Kadlekai Time Tracking

2. Generate an API Token

From the Kadlekai Rails app:

bin/rails runner "
user = User.find_by(email: '[email protected]')
auth_service = AuthenticationService.new(current_user: user, request: nil)
result = auth_service.create_api_key_for_user(user, client_name: 'Codex', expiry_hours: 720)
puts 'Token: ' + result[:auth_token]
"

3. Install this skill

mkdir -p ~/.codex/skills/kadlekai-time-tracking
cp /path/to/kadlekai/codex/skills/kadlekai-time-tracking/SKILL.md \
   ~/.codex/skills/kadlekai-time-tracking/SKILL.md

URL drift fix: If the MCP server returns ENOTFOUND, an old install wrote kadle.ai to your config. Fix it:

sed -i '' 's|kadle\.ai|kadle.beskar.tech|g' ~/.codex/config.toml

Then restart Codex.

Available MCP Tools

Reconciliation Flow

Use when the user says "reconcile my worklogs", "log today's work", or "log this session".

Since Codex has no hook events, reconciliation is session-based: ask the user to describe what they did.

Steps

1. Gather session context — ask:

   "When did you start working today? What did you work on? Give me a start time, end time (or duration), and a brief description."

2. Check existing entries — call list_worklogs for today. For each returned entry, compare its [start_time, end_time] against the proposed new entry's time range. An overlap exists when: proposed_start < existing_end AND proposed_end > existing_start. Note any overlapping entries — they are used in step 5.

3. Suggest a project — call list_projects, then fuzzy-match against:

   • The current repo directory name (e.g. kadlekai → project named "Kadlekai")
   • Any project or task names mentioned by the user

4. Confirm before creating — if NO overlapping entries were found in step 2, present a summary and ask:

   "I'll log X hours to project Y — description: 'Z'. Does that look right?" Wait for explicit confirmation before calling create_worklog. If any overlap was found in step 2, skip this step and go directly to step 5.

5. Handle overlaps — if a new entry overlaps an existing worklog, show both and ask the user:

   "Entry A already covers [X minutes] of this time. What should I do?

   1. Keep existing 2. Replace existing 3. Split into two entries" Never auto-merge or auto-delete without user approval.

6. Create the entry — call create_worklog with confirmed params.

7. Record reconcile time — write {"last_reconcile_at": "<ISO timestamp>"} to ~/.codex/kadle/state.json.

Time Commands (Direct)

CRITICAL — direct timer commands: The commands "start timer", "stop timer", and "timer status" MUST call start_timer, stop_timer, and get_running_timer directly. NEVER route these through process_natural_language_command. The NL processor may mis-classify them (e.g. "timer status" → intent: start_timer) and take the wrong action. process_natural_language_command is only for free-form commands that do not match the explicit patterns in the table above.

Behavioural Rules

SAFETY — delete: Before calling delete_worklog, you MUST present the worklog details and ask: "Are you sure you want to delete this worklog? (yes/no)" Then wait for an explicit "yes" or "confirm" in the user's reply. Never call delete_worklog without this step.

SAFETY — overlaps: Before calling create_worklog or update_worklog, verify the time range against existing entries using: new_start < existing_end AND new_end > existing_start. If any overlap exists, STOP — do NOT proceed to confirmation. Present both entries with the overlap duration in minutes, then ask: "Keep existing / replace existing / split?" Never create, update, or delete worklogs to resolve an overlap without an explicit user instruction in the same turn.

RECOVERY — orphaned timer: If start_timer fails with a conflict or "only one running worklog" error despite get_running_timer returning no active timer:

1. Call list_worklogs with status: "running" to find any orphaned timers.
2. If entries are returned, show them and ask: "Found a running timer that wasn't visible — stop it first?"
3. If list_worklogs also returns empty (both endpoints disagree with the API error), tell the user: "The server reports a running timer but neither get_running_timer nor list_worklogs can see it — this is a backend visibility bug. Wait a moment and try again, or contact support if it persists." Do NOT retry start_timer automatically.

• Never use process_natural_language_command for start/stop/status timer commands. Always call get_running_timer, start_timer, or stop_timer directly.
• Always confirm project before creating or updating a worklog. Never assume.
• Always confirm on overlaps — never auto-merge, auto-split, or auto-delete.
• Keep descriptions terse — ≤ 100 chars. Prefer issue refs like "Fix #42" over prose.
• Omit timezone — the server uses the user's stored preference.
• Running timer check — before starting a new timer, always call get_running_timer first and warn if one is already active.
• State file — write reconcile timestamps to ~/.codex/kadle/state.json (create dir if needed).