gtd-todoist-health

name: gtd-todoist-health description: Read-only audit for a GTD-over-Todoist agent harness. Use when verifying that the GTD AGENTS section, three GTD workspace skills (gtd-inbox-triage, gtd-daily-review, gtd-weekly-review), reminder-only cron jobs, and Todoist scaffold stay consistent with the deployment plan in this repo. Does not read real task content.

GTD Todoist Health

Overview

Use this skill to audit a live GTD-over-Todoist agent harness without reading the user's real Todoist task content. The audit covers:

• the AGENTS.md GTD section
• three GTD workspace skills
• reminder-only cron jobs
• runtime skill visibility
• optional Todoist scaffold name drift

Scope Guard

• Target is an agent runtime host reachable over SSH and a workspace directory (e.g. ~/.openclaw). Both are passed in at invocation time; nothing is hard-coded.
• Always run read-only checks.
• Do not restart the agent runtime, edit files, disable cron, clear sessions, or mutate Todoist during a health check.
• Do not run td inbox, td today, or td task list by default.
• Treat Todoist scaffold inspection as opt-in only. Use -IncludeTodoistScaffold only when the user explicitly wants scaffold verification and still avoid task content.
• Do not inspect other agent profiles unless the user explicitly changes scope.

Workflow

1. Read local project context before touching the server:

   • AGENTS.md
   • any local planning or findings notes

2. Run the wrapper from the repository root:

   .\scripts\run-gtd-health.ps1 -HostName "<user>@<host>" -StateDir "~/.openclaw"
   

3. If the wrapper is unavailable, stream the probe script over SSH directly:

   Get-Content -Raw -Encoding UTF8 .\scripts\gtd_health_check.sh | ssh <user>@<host> "tr -d '\r' | bash -s"
   

4. Interpret the output by layers:

   • P0: main state, workspace files, openclaw config validate (or your runtime's equivalent)
   • P1: AGENTS.md GTD contract
   • P1: three GTD workspace skills and runtime visibility
   • P1: five GTD reminder-only cron jobs
   • P2: stale scaffold residue, unsafe Todoist command examples, optional Todoist scaffold drift
   • P2: manual semantic GTD review that deterministic scans cannot prove

5. Report with:

   • Overall status: clean, issues, or residual-risk
   • Blocking red flags
   • Evidence by layer
   • Manual GTD semantic review notes
   • Recommended next action

Required Checks

Always include these checks:

• <workspace>/AGENTS.md exists.
• These workspace skills exist:
  • gtd-inbox-triage
  • gtd-daily-review
  • gtd-weekly-review
• The agent runtime config validates (e.g. openclaw config validate).
• Runtime skill visibility shows all three GTD skills as eligible=true, modelVisible=true, commandVisible=true, and sourced from the workspace.
• AGENTS.md includes GTD Capture, Clarify, Organize, Reflect, Engage.
• GTD mutation rules require id:<task_id> or Todoist URL and warn that td task update --labels replaces all labels.
• GTD skills require proposal / confirmation before mutation.
• gtd-inbox-triage covers clarify classification, waiting-for, project routing, hard due dates, label preservation, and MIT confirmation.
• gtd-daily-review covers hard landscape, no soft-date drift, label preservation, and closing suggestions.
• gtd-weekly-review covers Get Clear, Get Current, Get Creative, Next Actions, Waiting For, Someday/Maybe, and Horizons.
• Five GTD cron jobs are reminder-only:
  • gtd-inbox-reminder-0900
  • gtd-inbox-reminder-1300
  • gtd-inbox-reminder-1900
  • gtd-daily-review-reminder-2200
  • gtd-weekly-review-reminder-sun-1315
• Each GTD cron is enabled=true, sessionTarget=isolated, delivery.mode=announce, delivery target matches the configured Telegram user / bot, payload.thinking=off, payload.lightContext=true, payload.toolsAllow=["read"], and payload.timeoutSeconds within a safe window (default 600).
• Stale scaffold scans find no invalid td mutation examples and no soft-date drift.

Manual Semantic Review

The script provides evidence, not the final judgment. After running it, manually review whether:

• Clarify still means deciding what an item is before organizing it.
• Calendar dates are reserved for hard landscape items, not wishful "do this someday" dates.
• Reference is treated as non-actionable material, not as an orphaned task bucket.
• Waiting For items include who, what, and follow-up context.
• Projects are multi-step outcomes with at least one next action.
• Weekly Review reviews the whole action system, not only projects.
• Engage recommendations are limited and context-aware, not a raw task dump.
• The harness avoids automated GTD decisions from cron and keeps user discussion in the main chat session.

Optional Todoist Scaffold Check

Use only when the user explicitly asks to include Todoist scaffold, and do not inspect task content:

.\scripts\run-gtd-health.ps1 -HostName "<user>@<host>" -IncludeTodoistScaffold

This checks only the project, label, and filter names expected by the GTD scaffold defined in ../deployment-plan.md.

Clean Standard

Report clean only when all required checks pass, the script exits cleanly, and the manual semantic review finds no contradictions. Report residual-risk when deterministic checks pass but semantic uncertainty remains. Report issues when any required check fails.

Fix Policy

• Apply no fixes during the first health pass.
• For mechanical fixes, propose exact files and changes first.
• For live workspace changes, back up remote files before editing and validate after editing.
• Restart the agent runtime only when a workspace or config change needs to be loaded, and only after the user asks for the fix.

Resources

• scripts/run-gtd-health.ps1: Windows wrapper that finds gtd_health_check.sh, sends it to the remote host over SSH, and runs it read-only.
• scripts/gtd_health_check.sh: remote read-only GTD harness probe.