ulw-loop

Use this skill when the user asks for ulw-loop, ulw, durable goal execution, evidence-led work, manual QA, or checkpointed long-running delivery.

This skill is compact by design: the run contract below is the whole bootstrap. references/full-workflow.md and references/define-goal.md carry the full doctrine; open a section only when the phase you are in needs it.

Run contract

1. Create goals from a JS eval cell: agentToolkit.createGoals({ brief }) after the SDK import below. The SDK binds this session from the host env, so never pass a session id or a plan path. If the envelope reports ULW_LOOP_PLAN_EXISTS_COMPLETE (this session's aggregate is already complete), unrelated new work needs a fresh session; createGoals({ brief, force: true }) is only for deliberately overwriting completed evidence.
2. Register the aggregate objective from the returned handoff with create_goal, shaped by references/define-goal.md. Goal creation is NEVER skipped.
3. Mirror every atomic step into the live todo checklist: one granular step per action, exactly one in_progress, transitions marked the instant they happen.
4. Treat each goal as a phase: create its own worktree off the integration base; dispatch its dependency-ordered lanes as ONE workflow run (read the mass-ulw skill first; ordering-free lanes stay a task batch); verify every criterion with real-surface evidence; land the worktree on the integration base at agentToolkit.checkpoint({ goalId, status: "complete", evidence }) per the repository's flow (direct merge or merged PR); define the next goal's run from what this one proved. Tests alone never prove done. When a mass-ulw pointer accompanies this skill, this contract still owns goals, criteria, evidence, and checkpoints.
5. Stop when the goal's WHEN-TO-STOP line holds with evidence in hand.

When the injected ultrawork directive accompanies this skill, its goal/notepad/todo bootstrap is subsumed by this contract: the loop SDK owns goal state and the loop ledger is the notepad. Do not create a second one.

The SDK: one import, then method calls

Every ulw-loop operation runs inside a JS eval cell through the SDK the extension publishes at OMO_AGENT_TOOLKIT_SDK_ROOT. There is no omo_agent_toolkit tool and no CLI to spawn on Senpi.

const { agentToolkit } = await import(`${env("OMO_AGENT_TOOLKIT_SDK_ROOT")}/sdk.js`)
print(await agentToolkit.status())

Rules that keep it working:

• JS cells only. From a py, rb, or jl cell, run a separate eval with language js.
• Import once per kernel lifetime. agentToolkit stays bound in later cells; re-import only after a kernel restart or a ReferenceError: agentToolkit is not defined.
• Every call resolves to an envelope, never a throw: { ok: true, operation, result, nextActions, warnings? } or { ok: false, operation, error: { code, message }, warnings? }. nextActions are things to do next; warnings are facts to know (a fallback the binder took, a driver objective that differs). Read nextActions before deciding the next step, and branch on error.code, not on the message text.
• Never pass a session id or plan path. The SDK binds PI_SESSION_ID and PI_SESSION_CWD from the host env on each call; state lives under .omo/ulw-loop/<session-id>/. When PI_SESSION_CWD is missing (seen after an extension reload restarted the kernel), the binder falls back to the cwd recorded in the PI_SESSION_FILE header, then to process.cwd(), and every envelope carries a warning naming the fix: env("PI_SESSION_CWD", "<session cwd>"). status().result.binding shows { cwd, cwdSource, sessionId, goalStorePaths }, so check it after any kernel restart and re-pin the env when cwdSource is not PI_SESSION_CWD.
• The driver snapshot is filled automatically from this session's goal store. Pass codexGoalJson only to override it.

Methods (argument fields are exact):

Non-Negotiables

• Write loop state only through the SDK; it lives under .omo/ulw-loop/<session-id>/ and is never hand-edited. Mutations are serialized across processes by the session's .state.lock, so parallel recordEvidence calls from workers are safe.
• Register goals up front, shaped by references/define-goal.md (agentToolkit.createGoals({ brief }), then create_goal from the returned handoff), and mirror every atomic step into the live todo checklist: one ultra-granular step per action, exactly one in_progress, transitions marked the instant they happen.
• After any compaction or context loss, re-read brief + goals + ledger FIRST plus agentToolkit.status() (re-import the SDK if the kernel restarted; help() lists every method with its argument fields), confirm result.binding.cwdSource is PI_SESSION_CWD and re-pin it with env("PI_SESSION_CWD", result.binding.cwd) when it is not, then resume; never re-plan from scratch.
• If createGoals answers ULW_LOOP_PLAN_EXISTS_COMPLETE, this session's aggregate is already done: start unrelated new work in a fresh session instead of steering or forcing the completed state. Use force: true only to intentionally overwrite completed evidence.
• Every success criterion needs observable evidence from a real surface: a channel (terminal/TUI via the xterm.js web terminal, HTTP, browser, computer-use) or, for CLI- or data-shaped criteria, an auxiliary surface (CLI stdout, DB diff, parsed config dump).
• Evidence is bound to the tree it was captured at (git rev-parse --short "HEAD^{tree}"); it goes stale only when tracked content changes — a rebase or amend that keeps the tree identical keeps it valid. When the tree differs, re-run at the current HEAD and re-record, never relabel or regenerate. Record only after cleanup receipts exist.
• Delegate code edits, test writes, fixes, and QA execution to right-sized omo-senpi subagents through the native task tool or through workflow nodes when the phase's lanes carry ordering.
• Use git-master for git-tracked edits: inspect recent and touched-path commit history, then commit each verified work unit atomically in the repository's observed language, scope, and message style with only that unit's files staged. Never carry verified units into a later omnibus commit.

Team mode: decide it, do not default to it

Solo execution with parallel background task workers is the default: fan independent units out in one batched spawn, each routed to the category (or configured subagent_type) that fits it, with scopes cut so no two workers write the same files. A team (team_create) adds per-member briefing, shared-state, and relay overhead, so it must be paid for by the work's shape. Decide ONCE, when the plan's work units are known, and record the verdict plus its reason in the notepad.

Stand up a team when BOTH hold:

1. The units' scopes overlap in a way you cannot cleanly cut. They touch the same module, contract, or migration, so one unit's discovery changes what another should do. Fire-and-forget workers cannot exchange that mid-flight; teammates can, because the lead relays it.
2. Running them at the same time actually finishes sooner. The units are each substantial and none is merely waiting on another's output. Two units where the second only consumes the first's result are a sequence, not a team.

When the units are genuinely independent — separate files, no shared contract — spawn parallel background task workers instead and avoid the team coordination overhead entirely. When the work is one cohesive unit, do it yourself. Overlap alone is not enough: near-identical units that would collide on the same lines are faster done in sequence by one worker.

Under team mode, isolate and land per unit:

• One git worktree per member, never a shared checkout — concurrent members editing one working tree corrupt each other's diffs and evidence. Give each member its own branch off the base and its own worktree path.
• Merge per work unit, as each unit is verified. A member's unit lands when its own evidence is captured and its gates are green; it does not wait for the slowest sibling. Integrate each merged unit back into the base the others branch from, so overlapping members rebase onto real merged work rather than guessing at it.
• Conflicts are the lead's job. When two members' units touch the same lines, the lead decides the order they land and tells the later member what changed; members never resolve a sibling's conflict blind.

Native Senpi Task Contract

Senpi already exposes its real subagent spawn surface through the omo-senpi task component. Use it directly. Do not route delegation through external app-server threads or another harness.

Every worker prompt starts with TASK: and names DELIVERABLE, SCOPE, VERIFY, and STOP WHEN. Put requested skill names and all required context inside prompt; children do not inherit interview context automatically.

Driver goal lifecycle

The session goal you create with create_goal is a DRIVER the loop instructs, never a gate. A checkpoint accepts an optional driver snapshot, records it verbatim, and never rejects on its status or objective; the advice comes back in nextActions. The snapshot is filled from this session's goal store automatically, so pass codexGoal only to override it. If the driver was completed early, the advice tells you to create_goal again with the plan's objective verbatim - a completed goal is replaced, not reused. If the driver is paused or usage/budget limited, the advice tells you to resume it. A different objective is a warning, not a refusal, and it is reported once: the first checkpoint under that driver records the objective in the plan (acknowledgedDriverObjectives) and later calls stay quiet; status().result.driver shows the relation at any time. create_goal is advised only when the goal store really holds no goal.