feat(adaptation): enhance platform detection and cache path handling …

[mikij]

…for OpenCode/KiloCode

What / Why / How

• I want to add new argument to CLI for upgrade so it will make OpenCode/KiloCode plugin development/testing much easier allowing it to update itself from local git branch. CLI should support local upgrade (context-mode upgrade local) which copies current git branch instead of cloning latest GH published version.
• Plugin needs to support new cache folder location. OpenCode silently changed (not documented) where is plugin stored inside of .cache folder (before .cache/opencode/node_modules/context-mode, now .cache/opencode/packages/context-mode@latest/node_modules/context-mode), so doctor or upgrade will not properly work for new versions of OpenCode/KiloCode. KiloCode also has a bug that uses OpenCode folder location instead of own. So `context-mode upgrade/doctor' will not work until that is resolved for KiloCode.

This fails in doctor:

■  Hook script exists: FAIL — not found at /Users/mickey/.cache/kilo/packages/context-mode@latest/node_modules/context-mode/hooks/pretooluse.mjs

Addition:

1. OpenCode now has SessionStart (via experimental.chat.messages.transform) — no longer "not yet available"
2. Both OpenCode and KiloCode now have high continuity (all 4 hooks working)
3. The AGENTS.md copy step is now optional rather than required for routing awareness

Affected platforms

• Claude Code
• Cursor
• VS Code Copilot (GitHub Copilot)
• JetBrains Copilot
• Gemini CLI
• Qwen Code
• OpenCode
• KiloCode
• Codex CLI
• OpenClaw (Pi Agent)
• Pi
• Kiro
• Antigravity
• Zed
• All platforms

Test plan

Checklist

• Tests added/updated (TDD: red → green)
• npm test passes
• npm run typecheck passes
• Docs updated if needed (README, platform-support.md)
• No Windows path regressions (forward slashes only)
• Targets next branch (unless hotfix)

Cross-platform notes

Our CI runs on Ubuntu, macOS, and Windows.

• If touching file paths, verify forward-slash normalization on Windows
• If touching hook paths, verify no backslash separators
• Use path.join() / path.resolve(), never hardcode / separators
• Use event-based stdin reading — readFileSync(0) breaks on Windows
• Use os.tmpdir(), never hardcode /tmp

[mikij]

@mksglu this is test run result from main

[mikij]

from next

[mikij]

Opened PR today. It is approved so hopefully it will be merged soon.

[mikij]

It was merged a several minutes ago. But it will be part of some 7.2.34+ version of KiloCode (latest is 7.2.33 and it was released just before my commit was merged)

[mksglu]

Mickey — real work in here. Thank you. Let me give you a piece-by-piece take so you can see exactly what we're keeping and why.

What you got right (adopting into the architectural PR):

1. OpenCode/KiloCode cache path fix — the packages/context-mode@latest/ cache layout change is a real, current bug. Users on OpenCode and KiloCode are hitting it. We're adopting your fix verbatim.
2. OpenCode SessionStart via experimental.chat.messages.transform — this is the correct hook surface for OpenCode and we didn't have it before. We're adopting this hook integration.

What we're not adopting (and why — happy to discuss):

3. upgrade local subcommand — redundant with the existing symlink workflow. CONTRIBUTING.md already documents npm link for local development, and adding a CLI subcommand for it adds surface area without solving a new problem. Open to revisiting if you've hit a case where the symlink workflow falls short.
4. tool-naming.ts + routing-block.ts extraction — premature abstraction. Both modules have a single consumer right now and no isolated tests. Extracting them now locks in an interface before we know the second use case. Happy to extract these the moment a second consumer appears.
5. Bundle artifacts in the PR — these should be CI-generated on release. We've intentionally kept hand-built bundles out of the source tree to avoid diff noise on every commit.

Why we're not merging this directly: your PR ran into the same root cause as #379 and #370 — v1.0.100's "Unified Persistent Memory" hardcoded ~/.claude/ across searchAutoMemory, extract.ts, prior-session loading, and ctx_search timeline. Rather than merge 3 platform-specific PRs (Codex, Qwen, OpenCode/KiloCode), we're doing one architectural fix: HookAdapter gets getConfigDir(), getInstructionFiles(), getMemoryDir(), and every adapter (all 14) overrides with its own conventions. Your OpenCode pieces drop into that structure cleanly.

You'll be Co-Authored-By on the architectural commit and called out in the v1.0.101 release notes. When the PR lands this week, I'd love your review — especially on the OpenCode adapter to confirm we got the cache path and SessionStart wiring right.

Closing in favor of the unified fix. Thanks for digging into this.

[mikij]

@mksglu Have a question about this function:

function getPlatform(): AdapterPlatformType {
  const [ kiloEnv ] = PLATFORM_ENV_VARS.filter(p => p[0]==="kilo");
  const [platform, vars] = kiloEnv;
  for (const _var of vars){
    if (process.env[_var]) {
      return platform;
    }
  }
  
  return "opencode";
}

I think this is more robust way to check for KiloCode platform and as you did not mention this, I want to ask did you miss it or also did not want to include for some reason?

[mksglu]

@mksglu Have a question about this function:

function getPlatform(): AdapterPlatformType {
  const [ kiloEnv ] = PLATFORM_ENV_VARS.filter(p => p[0]==="kilo");
  const [platform, vars] = kiloEnv;
  for (const _var of vars){
    if (process.env[_var]) {
      return platform;
    }
  }
  
  return "opencode";
}

I think this is more robust way to check for KiloCode platform and as you did not mention this, I want to ask did you miss it or also did not want to include for some reason?

checking

[mksglu]

@mikij — you were right. Landed the full fix on `next` as commit 4e75d11 with you credited as Co-Author.

What ended up in the fix

Your point about `getPlatform()` (src/opencode-plugin.ts:80-82) was the surface. Pulling the thread surfaced a deeper problem: the canonical PLATFORM_ENV_VARS list itself had drift — half the entries were unverified placeholders, several platforms were entirely missing, and the plugin paradigm's fallback to `opencode` was blind (didn't actively check OPENCODE env vars either).

Audited every entry against the platform's own runtime source code. Highlights:

Removed / corrected (no source-line evidence):

• kilo: dropped bare KILO — Kilo-Org/kilocode packages/opencode/src/index.ts:140 only sets KILO_PID, never bare KILO.
• jetbrains-copilot: dropped IDEA_HOME, JETBRAINS_CLIENT_ID — neither verifiable in any JetBrains source repo.
• qwen-code: dropped QWEN_SESSION_ID — 0 hits in QwenLM/qwen-code.
• openclaw: removed entirely from the env-var tier — runtime never sets OPENCLAW_HOME/OPENCLAW_CLI. Detection still works via the existing ~/.openclaw/ config-dir tier.

Added (with verified source citations):

• antigravity: ANTIGRAVITY_CLI_ALIAS — verified in google-gemini/gemini-cli packages/core/src/ide/detect-ide.ts (Google's own canonical IDE detection map).
• zed: ZED_SESSION_ID, ZED_TERM — verified in zed-industries/zed crates/terminal/src/terminal.rs::insert_zed_terminal_env().
• pi: PI_PROJECT_DIR — confirmed by our own consumers at src/pi-extension.ts:154 and src/server.ts:153.

Reordered fork pairs so collision detection works:

• kilo before opencode (Kilo sets OPENCODE=1 because it's an OpenCode fork).
• cursor + antigravity before vscode-copilot (both inherit VSCODE_PID).

getPlatform() itself rewritten exactly along your line — iterates PLATFORM_ENV_VARS instead of hardcoding any env var. Symmetric: actively checks BOTH platforms' env vars instead of blind fallback to opencode. Filtered to kilo+opencode so a stray CLAUDE_PROJECT_DIR can't leak into the plugin's platform decision.

function getPlatform(): AdapterPlatformType {
  for (const [platform, vars] of PLATFORM_ENV_VARS) {
    if (platform !== "kilo" && platform !== "opencode") continue;
    if (vars.some((v) => process.env[v])) {
      return platform as AdapterPlatformType;
    }
  }
  return "opencode";
}

Tests

• tests/adapters/detect.test.ts: removed 5 broken assertions (for unverified env vars); added 4 for new platforms + a KILO_PID + OPENCODE both set → kilo wins fork-collision assertion.
• tests/adapters/detect-config-dir.test.ts: priority chain rewritten as fork-collision tests (KILO/OPENCODE, CURSOR/VSCODE, ANTIGRAVITY/VSCODE, CURSOR/CODEX).

451/451 adapter + plugin tests pass on next. Typecheck clean.

Thanks again for surfacing this — both the immediate bug and the broader audit it triggered. Happy to credit you for any follow-up contributions.