 and timeout > 0 (2026-04-03)
+- Power control API rate limiting: 5s minimum between power cycles per SBC (HTTP 429) (2026-04-03)
+- Health check power probe using wrong function signature (2026-03-28)
+  - `get_controller(sbc.power_plug)` → `PowerController.from_plug(sbc.power_plug)`
+- Monitor daemon ping failing under systemd sandbox (2026-03-28)
+  - Removed `NoNewPrivileges=yes`, added `AmbientCapabilities=CAP_NET_RAW`
+- Kasa unclosed aiohttp sessions on authentication failure (2026-03-28)
+  - `_get_device()` now disconnects device in all error paths
+- USB serial discovery returning root hub path instead of device-specific paths (2026-03-28)
+- Authentication system (2026-03-06)
+  - Session-based web login with username/password
+  - API key authentication via `X-API-Key` header for REST endpoints
+  - CSRF protection on all web POST/PUT/DELETE forms
+  - Auth disabled by default — existing deployments unaffected
+  - `/api/health` remains open without authentication for monitoring tools
+  - `AuthConfig` and `UserConfig` dataclasses in config module
+  - Login page template with dark theme styling
+  - Logout button in navigation bar when logged in
+  - `labctl user` CLI command group:
+    - `labctl user hash-password` — generate werkzeug password hash
+    - `labctl user generate-key` — generate random API key
+    - `labctl user add <username>` — interactive user creation with YAML output
+    - `labctl user verify <username>` — verify password against config
+  - Auth section in `config/labctl.yaml.example`
+  - 14 authentication integration tests (188 total)
+  - `[tool.setuptools.package-data]` in pyproject.toml for templates and static files
+- Project documentation structure
+  - ARCHITECTURE.md - System design and component specifications
+  - IMPLEMENTATION.md - Milestone-based TODO lists
+  - AGENT_RULES.md - Development process and standards
+  - STATUS.md - Current project state tracking
+  - CHANGELOG.md - This file
+- M1.1 Project Setup (2025-12-31)
+  - Created project directory structure (src/labctl, tests, scripts, config, docs)
+  - Added pyproject.toml with dependencies (click, pyyaml, requests)
+  - Created virtual environment with dev tools (pytest, black, isort, flake8)
+  - Installed ser2net system package
+- M1.2 udev Rules System (2025-12-31)
+  - Created `scripts/discover-usb-serial.sh` for USB device enumeration
+  - Created `scripts/generate-udev-rules.py` for rule generation from YAML mapping
+  - Created `scripts/install-udev.sh` for udev rules installation
+  - Created `docs/HARDWARE_MAP.md` with USB topology documentation
+  - Symlinks now created under `/dev/lab/` for deterministic device access
+- M1.3 ser2net Configuration (2025-12-31)
+  - Created `src/labctl/serial/ser2net.py` for config generation
+  - Support for baud rate, parity, local-only connections
+  - Generated config for 5 ports on TCP 4000-4004
+  - All ports verified accessible via TCP
+- M1.4 Basic CLI Structure (2025-12-31)
+  - Created `src/labctl/cli.py` with Click framework
+  - Implemented `labctl ports` - lists configured serial ports with TCP mappings
+  - Implemented `labctl connect <port>` - connects via TCP or direct serial
+  - Added 8 CLI integration tests
+- M1.5 Configuration File (2025-12-31)
+  - Created `src/labctl/core/config.py` with Config dataclasses
+  - YAML config loading from ~/.config/labctl/config.yaml
+  - Environment variable overrides (LABCTL_DEV_DIR, LABCTL_BASE_TCP_PORT, etc.)
+  - Added `-c/--config` CLI option
+  - Added 17 unit tests for config module
+  - **Milestone 1 Complete!**
+- M2.1-2.5 Data Layer (2025-12-31)
+  - Created `src/labctl/core/database.py` with SQLite schema
+    - Tables: sbcs, serial_ports, network_addresses, power_plugs, status_log, audit_log
+    - Foreign keys with cascade delete
+    - Schema versioning support
+  - Created `src/labctl/core/models.py` with data models
+    - SBC, SerialPort, NetworkAddress, PowerPlug dataclasses
+    - Status, PortType, AddressType, PlugType enums
+    - from_row() methods for database deserialization
+  - Created `src/labctl/core/manager.py` with ResourceManager
+    - Full CRUD operations for SBCs
+    - Serial port, network address, power plug assignment
+    - Audit logging for all operations
+  - Added CLI SBC management commands
+    - `labctl list` - tabular output with filters (--project, --status)
+    - `labctl add <name>` - create SBC with options
+    - `labctl remove <name>` - delete with confirmation
+    - `labctl info <name>` - detailed SBC view
+    - `labctl edit <name>` - update properties
+  - Added CLI port assignment commands
+    - `labctl port assign <sbc> <type> <device>` - with auto TCP port
+    - `labctl port remove <sbc> <type>`
+    - `labctl port list` - all port assignments
+  - Added 27 new tests (8 database, 19 manager)
+  - Total: 62 tests passing
+- M2.6-2.7 Network and ser2net Commands (2025-12-31)
+  - Added `labctl network set <sbc> <type> <ip>` with --mac, --hostname options
+  - Added `labctl network remove <sbc> <type>`
+  - Added `labctl ser2net generate` - generates config from database
+    - Supports --output file and --install flag
+  - Added `labctl ser2net reload` - restarts ser2net service
+  - **Milestone 2 Complete!**
+- M3 Power Control (2025-12-31)
+  - Created power module `src/labctl/power/` with controller framework
+    - PowerController ABC with power_on, power_off, power_cycle, get_state
+    - PowerState enum (ON, OFF, UNKNOWN)
+    - Factory function get_controller() for plug type dispatch
+  - Implemented TasmotaController for Tasmota smart plugs
+    - HTTP API client with multi-relay support
+  - Implemented ShellyController for Shelly devices
+  - Implemented KasaController stub for TP-Link Kasa (requires python-kasa)
+  - Added CLI plug commands
+    - `labctl plug assign <sbc> <type> <address>` with --index option
+    - `labctl plug remove <sbc>`
+  - Added CLI power commands
+    - `labctl power on/off/cycle/status <sbc>`
+    - `labctl power-all on|off` with --project filter
+  - Added 19 power module tests (81 total)
+  - **Milestone 3 Complete!**
+- M4 CLI Completion (2025-12-31)
+  - Added `labctl console <sbc>` - connect to SBC's serial console by name
+    - Looks up port from database, supports --type for jtag/debug
+  - Added `labctl ssh <sbc>` - SSH to SBC by name
+    - Looks up IP from database, uses configured ssh_user
+    - Supports --user override
+  - Added `labctl status` - overview with color-coded status and power state
+  - Added `labctl export` - export SBCs to YAML/JSON
+  - Added `labctl import <file>` - import SBCs with --update option
+  - Added `labctl completion` - shell completion for bash/zsh/fish
+  - **Milestone 4 Complete!**
+- M5 Web Interface (2025-12-31)
+  - Created Flask web module with application factory pattern
+  - Implemented REST API endpoints:
+    - `GET/POST /api/sbcs` - list and create SBCs
+    - `GET/PUT/DELETE /api/sbcs/<name>` - CRUD operations
+    - `GET/POST /api/sbcs/<name>/power` - power status and control
+    - `GET /api/ports` - list serial ports
+    - `GET /api/health` and `/api/status` - system status
+  - Created web dashboard with dark theme:
+    - Dashboard page with SBC card grid
+    - SBC detail page with all information
+    - Power control buttons (on/off/cycle)
+    - Status indicators and responsive layout
+  - Added `labctl web` command with --host, --port, --debug options
+  - **Milestone 5 Complete!**
+- M6 Multi-Client Serial Access (2026-01-01)
+  - Created serial proxy module `src/labctl/serial/proxy.py`
+    - SerialProxy class with asyncio-based fan-out architecture
+    - First-writer-wins write lock policy
+    - Multiple readers can view console output simultaneously
+    - ProxyManager for managing multiple proxy instances
+  - Added session logging
+    - SessionLogger class for timestamped traffic logs
+    - Logs stored in ~/.local/share/labctl/logs/
+    - Format: [timestamp] [direction] data
+  - Added ProxyConfig to config.py
+    - Configurable port range (default 5000-5099)
+    - Write policy selection (first, all, queue)
+    - Max clients per proxy, idle timeout
+  - Added CLI proxy commands
+    - `labctl proxy start <sbc>` - start proxy for SBC (foreground)
+    - `labctl proxy list` - list running proxies (placeholder)
+    - `labctl sessions [sbc]` - list connected clients (placeholder)
+  - Added web console with xterm.js
+    - WebSocket bridge for browser-to-proxy communication
+    - Console page at /sbc/<name>/console
+    - Full terminal emulator in browser
+  - Added 26 proxy tests (143 total)
+  - **Milestone 6 Complete!**
+- M7 Monitoring and Health (2026-01-01)
+  - Created health check module `src/labctl/health/`
+    - HealthChecker class with ping, serial, and power checks
+    - CheckResult and HealthCheckSummary dataclasses
+    - Automatic status determination based on check results
+  - Added alerting framework
+    - Alert and AlertLevel for structured notifications
+    - AlertHandler ABC with pluggable handlers
+    - LogAlertHandler for file-based alerts
+    - ConsoleAlertHandler for terminal output
+    - EmailAlertHandler and SlackAlertHandler stubs for future
+    - AlertManager for dispatching to multiple handlers
+  - Created monitoring daemon
+    - MonitorDaemon class for periodic health checks
+    - Configurable check interval
+    - Automatic status updates on check results
+    - Alert triggering on status/power changes
+  - Added status tracking
+    - log_status() method writes to status_log table
+    - get_status_history() for retrieving history
+    - cleanup_old_status_logs() for retention management
+  - Added HealthConfig to configuration
+    - check_interval, ping_timeout, serial_timeout
+    - status_retention_days, alert_log_path
+    - alert_on_offline, alert_on_power_change options
+  - Added CLI commands
+    - `labctl health-check` - run health checks on SBCs
+      - --type ping|serial|power|all filter
+      - --sbc to check single SBC
+      - --update to update status in database
+    - `labctl monitor` - start monitoring daemon
+      - --interval to override check frequency
+      - --no-update, --no-alerts options
+  - Added web endpoints
+    - GET /api/sbcs/<name>/history - status history
+    - GET /api/health/check - run health checks via API
+    - Status history page at /sbc/<name>/history
+  - Added 28 health module tests (171 total)
+  - **Milestone 7 Complete!**
+- Deferred Items Implementation (2026-01-01)
+  - CLI Enhancements:
+    - `labctl log <sbc>` - capture serial output to timestamped file
+      - `--follow/-f` for continuous output
+      - `--lines/-n` to capture N lines then exit
+      - `--timestamp/--no-timestamp` for timestamp prefixes
+    - `labctl status --watch` - continuous status update display
+    - `labctl port list --unassigned` - show unassigned /dev/lab/* devices
+    - `-q/--quiet` global flag for suppressed output
+    - Command aliases: ls=list, rm=remove, delete=remove, show=info
+  - Web Interface:
+    - SBC edit form on detail page
+    - Port/network/plug assignment forms
+    - Settings page with configuration display
+    - Settings link in navigation
+    - GET /api/sbcs/<name>/console/info endpoint
+    - GET /api/sbcs/<name>/uptime endpoint
+  - Infrastructure:
+    - Systemd service files (labctl-monitor.service, labctl-web.service)
+    - Service installation script (scripts/install-services.sh)
+    - Session log rotation with size-based rotation and compression
+    - Old log cleanup based on retention period
+  - Health Monitoring:
+    - Uptime tracking with current uptime and 24h percentage
+    - Serial probe with string/response verification
+    - `serial_probe()` method for active serial testing
+
+### Changed
+- Moved documentation files to docs/ folder (AGENT_RULES.md, IMPLEMENTATION.md, DECISIONS.md)
+- Fixed README.md and CLAUDE.md paths to reference docs/ folder
+
+[Unreleased]: https://github.com/johnjezl/Embedded-Lab-Control/compare/v0.2.0...HEAD
+[0.2.0]: https://github.com/johnjezl/Embedded-Lab-Control/releases/tag/v0.2.0