mkly

Drive a development goal through durable task creation, isolated implementation, review, merge or follow-up fixes, and cleanup in bounded rounds. Use for end-to-end or loop-until-done requests.

Claim and implement one queued Taskwarrior development task in its isolated worktree and Crabbox, then produce a local review branch without completing or merging the task.

Review and finish one implemented Taskwarrior task by verifying its local review branch, fixing the findings on it, then merging, or creating durable follow-up tasks when a major blocker stops the merge, and cleaning resources.

Create or refine durable, implementation-ready Taskwarrior development tasks without claiming or executing them. Use for filing, queuing, or decomposing development work, including follow-up findings.

Stop and ask when development is blocked by broken tooling, permissions, services, credentials, or harness infrastructure. Use after two distinct environmental repair attempts fail, or before bypassing the environment.

Maintain a shared, searchable project knowledge archive in Maildir form. Use to post durable decisions, questions, findings, and threaded context that any number of development agents can discover with notmuch. Also use when blocked, escalating, or abandoning…

Claim one +PLAN decomposition task, read its attached plan artifact, create the finer-grained follow-up tasks it describes, and finalize it as decomposed. Use for draining the plan queue; never implements code.

Research current external technical information and synthesize authoritative guidance. Use when Codex lacks confidence about third-party APIs, syntax, CLI flags, error messages, versioned behavior, or documentation; consult sources before guessing or invoking…

Run SessionBoard eval scenarios by driving a real browser yourself via the sbek MCP server. Use when asked to evaluate, test, or gather evidence on a submission URL, or when asked to run sbek scenarios in-session (no API key needed).

Grade a completed sbek run — read the evidence and screenshots, write per-area judgements, and produce the score report. Use after browsing scenarios, or when asked to judge, grade, or score a SessionBoard eval run (no API key needed).

Use the saas CLI to manage The Last SaaS collections, records, access rules, members, notifications, and files on behalf of an authenticated user.

Drive this machine's headless desktop - launch programs, read the screen, and send input. Use for GUI applications that have no command-line equivalent.

Use when designing or evaluating prompts that need distribution-faithful sampling, probabilistic instruction following, mixed-strategy actions, random-seeming choices, or more diverse open-ended LLM outputs using String Seed of Thought prompting.