Sks Async

Shikanime Org Parallel Streams

Decompose a multi-unit change into parallel, isolated streams; land each as an independent PR or stacked chain. Distills unlazy depth-tree delegation fan-out onto jj's commit DAG as plain gh pr merges. Core splitting component of sks-dev-workflow.

When to Use

• Several units, some independent.
• Parallel agents (delegate_task fan-out) must not share a working copy.
• Shape: B needs A → depth (stack); C needs A AND B → join.

Procedure

1. Tree before work — decompose vs issue ledger; record tree in plan/todo with leaf gates fixed BEFORE fan-out (contracts before delegation).

2. Fan out — ALWAYS a NEW jj workspace; never an existing one (default is for trunk observation). One workspace per unit, rooted at trunk; name <repo-name>.<unit> (repo-global, dot-qualified):

   jj workspace add ../<repo-name>.<unit> --name <repo-name>.<unit>
   

   New workspace's working copy is a child of @; for depth > 1 root with jj new <parent>.

3. Work each stream in its dir; commit per sks-commit — every commit carries the trailer:

   jj describe -m "<subject>" -m "Co-authored-by: Automata <automata@shikanime.studio>"
   

4. Land (push to origin, PRs with --head <org>:<branch>; see sks-dev-workflow):

   • Independent unit → own bookmark + standalone PR (or single-member stack).

• Dependent chain → one bookmark per link. Example:

jj bookmark set <next> -r <next>
gh pr create --repo <org>/<repo> --base main --head "<org>:<next>" \
  --title "<subject>" \
  --body "$(cat <<'EOF'
## What
<brief bullet scope of this chain link>

## Why
<linked issue: why now>

## References

Related: <issue URL>
EOF
)"

• PR↔issue linkage per sks-pr: Related: <issue URL> by default.
• Before opening each PR, run the sks-pr step 2b duplicate/stack check — an existing open PR covering the unit means push to it or stack on it, never a second PR for the same change.

5. Verify bottom-up — each leaf's checks run IN its workspace; dispatcher re-runs them (subagent self-reports aren't evidence). Retire with jj workspace forget <name>; refresh idle with jj workspace update-stale.

Fan-out via delegate_task

Each child: workspace path, unit gates, commit shape (plain English + Automata co-author trailer). Parent re-verifies every gate via terminal in each workspace before reporting done. Dispatch delegate_task(tasks=[]): one task per leaf, goal carries the contract; never bundle two leaves (defeats isolation). Skeleton: references/sks-async-delegate.md.

Verification

jj workspace list && jj log -r 'all()' --limit 20   # tree shape on screen
gh pr list --repo <org>/<repo> --state open          # PRs exist per leaf

DAG matches planned tree; every leaf has a PR linked to its issue; every gate has in-workspace evidence.

See also

• sks-dev-workflow — parent; run its assumption-validation gate BEFORE fan-out.
• sks-commit / sks-pr — commit shape (co-author trailer) and PR linkage.
• cpn-dev-workflow — same fan-out for console module migrations.
• Model, pitfalls, dispatch skeleton: references/sks-async-delegate.md.