devai

Stress-test a plan, decision, or idea with a context-aware interview. Use for /ctx-grill, "grill this plan", "stress-test this plan", or when the user wants their thinking stress-tested before committing. Reads durable contexts first (never asks what they answer), asks direction-changing questions one at a time with a recommended answer, stops at diminishing returns, and files resolved decisions back. Interactive only — never invoke under CI, loops, or other non-interactive runs.

Deep research grounded in the project's durable context system. Use for /ctx-research, "ctx research", "research X starting from our contexts". Reads existing contexts first, sharpens the question to the real knowledge gap, fans out multi-agent research, verifies key claims, and files results back.

Dispatch the current task to multiple subagents — after checking that fan-out is actually justified. Use for /fan-out, when the user says "multi-agent this" or "fan out", or when a task decomposes into genuinely independent subtasks. Decomposes along context coupling, uses handoff discipline for spawn contracts.

Generate a minimal-sufficient handoff when another agent or session will take ownership of a task or continue unfinished work. Use for /handoff, "hand this to the next agent", "let another session continue", session transfer, task ownership transfer, or restarting in a fresh context when the current session has degraded (a handoff to your future self). Reuses durable artifacts by reference instead of copying them, and redacts secrets.

One-shot answer-only turn — the user is asking a question, not assigning work. Invoked as /just-ask with the question attached (e.g. "/just-ask why does this test flake"). For that turn only, investigate read-only and reply with an analyzed answer; do not edit files, run state-changing commands, write a plan, or start a fix. The next message returns to normal. Meant for explicit invocation; also applies when the user says "just answer, don't do anything", "just explain, don't fix it".

Execute a spec'd goal in one long autonomous run — hours of agent work pushing a whole acceptance ledger, with humans only at the two gates, spec before and review after. Use for /long-horizon, "make this a long-horizon goal", "run this across sessions", or when resuming work in a repo that already has a `.long-horizon/` workspace. Admission is interactive and iterative; the run is fan-out-driven and side-effect-bounded by an authorization envelope; progress is evidence-gated; a live report.html shows the ledger as a DAG.

Low-bandwidth collaboration mode — response protocol for when the user's attention or mental capacity is depleted. Activate on /lowband or $lowband, or when the user says "too tired, keep it short", "brain is fried", "low bandwidth". Stays active every turn until the user says "back to normal" (or the equivalent in their language) or invokes with off. Work capability is not degraded; only presentation is compressed and decisions are filtered.

Write replies and work reports a person can take in on one read — one claim per sentence, action in the verb, uncertainty attached to the claim it belongs to. Use for /readout, "make your reports readable", "your output is hard to read", "stop writing like an AI", or when a run's reporting style needs fixing. The body below is standalone prose — paste it into AGENTS.md or CLAUDE.md to apply it every turn instead of on invocation.

Turn the current conversation's conclusions into a well-formed context document filed in the project's context system. Use for /to-ctx, "file this into contexts", "turn this conclusion into a context doc". Finds the placement rules (e.g. CONTEXT_MAP.md), writes in the project's house style, and reports the path.