Task Review

Task Review (Weekly Refresh)

The single ritual of Task System v3: a ~10 minute conversation that rebuilds the +active working set, catches what slipped, prunes decay, and records weekly focus. The design contract is projects/task-system-v3.md in the notes repo.

The user makes decisions; the agent does all mechanics. The user never scans, sorts, dedups, or moves anything.

Principles (non-negotiable)

1. Brain dump first. The user states what matters next week BEFORE seeing any list. Never show leftovers, backlog, or Jira before the brain dump. This forces "what does my brain think matters?" instead of copy-paste rollover.
2. Constant re-entry cost. A refresh after a 5-week gap is the same 10 minutes as after 1 week. Compress, summarize, batch. Never inflate the session because of a gap, and never mention the gap as a problem.
3. No guilt. Lead with what got done. Frame overdue/stale items forward-looking (reschedule? drop?). Never report how big the backlog is unless asked.
4. Skipping is free. If the user declines or abandons a refresh midway, finish nothing silently: state what was and wasn't done, and leave the system consistent (sync, push what was written).

Prerequisites

• All task CLI operations follow the taskwarrior skill and its taskwarrior.local.md (capture rule, decay policy, due policy).
• Notes repo location: $ZK_NOTEBOOK_DIR if set (the environment decides; never override it), else ~/Documents/rahlir-notes (user machine convention). Do not guess other paths.
• If the repo is missing at that location:
  • If you are able to clone from github (requires Github SSH access on the device), ask the user for approval and clone with: git clone git@github.com:Rahlir/rahlir-notes.git <location>
  • Otherwise, report that the repo is missing and continue with degraded refresh.
• If the repo is unavailable (not seeded, no auth, clone declined), run a degraded refresh: all taskwarrior steps proceed normally; skip daily-note mining (step 3) and the focus.md write (step 10). Like the Jira pass, state explicitly that these were skipped, and note that the focus entry should be written at the next refresh on a device with the repo.
• All writes to the notes repo follow the repo's AGENTS.md (git hygiene; agent commits/pushes there are pre-approved within its rules).

Refresh procedure

Run steps in this order. Steps 4-7 are conversational; keep each to a few lines.

1. Sync & pull. task synchronize; git pull --rebase in the notes repo. Surface failures, then continue with what is available.
2. Establish the window. Last refresh date = date of the top section in focus.md (fallback: git log for focus.md; first ever refresh: ~2 weeks). The window is [last refresh, now].
3. Gather silently (no output yet):
   • Completions in the window: task end.after:<date> completed export
   • Current +active leftovers, backlog (task shelf), upcoming hard deadlines (task due.before:<now+3wk> list)
   • Daily notes in the window (daily/YYYY-MM-DD.md): scan for finished work matching open tasks and for new commitments that match no existing task
   • Stale candidates per the decay policy (~8 weeks untouched)
4. Momentum summary. 3-6 lines: what got done since last time (from completions + daily notes). Then ask for the brain dump: "What matters next week?"
5. Brain dump processing. Map the user's items onto existing tasks (these become +active); capture genuinely new ones (accept-first, dedup per local config).
6. Show what they forgot, in three short blocks, each with a quick keep/drop/backlog decision:
   • leftover +active tasks not mentioned in the brain dump
   • backlog items that look timely (deadlines approaching, related to brain-dumped work, or repeatedly mentioned in daily notes)
   • hard deadlines inside the next ~3 weeks
7. Stale batch. One list, one question ("drop these 5?"). Approved items are deleted; the rest are left untouched. Never auto-delete.
8. Jira pass - only where Jira is reachable (work laptop). Use the jira skill: diff assigned open issues against pending tasks; propose additions (as user-shaped tasks, annotated with the issue key) and completions/closures. Always state explicitly whether this pass ran or was skipped and why.
9. Apply. Rebuild the +active set (bulk tag add/remove for the agreed set is pre-approved; no per-task confirmation theater). Sweep leftover +next tags: one line ("still preempting: X, Y?"); clear any not reconfirmed, so the flag stays scarce. Mark confirmed completions done. Run task synchronize.
10. Write focus.md (format below). The prose must go through the user; it is a personal piece of writing, not generated text they never see:
    • Briefly interview the user about the week's focus (1-2 questions: what's the through-line, why does it matter). The brain dump and decisions provide the raw material; the interview provides the framing.
    • Draft the section from their answers and show it for review. Apply their edits until they approve.
    • Prepend the approved section; append the outcome line to the previous section. Commit and push per the repo's AGENTS.md.
11. Close. One screen: the new active set, hard deadlines, and what passes ran (Jira yes/no). Done.

Mid-refresh, the user may digress (annotate something, ask about a project). Follow, then return to the sequence.

focus.md format

Single file at the notes repo root. Dated sections, newest on top. The agent prepends; never rewrites old sections except to append the one-line outcome.

---
tags:
  - focus
---
# Weekly Focus

## Week of [[daily/2026-06-15]]

Prose, 2-5 sentences, in the user's voice, drafted from the focus
interview and approved by the user. What matters and why; hard
deadlines in words.

## Week of [[daily/2026-06-08]]

(older entry, untouched)

_Outcome:_ shipped the selector; envelope decision slipped to web team.

• Section heading: ## Week of [[daily/<ISO date of that week's Monday>]] - a wiki link to that Monday's daily note, like the dated sections in project notes.
• The week prose is never written unilaterally: interview first, then user-reviewed draft (step 10). The _Outcome:_ line is exempt and stays agent-written.
• The _Outcome:_ line is appended to the previous section during step 10, based on the momentum summary. One line, factual, no grades.
• No task lists in this file. Tasks live in taskwarrior only.

Scheduled nudges (content rules)

Job scheduling lives chatbot-side; the message content follows these rules so nudges and follow-up conversations stay consistent.

Refresh nudge (Mondays):

• One message, actionable in place: "Ready for a ~10 minute refresh? We can do it right here."
• Never repeats within the week, never escalates.
• After a multi-week gap, get softer, not louder: "Been a few weeks - want a 5-minute catch-up? The backlog kept itself."
• If the user declines or ignores it: drop the subject until the next scheduled nudge.

Morning deadline check (daily):

• Query task due:today or +OVERDUE. If empty, send nothing.
• If non-empty: one short message listing the items, framed forward-looking. Overdue items get "still relevant - reschedule or drop?", never a red pile or a count of days overdue.
• Act on replies directly (reschedule, complete, drop) per the taskwarrior skill rules.

NEVER

• NEVER show lists before the brain dump (step 4 before 5-6).
• NEVER auto-delete stale tasks; the batch question is mandatory.
• NEVER inflate a refresh because of a long gap, or remark on the gap as a failure.
• NEVER write task lists into focus.md or any other note.
• NEVER claim the Jira pass ran when it was skipped.
• NEVER leave the session half-applied: if aborted, sync taskwarrior, push what was committed, and say plainly what remains undone.