chore: remove dead code and unify two repeated patterns

[mangit955]

Why

This started from a suspicion that the codebase carries too many comments. Measured, it does not: 3,983 comment lines against 19,572 non-test source lines, and almost all of them explain a decision rather than restate the line below. The ratio is unchanged by this branch — 3,981 / 19,552. About fifty comments that named what the next line already said were removed, and roughly the same number added saying why the code is shaped as it is.

The real problems were elsewhere, and most of them were not visible from reading:

• The engineering that decides this project's defaults was invisible where anyone would look for it. runtime/compaction.ts records a genuine A/B — compaction cuts peak prompt characters by 36–43% and is off, because implicit caching stopped entirely when it ran — and README.md described the product without mentioning any of it.
• tui/src/store/ui-store.ts solved the same problem two ways in one file. Three of its four modal flows settle through a private resolvePendingX helper; the edit flow predated it and repeated the resolver lookup, state clear and emit three times.
• config/sessions.ts spelled out the index read-modify-write at five call sites — the sequence that has already lost a row once, which is why summariesFor compares row count to files on disk and rebuilds.
• Dead code: an unreachable /exit branch in the prompt, and a commented-out console.log in the onDone callback.

What

Commit	Change
5149d5d	README.md states how the context work was measured, using the numbers already in runtime/compaction.ts — including that the harness reports cache rates for the original recordings only and a modified assembly cannot inherit them
44725bf	Removes the unreachable /exit branch. handler.ts returns handled: true for every /-prefixed input, unknown commands included, and /exit is registered at commands/slash/commands.ts:211
e4805da	Drops the comments that restate the code; hoists glob's dynamic import("fs") to the top of the file
e1c4156	ui-store.ts: the edit flow settles through one resolvePendingEdit like its siblings
7721587	config/sessions.ts: five index read-modify-writes go through one updateIndex

Two shapes are deliberate and worth review.

resolvePendingEdit takes "approved" | "rejected" | "cancelled" rather than a boolean. Cancelling rejects the waiting promise where declining resolves false, and editFile reads the difference: a rejection is reported as "Edit cancelled", a false as "Edit rejected ... Do not claim this edit was completed". Collapsing them would report an interrupted turn as a refusal the user never made.

updateIndex takes the missing-index behaviour as a required parameter rather than defaulting it, because the five call sites did not agree and the disagreement is load-bearing: saving rebuilds from the files so the new row joins the existing ones, pruning starts from empty, and removing does nothing at all — writing an index for a project with no sessions would create the directory lazy creation exists to avoid.

updateIndex does not close the window between the read and the write. Two processes still interleave; summariesFor is still what catches it. The alternative — a lock file or a single-writer queue — buys correctness this store does not need, since the count check already heals the case and costs one directory read.

config/sessions.ts was left as one file. Splitting it into modules is a reasonable next step and does not belong in the same diff as a behavioural extraction.

Verified

bun run verify --all on the tree as pushed:

  ok    docs lint  0.0s
  ok    docs surface  0.1s
  ok    type check  1.4s
  ok    tests  15.7s
4 gate(s) passed.

• Reverse-order sweep, bun test $(git ls-files '*.test.ts' '*.test.tsx' | sort -r): 1508 pass, 0 fail.
• Four concurrent bun test runs: 1508 pass, 0 fail in all four.
• Touched files in isolation: ui-store.test.ts 62 pass, sessions.test.ts 52 pass, app.overlay.test.tsx 11 pass, persistence.e2e.test.ts 16 pass.
• bun run replay:baseline before and after: identical — mean peak 126,563 characters, totals −44.4% over 932 iterations. No context change was intended and none happened.

Five tests were added, because mutation testing showed four behaviours this branch depends on were covered by nothing. Each mutation was confirmed applied by grep before the run, not assumed:

Mutation	Before	After
clearPendingEdit resolves false instead of rejecting	1503 pass, 0 fail	2 fail
Index "rebuild" fallback collapses to empty	1505 pass, 0 fail	1 fail
Index "skip" fallback collapses to empty	1505 pass, 0 fail	1 fail
Index "empty" fallback collapses to skip	—	4 fail

The ui-store case is the one worth knowing about. A test named "resolves a dismissed approval as declined, not as an error" looks like it covers the rejection path and does not — it covers dismissTopModal, which declines. The cancel path was reached only by overlay tests that swallow the rejection as teardown. It read as covered and was not.

The three index tests assert against the index file on disk rather than through listSessions, which heals a wrong index by rebuilding and would have hidden all three.

Not verified

• Linux. Everything above ran on macOS only. This branch touches config/paths.ts and config/sessions.ts, and CI runs both operating systems precisely because the config loader depends on path semantics that differ between them. That check happens on this PR, not before it.
• Interactive input. The /exit branch removal and the modal settlement change are both on the keyboard path. The suite drives the store and mounted components; no real terminal, no real keypress. Ctrl+C over a live diff is unexercised.
• Live providers. Untouched by this branch and unexercised by it.
• /exit itself has no test. Its removal rests on a static reading of handler.ts plus a green suite. If registerCommands() ever failed to run before a prompt was submitted, the deleted branch was a fallback and no longer is.
• The full per-file isolation sweep was run over the four touched files, not all 95.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
woop-code	Ready	Preview	Aug 5, 2026 2:23pm