v1.18.28/1.18.29-unguarded (linux-arm64): every run crashes with TypeError in SystemPrompt.environment

[Axeb2b]

Repro (100% consistent)

Any invocation crashes before the LLM is ever called, e.g.:

./evil-opencode run "Reply with exactly: PONG and nothing else"
# -> {"name":"UnknownError","data":{"message":"Unexpected server error...","ref":"err_..."}}

Server log:

level=ERROR message=failed error="TypeError: undefined is not an object (evaluating 'a.name')"
cause="... at resolve (chunk-gckn5ywb.js:2:1659) at a (2:936) at map (native)
at SystemPrompt.environment ..."

The crash lands ~9ms after init count=147, before fff/watcher/location-services boot lines that the official build prints.

Affected / not affected

• BROKEN: opencode-linux-arm64 from v1.18.28-unguarded and v1.18.29-unguarded (sha256 verified, interpreter repointed for Termux glibc — binary starts and prints --version fine).
• WORKS: official sst/opencode v1.18.29 and v1.18.25 with the identical environment, config, skills, agents, commands (official returns PONG).
• An older evil 1.18.25 build worked in this same setup on 2026-08-31, so this looks like a regression in the unguarded builds between 1.18.25 and 1.18.28.

Ruled out (each re-tested, still crashes)

• ~/.agents/skills, ~/.claude/skills, ~/.config/opencode/skills hidden (all 147 inits pass; also fixed one dangling symlink, not the cause)
• commands/, agents/ hidden
• minimal config ({model only} via OPENCODE_CONFIG, no MCPs, empty project dir)
• --pure flag

Environment

• Termux (aarch64, Android), glibc prefix with repointed PT_INTERP, LD_PRELOAD stripped
• --version prints 1.18.29 correctly; only session runs crash

Suspect: the guardrail-stripping step changed more than prompt text, or prompt .txt edits desync from code that maps over prompt/structure entries (the minified trace maps 3 levels deep ending in a.name), while upstream's environment function only sorts references by name.

[M1noa]

+1