[teamleaderleo]

Leo (2026-11-04): "Couldn't we architect stuff to make everything more amenable to hot swappage?"

The view shell's live code never changed what it drew. In the headless measurement (scripts/measure/live-code-shell.sh), every generation's badge was the same picture, grey with padding 8. Each patch loaded and replaced its functions (functions=4), but the drawing still ran the build's code. The measurement's "second patch redrew: yes" came from cmp on PNG bytes (2,036 vs 2,044 bytes for the same pixels).

Why. The executable is linked -interposable -no_fixup_chains, so its calls to its own Swift functions go through classic symbol pointers. One of them is __la_symbol_ptr for PreviewBadge.body (dyld_info -fixups: lazy-bind PreviewBadge.body). Generation 0 draws before any patch, so by then dyld has bound that pointer to the build's body, and dyld_dynamic_interpose does not rewrite a pointer that is already bound.

Fix. After the existing dyld_dynamic_interpose call, Interposer.rebind writes the patch's address into every __la_symbol_ptr and __got slot of the executable that names a replaced function, as fishhook does. __DATA_CONST is made writable for the write. The existing interpose call stays as it is.

Measured on a fleet Mac at this fix plus #35705. inject.sh was the only other change, to tar the PNGs out. The badge's rendered width, at 2× scale:

generation	edit	before this fix	with it
0	none	135 px, grey	135 px, grey
1	padding 8 → 24, red	135 px, grey	199 px, red (+64 = 2 × 16 pt × 2)
2	padding 24 → 40	135 px, grey	263 px, red (+64)
3	a new file's green dot (#35705)	135 px, grey	283 px, red, with the green dot (+20 = (6 + 4) pt × 2)
4	a new @State	refused	refused

I checked the PNGs by eye as well. Edit to pixels: 1.45 s for the first patch and 1.14 s for the next.

Follow-up: the measurement should compare decoded pixels or widths, not bytes. It is a separate PR because the fleet runs only the reviewed bytes of a measurement script.

Part of manaflow-ai/cmuxterm-hq#7066.

🤖 Generated with Claude Code
Agent session: agent session

---

^{Open in [code]smith · Need help on this PR? Tag @codesmith-bot with what you need.}

---

Summary by cubic

Fixes the view shell's live code so a patch actually changes what draws: previously every generation of the headless measurement drew the same badge, because dyld_dynamic_interpose does not rewrite a symbol pointer dyld has already bound (the build's view had been drawn before any patch).

• Adds Interposer.rebind, which after the interpose call writes the patch's address into every __la_symbol_ptr and __got slot that names a replaced function, as fishhook does.
• Makes __DATA_CONST writable for the write via mprotect.
• Measured result: the badge's padding and color changes from a patch now show in the rendered output, and a new file's drawing appears.

Migration

• None. The executable's existing -interposable -no_fixup_chains linking is required for the rebind to be possible; that is already in place.

^{Written for commit 267eac1. Summary will update on new commits.}

[github-actions]

cmux status _{(one comment, edited in place; each line is updated by the job that owns it)}

Review 267eac1: HQ feed card · gallery before/after runs · checks

App of this head (app.zip, kept 1 day)

gh workflow run reload-build.yml -R manaflow-ai/cmux --ref feat-cmux-next -f tag=pr35850 -f ref=lane/live-code-rebind-pointers -f platform=macos -f backend_mode=local

Shipped in:

• cmux NEXT nightly 1.0.0-nightly.4191131679101 (e01bbae5c9)

[teamleaderleo]

Review (codex, router-review): SHIP at 267eac1

Reviewer: codex gpt-6-astra high, fast tier.

Follow-ups: 2. They land as checklist items in this PR's one follow-up tracking issue in this repo (Follow-ups: #N ..., label follow-up; add to an open one before filing, never one issue per finding), linking this receipt; the lane that opened the PR fixes them forward.

SHIP at 267eac1

Visual verdict: no images; Leo tries the build

• [P2] [follow-up] [category:reliability] [finding:F1] [outcome:pending] [Prototypes/ViewShell/Sources/ViewShell/LiveCode.swift:156] rebind ignores failed mprotect calls and apply ignores its write count, so a patch can report success while some pointers still target old code. Surface failures or verify all required slots were rebound.
• [P2] [follow-up] [category:ux] [finding:F2] [outcome:pending] [Prototypes/ViewShell/Sources/ViewShell/LiveCode.swift:156] No after screenshot or gallery artifact was available; attach one so the live-render visual claim is reviewable.

Follow-ups:

• F1: Handle partial or failed pointer rebinding.
• F2: Provide an after visual artifact.

[teamleaderleo]

Pictures from the headless measurement (--live-render, badge at 2×):

	picture
before this fix, generation 2 (padding 40, red, expected)
with it, generation 1 (padding 24, red)
with it, generation 2 (padding 40)
with it and #35705, generation 3 (a new file's dot)
Agent session: agent session