kortix-goals

Own an outcome until it is verifiably done, across days or months: a goal folder in Files (`goal.json`, `log.md`, `notes.md`, `docs/`, `evidence/`) run with the `kortix goals` CLI, a daily heartbeat r

SKILL.md

GitHub
<skill name="kortix-goals"> <overview> A goal is an outcome you own until it is true, not a task you attempt once. "Make sure my Rho account is fully closed" and "secure EUR 2.5M of EIC Accelerator funding" are goals. The user gives the objective once and walks away. You plan, act, wait, chase, and verify until it is done.

A goal is a folder in Files. Sessions come and go; the folder is the goal's memory and the only state that counts. A new session with no transcript must be able to continue from the folder alone.

You run a goal with the kortix goals CLI. It writes the parts of goal.json that the platform depends on and refuses moves the rules do not allow. Do not write those parts by hand (see <goal-json>). </overview>

<how-it-keeps-going> Nothing runs between your turns. Three things wake you, and each wake is one turn:
  1. The heartbeat: a reminder on this session, armed by kortix goals new, that fires daily with the prompt "Goal wake: <folder>. Load the kortix-goals skill and run one wake."
  2. A reply: a person answering in a thread bound to this session.
  3. Your own next step: anything you queue for yourself.

After every turn, the sandbox runs kortix goals check --enforce --quiet for every goal this session coordinates. A broken rule (for example no heartbeat while the goal is open) sends you a [GOAL CHECK] message with the fix, at most 3 per goal in 6 hours. Fix exactly what it names, then end the turn.

If this session's sandbox fails, the next heartbeat starts a new session, moves the heartbeat and the Slack threads to it, and tells it which session failed. If that is you, you are the coordinator now (see <gotchas>).

If this session is deleted, the goal stops waking. Never delete a coordinator session of an open goal. </how-it-keeps-going>

<commands> | Command | What it does | | --- | --- | | `kortix goals new "<objective>" [--shared] [--slug <slug>]` | Creates `Goals/<slug>/` with `goal.json`, `log.md`, `notes.md`, `docs/`, `evidence/` in `/drives/me` (or `/drives/company` with `--shared`); marks this session unattended; arms the daily heartbeat (first fire in 24 hours) and records its id; installs the after-turn check; sets status `draft`. Prints the folder. Refuses when the folder already holds a goal. | | `kortix goals transition <to> --reason "<why>" [<folder>]` | The only way to change `status`. Allows only the edges in `<lifecycle>`, checks each guard, appends the move to `transitions` in `goal.json` and to `log.md`. `done` and `abandoned` also delete the heartbeat. A refusal names every missing condition; exit 1. | | `kortix goals check [<folder>]` | Checks every goal this session coordinates (or one folder). Prints each broken rule with its fix; exit 1 when any is broken. | | `kortix goals ls` | Every goal in the mounted folders: status, criteria met, open asks, next wake. | | `kortix goals status [<slug>]` | One goal: objective, status, criteria met, summary, open asks, next wake. |

Add --json to any of them for machine-readable output.

new needs three project features on: goals, volumes (Files) and reminders. When one is off it refuses and prints the kortix projects features enable … command. Run it if you may (a project admin can); otherwise tell the user which feature to turn on. </commands>

<lifecycle> ``` draft ──▶ active ──▶ waiting_on_user ──▶ active ──▶ waiting_external ──▶ active ──▶ verifying ──▶ done └──▶ active (a criterion FAILed) any of draft, active, waiting_*, verifying ──▶ abandoned (the user said stop) ```
Move Guard checked by transition
draft → active 2 to 6 success_criteria; exactly one active heartbeat; this session is unattended
active → waiting_on_user at least one open ask item; every open ask has an owner and a due
active → waiting_external at least one open wait item; every open wait has a due
waiting_* → active none
active → verifying every criterion has met: true and an evidence path that exists in the folder; a session with role verifier is listed in sessions
verifying → done evidence/verification.md contains PASS and no FAIL; verified_by is the id of a verifier in sessions, not a coordinator
verifying → active none
→ abandoned the goal is open

There is no move from waiting_on_user to waiting_external or back: go through active. </lifecycle>

<goal-json> `kortix goals new` writes this file. Keep its shape. Times are UTC ISO 8601.
{
  "version": 1,
  "objective": "Rho business account fully closed",
  "status": "active",
  "summary": "Closure requested 2026-10-19. Waiting for Rho's written confirmation (due 2026-10-24).",
  "success_criteria": [
    { "id": "c1", "text": "Rho confirms the closure in writing", "met": false, "evidence": null },
    { "id": "c2", "text": "No charge on any linked account for 30 days after closure", "met": false, "evidence": null }
  ],
  "items": [
    { "id": "i1", "kind": "task", "title": "Request closure in the Rho app", "status": "done", "owner": "agent", "due": null },
    { "id": "i2", "kind": "ask", "title": "Approve the 2FA prompt", "status": "open", "owner": "user", "due": "2026-10-20T09:00:00Z" },
    { "id": "i3", "kind": "wait", "title": "Rho written confirmation", "status": "open", "owner": "Rho support", "due": "2026-10-24T17:00:00Z" }
  ],
  "sessions": [{ "id": "<session_id>", "role": "coordinator", "since": "2026-10-19T19:00:00Z" }],
  "reminder_id": "reminder.3f9a1c2e7b04",
  "verified_by": null,
  "next_wake_at": "2026-10-21T09:00:00Z",
  "transitions": [{ "from": null, "to": "draft", "at": "2026-10-19T19:00:00Z", "reason": "goal created", "by": "<session_id>" }]
}
Field Who writes it Rule
version CLI always 1
status, transitions CLI only (transition) never edit by hand: check reports a status that differs from the last transition
objective CLI, then you a true-or-false outcome
success_criteria you 2 to 6; met: true only with an evidence path (relative to the folder) to a file that exists
items you kind: task (work), ask (something only a person can give), wait (an outside event). status: open, doing, done, dropped. Open asks need owner and due; open waits need due
sessions CLI adds the coordinator; you add workers and verifiers roles: coordinator, worker, verifier
reminder_id, next_wake_at CLI, then you when you re-arm reminder_id must be the one active heartbeat
verified_by you, after the verifier answers the verifier's session id
summary you one paragraph: where the goal is now

goal.json holds only these fields, and items only id, kind, title, status, owner, due. check reports any other field. Everything else goes in notes.md: the user's original request, the constraints, one thread per person, and what you learned. kortix goals new creates those sections.

Read the file just before you write it, and write it with valid JSON. </goal-json>

<start> When the user gives an objective:
  1. kortix goals new "<the objective, as an outcome>" [--shared]. Use the user's own folder unless they say team or company.
  2. Write 2 to 6 success criteria into goal.json that someone else can check. Each one names its proof.
  3. Write the constraints under ## Constraints in notes.md: the user's words, plus the obvious ones (no spend above what the user named, nothing that harms their other accounts). Ask only when a wrong guess is expensive or irreversible.
  4. Add the first items, then kortix goals transition active --reason "criteria written".
  5. Act on the first item in this turn.
  6. kortix goals check, then end the turn with three lines to the user: the outcome, the criteria, the first thing you need from them (or "nothing for now").</start>
<wake> Every turn on a goal, whatever woke you:
  1. Orient. Read goal.json, the last 40 lines of log.md, notes.md. Look for what changed: new files in docs/, replies in your threads, finished workers.

  2. Advance. Do the highest-leverage open item. Mark finished items done. Add new items. A wait whose due passed: check it, then follow up.

  3. Chase. For every open ask past its due, message its owner (see <chase>), then set a new due (the next day, as a rule).

  4. Move. When the situation changed, kortix goals transition <to> --reason "…": waiting for a person, waiting for an event, back to active, or verifying.

  5. Save. Update summary and items. Append one line per action to log.md. Put durable facts in notes.md.

  6. Re-arm when the next wake must be earlier than the heartbeat. The heartbeat fires daily. For an earlier wake (a due today, a worker you started), replace it:

    kortix reminders rm <reminder_id>
    kortix remind "Goal wake: <goal folder>. Load the kortix-goals skill and run one wake." --at <ISO time> --every 1d --name "goal <slug>"
    

    The prompt must contain the goal folder path: that is how check finds the heartbeat. Write the new id into reminder_id and the time into next_wake_at.

  7. kortix goals check. End the turn only when it passes.

If kortix remind answers feature_disabled, run kortix projects features enable reminders. If that is refused, tell the user the goal cannot wake on its own. </wake>

<chase> You depend on people. Act like a colleague who needs the thing to do the job: specific, polite, persistent.

Never ask with the question tool. kortix goals new marks this session unattended, so the platform answers a question tool call at once with "nobody can answer". Write the question in plain text, send it on a channel, record the ask, and end the turn. The answer arrives as a new message and wakes you.

  • Ask once, precisely: what you need, why it unblocks the goal, how to send it ("reply with the PDF", "drop it in <goal folder>/docs/").
  • Use the channels this project has (kortix connectors ls). Use the one the person answered on before, else the one they use most. Never invent one.
  • One thread per person per goal. A reply reaches you only in a thread bound to this session. On Slack, the first slack send --channel <DM or channel> binds its thread; read thread_binding in the output (the kortix-slack skill). Write the person, channel and thread under ## Threads in notes.md, and send every later message to that person with --thread. Tell them once to reply in that thread.
  • Raise the pressure with each chase, recorded in log.md: first a reminder, then the consequence of the delay ("the EIC cut-off is 2026-11-04; without the financials I cannot submit"), then a short daily ping. Stay professional.
  • A reply or an upload resolves the ask: set it done and record where the answer is.
  • Chase outside parties (a vendor, a bank, an agency) the same way, on the channel you used to reach them.</chase>
<autonomy> - Act on anything inside the goal's constraints (`notes.md`) without asking. Report it in `log.md` and `summary`. - A constraint is a hard limit. An action outside it becomes an `ask`. - Do not ask for what you can find: search Files, the connectors, the web and earlier goal folders first. - Credentials go in the Secrets Manager, never in the goal folder. Ask for a missing one with a setup link (the `kortix-system` skill). </autonomy> <workers> Give parallel or heavy work to worker sessions. The coordinator owns `goal.json`; a worker does one item.
kortix sessions new --meta unattended=true --prompt "Work item <id> of the goal in <goal folder>. Load the kortix-goals skill. Do only that item. Put outputs in docs/ or evidence/, append one line to log.md, do not edit goal.json, and end with a one-paragraph result." --json

Add the worker to sessions with role worker. On a later wake, check it with kortix sessions wait-for <id> --timeout 5 (exit 0 = finished, 1 = its turn failed, 124 = still working). While a worker runs, re-arm the heartbeat to wake you in 15 minutes. </workers>

<done> A goal is done when every criterion is met with proof in `evidence/` and a session that did none of the work confirms it. You never judge your own work.
  1. Set every criterion met: true with its evidence path.

  2. Start the verifier and add it to sessions with role verifier:

    kortix sessions new --meta unattended=true --prompt "Verify the goal in <goal folder>. Load the kortix-goals skill. You did none of the work. For each success criterion in goal.json, check its evidence yourself. Write PASS or FAIL per criterion with the reason to evidence/verification.md. Do not edit goal.json." --json
    
  3. kortix goals transition verifying --reason "all criteria have evidence".

  4. Wait for it in this turn: kortix sessions wait-for <id> --timeout 900. Then read evidence/verification.md.

  5. All PASS: set verified_by to the verifier's session id, then kortix goals transition done --reason "verified" (this also deletes the heartbeat). Send the user one message: the outcome and where the proof is.

  6. Any FAIL: kortix goals transition active --reason "<what failed>", add items for the gaps, and continue.

The goal also ends when the user says stop: kortix goals transition abandoned --reason "<the user's words>". It never ends because it is hard, slow or blocked. A blocked goal waits and chases. </done>

<status> When anyone asks how a goal is going: `kortix goals status <slug>` (or `kortix goals ls` for all). Answer with the status, the summary, what waits on them, and the next wake. </status> <gotchas> - **You were started after a failed coordinator.** Your first message names the failed session. Add yourself to `sessions` with role `coordinator` (keep the old entry), then run one wake. The heartbeat and the Slack threads already moved to you. - **One writer for `goal.json`.** Only the coordinator writes it. A clash leaves a `(conflict ...)` copy beside it: merge the two, then delete the copy. - **The user's edits win.** A field the user changed in Files is the truth. - **Absolute dates.** Write "2026-11-04", never "next Tuesday". - **No secrets in the folder**, and no customer data beyond what the goal needs in a shared folder. </gotchas> </skill>