Imported from Generous-Corp/pulp (.agents/skills/forge-app-delivery/SKILL.md). Install upstream with npx skills add Generous-Corp/pulp --skill forge-app-delivery. Copyright stays with the author.

Delivering a Forge app

A Forge app is not one binary. It is a shell compiled in a different repository, a Python generator, a Rack plug-in, three plug-in formats, an uninstaller and a toolchain the user's machine fetches for itself. Every delivery failure recorded here came from one of those parts being absent, stale, or present but unreachable while every check reported success.

Read this before packaging anything, and before concluding that a feature "works".

The rule that would have saved the most time

Verify identity, not size, and never a signature. In a single day this project shipped:

• a 292-byte package that signed, notarized, stapled and passed Gatekeeper while containing nothing (staged with symlinks instead of ditto copies)
• three plug-in bundles of 72 KB that installed cleanly and contained no code (CMake creates Contents/MacOS at configure time, so -e is true for an empty husk)
• a 76 MB, four-payload, correctly-sized, notarized installer containing the wrong application entirely

The third is the important one. Size caught the first two and was useless for the third, because the payloads were the right size. Only identity distinguishes a correct build from a plausible one. Pick a string that exists in the current build and cannot exist in a stale or wrong one, and assert it in the packaging script:

hits=$(strings "$binary" 2>/dev/null | grep -cF "$SHELL_MARKER" || true)

Then expand the finished package and check the payload, rather than trusting the script that made it:

pkgutil --expand "Foo.pkg" out      # --payload-files does NOT recurse into
                                    # nested component payloads

The seam: the shell is built in the other repo

Forge app shells live in forge-seam/, are copied into a throwaway Forge worktree by forge-seam/populate.sh (/tmp/forge-cur), compiled there, and copied back by forge-seam/sync.sh.

Consequences that have each cost a cycle:

• /tmp is cleared. Run forge-seam/sync.sh before finishing any session that touched a shell source, or the work is gone.
• Changes to Forge-repo files must travel in forge-seam/patches/ or they are lost on the next repopulate.
• Clangd diagnostics on forge-seam/*.cpp are noise. Those files only resolve their includes inside the Forge worktree. A wall of "no type named string in namespace std" means you are reading the file outside its build, not that the file is broken.
• There may be two apps of the same name. Forge Modular exists both as the real shell (Forge worktree) and an older examples/<app>/app/src/shell.cpp. A packaging script that hardcodes one path cannot be pointed at the other by any argument. Print which artifact you selected — a silent choice between two same-named apps is how the wrong one shipped.

Push capability down, then prove every Forge descendant

Forge Modular, Sequencer, Instrument, MIDI, and FX share Pulp infrastructure but do not share one product path. A fix discovered in one app therefore starts with two questions: can the capability or invariant live in Pulp, and which Forge descendants consume the affected seam? Put reusable behavior in Pulp or the common Forge layer rather than copying it between apps. Then record every descendant as one of:

• inherited by construction, with the shared source and a contract test named;
• adapted for a product-specific host or generator, with that adapter tested; or
• not applicable, with the differing path identified.

Never infer product acceptance from common ancestry. Modular still needs its module/Rack/standalone/DAW ladder; Sequencer needs its own timeline, MIDI, transport, persistence, and control-surface ladder. The durable win is one shared repair plus narrow product-specific proofs, not identical code or a single sibling's green build standing in for the rest.

A Modular-only release starts from Forge, not the historical seam base

The supported Modular-only signed/notarized installer entry point is examples/forge-modular/release-package.sh. Its product binaries come from a clean detached current Forge checkout and its Release build; do not reconstruct them from forge-seam/patches/BASE, which records the historical patch seam and is not release source authority.

The release driver deliberately passes exactly Forge Modular AU, VST3, CLAP, and Standalone to the shared build_combined_installer.sh primitive. It rejects different Forge source snapshots, Pulp SDK refs, toolchain stamps or toolchain content across those four bundles, requires the runtime-resolved Rack saved-patch decoder and checks both bundled helper architectures, then expands the finished PKG and refuses any FX, Instrument, MIDI, Sequencer, or other product bundle. Notarization is not optional on this path. Local packaging experiments that need an unsigned artifact continue to use the older development package.sh; they are not a release claim.

Ship the runtime, not just the binary

A Forge app runs a Python generator. If the installer carries the app and not the generator, it installs and does nothing.

This shipped. The package had the app, three plug-in formats and the Rack modules, and zero Python. It went unnoticed because the build machine had the tools seeded into ~/Library/Application Support/<app>/tools/ by an earlier manual step, and that path wins the lookup.

Two halves are required and either alone still ships a dead app:

1. Package staging copies the tools into the bundle before signing, so the signature covers them.
2. The app's tools_dir() looks inside its own bundle (walk up from _NSGetExecutablePath to Contents, then Resources/tools/…), after the Application Support copy so a user-replaceable copy still wins.

The same boundary applies to native helper executables shipped inside that toolchain. Build them at package time for the package's declared architecture, not at end-user install time and not implicitly for the packaging host. A user may have no compiler, and an arm64 packaging host may be producing an x86_64 artifact. Verify the helper's file identity, requested Mach-O slice, mode and system-only linkage before signing. Execute it only when its architecture is native to the verification host; requiring a foreign slice to run silently turns Rosetta into a packaging dependency. Cross-architecture verification must remain structural, while a native package lane exercises the helper against a real fixture.

Signed installer staging also has a hard precondition: Pulp's unattended signing doctor must pass before the first production codesign. A missing or failed preflight terminates packaging; never skip it or turn it into a warning, because login-keychain fallback can open a GUI password prompt and wedge an agent/SSH/CI session. The shared doctor owns dedicated-keychain repair, full partition authorization, identity-hash selection, and the real timestamped probe.

Test from a clean state. Move the seeded directory aside before testing, or you are exercising the fallback that hides the bug:

mv ~/Library/Application\ Support/<app>/tools{,.bak}

An installer must be able to update what it installs

The app's working copy of the generator lives in Application Support and the shipped one lives in the bundle. Preferring Application Support unconditionally means a toolchain written by an older release shadows every fix a newer one ships — and it fails silently, because the shadowed script is old enough to reject a subcommand the new app calls.

That is exactly what happened: library_catalog.py index did not exist in the August 1st copy, so the app asked for a library index on every launch, the script printed its usage, exited 2, and four days passed with a 200-plugin index nobody could explain.

• Compare a version stamp written at package time, never an mtime. Every path here is a copy and a copy rewrites mtimes. package.sh writes Contents/Resources/tools/rack/VERSION (version, then packaged-at); choose_toolchain() prefers the bundle only when it is strictly newer, so an equal stamp leaves the installed copy in charge and hand-editing it still works for development. An unstamped directory is the oldest thing there is, which is what every pre-stamp machine looks like.
• install_toolchain.sh must not strip that stamp. It rsync --deletes from its source, and a source checkout has no VERSION — so a developer's install would make the destination look older than the release and lose to it forever. Exclude VERSION when the source lacks one.
• Anything spawned in the background must record its exit status, somewhere the app reads. library_index_command() writes runs/library-status; the settings row turns that into "the refresh failed (exit 2), see …". A log nobody opens is the same as no report at all.

The app must be able to say what it is

package.sh --version named the .pkg and nothing else, so an installed 0.12.7 answered CFBundleShortVersionString 0.11.0 and 12.6 was indistinguishable from 12.7 on the machine. Stamp the version into the app and all three plug-in bundles (staged copies, before signing, so the signature covers it), then read it back out of the expanded package and refuse the release when it disagrees.

Ship a details surface with it. The field that matters most is the live toolchain path — had it been visible, a day of shadowed fixes would have been obvious in seconds. Version, packaged date, that path and its stamp, index count and age, Rack SDK location, and whether a VCV sign-in was found — never the token. A Label cannot be selected with a mouse, so a Copy button is the affordance; promising selectable text you do not have is the same kind of claim as an installer promising modules it lacks.

The settings pane does not clip, it collapses

Forge's settings card is a fixed 660 tall and the Permissions pane was not a scroll view. A pane taller than the card therefore did not clip — flex shrank whatever it could, captions collapsed to zero height, and every row drew on top of the one below it. Adding one product row made it unmistakable; it was already true of the built-in ones.

There is a second, independent cause worth knowing anywhere in Pulp: Yoga's measure callback asks a Label for its INTRINSIC width first, and a paragraph's intrinsic width is the whole thing on one line — so the height it measures is one line however narrow the label is bounded to. Reserve the real height explicitly with label.measured_height(bound) when you bound a multi-line label's width.

Present but unreachable is the most expensive defect class

Every one of these was a finished feature that behaved exactly like a missing one, and each was found by a user rather than a test:

Before claiming a capability works, trace it end to end from the surface the user touches. "The function exists" is not the claim being made.

The most recent entry is the subtlest, because every part of it behaved as specified: naming something did not fetch it. A prompt that named a maker expanded correctly into the model's brief, the model reached for that maker four times, was told each time that the plugin was not installed, substituted something else and said so honestly, and the count at the end read "0 module(s) drawn from this maker". The download machinery worked and had one trigger — a missing-CAPABILITY gap in preflight. A mention was not a trigger at all. If a surface lets somebody name a thing, naming it has to guarantee it is there before the thing that consumes the name runs. Bound what a category fetches (a maker is a preference, so rank by the request and cap it) and keep the exactness for what was named outright.

Two design rules fell out of this:

• One resolver per resource. Three components disagreeing about where the Rack SDK lived meant fetching it correctly still failed. If you add a second fetcher, look for the first one — it usually exists.
• Cost is friction, not price. Ranking modules by premium put everything free ahead of the 70 premium plugins the user had bought. Sort by what it takes to obtain a thing (installed → owned → free → unavailable), and never conflate "premium" with "not owned". Downloading something already paid for is not a purchase.
• Measure entitlement, never infer a tier. The library API exposes what an account owns and not its subscription level. Ownership is the better signal anyway: it is correct for a user on any plan who bought modules individually.

The bundle is read-only, and a copy of it inherits that

The generator's working copy is laid down by install_toolchain.sh from Contents/Resources, and an installed bundle is root-owned and sealed. Two consequences, both of which stopped the first build on a genuinely clean machine dead:

• rsync -a reproduces the source's modes. The module pack arrived r--r--r-- and the panel emitter died on PermissionError: .../res/ATT.svg — after installing everything and verifying nothing. A copy that is going to be rewritten must be made writable explicitly.
• macOS's rsync is openrsync, which ACCEPTS --chmod and ignores it. No error, no warning, not one mode bit changed. Set the modes afterwards with chmod -R u+rwX, by a tool that does what it says.
• A failed first install leaves an unwritable tree that every later attempt also fails on, so reclaim the destination (chmod -R u+rwX "$DEST") before copying.

Simulate this by making the staged bundle read-only (chmod -R a-w) before testing the install path. A writable copy of the app tests nothing.

A test program in the module pack breaks the behavioural gate

examples/forge-modular/src/ holds the modules AND test_portmap_merge.cpp, which has its own main. It compiled into the plugin dylib harmlessly for months. The behavioural gate links those same objects beside its main, so every module build ended in duplicate symbol '_main' — three attempts, three model calls, several minutes — and the gate had therefore never passed for any generated module.

Two rules fall out of it:

• Anything that links the pack's objects must exclude the standalone programs. generate.py's sources() and the CMake glob both skip _* and test_*.
• Print the tail of a link failure, not the lines containing error:. The only error: line a linker emits is "linker command failed", which says nothing; the symbol is in the lines above it. The message read as a blank refusal for exactly as long as that filter existed.

Signals that have lied

• Exit code 0 from a backgrounded launcher means the launcher exited, not that the work succeeded. Wait on the process: until ! pgrep -f "<cmd>"; do sleep 10; done
• grep -qF under set -o pipefail exits on first match, SIGPIPEs the upstream command, and fails the pipeline — so a binary that does contain the marker is rejected for containing it. Count instead; it drains the stream.
• find … | head -1 in Contents/MacOS returns libwgpu_native.dylib, which is copied in beside the executable and sorts first. Prefer the file named after the bundle.
• zsh aborts the whole command when any glob matches nothing, so ls a/*.x b/*.x reports nothing found even when a/ has matches. That produced a false "the Rack modules are missing" conclusion.
• A test satisfied by nothing happening. A cache test asserting "fewer than 16 builds" passed while measuring zero activity. Assert the work happened (REQUIRE(cache.stats().builds > 0)) before asserting it was cheap.
• sample and other profilers fail silently over SSH (TCC is per-process).
• A CRASHED SUBPROCESS READ AS A VERDICT. The audibility gate segfaulted loading third-party Rack plugins. returncode != 0 was taken to mean "this patch makes no sound", so six generations in a row ended "gave up after 3 attempts" with an empty explanation and nothing anywhere saying a process had died. A negative return code is a signal, not an answer: check returncode < 0 separately and say which signal and what it was loading. The same shape applies to any gate that shells out. Naming it is only half. The retry context still said "structurally valid but SILENT when run" whatever had happened, so the model was sent to fix a fault nobody had measured and the patch was discarded at the end anyway. A check that could not run must not feed the verdict path at all: keep the artifact, say the doubt out loud.
• A HARNESS THAT STANDS IN FOR A FRAMEWORK MUST DO WHAT THE FRAMEWORK DOES, IN ITS ORDER. Both of the gate's crashes were that, and both were found from a real backtrace (~/Library/Logs/DiagnosticReports/*.ips, or lldb with settings set target.env-vars DYLD_LIBRARY_PATH=… — the env var is stripped from a debugged process, so a run under lldb otherwise dies in dyld and looks like a different bug):
  • EXC_BAD_ACCESS at 0x10 in a module's constructor. APP is rack::contextGet() and is null until something calls contextSet(); Context::engine sits at offset 0x10, so any module reading the sample rate while being built dies. Bogaudio's base module does it for all 111 models.
  • EXC_BAD_ACCESS at 0x0 in a module's process(). A constructed module is not a running one. Rack sends onSampleRateChange then onAdd before it ever calls process(), and modules allocate their DSP buffers there — CV funk's Alloy sizes a delay line in it, so the harness read through a null pointer with a zero ring mask. Fixing the first uncovered the second, and each survives the other's fix, so one plugin is not a proxy for the other in a regression test.
• The staging output directory may be left read-only. Testing the install path with chmod -R a-w (above) leaves --out unwritable, and the next package.sh run dies in a wall of rm: Permission denied that reads like a packaging bug. chmod -R u+rwX "$OUT_DIR" before removing it.

Every test ships broken-on-purpose once. A test that has never failed has not been tested. State the mutation and its result when reporting.

Installer text

• Do not hard-wrap the licence/Read Me text. macOS rewraps it to the pane width and pre-wrapped lines come out ragged. One long line per paragraph, blank lines between, indented list items preserved.
• Quote every path you print. App names contain spaces, so /Applications/Forge Modular.app/... unquoted is read as far as Forge and reports "no such file or directory" — which reads as a missing uninstaller rather than a quoting mistake. Show the quoted form and say why.
• Say where the uninstaller is. "An uninstaller ships inside the app" without a path means nobody runs it.
• Do not promise an interaction you did not build. The pane said the app "will offer to download… when you say yes"; what exists is announce-and-fetch — fetch_sdk.ensure() says what it is about to do and then does it, and nobody is asked. The sentence now describes that. Softening copy is only half the fix, because it is correct only while the code stays announce-and-fetch: check_installer_promises() in tools/rack/test_patch.py asserts BOTH halves, driving ensure() with stubs to prove the order is announce-then-fetch and then scanning the pane for consent wording. Build the prompt later and that check is what says the pane has to be rewritten with it. Note the setting that governs it, auto_fetch_sdk, has no control in Settings (settings_choices() exposes module source, downloads, the time limit, the index refresh and the about pane, and not this one) — so do not write copy telling anybody they can switch it off there.
• Mark the payload others depend on required (enabled="false" selected="true"). For Forge Modular that is the app, because the Rack modules and the uninstaller live inside its bundle.

Licence boundaries that must not drift

• The Rack SDK is GPLv3 and VCV's. It is never in a shipped artifact. The user's machine fetches it, which is not redistribution.
• A module built against that SDK inherits GPLv3. Fine for personal use; the licence follows if it is distributed.
• The .vcvplugin is the only artifact linking the SDK. Keeping it a separate payload from the app preserves that boundary; merging them blurs it.

A Forge shell is a bigger view tree than a plugin editor

Two per-frame walks in Pulp were written for a plugin editor's tens of views and each asked libc++abi a question per node: needs_continuous_frames tried six dynamic_casts (three of them through multiple inheritance, so each miss walked __vmi_class_type_info), and the host-parameter pump asked "is this a DesignFrameView". Sampled on an idle Forge Modular window on an M3 Ultra those two were ~29% of the process — more than the Skia drawing and the Yoga layout they were gating. Both are now a virtual call and a bool.

The general rule: anything that runs over the whole view tree every frame must not use RTTI. A shell's tree is thousands of views, and 120 Hz multiplies whatever a node costs by four or five orders of magnitude.

Measuring it: no PULP_TRACING in these builds, so sample <pid> 10 -f out.txt. It fails silently over SSH (TCC is per-process), so run it from a window on the machine. Read it by attributing each libc++abi run to its nearest non-libc++abi ancestor, or the cost hides inside dyn_cast_slow where no Pulp symbol appears. Thread count is not a smell by itself: an idle window here was 20 threads, all accounted — main, CVDisplayLink, NSEventThread, four CoreAudio (caulk* + IOThread.client, because the standalone opens a device), eight Dawn AsyncWorkerThreadPool workers, one BackgroundJobService, four libdispatch workers serving Metal/CoreAnimation queues.

Notes for future Forge apps

Sequencer-specific notes go here when that build starts. The parts above are app-agnostic: the seam, the staging, the identity check and the wiring-gap class apply to any Forge shell. What tends to differ per app is the generator's external toolchain (Forge Modular needs the Rack SDK and a C++ compiler) and what "the runtime" means for it — enumerate that first, because it is the thing most likely to be left out of the installer.

Forge evidence in DPR trials

The A4 DPR matrix requires the exact Forge Modular native shell, not a Pulp example standing in for it. Use a fresh Forge worktree, record both Forge and Pulp SHAs, and preserve binary identity with the result artifacts. Exercise the same logical fixture at each requested DPR and mode; a rebuilt or substituted payload is a different trial. Do not publish a scale-policy conclusion from a planned/synthetic run or before the A2T trace and A3 budget dependencies exist.

Register each real Forge shell with tools/scripts/gpu_dpr_runner.py as its scenario-specific executable adapter. The adapter receipt must bind the exact Forge/Pulp SHAs and binary; missing native, DAW, or browser legs stay explicit dependencies and must not be replaced with a Pulp example.

For terminal v2, initialize only after the fixed A2T, A3 product-policy, and A3 runtime receipts are terminal on protected Pulp main. The runner snapshots the actual Forge adapter, exact executable product, producer/product/fresh-process identities, and all eight artifacts for each original/repeat cell. Do not pass a Forge-authored manifest, draft result, or disposition to finalize-v2; it derives them from the terminal dependencies and 168 accepted nonce receipts. Protected publication is a separate fixed-path Git-blob verification step.

The Forge producer must also use the shared instrument-validity contract: each metric declares measured/derived/unavailable provenance; timer samples carry a detectable known-extra-work calibration; the expected logical point/target comes from the frozen scenario while Forge reports the actual event and hit; and two same-content captures supply numeric similarity, text, and stroke oracles. Preserve any older self-satisfying or constant-valued receipts as SUPERSEDED/NONCOUNTED, not as partial terminal cells.

Forge GPU-health canary

Forge shells may reuse ControlGpuHealthViewAdapter by supplying their real back-buffer capture and GpuSurface callbacks. The provider is Pulp-owned and bounded, but a Forge canary still needs exact Forge/Pulp binary identity and a correlated trace. Forge is a visible role: declare native-compositor-presentation and supply an independent compositor timestamp; capture completion is only an upper bound. Missing causal instrumentation stays nullable and named rather than becoming fabricated Vellum evidence. Leave an unratified startup budget unverified; never substitute a Pulp example for the requested Forge product. Exercise the receipt-producing seeded blank control documented in docs/validation/gpu-first-visible-a3-acceptance.md before accepting a real-machine result.

For the A3 terminal campaign, invoke the exact Forge-shell adapter through gpu_first_visible_a3_campaign.py run-role --role forge. Its identity JSON must use plugin_format: standalone and bind exact Pulp and Forge revisions; a DAW plugin format or Pulp example is rejected even if its pixels match. The adapter owns the real 10-cold/10-warm shell lifecycle and native presentation source. The runner owns the immutable adapter/budget snapshot and closed artifact validation. Preserve timeout, unavailable shell delivery, and missing post-Vellum instrumentation as explicit dependencies rather than substituting another executable.

For the executable handoff, use gpu_first_visible_a3_external_adapter.py with the checked-in gpu_first_visible_a3_forge_producer.py. Configure clean exact-revision Pulp and Forge roots, the exact app executable for both product and host identity, the product-specific lifecycle driver, checked-in source-bound trace analyzer, embedded-build verifier, and the digest-bound Forge build attestation/receipt. Configure a reviewed Pulp- or Forge-owned source-build driver as well; it must reproduce the executable and complete .app tree from clean exact-revision sources without receiving the measured path. Both trees must bind CFBundleExecutable, CFBundleIdentifier, and CFBundleName to the request. The producer pins and rehashes those inputs; the driver must resolve to reviewed source at the declared clean Pulp or Forge revision. It requires 10+10 reopen/reset rows, producer-observed live executable/start identities, independent native presentation, terminated owned Forge PIDs, and named replay on the challenged trace-host PID. A missing driver, provenance, dirty/wrong source head, different Forge build, or Pulp demo stays nonterminal or fails closed.

The exact Forge executable must carry the canonical embedded build marker. Its role campaign cost includes active product spans, but terminal A3 also requires the separate four-state pre-change/compile-out/compiled-in-idle/active product overhead receipt with zero xruns and audio-thread trace events. Keep the analyzer's structural unverified result distinct from the campaign budget verdict; neither an offline A2T classification nor a passing Forge role waives the overhead control.

Run that control through gpu_first_visible_a3_trace_producer_overhead.py collect-state with the exact Forge executable and reviewed Forge lifecycle driver. Each state requires 55 live challenges; the source-built .app and measured executable identities remain exact, and idle/active reuse identical bytes. Active binary Perfetto replay covers both the health producer and the complete b4ba exact 20-signature state/render/js inventory from the A3 guide. Acquire/submit/present are mandatory; all other signatures are counted and an unobserved one stays not-covered, not zero-cost. Require zero xruns/audio-thread producer events. A direct driver result or Pulp example is nonterminal. Also pin the candidate-relative state_build_driver. The collector exports the exact source row and default-deny rebuilds it without the measured .app, ambient build output, or network, then byte-compares the rebuilt executable and checks its tracing sentinel. Retain source archive, closed build request/receipt, product, logs, and toolchain snapshots; a pre-existing Forge bundle cannot satisfy this proof.

Forge A3 v2 campaigns must retain producer/sample provenance and exact trace/analyzer artifacts. Terminal validation reruns the prepared analyzer and binds the result to the Forge revision, role, host process, evidence cohort, categories, and capture completeness.