git:20260908.e9ac98d to git:20260909.cac77c1
23 added, 15 removed. Audit A to A.
---
name: failure-desktop-runtime-env-name
- description: Diagnose strict desktop staging failures caused by confusing the model-server runtime variable with the packaging binary variable. Use whenever desktop/prepare.mjs rejects a required llama-server runtime.
+ description: Diagnose desktop staging failures caused by confusing the engine-source variable with model-runtime variables, or by assuming an AIDE_LLAMA_SERVER_BINARY override exists. Use whenever desktop/prepare.mjs skips engine staging, when a staged stack reports a missing llama-server runtime, or when a skill claims AIDE_REQUIRE_MODEL_RUNTIME or AIDE_LLAMA_SERVER_BINARY drive staging.
---
- # Desktop Runtime Environment Names
+ # Desktop Runtime Environment Names (verified against desktop/prepare.mjs, 2026-09-10)
- ## Contract
+ ## Contract (READ desktop/prepare.mjs FIRST — do not trust skill memory)
- - `AIDE_REQUIRE_MODEL_RUNTIME=1` enables the fail-closed staging requirement.
- - `AIDE_LLAMA_SERVER_BINARY` supplies the verified executable to
- `desktop/prepare.mjs`.
- - `AIDE_LLAMA_SERVER` is a separate serving/runtime setting and does not supply
- the desktop staging input.
+ - `AIDE_ENGINE_SOURCE` (default `E:\llama-cpp`) is the ONLY engine input to
+ `desktop/prepare.mjs` (line 31). Engine files are staged via an explicit
+ allowlist (line 39): `llama-server.exe` + `llama-server-impl.dll`, `llama.dll`,
+ `llama-common.dll`, `ggml-base.dll`, `ggml.dll`, `ggml-rpc.dll`,
+ `ggml-rpc-server.exe`, `libomp140.x86_64.dll`, `mtmd.dll`, plus `ggml-cpu-*.dll`.
+ - `AIDE_LLAMA_SERVER_BINARY` and `AIDE_REQUIRE_MODEL_RUNTIME` DO NOT EXIST in
+ desktop/prepare.mjs. Historical skills that name them are stale — correct the
+ skill, do not search for the variable.
+ - `AIDE_LLAMA_SERVER` is a serving/runtime setting, unrelated to staging.
+ - The `E:\llama-cpp\llama-server.exe` is a 9216-byte THIN LAUNCHER; the payload
+ lives in the sibling DLLs. Staging only the exe silently breaks model spawns.
## Procedure
- 1. Stop at the staging failure and read `desktop/prepare.mjs` instead of
- guessing variable names.
- 2. Verify the binary path exists and is the intended local executable.
- 3. Run strict preparation with `AIDE_REQUIRE_MODEL_RUNTIME=1` and
- `AIDE_LLAMA_SERVER_BINARY=<verified-path>`.
- 4. Run `desktop/verify-prepare.mjs` and record whether the runtime is staged.
- 5. Do not weaken strict mode or silently substitute the serving variable.
+ 1. Stop at "engine source absent - skipping engine staging": the env var is
+ `AIDE_ENGINE_SOURCE=<llama.cpp build dir>`. Verify the dir exists and
+ contains `llama-server.exe` AND its DLLs.
+ 2. After prepare, verify the staged copy: `desktop/resources/runtime/` must
+ contain `llama-server.exe` plus the allowlisted DLLs and `node.exe`.
+ 3. Run `desktop/verify-prepare.mjs` and `scripts/desktop-battery.mjs` (expect
+ 12/12) to prove a model-capable staged stack.
+ 4. Do not weaken strict mode, do not substitute the serving variable, and do
+ not copy `AIDE_LLAMA_SERVER_BINARY` into new scripts.