@toolkit-cli/toolkode 1.9.0 → 1.12.0

package/README.md

@@ -35,6 +35,52 @@ toolkode --prompt "fix the failing tests"
 toolkode serve
 ```
 
+## What's new in v1.12 "Cortex"
+
+**Cross-tool memory import.** Toolkode now reads memories from other AI coding tools automatically — matching their path resolution, 200-line/25KB truncation, and git-root-aware directory layout. Your memories show up in every session without importing anything. Switching tools shouldn't mean losing your brain.
+
+**Permission modes.** Choose how toolkode handles tool approvals per session: Default, Accept Edits, Plan, Auto, Bypass, or Don't Ask. Switch on the fly with Ctrl+1-6 in the permissions dialog, or set a default in `toolkode.json`. Auto mode classifies each tool call by risk level and auto-approves safe operations. Plan mode blocks all writes until you switch back.
+
+**Navigator sidebar.** Press `<leader>1` to open a left-side navigator showing your recent sessions and active pairs. Collapsed mode shows just a thin gutter; expanded gives you a full session list with counts.
+
+**`/pair` command.** Start, approve, reject, and manage pair sessions right from the prompt. `/pair start fix the login bug` kicks off a builder/reviewer loop without touching the dashboard. Onboarding messages guide you if you forget the syntax.
+
+## What's new in v1.11 "Pair"
+
+**Your AI agents now review each other's work.** Pair Mode introduces a builder/reviewer loop where two LLM agents collaborate on a task — one builds, one reviews, and the loop continues until the reviewer approves or you escalate. Think code review, but both participants are tireless and available at 3am. Unlike certain other tools, ours was built on purpose and not accidentally pushed to npm at midnight.
+
+**Live model reassignment.** Mid-task, your builder hitting rate limits? Provider lagging? Model not cutting it? Open the switch dialog (Ctrl+P > "Reassign pair builder"), pick a new provider and model, choose when to swap (next iteration, after current step, or immediately), and the new model picks up right where the old one left off. The system even suggests upgrades automatically when it detects repeated failures — because your AI should be smart enough to know when it needs a smarter AI.
+
+**97 tests. 24 blocked command patterns. 0 trust in LLM output.** The pair validator runs every command the builder claims to need, but it won't let an LLM `rm -rf /` your machine, `sudo` anything, pipe curl to bash, or sneak in a `python -c` backdoor. We block 24 dangerous patterns including exotic interpreters like awk, php, lua, and osascript. The builder is semi-trusted. Semi.
+
+**Sidebar toggle.** Click the `>>` caret in the top-right corner to hide the sidebar. Click `<<` to bring it back. Revolutionary, we know. But hey, while Anthropic was busy leaking Claude Code's source to the entire internet, we were up at 3am making sure a chevron rendered correctly in a terminal. Priorities.
+
+**Agent inspector upgrades.** The sub-agent modal now adapts to your terminal width instead of overflowing, and you can press W to watch an agent's live session. Because staring at a loading spinner is so 2025.
+
+> **A note on timing:** We apologize for any rough edges in this release. We had planned a more leisurely ship date, but when Anthropic accidentally published Claude Code's entire codebase to npm for the world to browse, we figured we should probably stop polishing and start shipping before someone else builds what we've been building. Nothing like a competitor's security incident to motivate a release. Thanks for the adrenaline, Anthropic. We'll send coffee.
+
+### Pair Mode at a glance
+
+```
+/pair start "add input validation to the API"
+
+Builder: implements the change, writes tests, produces HANDOFF.md
+Validator: runs the test commands, checks for TODOs, verifies artifacts
+Reviewer: audits the work, scores it, approves or rejects with reasons
+Loop: rejected work goes back to builder with REWORK.md instructions
+
+Press m/n to reassign builder/reviewer models on the fly
+Press w to accept system suggestions when the model is struggling
+Press p to manually approve, x to reject
+```
+
+### Fixes
+
+- Agent inspector no longer overflows on terminals narrower than 140 columns
+- Rejection toast now shows as warning (amber) instead of error (red) — rejecting is an action, not a failure
+- Event log text is truncated to prevent layout breakage
+- Pair panel in sidebar shows builder/reviewer model assignments
+
 ## What's new in v1.5
 
 **Your conversations last longer and remember more.** We completely rebuilt how Toolkode manages context. Long sessions no longer lose track of what you asked for, what files you were editing, or what errors you hit. When the context window fills up, Toolkode summarizes the older history while keeping your recent work intact — instead of compressing everything at once.

package/package.json

@@ -12,9 +12,12 @@
   "scripts": {
     "postinstall": "bun ./postinstall.mjs || node ./postinstall.mjs"
   },
-  "version": "1.9.0",
+  "version": "1.12.0",
   "license": "SEE LICENSE IN LICENSE",
   "optionalDependencies": {
-    "@toolkit-cli/toolkode-darwin-arm64": "1.9.0"
+    "@toolkit-cli/toolkode-linux-arm64": "1.12.0",
+    "@toolkit-cli/toolkode-linux-x64": "1.12.0",
+    "@toolkit-cli/toolkode-darwin-arm64": "1.12.0",
+    "@toolkit-cli/toolkode-windows-x64": "1.12.0"
   }
 }