CLI 0.9.0: CI checkout names and rewrite coverage

CI receipts name the checkout they ran in (repository and branch names, never a path): kibble run reads the working directory's remote, then the runner's variables, then the directory name; kibble ci collect reads the session records. Codex file changes carry no rewrite marker, so userModified is reported unavailable for Codex instead of zero. A rewrite now counts only on the edit it belongs to. Requires a server that accepts the workspace field (app.usekibble.com does).

CLI 0.8.0: experimental Cursor collection

Experimental Cursor collection joins the existing Codex, Claude Code, Copilot and CI collectors.

• Supported Cursor hooks capture future usage and tool metadata while preserving other hooks. Automatic collection policy installs or repairs these hooks; manual collection can use kibble cursor install.
• Read-only local Cursor metadata supplies observed capability selections and completed MCP calls, with replay guards and explicit unavailable metrics.
• kibble cursor sync explicitly downloads account totals through tokscale. Subsequent pushes reuse the saved snapshot. Credentials and the original account ID remain local.
• Account totals replace only covered Cursor days across a member’s laptops and model aliases, preventing duplicate spend. Unsupported or incomplete days remain unchanged.
• kibble cursor account-disable stops snapshot uploads. kibble cursor uninstall removes Kibble’s hooks while retaining other hooks and collected history.

The matching server migration is deployed before this release. Existing CI support and non-expiring, rotatable CI credentials are retained.

Cursor support is experimental: independent live accuracy, complete child accounting, archived history and automatic capability invocation remain unverified. See the collector README for coverage and limitations.

npm publication is gated by all nine Linux, macOS and Windows checks on Node 20, 22 and 24. The npm auto tag remains unchanged.

npm install -g @usekibble/cli@0.8.0
kibble cursor status

CLI 0.7.0: CI and cloud agent collection

Collect Codex and Claude Code usage from CI jobs and cloud runners.

• kibble run --receipt usage.json --upload -- codex exec ... captures ephemeral runs as counts-only receipts.
• kibble ci collect imports ordinary saved Codex or Claude Code sessions from an isolated job directory.
• kibble ci upload retries saved receipts without rerunning the agent. Delivery is idempotent and preserves failed-run usage.
• Owners create named CI connections in Settings, with optional team attribution. Tokens never expire and can be rotated or revoked at any time.
• Receipts exclude prompts, replies, tool arguments and paths. Unknown usage and cost remain explicit.

The matching server and migrations are deployed. CI estimates are shown separately from reconciled laptop and vendor totals.

The npm publish job runs only after all nine Linux, macOS and Windows checks on Node 20, 22 and 24 pass for this release commit. The separately approved npm auto tag is unchanged.

Install after publication:

npm install -g @usekibble/cli@0.7.0

See the CI setup and usage guide for stream capture, saved sessions, delivery retries and platform limitations.

CLI 0.6.1: VS Code collection accuracy

Improves VS Code/Copilot metadata collection: deduplicates parent and subagent tool calls, reads explicit terminal outcomes, recovers uniquely paired invocation timings, and discovers configured local skills, prompts and plugins with enablement settings. Missing measurements remain unavailable instead of becoming zero. Saved cancellation state alone no longer counts as a confirmed interruption.

No runtime hooks, telemetry enablement, editor settings changes, or content reporting. Full agent-loop token totals and automatic skill attribution remain limited by saved metadata.

The matching server migration is deployed. All nine Windows, macOS and Linux checks on Node 20, 22 and 24 passed. The separate automatic-update tag is not promoted by this release.

v0.6.0

Adds local GitHub Copilot CLI and VS Code/Insiders collection on macOS, Windows and Linux. Reads persisted numeric counters and allowlisted activity metadata, never sends prompts, replies, tool arguments, paths or credentials. Copilot CLI and editor rows merge before ingest, with replay and migration deduplication. Preserves the shared Codex decoder, broader client coverage and opt-in updater. All nine platform checks passed. Coverage limits: VS Code inline completions, cloud coding-agent sessions and aggregate-only Copilot OpenTelemetry history are not collected; older multi-round VS Code records may retain only the final call and warn when totals are incomplete. Estimated model-list cost is not a GitHub invoice. Install or upgrade with npm install -g @usekibble/cli@0.6.0. The separate automatic-update tag is unchanged.

v0.5.1

Sends usage as soon as you log in, instead of waiting for the first hourly push.

• kibble login uploads immediately before installing the startup job when the organization has automatic collection on. A failed push keeps the link and schedules retries rather than losing the connection. Policy-off login stays manual.
• First sync imports today and the preceding 29 UTC days, wherever supported logs still exist on the machine.
• Codex capability reporting narrows to what is actually observable: explicit skill selections and recognized command submissions. Automatic skill reads and skill cost attribution remain unobserved, so a zero does not prove non-use.

No collector fields change. What leaves a machine is still counts and names, never prompts, file contents, tool arguments or paths.

v0.5.0

Kibble 0.5.0 adds explicitly authorized team reporting and opt-in automatic CLI updates, with fixes for collection configuration and usage labels.

• Run kibble login --reporting to approve role-scoped reads, then use kibble usage --scope team|org, --team and --list-scopes. Personal reads remain the default; the server checks current roles and assignments on every request.
• Interactive login asks once about automatic updates. Headless setup requires --auto-update. Verified managed releases, health probes, recovery and rollback preserve collection and credentials. Help and version commands do not check for updates.
• Keep CLI output in English. Refresh the agent instructions with kibble skill install to use scope discovery and the corrected estimate labels.
• Resolve Claude cloud/API configuration before stale subscription metadata, and respect the same custom Codex home for billing plans and collection.
• Correct usage language: Claude Analytics supplies vendor estimates, not invoice charges. Vendor source priority and actual billing certainty are distinct.
• Retain Windows lock safety and register scheduled tasks through XML to support managed paths with spaces and long command lines. Node 20, 22 and 24 are supported.

Self-hosted servers must deploy the matching server and migrations through 0020_reset_unverified_reporting before using the new reporting grant. Existing reporting devices must approve kibble login --reporting again; collection and personal reads continue. Existing CLI installations need one manual upgrade to acquire the updater: npm install -g @usekibble/cli@0.5.0.

Validation: all nine Windows/Linux/macOS and Node 20/22/24 CI jobs passed, including native loading, managed installation and real scheduler execution. Local raw Claude accuracy had 0% token delta for September 4. Server regression checks, HTTP auth/ingest/reporting checks, typecheck, lint, build and bilingual validation passed.

Automatic updates use the separately approved npm auto tag. Publishing to latest alone does not approve a release for automatic installation.

v0.4.0

This release improves Claude Code and Codex collection accuracy and adds observed Codex skill and command usage.

• Collect all supported agents automatically, with overlap excluded between parsers. The --source option has been removed; use plain kibble push.
• Correct Codex token snapshots, cache accounting and pricing, and deduplicate replayed sessions across daily, repository and model metrics.
• Record explicit Codex skill selections, recognized command submissions and MCP calls. Automatic shell-based skill loads and commands omitted from history remain unavailable.
• Add model activity and compaction metrics, and improve Claude Code response, capability and activity accounting.
• Improve push recovery, credential handling, locking, scheduler repair and transcript read failures.
• Make the public CLI build independently of the private monorepo.

Self-hosted servers must update their ingest schema and apply the model activity migrations before upgrading to this CLI version.

Validation: standalone build, collector regression fixtures, raw Claude transcript accuracy comparison and workspace typecheck passed.