1.3.5

Fixed

• thinr_ls describes what "/" actually resolves to: the agent process's
  working directory, which is normally the filesystem root. 1.3.4 said it was
  a base directory "usually /var/log", which was a stale service definition on
  one device rather than how installs behave.

1.3.4

Fixed

• thinr_script_delete resolves the script by the name the agent reports, so
  both battery and battery.sh remove the same file. It used to build the
  path from the argument verbatim, which meant following its own description
  ("without extension") returned a bare 404. A name that matches nothing now
  lists the scripts that are registered.

Changed

• thinr_script_write refuses a name that would replace one of the device's
  built-in resources (monitoring, system, cmd, update, agent) unless
  force: true is passed. Installing monitoring.sh used to shadow the real
  monitoring resource silently. Its description now also states that the
  script runs with the agent's privileges, root in a system install.
• thinr_devices shows when each device last changed state: connected since
  for a live one, offline since for a dead one. The timestamp was already in
  the response; finding out whether something dropped minutes or months ago
  meant a trip to the monitoring history.
• thinr_ls explains that / resolves to the agent's base directory while
  every other absolute path is literal, and a listing of / now says which
  directory it came from. Nothing is confined to that directory, so reading a
  / listing as the root filesystem was an easy wrong conclusion to draw.
• thinr_profiles states that profiles are not accounts, and that user
  addresses another account on the same server.

1.3.3

Changed

• MCP tools take the target device as device_id. A tool argument named
  device never reaches this server through Claude Desktop's remote-devices
  bridge, which claims that name for its own call routing and strips it, so
  every tool needing a device failed with device is required however the
  caller spelled the call, while tools without one worked. Reproduced on one
  machine with one binary: a direct JSON-RPC call succeeded, the same call
  through Claude Desktop arrived without the field, and renaming the argument
  fixed it with nothing else changed. device is no longer accepted.

1.3.2

Fixed

• Device properties and device status now honour the user argument.
  thinr_property_get, thinr_property_set and the playbook property
  actions accepted it but addressed the active profile's own account, the
  same gap fixed for resources in 1.3.1.

Changed

• Account-scoped API paths are built in one place (lib/paths.js): the
  /v{1,2,3}/users/{account} prefixes and the device prefix, all resolving
  the account through a single resolveUser. Five modules carried private
  copies of the same three lines and four more interpolated the account by
  hand. A test fails the build if any module starts doing that again, which
  is what keeps the account from being silently dropped: omitting it
  produces a URL that works against your own account and only breaks when
  acting on behalf of another.
• Not-found errors from properties and status name the account they looked
  in, matching the resource helpers.

1.3.1

Fixed

• Device resources now honour the user argument, so an admin acting on
  another account reaches that account's devices. thinr_resource_list,
  thinr_resource_call and the thinr_script_* tools accepted the
  argument but addressed the active profile's own account, which answered
  "Device not found"; playbook resource and script actions had the same
  gap. Not-found errors now name the account they looked in.

1.3.0

Added

• thinr login — re-authenticate the current profile in place, without
  deleting and recreating it. Renews an expired session in one step and
  defaults the server prompt to the profile's existing server.

Changed

• Session-expiry errors now point to the working command: "Session
  expired. Run thinr login to sign in again." (previously suggested
  thinr, which opens the dashboard instead of re-authenticating).
• More robust token refresh for WebSocket sessions: the interactive
  console and streaming exec now refresh an expired access token
  before the handshake and retry once on a 401, matching the
  refresh-on-401 behaviour of HTTP calls.

1.2.0

Added

• thinr with no arguments now launches the interactive dashboard
  directly when stdout is a TTY (and a configuration exists), so the
  fleet view is one keystroke away. Non-TTY contexts (pipes, scripts)
  fall back to --help. First-time users still get the auth flow,
  then drop straight into the dashboard.
• TUI dashboard: alarms panel on the overview tab, with rule- and
  instance-level views, severity colouring and live updates over the
  same event stream the rest of the dashboard already uses.
• thinr device delete <deviceId> — remove a device record from the
  platform. Interactive confirm by default, --yes to skip in
  scripts, idempotent on a missing device. Useful after re-provisioning
  an agent under a new hostname leaves the old ip-… record offline.
• MCP: thinr_device_delete — same operation for AI clients.
• MCP: thinr_push and thinr_pull — upload a local file to a
  device and download a remote file to local disk, respectively.
  Rounds out filesystem parity with the CLI so an agent can move
  binary or large payloads that don't fit inline in thinr_write.
• MCP: alarms tools (thinr_alarm_instances,
  thinr_alarm_instance_get, thinr_alarm_instance_stats,
  thinr_alarm_instance_update, thinr_alarm_instance_delete,
  thinr_alarm_rules, thinr_alarm_rule_read, thinr_alarm_rule_write,
  thinr_alarm_rule_delete) — list, inspect and manage alarm
  instances and rules from AI clients, with state/severity parsing
  shared with the dashboard.
• MCP: product profile config tools (buckets, properties, API
  resources) and access-token tools (user-level and device-level).
• thinr product exec <productId> <command...> — run a shell command
  in parallel on every active device of a product, with bounded
  concurrency (-c, --concurrency, default 10), per-device timeout
  (--timeout, default 30 s), --fail-fast, asset-group filter
  (-g, --group) and an -a, --all switch to include offline
  devices. Live progress shows a single spinner with done/total
  counters; the final render uses a cli-table3 table per device
  plus a compact Output: section and a one-line totals bar.
• thinr_product_exec MCP tool — the same fan-out as the CLI
  subcommand, returning a consolidated text report so an AI agent
  can roll out a change across a whole fleet in a single call
  instead of looping thinr_exec per device.
• lib/concurrency.js#runPool — tiny helper used by both entry
  points; keeps at most N workers in flight and returns a parallel
  results array. runPool plus an internal try/catch in each worker
  keeps fleet-wide errors localised per device.

Changed

• Unified filesystem verbs across CLI, MCP and playbooks so the same
  eight nouns mean the same thing everywhere: read / write
  (inline content) · push / pull (local ↔ remote file transfer)
  · ls · mkdir · rm · mv.
• CLI: thinr device cat is now thinr device read (cat kept as
  a hidden alias). New thinr device write <deviceId> <path> [content] accepts the payload as an argument or on stdin.
• MCP: thinr_delete renamed to thinr_rm, thinr_move renamed to
  thinr_mv. No backwards-compatibility aliases.
• Playbook actions: delete → rm, move → mv. The write
  action now only accepts inline content; use the new push
  action to upload a local file (source, destination). New
  pull action downloads a remote file to the local disk.
• cli-table3 added as a direct dependency to render the
  product-exec summary.

Removed

• thinr dashboard subcommand. The dashboard is now the default for
  bare thinr; the explicit subcommand had not shipped to anyone yet.

1.2.0-alpha.1

Added

• thinr product exec <productId> <command...> — run a shell command
  in parallel on every active device of a product, with bounded
  concurrency (-c, --concurrency, default 10), per-device timeout
  (--timeout, default 30 s), --fail-fast, asset-group filter
  (-g, --group) and an -a, --all switch to include offline
  devices. Live progress shows a single spinner with done/total
  counters; the final render uses a cli-table3 table per device
  plus a compact Output: section and a one-line totals bar.
• thinr_product_exec MCP tool — the same fan-out as the CLI
  subcommand, returning a consolidated text report so an AI agent
  can roll out a change across a whole fleet in a single call
  instead of looping thinr_exec per device.
• lib/concurrency.js#runPool — tiny helper used by both entry
  points; keeps at most N workers in flight and returns a parallel
  results array. runPool plus an internal try/catch in each worker
  keeps fleet-wide errors localised per device.

Changed

• cli-table3 added as a direct dependency to render the
  product-exec summary.

1.1.0

Major release: MCP server for AI clients, subcommand-first CLI, JSON
output, multi-profile configuration, product-level fan-outs, and a
consistent error taxonomy across both interfaces.

Added

• MCP server (thinr mcp) exposing ~26 typed tools over stdio for AI
  clients (Claude Code, Claude Desktop, and any other MCP host). Covers
  discovery, shell exec, filesystem (read/write/ls/mkdir/delete/
  move), typed resource calls, properties, scripts (device- and
  product-level), products, monitoring and agent updates, plus a profiles
  lookup. MCP errors carry the same { code } taxonomy as the CLI JSON
  mode (exposed on _meta) so clients can pattern-match without parsing
  human strings.
• Subcommand-first CLI UX: thinr device <action> <deviceId> /
  thinr product <action> <productId> replaces the previous top-level
  flag shape. Matches kubectl / git idioms.
• thinr device list [pattern] with case-insensitive regex filtering on
  the server side (replaces the old thinr devices).
• thinr device exec <deviceId> <command...> with streaming stdout/stderr,
  SIGINT cancellation, --json envelope, and --legacy one-shot fallback
  for older agents.
• thinr device update check|apply for managing agent upgrades.
• thinr device resource with per-resource input/output schema
  introspection, and thinr product resource / thinr product property
  for fleet fan-outs.
• Multi-profile configuration store (~/.config/thinr-cli/config.json):
  one profile per server, selected via --profile, the THINR_PROFILE
  env var, or a per-call MCP parameter. Legacy single-file configs are
  detected and migrated in-memory on read.
• thinr profile list|current|use|delete for managing configured
  environments.
• --json / -j output envelope on every data-producing command with a
  stable { code } taxonomy: not_configured, not_found,
  unauthorized, server_error, network_error, input_error,
  timeout, cancelled, error.
• -u, --user admin impersonation flag plumbed through every
  device-scoped operation (CLI + MCP).
• THINR_INSECURE=1 env var to accept self-signed TLS certificates
  against non-localhost hosts (localhost / 127.0.0.1 are always
  accepted).
• Device property setter via PUT (setDeviceProperty in the CLI,
  thinr_property_set in MCP).
• thinr_device_info MCP tool surfaces server-side product, asset group
  and description alongside the agent's own system_info.
• Refresh-token flow on 401s: the CLI transparently swaps expired access
  tokens for a fresh one when a refresh_token is on record, and fails
  cleanly (with an actionable unauthorized error) when it isn't.

Changed

• MCP server internals: tool schema and handler now paired in a
  registry, so they cannot drift apart. Tools grouped by area under
  lib/mcp/tools-*.js, with server.js reduced to a thin bootstrap.
• CLI device command split into one file per subcommand under
  commands/device/ (previously a single 17 KB module).
• lib/device.js renamed to lib/devices.js (plural: operations on the
  fleet), paired with the low-level lib/device-api.js.
• Interactive handlers (auth / console / proxy) no longer call
  process.exit from lib/*; the CLI entry owns the exit code.
• Centralised API error classification in lib/errors.js — one
  { message, code, status } contract consumed by both CLI JSON mode
  and MCP _meta.
• Re-thrown auth errors chain the original cause for better
  debugging.
• Bad -i key=value values surface as a clean Commander
  InvalidArgumentError instead of a stack trace.
• Node.js engine requirement bumped to >=18; .nvmrc pinned to 20.
• The MCP server's initialize response now reports the CLI's
  package.json version in serverInfo.version (was hardcoded).
• Runtime dependencies refreshed, including majors: @inquirer/prompts
  7 → 8, conf 14 → 15, open 10 → 11, ora 8 → 9. All npm audit
  advisories cleared (0 vulnerabilities on install).

Removed

• thinr devices (use thinr device list instead).
• thinr device mcp <deviceId> (use thinr mcp with per-call device
  parameter).
• thinr device env launcher.

Tooling

• ESLint 10 (flat config) + Prettier 3 with eslint-config-prettier
  bridge. Scripts: npm run lint / lint:fix / format /
  format:check.
• TypeScript checkJs via per-file // @ts-check opt-in; 21 files
  under lib/mcp/ and commands/device/ pass type-check today. Script:
  npm run typecheck.
• Unit tests with node --test (zero deps): MCP registry invariants +
  error helpers. Script: npm test.
• Publish pipeline migrated from a long-lived NPM_TOKEN secret to npm
  Trusted Publishing (OIDC), with --provenance attestations on every
  publish. GitHub Actions bumped to checkout@v5 / setup-node@v5.

1.1.0-alpha.2

Major release: MCP server for AI clients, subcommand-first CLI, JSON
output, multi-profile configuration, product-level fan-outs, and a
consistent error taxonomy across both interfaces.

Added

• MCP server (thinr mcp) exposing ~26 typed tools over stdio for AI
  clients (Claude Code, Claude Desktop, and any other MCP host). Covers
  discovery, shell exec, filesystem (read/write/ls/mkdir/delete/
  move), typed resource calls, properties, scripts (device- and
  product-level), products, monitoring and agent updates, plus a profiles
  lookup. MCP errors carry the same { code } taxonomy as the CLI JSON
  mode (exposed on _meta) so clients can pattern-match without parsing
  human strings.
• Subcommand-first CLI UX: thinr device <action> <deviceId> /
  thinr product <action> <productId> replaces the previous top-level
  flag shape. Matches kubectl / git idioms.
• thinr device list [pattern] with case-insensitive regex filtering on
  the server side (replaces the old thinr devices).
• thinr device exec <deviceId> <command...> with streaming stdout/stderr,
  SIGINT cancellation, --json envelope, and --legacy one-shot fallback
  for older agents.
• thinr device update check|apply for managing agent upgrades.
• thinr device resource with per-resource input/output schema
  introspection, and thinr product resource / thinr product property
  for fleet fan-outs.
• Multi-profile configuration store (~/.config/thinr-cli/config.json):
  one profile per server, selected via --profile, the THINR_PROFILE
  env var, or a per-call MCP parameter. Legacy single-file configs are
  detected and migrated in-memory on read.
• thinr profile list|current|use|delete for managing configured
  environments.
• --json / -j output envelope on every data-producing command with a
  stable { code } taxonomy: not_configured, not_found,
  unauthorized, server_error, network_error, input_error,
  timeout, cancelled, error.
• -u, --user admin impersonation flag plumbed through every
  device-scoped operation (CLI + MCP).
• THINR_INSECURE=1 env var to accept self-signed TLS certificates
  against non-localhost hosts (localhost / 127.0.0.1 are always
  accepted).
• Device property setter via PUT (setDeviceProperty in the CLI,
  thinr_property_set in MCP).
• thinr_device_info MCP tool surfaces server-side product, asset group
  and description alongside the agent's own system_info.
• Refresh-token flow on 401s: the CLI transparently swaps expired access
  tokens for a fresh one when a refresh_token is on record, and fails
  cleanly (with an actionable unauthorized error) when it isn't.

Changed

• MCP server internals: tool schema and handler now paired in a
  registry, so they cannot drift apart. Tools grouped by area under
  lib/mcp/tools-*.js, with server.js reduced to a thin bootstrap.
• CLI device command split into one file per subcommand under
  commands/device/ (previously a single 17 KB module).
• lib/device.js renamed to lib/devices.js (plural: operations on the
  fleet), paired with the low-level lib/device-api.js.
• Interactive handlers (auth / console / proxy) no longer call
  process.exit from lib/*; the CLI entry owns the exit code.
• Centralised API error classification in lib/errors.js — one
  { message, code, status } contract consumed by both CLI JSON mode
  and MCP _meta.
• Re-thrown auth errors chain the original cause for better
  debugging.
• Bad -i key=value values surface as a clean Commander
  InvalidArgumentError instead of a stack trace.
• Node.js engine requirement bumped to >=18; .nvmrc pinned to 20.

Removed

• thinr devices (use thinr device list instead).
• thinr device mcp <deviceId> (use thinr mcp with per-call device
  parameter).
• thinr device env launcher.

Tooling

• ESLint 10 (flat config) + Prettier 3 with eslint-config-prettier
  bridge. Scripts: npm run lint / lint:fix / format /
  format:check.
• TypeScript checkJs via per-file // @ts-check opt-in; 21 files
  under lib/mcp/ and commands/device/ pass type-check today. Script:
  npm run typecheck.
• Unit tests with node --test (zero deps): MCP registry invariants +
  error helpers. Script: npm test.