Keep an open PR moving to merge-ready - fix failing CI, resolve merge conflicts, and handle review comments rigorously (verify each before acting, push back when it's wrong). Use for "babysit PR 142", "get it green", "address the review comments", "check on my PR". Opening a PR does not start it.
Drive the installed code-review-graph CLI - build or update the repo's code graph, check status, export an interactive view, and get the impact (blast radius) of a diff with detect-changes. Use in setup (build), recon (what does this change touch), before a PR (impact summary), and for "graph the code", "what calls X", "what does this touch".
Draw an architecture, sequence, data-flow, or lifecycle/state diagram that reflects the real code. Grounds the picture in source first (how, code-graph), then renders it with the installed archify skill (HTML/SVG/PNG); falls back to a Mermaid block when archify isn't installed. Use for "diagram this", "draw the architecture", "show the request flow", "visualize X".
Shape a fuzzy idea or stress-test a plan by interviewing the user in rounds until every decision is settled, then write the decisions down as the ticket work runs from. Use for "brainstorm", "grill me", "stress-test this plan", "help me think this through", or when work finds no falsifiable definition of done.
Explain how something in the codebase works, from source, with file:line citations - at the level of a senior engineer onboarding onto a subsystem. Use for "how does X work", "walk me through Y", "what happens when Z", and before any nontrivial change or architecture decision.
Run the runs or decisions tickets you list at bedtime through work while you sleep - gates answered for this session only, PRs merged only where the [overnight] policy allows, and a report in the morning. Use for "run my listed work while I sleep", "get these done overnight", "finish them while I sleep".
Find best-practice skills and docs for the repo's stack and link them in .sstack/project.toml [practices] so dispatched children read them. Use for "best practices for this stack", "link skills to this repo", "what conventions apply here", or from setup.
Turn the user's corrections and the agent's dead ends into lessons so mistakes aren't repeated, and turn repeated lessons into structural edits (a skill, AGENTS.md, a check) instead of more notes. Use on "reflect", "/reflect", "remember this", after the user corrects you, or at the end of a task that hit a dead end.
Adversarial review of a working diff or a PR ("review", "review 142", "are we sure?", "check this before I merge"). Reviewers on a different model family try to break the change; only findings that reproduce against current code survive. Reports findings with evidence; never fixes. Also used by work's review step and for plan critique.
Make a repo agent-ready for sstack, new or existing. Audits the repo and the machine read-only, shows one table of recommended changes, and writes only the rows the user picks - .sstack/product.md (what the app is, linking deeper docs), .sstack/project.toml, the verify commands, the code graph, and one AGENTS.md row. Use for "set up this repo", "onboard", "make this repo sstack-ready", or when work finds no .sstack/product.md on a sizeable task.
Entry point for any engineering request (build, fix, refactor, speed up, review, explain, set up a repo, diagram, "what went wrong") and for "/sstack". Picks the right sstack skill or playbook, loads repo context and accepted lessons, and enforces gates, verification and delegation rules. Use before starting engineering work.
Fan one question out to N independent children (role scout, or another role) through dispatch, then aggregate their PASS/ISSUES/BLOCKED reports into one result. Use for coverage sweeps ("check every service for X"), races (same brief, several models), or recon that needs many sources. Not for edits to a shared working tree.
Pre-PR cleanup of a diff and of prose (PR description, commit message, docs, replies). Removes AI code slop and needless comments without changing behaviour, then rewrites text to drop AI patterns. Use before every PR, on "deslop", "unslop", "clean this up", or when writing prose others will read.
Find out why code is the way it is - the forces, decisions and incidents behind its shape - from git history, PRs, tickets, docs and chat, with every claim tagged by confidence. Use for "why do we do X", "why was this added", "is this still needed", or before changing code whose purpose is unclear.
Run any engineering task (feature, bug fix, refactor, perf, multi-phase) through the repo's resolved workflow steps - recon and report, plan with an answer key, implement (local or routed through dispatch), test/fix with review, PR on the user's yes. Ticket ID optional. Use for "work this", "work UNI-142", "fix this bug", "build X", "phase 0 this", or when the sstack router picks a playbook.