m8m-task-builder
Author reusable M8M Task packages for server installation, or persist and manage the current live M8M Task's instructions, progress, waits and final JSON.
SKILL.mdskills/m8m/m8m-task-builder/SKILL.md
GitHubM8M task builder
Choose the context
- Reusable authoring or delivery: When creating or changing a Task for installation on a server, author its reusable source locally and use the dedicated server installer. Follow Local authoring below and read server delivery when exporting or delivering.
- Current live Task: When working in an existing M8M chat and saving its
goal, progress, wait or result, follow Server runtime below. Its room-scoped
m8m_workspaceoperations manage this live Task.
A local authoring request does not become a live Task registration because the
skill is called Task Builder. Available room tools and the user's requested
operation determine the runtime branch. Keep a supplied live Task identity when
continuing its work. Independent reusable authoring can happen outside an M8M
room and does not require m8m_workspace.
Local authoring
Preserve the goal, complete master prompt and a self-contained JSON object output schema. Retain workflow references, required connections/tools/resources, external effects and any declared model/reasoning configuration. References may be semantic names, existing exact pins or explicit installation member references; the installer resolves them against the destination owner's actual catalog. Do not replace complete instructions with a synopsis.
Use the bundled <loaded skill directory>/scripts/m8m_local_bundle.py task to
export a reusable template. Its --help describes goal/master-prompt/schema
files, workflow references, resource requirements and runtime configuration.
The helper preserves parsed source schemas; the server validates its prepared
output. Attach referenced implementation, assets and resource context separately
or collect them with bundle. A template JSON is not automatic dependency closure.
Exclude live room/thread IDs, waits, progress, ledger rows and prior outputs. Reusable ledger behavior belongs in the master prompt: work-unit rule, ordering, selection and completion obligations. A real ledger is created later in its launched server Task using the installed Ledger Builder.
Deliver new and changed reusable definitions through m8m-server-installer.
The installer installs an immutable Task template. An authenticated launch
creates a fresh live Task through normal admission; exporting or installing a
template does not start that Task or authorize additional external actions.
Server runtime
Apply this section only in the current live M8M Task room. The host injects these instructions without requiring filesystem access; local export/transport instructions above do not grant filesystem or installer tools to this room.
One chat is one task in the existing Codex session. Understand ordinary user messages directly; do not classify messages through another agent or require an input schema. Answer related questions and incorporate relevant information. For independent new work, explain briefly that it belongs in another chat and keep the original task's progress.
Use only the available m8m_workspace tools. Read read_task first. If no task
is registered, save the user's goal through register_task with a complete
master_prompt, workflows bindings from the actual inspected catalog (or {}
for a simple task), and a self-contained JSON object output_schema. Bindings
contain import_id, release_id, expected_release_digest; the server records
the source message. Registration stores instructions; it does not authorize
additional external actions. Do not register a separate task within this room.
When the same goal needs a newly available workflow, inspect it and use
bind_task_workflows with {workflows:{alias:binding}} from the current user
conversation. Pause ledger dispatch first. Existing aliases and releases stay
fixed; this adds bindings without replacing the goal or full master prompt.
Use checkpoint_task only when useful progress or a wait must survive the turn.
For batches, generation queues or repeated work use the installed
m8m-ledger-builder. Keep the same Task and its overall goal. The ledger owns
item membership and action/result references; do not duplicate its rows or set
its completion in Task context. The server maintains context.ledger.
current_step is ordinary text and context is a small JSON patch with useful
facts or references. A new waiting object contains reason, match (string
values), on_wake and remaining_checks (a finite positive count). The server
returns its wait_id; never manufacture or replace server-owned IDs. Resolve
a wait with waiting: null. To stop progress, set status to blocked or
cancelled as appropriate. Do not mark cancelled without the user's request.
For future follow-up use the installed m8m-schedule-builder skill and the existing schedule tools. A saved wait alone does not create a timer. After waking, read current state and actual results before deciding what is still missing. Keep the same wait and edit its saved reminder for a further check; the server decrements its remaining check allowance when a reminder is queued.
Read actual workflow results before finish_task. Save only the final business
JSON matching the output schema; asset values must be uploaded server URLs.
Never use intentions, local filenames, or a notification's text as proof of a
completed external operation. Workflow execution remains in the existing tools.
Session-reminder turns expose progress and reminder tools only. Resolving a wait allows an already-authorized active ledger to continue in a separate server ledger_continue turn in the same session. Each such turn may start one saved ledger action. The native runtime still owns execution and recovery; ordinary run_result summaries remain read-only. External email ingress requires its own installed connector. Never claim that saving a ledger installed that capability.