| name | collapse-pass |
| description | Run a continuous collapse-and-simplify pass that surgically removes indirection failing to earn its boundary. Use when the user says 'collapse pass', 'simplify pass', 'reduce indirection', 'shrink the surface', 'find what to delete', when asking to audit a package for dead abstractions, or when the goal is a sequence of small refactor commits that delete more than they add. Pairs with code-audit (smell catalog), refactoring (per-change mechanics), one-sentence-test (cohesion gate), cohesive-clean-breaks (deep redesigns), approachability-audit (first-read sanity), and post-implementation-review (second-read after each commit). |
| metadata | {"author":"epicenter","version":"1.0"} |
Collapse Pass
A collapse pass is a session-long sequence of small commits that each delete one piece of indirection. Every commit must shrink the public surface, the file count, the call-graph depth, or the first-read effort. If a commit moves none of those needles, revert it and find a deeper smell.
Related skills: code-audit lists the codebase-specific smell categories with grep patterns. refactoring owns the per-change mechanics (caller counting, inlining, surgical commits). one-sentence-test is the cohesion gate for each candidate file. cohesive-clean-breaks covers the deeper redesigns when a collapse won't fit in one commit. approachability-audit checks the diff from a stranger's perspective. post-implementation-review is the second-read protocol after each commit.
References
Load on demand:
Operating principle
When a library refuses your model, treat the refusal as information about the model, not as friction to route around. If a "simplification" requires reimplementing a library's public surface, stop and delete the model instead.
Per-iteration ritual
For each candidate file or symbol family:
-
Pick one target. Count non-test callers exactly with rg. If zero, the candidate is a dead-export collapse. If one, it's an inline-the-helper collapse. If many, find a narrower target.
-
Mentally inline every helper, wrapper, prop, and indirection layer into its callers. Read the inlined result as a stranger would.
-
Run the one-sentence test on the file. Write one concrete sentence describing what it does today. If the sentence drifts or grows "or" clauses, name the ambiguity; that ambiguity is usually the smell.
-
Surface the finding BEFORE editing, in this exact shape:
Finding N: <smell>
Inline check: <what mental inlining showed>
Fix: <proposed change>
What stays the same: <visible behavior, durable strings, blob layout>
-
Apply mechanical, low-risk findings only. For each:
bun test on impacted packages
bun run typecheck on impacted packages
- One conventional-commit per logical simplification, citing
file:line and naming what collapsed
-
Re-grep the removed/renamed symbol. Sweep stragglers (stale JSDoc, dead re-exports, orphaned imports) in a follow-up chore: straggler sweep commit.
-
Move to the next file.
The anti-cosmetic gate
After each commit, at least one must be true:
- Public API surface shrank (one fewer exported name)
- File count shrank (single-function file folded into its caller)
- Call-graph depth shrank (one indirection layer removed)
- A future first-read got measurably easier
If none is true, the change was cosmetic. Revert and find a deeper smell.
Implementation Gate
When a collapse pass follows a fresh implementation, do not limit the review to
symbols that existed before the change. New code is often the easiest place to
remove indirection. Check every new helper, component, wrapper, prop callback,
options object, and file split before declaring the implementation done.
Use this quick table before staging:
boundary callers earns itself by
WidgetHost 1 owns resource and widget lifetime
WidgetView 1 no, only passes a stable handle
One-caller boundaries can stay when they isolate a lifecycle, unsafe boundary,
public contract, or long imperative phase. They should collapse when they only
pass through stable handles, callbacks, or values that the caller already owns.
Pause and surface to the user
Stop and ask before:
- Changing any string from
references/never-touch.md
- Deleting a public exported name with zero in-repo callers but plausible external CLI/SDK consumers
- Collapsing two files where one's JSDoc documents a non-obvious invariant
- Merging packages or moving exports across package boundaries
- Changing a function signature that crosses a published package boundary
Stop conditions
Stop when any of the following is true:
- Three consecutive files yield no findings
- Remaining findings all require product input (renaming a public capability, splitting a package, adding a tenant axis)
- A typecheck or test regression cannot be resolved in one follow-up commit
- A configured checkpoint budget is reached (e.g. 8 checkpoints)
At stop, deliver the final report from references/report-format.md.
Pass parameters worth declaring
A goal that invokes this skill should say:
- Scope: which packages and which apps
- Stop condition: "three no-finding files" or "N checkpoints" or "queue empty"
- Citation requirement: whether library refusals must be backed by a deepwiki citation against the upstream repo
- Starting target: usually the narrowest surface first (e.g.
packages/auth before apps/api)
Everything else (the ritual, gate, finding format, never-touch list, report shape) is in this skill.