/s:workspaceone job, many repos
shipd boardlive delivery status
/s:epicdecompose big features
/s:autopilotepics, shipped unattended
/s:askdecisions, made once
$ /s:workspace clone git@…/ws-docs-linking.gitone clone stands up a cross-repo job
workspace repo — tracked
the job itself, travels with git clone
.shipd-config.json — manifest: members, clone urls, focus
.shipd/wiki/ — the job's LLM wiki, auto-committed
initiatives/ — goals & per-project context
MEMBER REPOS — ONE JOB ACROSS THREE CODEBASES, NEVER TRACKED, NO SUBMODULES
documents/ ★focus repo · worktree of your local cloneinstant
tasks/ clone --reference, borrows local objectsseconds
incentives/ full clone from the manifest urlfresh machine only
cheapest rung first, per machine · drift is reported, never "repaired"
/s:ask answers from:job wiki→base wiki→spec surfaces→queued for you· /s:teach promotes job-independent answers to base — every future job inherits them
$ shipd boardthe delivery board, full-screen in your terminal
77 specs · 13 epics · 78 shipped this week — specs grouped by epic, epics tagged to initiatives · shipd metrics turns the archive into throughput numbers · every read verb takes --json for your own dashboards
$ /s:epic a full export CLI — json, csv, streaminga feature, decomposed before anything is planned
Bigger than one change? /s:epic investigates the codebase, asks one batched round of what it can't infer, records the epic's binding Decisions and Design, and emits a stub table of member changes with complexity ratings — then stops. Members are planned later, one at a time, via /s:plan, each inheriting the epic's decisions. And when the feature needs grounding first, /s:research turns the open question into a cited report — bounded sub-questions, web search, a numbered source list — that the epic links as its evidence base.
EPIC: export-cli · MEMBERSCOMPLEXITYSTATE
export-json-flagsimpleplanned
export-csv-writersimplestub
streaming-exportmoderatestub
export-auth-scopescomplexstub
decisions are made once, at the epic — member plans cite them instead of re-asking · epics tag to exactly one initiative, so cross-repo goals roll up on the board · approve the epic and /s:autopilot takes it from here
$ /s:autopilot export-clian approved epic, delivered unattended
Autopilot picks up an epic's unplanned members and drives each to a shipped PR — plan → gate → build → auto-merging PR, in risk-ascending order, one worktree and branch per member. Every gate still runs; nothing is skipped because nobody is watching.
MEMBER (risk-ascending)RISKOUTCOME
export-json-flaglowPR #212 merged
export-csv-writerlowPR #213 merged
streaming-exportmedium● building 4/6
export-auth-scopeshighparked · resume pointer
open decisions go to the oracle and proceed on its recommendation — questions park for you, they never block the run · under pr-mode: draft, members land as draft PRs for human review instead of auto-merging
$ /s:ask which retention window should the queue enforce?the ask-first oracle — you are the last rung, not the first
1 · read — codebase & specs→2 · oracle — memory · wiki · answered queue→3 · you — asked once
ANSWER
one cited position, quoted verbatim from the wiki — nobody is interrupted
Cited: [[logging-conventions]]
INSUFFICIENT
nobody has decided this — filed as a question for a person, recommendation attached
Queued: q-queue-retention
your answer is written back to the queue — the next agent to hit that decision gets an ANSWER, and /s:teach settles it into the wiki. never answers from model knowledge: definitive evidence, or nothing.
not just a command — /s:plan consults the oracle internally before ever asking you, and an /s:autopilot run proceeds on its recommendation instead of blocking, parking the question for you to answer later.
