[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M75CAB33E41TF3CDRKD5YYHC":3,"skill-same:01M75CAB33E41TF3CDRKD5YYHC":47},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":26,"commitSha":27,"skillDirectory":28,"forksCount":29,"repositoryHtmlUrl":30,"repositoryHomepageUrl":31,"ownerHtmlUrl":32,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":33,"install":34,"resources":46},"01M75CAB33E41TF3CDRKD5YYHC","plandesk-timebox",null,"Paces a long run in fixed intervals over a work list: works items until the interval expires, verifies what actually completed, reports at every boundary, and starts the next interval while work remains. Chain it onto another skill (\"\u002Fplandesk-timebox 25m \u002Fplandesk-foreman next\") or run it over a list. Use when asked to timebox, pomodoro, work in sprints or intervals, check in every N minutes, or report progress on a schedule.","asyncdotengineering\u002Fplandesk","asyncdotengineering","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F130723569?v=4","01M75C95Y312DEGHZMVT0W3FPB","Asyncdot Engineering","01M75C95Z7AWGPJCAY38XYPX0W",16,"MIT","https:\u002F\u002Fgithub.com\u002Fasyncdotengineering\u002Fplandesk\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fplandesk-timebox\u002FSKILL.md",[],[],"2026-11-12 13:55:43",false,"---\nname: plandesk-timebox\ndescription: \"Paces a long run in fixed intervals over a work list: works items until the interval expires, verifies what actually completed, reports at every boundary, and starts the next interval while work remains. Chain it onto another skill (\\\"\u002Fplandesk-timebox 25m \u002Fplandesk-foreman next\\\") or run it over a list. Use when asked to timebox, pomodoro, work in sprints or intervals, check in every N minutes, or report progress on a schedule.\"\nuser-invocable: true\nargument-hint: \"[\u003Cinterval, e.g. 25m>] [\u003Ca skill invocation, or a list of work items>]\"\n---\n\n# Work in timeboxes\n\nA pacing posture, not a task. It wraps a run in fixed intervals so a long\nsession surfaces on a cadence instead of disappearing for an hour and coming\nback with a wall of diff.\n\nThe work list is **yours, not the board's**. Timebox drives what you handed it —\na list of things you want done, in the order you want them. It can be pointed at\nthe board when you ask for that explicitly, but it never goes shopping for work\non its own.\n\n## How to chain it\n\n```\n\u002Fplandesk-timebox 25m \u002Fplandesk-foreman next     # a box per board item\n\u002Fplandesk-timebox 50m                            # bare: paces the list you give it\n\u002Fplandesk-timebox 25m clear the board            # explicit opt-in to board-driven work\n\u002Fplandesk-autonomy \u002Fplandesk-timebox 25m \u002Fplandesk-foreman next\n                                                 # no permission pauses, plus a surfacing cadence\n```\n\nStacking with [autonomy](..\u002Fplandesk-autonomy\u002FSKILL.md) is the common pairing and the two\ndo different jobs: autonomy removes the pause between steps, timebox adds a\nrhythm and a report. Neither grants a permission the wrapped skill didn't\nalready have.\n\n## Harness primitives — use what the runtime gives you\n\nTimebox is harness-neutral. Prefer the scheduler and shell tools your runtime\nexposes; fall back to the portable stamp when it does not. **Never poll** — do\nnot schedule short wakeups to check on background work the dispatch monitor\nalready watches; that is wasted work ([heartbeat.md](..\u002F..\u002Ffactory\u002Fheartbeat.md)).\n\n| capability | Cursor | Claude Code | Codex | Pi | OpenCode |\n| --- | --- | --- | --- | --- | --- |\n| work-list scratchpad | `TaskCreate` \u002F `TaskList` \u002F `TaskUpdate` | same harness task tools | `.agents\u002Ffactory\u002Fruns\u002Ftasks-\u003Ctask>.md` | `TaskCreate` (pi-loop fallback) or pi-tasks | session notes \u002F plugin tasks |\n| background command | Shell `run_in_background` + `Await` | Bash `run_in_background` | background flag if present | `MonitorCreate` (pi-loop) | `\u002Fbackground` (monitor plugin) |\n| one-shot resume | backgrounded `sleep` sentinel (see Cursor specifics) | `ScheduleWakeup` | — (use external cron) | `schedule_loop_wakeup` \u002F `LoopCreate` (extensions) | `ScheduleWakeup` (opencode-routines) |\n| recurring loop | — | `\u002Floop` + `ScheduleWakeup` | `.codex\u002Fautomations\u002F*.toml` | `LoopCreate` \u002F `\u002Floop` (pi-loop) | `LoopCreate` \u002F `\u002Floop` (routines plugin) |\n| portable clock | `date +%s` at item boundaries | same | same | same | same |\n\n**Common pattern that works everywhere:**\n\n1. **List on the harness** — `TaskCreate` one task per work item (lead session) or\n   `.agents\u002Ffactory\u002Fruns\u002Ftasks-\u003Ctask>.md` (delegated worker). Add a meta task `timebox:state`\n   holding box number, `box_start` epoch, and interval seconds.\n2. **Clock at boundaries** — after each item, read elapsed with `date +%s` against\n   `box_start`. This is the universal fallback and always binds.\n3. **Scheduler when present** — if the runtime exposes `ScheduleWakeup` (or Pi \u002F\n   OpenCode loop tools), use it only to **resume this run across turns** after\n   a checkpoint — not to re-fire an expensive slash command verbatim.\n4. **Background through the harness** — dispatch workers with `run_in_background`\n   (never `&` \u002F `nohup`); completion is the **result file**\n   (`.agents\u002Ffactory\u002Fruns\u002Fresult-\u003Ctask>.json`), watched by the monitor armed at dispatch —\n   `Await` the monitor shell, and treat the harness's own exit notification as\n   a hint only. See [protocol.md](..\u002F..\u002Ffactory\u002Fprotocol.md).\n\n**Cursor specifics.** There is no scheduler tool, but there is a working box\ntimer: at each box start, background a sentinel shell —\n\n```bash\nsleep {interval_seconds} && echo \"TIMEBOX: box {n} boundary\"\n```\n\n— with `run_in_background` (block_until_ms 0). Its completion notification is\ndelivered even after your turn ends and **resumes the run**, so a box boundary\ncan wake you instead of relying on you noticing the stamp. This is the\ndocumented sleep-and-echo self-notification pattern, not polling: one shell\nper box, armed once, disarmed (kill the pid) when the list empties early. The\n`date +%s` stamp at item boundaries stays the authority on whether the box has\nactually expired — the sentinel is the wakeup, not the clock.\n\n**Claude Code specifics.** `ScheduleWakeup` clamps `delaySeconds` to 60–3600.\nPass a plain continuation prompt (\"Continue timebox box 3; pick up item X from\nTaskList\"), never a `\u002F`-prefixed slash command — re-firing those duplicates\nexpensive work. End the loop with `stop: true` when the list is empty. Foreground\n`sleep` for the full interval is blocked; use the scheduler or keep working and\ncheck the stamp at item boundaries.\n\n**Codex specifics.** No in-session scheduler in mainline Codex today. Pace with\n`date +%s` inside the session, or stand up `.codex\u002Fautomations\u002F*.toml` \u002F\nexternal cron for durable recurring runs outside this skill.\n\n**Pi \u002F OpenCode specifics.** Loop extensions (`pi-loop`, `opencode-routines`)\nare the richest schedulers — prefer fixed-interval `LoopCreate` over self-paced\nre-scheduling when you want a predictable pomodoro cadence. Coalesced delivery\nwhile busy is correct: the box boundary fires when the in-flight item finishes.\n\n## Set up the run\n\n1. **Get the work list.** Ask for it if it wasn't given: a list of items, a file\n   to work through, or a scope of changes. Write it down as harness tasks\n   (`TaskCreate`) — one per item, plus `timebox:state` (box number, start epoch,\n   interval). If the user said \"clear the board\" or named the board explicitly,\n   the list is `get_next_task` — but only then.\n2. **Set the interval.** Default 25 minutes when unspecified. Parse user input:\n   `25m`, `1h`, `90 minutes` → seconds once at setup. \"Work for an hour\" means\n   one 60-minute box, not three of twenty.\n3. **Stamp the clock.**\n   ```bash\n   date +%s    # box_start; persist in timebox:state and\u002For TaskUpdate metadata\n   ```\n   If `ScheduleWakeup` \u002F loop tools are available and the run may span multiple\n   turns, also record `box_end = box_start + interval_seconds` so the wakeup\n   prompt knows which box to report.\n4. **Say the plan back** in one line: the interval, the item count, and what\n   happens at the first boundary. A run nobody can predict is a run nobody can\n   interrupt at a good moment.\n\n## The cycle\n\n```\nbox_start = read from timebox:state (or date +%s on first box)\narm the box timer if the harness has one    # Cursor: sentinel sleep; Claude: ScheduleWakeup\nloop:\n  item = next pending harness task (skip timebox:state)\n  if item is null:\n    stop — the list is done; final report; disarm timer \u002F ScheduleWakeup stop\n  work(item)                       # the wrapped skill's own procedure, in full\n  verify(item)                     # its own completion check — exit codes, not claims\n  TaskUpdate(item, completed)      # or mark Done in .agents\u002Ffactory\u002Fruns\u002Ftasks-*.md\n  commit\u002Fcheckpoint(item)          # leave nothing half-applied\n  if (date +%s) - box_start >= interval_seconds:\n    checkpoint_report()            # see the template below — then KEEP GOING\n    increment box number in timebox:state\n    box_start = date +%s\n    re-arm the box timer for the new box\n  continue loop                    # same turn — the report is not a turn boundary\n```\n\n**Inside a box with a background dispatch:** background the worker per\n[protocol.md](..\u002F..\u002Ffactory\u002Fprotocol.md), keep working on what you can in the\nforeground, and `Await` (or watch `.agents\u002Ffactory\u002Fruns\u002Fresult-\u003Ctask>.json`) for completion —\ndo not schedule timer polls for it. The box timer and the worker timer are\nindependent; an expiring box still lets the in-flight item finish first.\n\n## The boundary rule — the timer is a checkpoint, not a kill signal\n\n**When the interval expires mid-item, let the item finish.** Then verify,\ncommit, mark the harness task complete, and only then checkpoint.\n\nThis is the whole design decision, and the reason is concrete: cutting a run\nmid-item strands work in exactly the state that is hardest to recover — a worker\npart-way through a change, nothing staged, nothing verified, and a report that\ncannot honestly say what landed. A box that runs four minutes long costs four\nminutes. A box that cuts clean through a dispatch costs the item.\n\nSo the interval governs *how often you surface*, never *where the work stops*.\nIf an item is so large that it routinely blows the box, that is a sizing\nproblem — say so at the checkpoint and suggest splitting it, rather than\nshrinking the box until it cuts.\n\n## The checkpoint report\n\nAt each boundary, keep it short enough to read on a phone:\n\n```\nBox N (25m) — 3 items, 2 done, 1 carried\n\n  done      \u003Citem> — \u003Cwhat proves it: command + exit code>\n  done      \u003Citem> — \u003Cproof>\n  carried   \u003Citem> — \u003Cwhere it stands, what is next>\n\nNext box: \u003Cwhat it will pick up>.  Remaining: \u003Ccount>.\n```\n\nReport what is *proven*, not what was attempted. An item without a verification\nresult is `carried`, not `done` — this is the check the interval exists to\nforce, and reporting an unverified item as done makes every later box's report\nworthless.\n\nThen continue into the next box **in the same turn — do not end your turn\nafter a checkpoint.** Ending the turn is how a run silently converts a report\ninto a permission request: in a harness with no scheduler, nothing wakes you,\nand the run dies with \"say continue for box 3\" — which is exactly the pause\nthe wrapper was invoked to remove. Never write \"say continue\", \"shall I\nproceed\", or any variant; the only legitimate turn-enders are the stop\nconditions below. The checkpoint is a surfacing moment, not a permission\nrequest: the human reads it if they are there and **interrupts** if they want\nto — interruption is their affordance, not something you solicit. Lane gates\nstill bind — see [lanes.md](..\u002F..\u002Ffactory\u002Flanes.md) and\n[autonomy](..\u002Fplandesk-autonomy\u002FSKILL.md).\n\n## Breaks\n\nAfter four boxes, take a longer checkpoint: re-read the work list (`TaskList` or\n`.agents\u002Ffactory\u002Fruns\u002Ftasks-*.md`) against what actually landed, drop items that are now moot,\nand say whether the remaining list still matches what the user wanted. Long runs\ndrift — the plan made sense ninety minutes ago and the fourth box is where that\nis worth checking.\n\nThis is the agent's version of the pomodoro break. It is not idle time; it is\nthe moment the list gets re-derived from reality rather than from memory.\n\n## When to stop instead of starting another box\n\n- The list is empty. Report and end — do not go looking for adjacent work. Disarm\n  any loop \u002F `ScheduleWakeup` with `stop: true` or `LoopDelete`.\n- Every remaining item is blocked, or gated on input you cannot obtain. Report\n  what is blocking and on whom.\n- Two consecutive boxes closed nothing. Something is wrong with the sizing, the\n  environment, or the plan; another box will not fix it. Say what you observed\n  and stop.\n- The user asked for a fixed number of boxes and you have finished them.\n\n## Gotchas\n\n- Timeboxing paces work; it does not authorize any. A wrapped skill's lane gates\n  and boundaries bind unchanged — see [autonomy](..\u002Fplandesk-autonomy\u002FSKILL.md).\n- **`date +%s` always works; schedulers are optional.** Do not fail to timebox\n  because `ScheduleWakeup` is absent — the stamp at item boundaries is the\n  portable core.\n- **Never poll background work on a timer.** The dispatch monitor watches the\n  result file and reports completion ([protocol.md](..\u002F..\u002Ffactory\u002Fprotocol.md));\n  heartbeat wakeups are for stall detection only ([heartbeat.md](..\u002F..\u002Ffactory\u002Fheartbeat.md)).\n- The clock is read at boundaries, so it is only as granular as your items. One\n  90-minute item inside a 25-minute box produces one box, not four — that is\n  correct behavior, and the checkpoint should name it as a sizing problem.\n- Do not let the box become a deadline the work is rushed to fit. Skipping\n  verification to land inside the interval defeats the entire point; the box is\n  a reporting rhythm, not a budget.\n- A checkpoint is not a turn boundary. If you catch yourself typing \"say\n  continue\" or \"ready for box N?\", delete it and start the next item — the\n  human interrupts when they want to; you never solicit it.\n- Harness tasks are the per-session scratchpad for the list. If the run is\n  board-driven, the board stays the source of truth and the harness list is\n  re-derived from it after any compaction.\n\n## References\n\n[autonomy](..\u002Fplandesk-autonomy\u002FSKILL.md) (the posture this most often stacks with);\n[foreman](..\u002Fplandesk-foreman\u002FSKILL.md) (the usual inner skill);\n[protocol.md](..\u002F..\u002Ffactory\u002Fprotocol.md) (background dispatch + stall detection);\n[heartbeat.md](..\u002F..\u002Ffactory\u002Fheartbeat.md) (why not to poll);\n[lanes.md](..\u002F..\u002Ffactory\u002Flanes.md) (gates a checkpoint never overrides);\n`.agents\u002Ffactory\u002Fhooks\u002F` (board re-anchor after compaction — re-read `TaskList`).\n","# Work in timeboxes\n\nA pacing posture, not a task. It wraps a run in fixed intervals so a long\nsession surfaces on a cadence instead of disappearing for an hour and coming\nback with a wall of diff.\n\nThe work list is **yours, not the board's**. Timebox drives what you handed it —\na list of things you want done, in the order you want them. It can be pointed at\nthe board when you ask for that explicitly, but it never goes shopping for work\non its own.\n\n## How to chain it\n\n```\n\u002Fplandesk-timebox 25m \u002Fplandesk-foreman next     # a box per board item\n\u002Fplandesk-timebox 50m                            # bare: paces the list you give it\n\u002Fplandesk-timebox 25m clear the board            # explicit opt-in to board-driven work\n\u002Fplandesk-autonomy \u002Fplandesk-timebox 25m \u002Fplandesk-foreman next\n                                                 # no permission pauses, plus a surfacing cadence\n```\n\nStacking with [autonomy](..\u002Fplandesk-autonomy\u002FSKILL.md) is the common pairing and the two\ndo different jobs: autonomy removes the pause between steps, timebox adds a\nrhythm and a report. Neither grants a permission the wrapped skill didn't\nalready have.\n\n## Harness primitives — use what the runtime gives you\n\nTimebox is harness-neutral. Prefer the scheduler and shell tools your runtime\nexposes; fall back to the portable stamp when it does not. **Never poll** — do\nnot schedule short wakeups to check on background work the dispatch monitor\nalready watches; that is wasted work ([heartbeat.md](..\u002F..\u002Ffactory\u002Fheartbeat.md)).\n\n| capability | Cursor | Claude Code | Codex | Pi | OpenCode |\n| --- | --- | --- | --- | --- | --- |\n| work-list scratchpad | `TaskCreate` \u002F `TaskList` \u002F `TaskUpdate` | same harness task tools | `.agents\u002Ffactory\u002Fruns\u002Ftasks-\u003Ctask>.md` | `TaskCreate` (pi-loop fallback) or pi-tasks | session notes \u002F plugin tasks |\n| background command | Shell `run_in_background` + `Await` | Bash `run_in_background` | background flag if present | `MonitorCreate` (pi-loop) | `\u002Fbackground` (monitor plugin) |\n| one-shot resume | backgrounded `sleep` sentinel (see Cursor specifics) | `ScheduleWakeup` | — (use external cron) | `schedule_loop_wakeup` \u002F `LoopCreate` (extensions) | `ScheduleWakeup` (opencode-routines) |\n| recurring loop | — | `\u002Floop` + `ScheduleWakeup` | `.codex\u002Fautomations\u002F*.toml` | `LoopCreate` \u002F `\u002Floop` (pi-loop) | `LoopCreate` \u002F `\u002Floop` (routines plugin) |\n| portable clock | `date +%s` at item boundaries | same | same | same | same |\n\n**Common pattern that works everywhere:**\n\n1. **List on the harness** — `TaskCreate` one task per work item (lead session) or\n   `.agents\u002Ffactory\u002Fruns\u002Ftasks-\u003Ctask>.md` (delegated worker). Add a meta task `timebox:state`\n   holding box number, `box_start` epoch, and interval seconds.\n2. **Clock at boundaries** — after each item, read elapsed with `date +%s` against\n   `box_start`. This is the universal fallback and always binds.\n3. **Scheduler when present** — if the runtime exposes `ScheduleWakeup` (or Pi \u002F\n   OpenCode loop tools), use it only to **resume this run across turns** after\n   a checkpoint — not to re-fire an expensive slash command verbatim.\n4. **Background through the harness** — dispatch workers with `run_in_background`\n   (never `&` \u002F `nohup`); completion is the **result file**\n   (`.agents\u002Ffactory\u002Fruns\u002Fresult-\u003Ctask>.json`), watched by the monitor armed at dispatch —\n   `Await` the monitor shell, and treat the harness's own exit notification as\n   a hint only. See [protocol.md](..\u002F..\u002Ffactory\u002Fprotocol.md).\n\n**Cursor specifics.** There is no scheduler tool, but there is a working box\ntimer: at each box start, background a sentinel shell —\n\n```bash\nsleep {interval_seconds} && echo \"TIMEBOX: box {n} boundary\"\n```\n\n— with `run_in_background` (block_until_ms 0). Its completion notification is\ndelivered even after your turn ends and **resumes the run**, so a box boundary\ncan wake you instead of relying on you noticing the stamp. This is the\ndocumented sleep-and-echo self-notification pattern, not polling: one shell\nper box, armed once, disarmed (kill the pid) when the list empties early. The\n`date +%s` stamp at item boundaries stays the authority on whether the box has\nactually expired — the sentinel is the wakeup, not the clock.\n\n**Claude Code specifics.** `ScheduleWakeup` clamps `delaySeconds` to 60–3600.\nPass a plain continuation prompt (\"Continue timebox box 3; pick up item X from\nTaskList\"), never a `\u002F`-prefixed slash command — re-firing those duplicates\nexpensive work. End the loop with `stop: true` when the list is empty. Foreground\n`sleep` for the full interval is blocked; use the scheduler or keep working and\ncheck the stamp at item boundaries.\n\n**Codex specifics.** No in-session scheduler in mainline Codex today. Pace with\n`date +%s` inside the session, or stand up `.codex\u002Fautomations\u002F*.toml` \u002F\nexternal cron for durable recurring runs outside this skill.\n\n**Pi \u002F OpenCode specifics.** Loop extensions (`pi-loop`, `opencode-routines`)\nare the richest schedulers — prefer fixed-interval `LoopCreate` over self-paced\nre-scheduling when you want a predictable pomodoro cadence. Coalesced delivery\nwhile busy is correct: the box boundary fires when the in-flight item finishes.\n\n## Set up the run\n\n1. **Get the work list.** Ask for it if it wasn't given: a list of items, a file\n   to work through, or a scope of changes. Write it down as harness tasks\n   (`TaskCreate`) — one per item, plus `timebox:state` (box number, start epoch,\n   interval). If the user said \"clear the board\" or named the board explicitly,\n   the list is `get_next_task` — but only then.\n2. **Set the interval.** Default 25 minutes when unspecified. Parse user input:\n   `25m`, `1h`, `90 minutes` → seconds once at setup. \"Work for an hour\" means\n   one 60-minute box, not three of twenty.\n3. **Stamp the clock.**\n   ```bash\n   date +%s    # box_start; persist in timebox:state and\u002For TaskUpdate metadata\n   ```\n   If `ScheduleWakeup` \u002F loop tools are available and the run may span multiple\n   turns, also record `box_end = box_start + interval_seconds` so the wakeup\n   prompt knows which box to report.\n4. **Say the plan back** in one line: the interval, the item count, and what\n   happens at the first boundary. A run nobody can predict is a run nobody can\n   interrupt at a good moment.\n\n## The cycle\n\n```\nbox_start = read from timebox:state (or date +%s on first box)\narm the box timer if the harness has one    # Cursor: sentinel sleep; Claude: ScheduleWakeup\nloop:\n  item = next pending harness task (skip timebox:state)\n  if item is null:\n    stop — the list is done; final report; disarm timer \u002F ScheduleWakeup stop\n  work(item)                       # the wrapped skill's own procedure, in full\n  verify(item)                     # its own completion check — exit codes, not claims\n  TaskUpdate(item, completed)      # or mark Done in .agents\u002Ffactory\u002Fruns\u002Ftasks-*.md\n  commit\u002Fcheckpoint(item)          # leave nothing half-applied\n  if (date +%s) - box_start >= interval_seconds:\n    checkpoint_report()            # see the template below — then KEEP GOING\n    increment box number in timebox:state\n    box_start = date +%s\n    re-arm the box timer for the new box\n  continue loop                    # same turn — the report is not a turn boundary\n```\n\n**Inside a box with a background dispatch:** background the worker per\n[protocol.md](..\u002F..\u002Ffactory\u002Fprotocol.md), keep working on what you can in the\nforeground, and `Await` (or watch `.agents\u002Ffactory\u002Fruns\u002Fresult-\u003Ctask>.json`) for completion —\ndo not schedule timer polls for it. The box timer and the worker timer are\nindependent; an expiring box still lets the in-flight item finish first.\n\n## The boundary rule — the timer is a checkpoint, not a kill signal\n\n**When the interval expires mid-item, let the item finish.** Then verify,\ncommit, mark the harness task complete, and only then checkpoint.\n\nThis is the whole design decision, and the reason is concrete: cutting a run\nmid-item strands work in exactly the state that is hardest to recover — a worker\npart-way through a change, nothing staged, nothing verified, and a report that\ncannot honestly say what landed. A box that runs four minutes long costs four\nminutes. A box that cuts clean through a dispatch costs the item.\n\nSo the interval governs *how often you surface*, never *where the work stops*.\nIf an item is so large that it routinely blows the box, that is a sizing\nproblem — say so at the checkpoint and suggest splitting it, rather than\nshrinking the box until it cuts.\n\n## The checkpoint report\n\nAt each boundary, keep it short enough to read on a phone:\n\n```\nBox N (25m) — 3 items, 2 done, 1 carried\n\n  done      \u003Citem> — \u003Cwhat proves it: command + exit code>\n  done      \u003Citem> — \u003Cproof>\n  carried   \u003Citem> — \u003Cwhere it stands, what is next>\n\nNext box: \u003Cwhat it will pick up>.  Remaining: \u003Ccount>.\n```\n\nReport what is *proven*, not what was attempted. An item without a verification\nresult is `carried`, not `done` — this is the check the interval exists to\nforce, and reporting an unverified item as done makes every later box's report\nworthless.\n\nThen continue into the next box **in the same turn — do not end your turn\nafter a checkpoint.** Ending the turn is how a run silently converts a report\ninto a permission request: in a harness with no scheduler, nothing wakes you,\nand the run dies with \"say continue for box 3\" — which is exactly the pause\nthe wrapper was invoked to remove. Never write \"say continue\", \"shall I\nproceed\", or any variant; the only legitimate turn-enders are the stop\nconditions below. The checkpoint is a surfacing moment, not a permission\nrequest: the human reads it if they are there and **interrupts** if they want\nto — interruption is their affordance, not something you solicit. Lane gates\nstill bind — see [lanes.md](..\u002F..\u002Ffactory\u002Flanes.md) and\n[autonomy](..\u002Fplandesk-autonomy\u002FSKILL.md).\n\n## Breaks\n\nAfter four boxes, take a longer checkpoint: re-read the work list (`TaskList` or\n`.agents\u002Ffactory\u002Fruns\u002Ftasks-*.md`) against what actually landed, drop items that are now moot,\nand say whether the remaining list still matches what the user wanted. Long runs\ndrift — the plan made sense ninety minutes ago and the fourth box is where that\nis worth checking.\n\nThis is the agent's version of the pomodoro break. It is not idle time; it is\nthe moment the list gets re-derived from reality rather than from memory.\n\n## When to stop instead of starting another box\n\n- The list is empty. Report and end — do not go looking for adjacent work. Disarm\n  any loop \u002F `ScheduleWakeup` with `stop: true` or `LoopDelete`.\n- Every remaining item is blocked, or gated on input you cannot obtain. Report\n  what is blocking and on whom.\n- Two consecutive boxes closed nothing. Something is wrong with the sizing, the\n  environment, or the plan; another box will not fix it. Say what you observed\n  and stop.\n- The user asked for a fixed number of boxes and you have finished them.\n\n## Gotchas\n\n- Timeboxing paces work; it does not authorize any. A wrapped skill's lane gates\n  and boundaries bind unchanged — see [autonomy](..\u002Fplandesk-autonomy\u002FSKILL.md).\n- **`date +%s` always works; schedulers are optional.** Do not fail to timebox\n  because `ScheduleWakeup` is absent — the stamp at item boundaries is the\n  portable core.\n- **Never poll background work on a timer.** The dispatch monitor watches the\n  result file and reports completion ([protocol.md](..\u002F..\u002Ffactory\u002Fprotocol.md));\n  heartbeat wakeups are for stall detection only ([heartbeat.md](..\u002F..\u002Ffactory\u002Fheartbeat.md)).\n- The clock is read at boundaries, so it is only as granular as your items. One\n  90-minute item inside a 25-minute box produces one box, not four — that is\n  correct behavior, and the checkpoint should name it as a sizing problem.\n- Do not let the box become a deadline the work is rushed to fit. Skipping\n  verification to land inside the interval defeats the entire point; the box is\n  a reporting rhythm, not a budget.\n- A checkpoint is not a turn boundary. If you catch yourself typing \"say\n  continue\" or \"ready for box N?\", delete it and start the next item — the\n  human interrupts when they want to; you never solicit it.\n- Harness tasks are the per-session scratchpad for the list. If the run is\n  board-driven, the board stays the source of truth and the harness list is\n  re-derived from it after any compaction.\n\n## References\n\n[autonomy](..\u002Fplandesk-autonomy\u002FSKILL.md) (the posture this most often stacks with);\n[foreman](..\u002Fplandesk-foreman\u002FSKILL.md) (the usual inner skill);\n[protocol.md](..\u002F..\u002Ffactory\u002Fprotocol.md) (background dispatch + stall detection);\n[heartbeat.md](..\u002F..\u002Ffactory\u002Fheartbeat.md) (why not to poll);\n[lanes.md](..\u002F..\u002Ffactory\u002Flanes.md) (gates a checkpoint never overrides);\n`.agents\u002Ffactory\u002Fhooks\u002F` (board re-anchor after compaction — re-read `TaskList`).\n",{"name":5,"description":7,"argument-hint":24,"user-invocable":25},"[\u003Cinterval, e.g. 25m>] [\u003Ca skill invocation, or a list of work items>]",true,"main","1e2d885d184029608df0aab985a4a1545c3a1a1a",".agents\u002Fskills\u002Fplandesk-timebox",1,"https:\u002F\u002Fgithub.com\u002Fasyncdotengineering\u002Fplandesk","https:\u002F\u002Fplandesk-docs.pages.dev","https:\u002F\u002Fgithub.com\u002Fasyncdotengineering","2026-10-04 12:26:19",{"source":8,"skillName":5,"commands":35,"prompt":45,"zipAvailable":25},[36,39,42],{"tool":37,"command":38},"npx","npx skills add asyncdotengineering\u002Fplandesk --skill plandesk-timebox",{"tool":40,"command":41},"bunx","bunx skills add asyncdotengineering\u002Fplandesk --skill plandesk-timebox",{"tool":43,"command":44},"pnpm","pnpm dlx skills add asyncdotengineering\u002Fplandesk --skill plandesk-timebox","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fasyncdotengineering\u002Fplandesk\u002Fblob\u002Fmain\u002F.agents\u002Fskills\u002Fplandesk-timebox\u002FSKILL.md",[],[48,55,61,67,72,78],{"skillNo":49,"name":50,"nameZh":6,"description":51,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":52,"categories":53,"keywords":54,"publishedAt":19},"01M75CB59TR0Q4BQ056JTQKRB3","connecting-agents","Agent conventions for Plan Desk MCP — task labels, docs, notes, files, artifacts, sharing, edges, and agent runs.","https:\u002F\u002Fgithub.com\u002Fasyncdotengineering\u002Fplandesk\u002Fblob\u002Fmain\u002Fapps\u002Fdocs\u002Fsrc\u002Fcontent\u002Fdocs\u002Fconnecting-agents\u002Fskill.md",[],[],{"skillNo":56,"name":57,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":58,"categories":59,"keywords":60,"publishedAt":19},"01M75CB3ZV1RWR3SJPG2PMBJRG",".plandesk","https:\u002F\u002Fgithub.com\u002Fasyncdotengineering\u002Fplandesk\u002Fblob\u002Fmain\u002F.plandesk\u002Fskill.md",[],[],{"skillNo":62,"name":63,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M75CB3428SD8HWHTBW857YXA","plandesk","https:\u002F\u002Fgithub.com\u002Fasyncdotengineering\u002Fplandesk\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fplandesk\u002FSKILL.md",[],[],{"skillNo":68,"name":5,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":19},"01M75CB277DZXMNDYF9FNDRN84","https:\u002F\u002Fgithub.com\u002Fasyncdotengineering\u002Fplandesk\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fplandesk-timebox\u002FSKILL.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":75,"categories":76,"keywords":77,"publishedAt":19},"01M75CB03K0H1H0Y5F4WR9TFZZ","plandesk-standup","https:\u002F\u002Fgithub.com\u002Fasyncdotengineering\u002Fplandesk\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fplandesk-standup\u002FSKILL.md",[],[],{"skillNo":79,"name":80,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M75CAXTZWTY17EY7T4SNA55A","plandesk-standdown","https:\u002F\u002Fgithub.com\u002Fasyncdotengineering\u002Fplandesk\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fplandesk-standdown\u002FSKILL.md",[],[]]