harness6

Use at the START of any task-execution session, or whenever picking up interrupted work, opening an epic to check on its children, or descending into a child ticket mid-session, to load ambient progress-state (blockers, drift, gotchas, where the last session left off) before touching code. Backed by a remote Graphiti temporal knowledge graph via the `memory` MCP server. Trigger on cold start, resuming a backlog ticket/epic, or picking up any work that a prior (possibly context-limited) session may have partially completed.

Use whenever you hit a blocker, discover drift between ticket intent and delivered result, reach a stopping point (session end, context-limit handoff, done-pending-verification), silently descope something, or resolve a previously-noted blocker/drift, to persist that ambient progress-state for the next cold start. Backed by a remote Graphiti temporal knowledge graph via the `memory` MCP server. Event-driven — not continuous per-step logging, not session-end-only.

Autonomously build a complete feature, bugfix, or project end-to-end from a single topic/initiative/epic — scope-review gating, nested autonomous design sessions, implementation dispatch, and autonomous PR review & merge, all tracked through a GitHub issue tree. Invoke explicitly with /autobot <topic>.

Use when the human asks to design, plan, or build a non-trivial feature, refactor, or change that needs upfront design (multiple files, architectural decisions, ambiguous requirements, "let's build/design/plan X"). Runs scope-review first and routes on its verdict: TASK grills to a single crisp PR, EPIC runs the full grill → spec → tickets → triage → docs-PR flow, INITIATIVE splits into epics and halts for child sessions. SKIP for: small bugfixes, single-file edits, well-scoped changes with clear requirements, or pure questions.

Initialize the shared harness6 infrastructure stack after installing the plugin. Resolve the plugin cache root from the active harness, choose a container runtime, scaffold infrastructure/.env without overwriting it, start Compose, register the bundled MCP servers, and set up memsearch.

Use when re-indexing project docs after content changes — user runs /memsearch-index or has added/edited docs since last index.

Use when setting up memsearch for a new project for the first time — user runs /memsearch-init or wants to enable semantic doc search in the current repo.

Use when running a semantic search against indexed project docs — user runs /memsearch-search or the auto-trigger fires on a docs-lookup question.

Use when a PR raised by a background team agent has merged — whether the user reports it, or the coordinating agent notices the MERGED state itself while checking on that agent (a task-completion notification mentioning the PR, an agent/task listing showing an agent with an open PR, or a `gh pr view`/`gh pr list` result showing MERGED). Handles worktree removal, agent shutdown, and task completion.

Expand the short answer just given into a detailed, evidence-backed Markdown report the user can read at their own pace. Use when the user asks for the full detail, a writeup, a deep-dive, or says they want to read something properly later — and as the escape hatch whenever the response-style line budget forced real detail out of a terminal answer.