Willian Phaiffer Cardoso

O que construiu

• ✦ Opened the v1.1.22 Linux testing with a model report (#1377): both install paths, npm and the clone, run clean in a fresh temporary HOME with identical doctors, the stray state.db fix of #1450 confirmed and the portable sqlite3 notice absent on Ubuntu 26.04 with Node 24;
• ✦ the report surfaced three real findings, the one that mattered most fixed the same day with his name on it (#1729, the egc doctor Summary now prints on a bare install so the done criterion is reachable) and the other two, the telemetry prompt ordering and two overstated egc init lines, tracked in #1730 and taken by him in #1736;
• ✦ then, on the same day, shipped his first three code PRs, each one a finding of the October review: getDartDeps reads only the dependency sections of a pubspec (#1733), the query_history schema default matches the tool description (#1734), and the 32 Antigravity skills whose description or origin carried an unquoted colon load as valid YAML again, with a test that parses every SKILL.md strictly (#1735);
• ✦ then made the fake Guardian CLI of the hook tests return the same response fields as the real engine, with a contract test that runs the same inputs through the fake and the real CLI and requires identical output (#1739);
• ✦ and took the Crush install target (#1491), mapping where it reads rules and skills, its hook surface and its MCP config on Linux, including the hook-shape trap that makes Crush refuse to start, before writing the adapter;
• ✦ then, on October 7, landed seven PRs in one day, each one verified locally before the merge: the installation guide says what auto-intuition and egc watch really need (#1751, closes #1666), direct tests for ModelResolver with full line coverage and more prompt-builder cases (#1758, closes #1691), the Crush install target itself, with the skills in the shared ~/.agents/skills root, the hooks and MCP servers in crush.json and the memory protocol in CRUSH.md, moving the supported tool count to 21 (#1744, closes #1491), the cost-tracking skill and /cost-report reading the files EGC really writes (#1757, closes #1701), the honest telemetry prompt, Token Crusher PATH status and dashboard launch line (#1736, closes #1730 and #1377), the EGC 2.0 pages marked as an unimplemented proposal (#1759, closes #1703) and the Junie row of the integration tiers following the official guidelines page (#1752, closes #1655);
• ✦ then, the next day, after two rebases, the OpenHands row documenting the global hooks file OpenHands reads when the project has none, the precedence of the project file and the fact that the two never merge, with the install comments aligned and a spec test that keeps the row and the modules honest (#1756, closes #1656);
• ✦ and, the same morning, the subsystem map and the runtime README describing main again, with the rows of files that no longer exist removed, agent.yaml and src/llm classified as active, a Schemas table naming which validator or reader covers each schema, and a docs test that fails when a named path disappears or a runtime file goes undocumented (#1776, closes #1764);
• ✦ then every schema in schemas/ validated against its data in CI, with the runtime map schema rewritten to match what discovery.js writes, the two registry schemas nothing used removed and the spec index naming what validates each schema, two findings of the correction plan closed in one PR after seven review threads answered in a single round (#1778, closes #1702);
• ✦ and the integration tiers spec saying protocol v9 and naming the Qwen Code hooks the adapter really registers, with two tests that read the version and the hook plan from the code so the lines fail when the implementation moves (#1783, closes #1657);
• ✦ and the first PR of phase 3 of the correction plan, the Antigravity skills installed under config/skills, the one directory the IDE, Antigravity 2.0 and the CLI all read, with the old copies retired on upgrade and a skill the person keeps there never overwritten, four review rounds in one afternoon and the strict move adopted after an strace proof on the real CLI (#1786, closes #1705);
• ✦ and, the same evening, the two leftovers of that move: the June 2026 links under antigravity-cli/skills that point into the retired skills/egc namespace, removed by the generic legacy-link migration whether they still resolve or already dangle, never touching what they pointed at, and a managed copy whose source changed in the same release that moved it, now retired by the hash of what EGC actually wrote instead of the current source, so a real edit by the person still survives (#1793, closes #1789);
• ✦ and, the next day, the Antigravity rules installed flat under config/rules with a trigger frontmatter, the one shape the IDE, Antigravity 2.0 and the CLI all read, closing the gap where none of the 109 rules ever reached Antigravity before (#1796, closes #1668);
• ✦ and, the next day, the Antigravity agents installed as subagents under config/agents and .agents/agents, the shape Antigravity actually reads, with tool names mapped to Antigravity equivalents and gemini-adapt-agents.js retired (#1801, closes #1669);
• ✦ then, since Antigravity stops running workflows on 19 October 2026, the 77 commands installed as skills Antigravity invokes the same way, with the Claude Code-only frontmatter keys dropped on copy (#1804, closes #1706);
• ✦ and, in the same session, VS Code Copilot's personal skills moved from ~/.github/skills, which VS Code never reads, to ~/.copilot/skills, with a symlink guard, an unreadable-state warning and the old copies retired on upgrade (#1811, closes #1672);
• ✦ and the Zed skills moved from ~/.config/zed/skills to the shared ~/.agents/skills, the only global directory Zed actually reads, with sibling-aware retirement so uninstalling Zed never touches what Codex still records there, and crush-home.js refactored onto the same shared helper instead of its own duplicate (#1812, closes #1654);
• ✦ and, through the night of 10 October, four more fixes against each tool's own docs: the Trae memory protocol written to ~/.trae/user_rules, where Trae reads global rules, with the old MEMORY.md retired only once the new file is in place (#1824, closes #1676);
• ✦ the CodeBuddy protocol in ~/.codebuddy/CODEBUDDY.md and the Token Crusher rewrite returned as modifiedInput, the field CodeBuddy applies (#1831, closes #1675);
• ✦ the Codex protocol in developer_instructions, the key Codex reads, a hand-written value kept with the segment appended once, the legacy key removed only when it held EGC text alone and the rewrite parsed back as TOML before anything is written, proven on Codex CLI 0.161.0 (#1833, closes #1708);
• ✦ and the engineering-auditor shipped to Kiro and Trae, with Warp and Aider recorded in a new unsupportedTargets field whose reasons must carry the docs link and the read date by schema (#1834, closes #1679);
• ✦ and, the same morning, one table of every install path per tool, 230 rows over the 29 adapters with a status per row (documented with the tool's docs link, EGC's own files, or unverified with the reason), a contract test that plans every adapter in every variant and fails on a path no row covers or a row no plan reaches, and the spec page generated from the table, with seven follow-up issues filed from what the table surfaced (#1835, closes #1714);
• ✦ and, on 9 October, the Warp skills installed as .warp/skills/<name>/SKILL.md, the layout Warp invokes as /<name> (#1814, closes #1673), and the Amp skills in the shared ~/.agents/skills and .agents/skills Amp discovers instead of ~/.amp/skills (#1820, closes #1671);
• ✦ then, the next evening, the memory protocol files replaced in one step through a temporary file renamed over the real file, a symlink kept pointing where the person put it, a link chain followed to its last target, a dangling link through a missing directory refused, a new file keeping the creation mask and a read-only file refused instead of replaced, with Windows teaching the test that .. in a link destination resolves lexically there (#1844, closes #1832);
• ✦ and, the same night, EGC reaching Cursor only where Cursor reads: nothing global written for a tool that keeps its user rules in the account, the protocol carried by the project rule the cursor target already generates, the closed blocks of older installs retired from cursor.rules and from an EGC-only ~/.cursor/rules file with backups while anything of the person's own stays and is reported, and the undocumented beforeFileEdit event dropped from the hooks seed and retired from existing installs, with seven bootstrap tests and two hook tests that fail on main (#1849, closes #1677);
• ✦ and, the same night, the Guardian on Kiro CLI 3.0: a standalone hook document in .kiro/hooks owned whole by EGC, one PreToolUse hook on the shell matcher running the adapter that already blocks with exit code 2, idempotent apply, drift reported by egc doctor, removed on uninstall, with the 2.x agent-config entry kept and the mesh notice sharing the whole-file helper (#1850, closes #1680);
• ✦ and, in the small hours, the installers building the MCP servers under NODE_ENV=production, with devDependencies asked for only where there is a src directory to build and a test that runs the real install helper and build block on a copy of the tracked tree (#1852, closes #1686);
• ✦ and the Guardian build that clears its output folder after the catalog prebuild and before tsc, so stale outputs of deleted sources stop accumulating, proven by a test that plants orphans in a copy of the tracked server and checks every file left has a source (#1855, closes #1664);
• ✦ and egc gain --history printing each run in local time with its numeric offset while the ledger stays in UTC, tested under two time zones (#1859, closes #1659);
• ✦ and the EGC Custom Rules for Codex installed at ~/.codex/AGENTS.md, the only global file Codex reads, with the native .agents children planned one by one, a link or a file of your own left alone and the old ~/.agents copy retired on upgrade (#1861, closes #1839);
• ✦ and the agent catalog for CodeBuddy installed at .codebuddy/rules/egc-agents.md, where every rule is loaded as project memory, with generated context files retired on upgrade by the sha256 of their template so an edited copy stays (#1872, closes #1840);
• ✦ and Cursor custom commands installed as skills at .cursor/skills/<name>/SKILL.md, the way the Cursor docs now describe them, with disable-model-invocation kept so a command runs only when invoked, a bundled skill of the same name winning, a skill of your own never overwritten and the old .cursor/commands copies retired while they still hold the bytes EGC wrote (#1889, closes #1837);
• ✦ and the Cline hook path confirmed against the extension source at a pinned commit, the unverified row of the table now documented and held by a contract test (#1892, closes #1838);
• ✦ and, on 12 October, the Trae CN project files installed under .trae, where the CN edition reads them (#1894), the EGC servers registered in the Qwen Code settings.json (#1898), a resolver that warns when it falls back and checks the Vertex AI project and location (#1899), three new install targets, Grok Build (#1900), Kilo Code CLI (#1901) and MiMo Code (#1907, closes #1498), and tool-call arguments parsed the same way in OpenAI, Ollama and Cohere with the provider named in the error (#1906);
• ✦ and, from that night into 13 October, two more install targets, Pi (#1909, closes #1499) and Droid by Factory (#1911, closes #1493), an npm run preflight that reports the lint findings a branch adds (#1921, closes #1904), and mcp-register.js split into seven modules under 350 lines (#1928);
• ✦ and, on 14 October, the installer's apply step split into six modules under 370 lines, applyInstallPlan and the same API re-exported from apply.js with no behaviour change, the twelfth of the seventeen files of #1712 (#1944)

Chegou em 3 de out. de 2026

Selos