Akisolu

What they built

Fixed the zh-CN README rendering after catching that the issue itself pointed at the wrong file, then became the resident Windows tester across five straight releases: the first mission report exposed that the Quick Start install was broken in every clean environment, the v1.1.17 report traced doctor false positives and cross-tool memory gaps back to the test instructions themselves, and the v1.1.18 report caught two real Windows-only crashes invisible to CI, a PowerShell quoting bug in the install script and a type error in the crusher shim, correctly diagnosing the quoting failure on sight; then authored the fix PR for both crashes, with a shared path-result normalization helper covering install and uninstall alike and a dedicated script extraction that survives PowerShell 5.1 quoting, tested end-to-end on a real Windows machine before and after review; the v1.1.20 mission then validated the non-breaking upgrade path on Windows end to end (global npm update, auto-update with every target preserved, in-place memory protocol migration from v2 to v4, doctor clean), the first report of the series with nothing left to fix; the v1.1.21 mission then put the state.db consolidation script through its first run on real, aged Windows state, dry run and apply, and her report caught that the doctor kept flagging the merged copies and that its hint omitted the arguments the script requires, both filed with her name on them (#1389, #1390), along with the npm EBUSY lock a running AI tool holds on the package folder during a global update; her follow-up checks the same night found the Antigravity MCP panel empty behind a 0-byte config file and an agent reading a plain Markdown state file it had written itself, which became #1395, plus #1394 (a DEP0190 warning present in her reports since July) and #1396 (OpenCode freezing on Windows), which she took on next; the five fixes from her report (#1391, #1392, #1393, #1397, #1398) landed on main the same night, and the sixth (#1401, the protocol bump that carries the corrected memory text to installs already in place, plus a doctor check for plain-text state files) closed #1395 the next morning; #1396 then took four rounds of bisect on her Windows and two log sets sent through a private channel, and the server-side one carried the mechanism: the installer had been copying the whole egc-universal package into OpenCode's config directory, where OpenCode loads every tool file on every request, so the freeze was EGC's after all, fixed in #1411 with the retirement of everything earlier installs had written there, and the same bisect exposed that the MCP registration had never reached OpenCode at all (#1405, fixed in #1406); she confirmed the repair on her own machine before the fix landed, the DCO Comment job that ends quietly when the triggering run has no result artifact (#1435), and the install.ps1 test helper that tells a timeout from a signal and retries a cold PowerShell start once instead of failing mute (#1447). The v1.1.22 two-tools mission (#1381) then proved on her Windows machine that a decision saved in one tool reaches the other, and her follow-up traced the two gaps it exposed: an OpenCode session that opened without the project memory (#1529) and a protocol that routed "remember this decision" to a tool whose rows never come back at session start (#1524); she fixed the second herself in #1531, which sends a decision to update_state so the next session sees it, across every copy of the protocol, and the first herself in #1530, so a new OpenCode session opens with the project memory again, the restore starting a Node found on an absolute PATH entry and never one from the project folder. Area steward for Windows and OpenCode since 14 September 2026, the first steward of the project; then took the OpenCode memory protocol issue of the hackathon review (#1674) and shipped #1724 through four review rounds in one evening: the protocol lives in egc-memory.md in the OpenCode config directory and is listed in the instructions of opencode.json under the same rules as the MCP registration, refused before the first write when the config leads outside the home, never editing an opencode.jsonc, and retiring the legacy file by its marker only once the new one is wired, with 67 bootstrap tests behind it and a linter lesson learned the honest way.

Joined Jul 19, 2026

Badges