@kimbho/kimbho-cli 0.1.6 → 0.1.8

package/README.md

@@ -22,7 +22,7 @@ From a packed tarball:
 
 ```bash
 npm run pack:cli
-npm install -g ./kimbho-kimbho-cli-0.1.5.tgz
+npm install -g ./kimbho-kimbho-cli-<version>.tgz
 ```
 
 ## Usage
@@ -54,6 +54,8 @@ Useful shell commands:
 /plan build a coding agent
 /run scaffold a SaaS starter
 /resume --execute --max-auto-tasks 2 --max-repair-attempts 2
+/approve
+/deny
 /review
 /quit
 ```
@@ -67,6 +69,7 @@ build a landing page for a small storefront
 That prompt now runs through the agent execution path, streams live task and tool activity while it executes, and then prints the saved session summary instead of returning a plain chat answer.
 During an active shell run, `Ctrl+C` now requests a pause instead of killing the shell; use `/resume` to continue the saved session.
 Shell runs now also show a compact live run board with active tasks, budgets, token usage, and the latest runtime note while execution is in flight.
+When a run pauses for an approval-required action, use `/approve` or `/deny` to resolve the pending request and continue the saved session.
 
 LM Studio example:
 
@@ -84,10 +87,13 @@ When resuming the current saved session, that saved session context is also inje
 After the agent changes workspace state, it now has to pass a verification step before the task can finish.
 If verification fails, Kimbho forces a repair pass before the next rerun and blocks the task after the configured repair budget is exhausted.
 If a non-debugger specialist burns through that budget, Kimbho automatically hands the task to `test-debugger` and keeps the same task id/dependency chain intact.
+The tool layer now enforces the configured sandbox and approval policy for high-risk actions: read-only sandboxes reject mutating tools, and destructive shell commands like `rm -rf` or `git reset --hard` are blocked under manual approval mode.
 When the workspace is a git repository with a valid `HEAD`, specialist tasks now run in isolated git worktrees and Kimbho reapplies the resulting diff back to the main workspace.
 If multiple specialist tasks are ready at the same time, Kimbho now schedules them concurrently in separate worktrees and surfaces their diff integration live in the shell.
 Those concurrent integrations are now serialized, so overlapping edits do not race generic patch application anymore.
 If a direct integration fails, Kimbho now replays the specialist diff on top of the latest workspace snapshot in a retry sandbox, reapplies a refreshed patch when that drift is mergeable, and preserves the original patch artifact when it is not.
+When the replay path still cannot land, Kimbho now writes an integration-conflict bundle with the relevant patch paths, conflicting files, failure output, and operator next steps.
+If that failure originates from a specialist task, Kimbho now rewrites the task into an `integrator` handoff instead of leaving it as a raw blocked patch failure.
 Kimbho also has managed background process tools and live HTTP verification: `process.start`, `process.list`, `process.logs`, `process.stop`, and `http.fetch`.
 Kimbho now also has browser-session tools for headless Chrome verification: `browser.open`, `browser.inspect`, `browser.click`, `browser.fill`, `browser.list`, and `browser.close`.
 Shell runs now surface model-usage telemetry when the provider returns token counts.