| name | wip-advance |
| description | Advance work-in-progress across many repos by fanning out a subagent tree that rebases every branch, resolves conflicts, tests, and fixes failures — looping while it makes progress and ending with a ✅/❌ tree report. Use when the user wants to advance, process, or sweep their WIP across projects (the manual `j g` → /git:conflicts → `j ta` → /build:fix loop, automated). Accepts optional include/exclude project filters. |
Advance work-in-progress
Automates the manual loop the user runs per repo today — j g (rebase every branch onto
UPSTREAM_REMOTE/UPSTREAM_BRANCH) → /git:conflicts on conflict → j ta (test every branch) →
/build:fix on failure → j test-fix to absorb the fix — and fans it out across every selected
project and branch as a tree of subagents.
The deterministic decisions (which projects, the DAG-deduped unit list per project, per-repo
concurrency, whether the baseline needs testing) come from the wip advance CLI. This skill is the
root (L1) of the tree: it turns that plan into project subagents, gates how fast they spawn by live
machine load, and renders the final report.
Resolve the wip binary
Do not rely on a wip alias on PATH — it is often stale (pointing at a pre-monorepo path). Use the
monorepo binary directly. This skill's base directory is <repo>/.claude/skills/wip-advance, so the
binary is <repo>/packages/cli/bin/run.js; set WIP="<repo>/packages/cli/bin/run.js" (resolving
<repo> from the injected base directory) and invoke it as "$WIP" advance …. Always read the
CLI's JSON from stdout only and discard stderr (… --json 2>/dev/null) — the CLI logs debug
output to stderr, and merging it (2>&1) corrupts the JSON.
Tool choice: mechanical vs. Claude
Use the mechanical command for anything deterministic; only spend a subagent where judgment over
file contents is required.
- Rebase, test, absorb-fix (
j test-fix), prune — mechanical (git, just --global-justfile …,
wip).
- Resolve merge conflicts —
/git:conflicts (fully qualified, never the conflicts shorthand).
- Fix a failing build/test —
/build:fix, or the /build:test-branch loop which already wraps
test → /build:fix → j test-fix → retest.
Compute the plan
Run "$WIP" advance plan --json 2>/dev/null, passing the user's filters as repeated --include /
--exclude globs (or a single project name as the positional argument). The output is one entry per
advanceable project: { project, dir, upstreamRef, upstreamRemote, upstreamBranch, concurrency, baseline: { sha, needsTest }, units: [{ id, branch, tipSha, chain, dependsOn, worktreeRequired }] }.
Discovery scans every repo under the projects dir, so always pass the user's filters to keep it
focused, and expect it to take a little while.
If the list is empty, report that there is nothing to advance and stop.
Fan out one subagent per project
Spawn an advance-project subagent per project, passing that project's full plan entry as JSON in
the prompt. Spawn in waves rather than all at once: before each wave run
"$WIP" advance admit --json 2>/dev/null and only start more projects while it reports ok: true;
otherwise wait for a running project to finish first. Always keep at least one project advancing so a
loaded machine still progresses.
Each advance-project subagent returns a project report node (the project plus a ✅/❌/🛑/⏭️/⚠️
result for every branch it touched).
Render the report
Assemble the project nodes under a single root node and render the tree with red/green emoji and
branch names, mirroring the subagent tree. Lead with anything that needs the user's attention —
projects where the baseline was broken and fixed (⚠️), and branches left stuck (🛑) with the failure
detail — then the fully-green projects.
