feat(kiro): add Kiro CLI harness for session capture and MCP install

[sumitvairagar]

Summary

Add a Kiro harness so sessions from Kiro CLI (kiro-cli) are captured into Hivemind shared memory, exactly like the existing Claude Code, Cursor, Codex, and Cowork harnesses.

Why Kiro

Kiro is an AI coding agent from AWS (free tier available). Its JSONL session format is already compatible with Hivemind's capture pipeline — no hooks or plugin API needed, just transcript tailing, the same approach the Cowork harness uses. It also installs MCP servers via a standard ~/.kiro/settings/mcp.json config file, so the Hivemind MCP server can be wired in with a single hivemind install kiro command.

I run Java/Spring workshops where attendees use kiro-cli. Hivemind would be a natural fit for shared team memory across those sessions, but there's no integration path today.

Technical approach

I've read the Cowork harness (src/mcp/cowork-ingest.ts) and the Kiro session format. They map cleanly:

Kiro JSONL format (~/.kiro/sessions/cli/<uuid>.jsonl):

{"version":"v1","kind":"Prompt","data":{"content":[{"kind":"text","data":"..."}]}}
{"version":"v1","kind":"AssistantMessage","data":{"content":[{"kind":"text","data":"..."},{"kind":"toolUse","data":{"toolUseId":"...","name":"...","input":{...}}}]}}
{"version":"v1","kind":"ToolResults","data":{"content":[{"kind":"toolResult","data":{"toolUseId":"...","content":[...]}}]}}

Mapping to Hivemind session rows:

Kiro kind	Hivemind type
Prompt	user_message (content from data.content[].data where kind=text)
AssistantMessage text blocks	assistant_message
AssistantMessage toolUse blocks	tool_call (tool_name, tool_use_id, tool_input)
ToolResults toolResult blocks	tool_result (tool_use_id, tool_response)

MCP installation path — Kiro reads ~/.kiro/settings/mcp.json:

{
  "mcpServers": {
    "hivemind": {
      "command": "npx",
      "args": ["-y", "hivemind", "mcp"]
    }
  }
}

Planned implementation

• src/kiro/kiro-ingest.ts — tails ~/.kiro/sessions/cli/*.jsonl, maps kinds to Hivemind rows, watermarks per-transcript, drains to the session queue. Mirrors cowork-ingest.ts.
• src/cli/install-kiro.ts — merges the Hivemind MCP server entry into ~/.kiro/settings/mcp.json, preserving existing servers.
• hivemind install kiro CLI command.
• startKiroIngestLoop() called from the MCP server on startup.
• Tests mirroring cowork-ingest.test.ts and install-cowork.test.ts.

Scope

Transcript tailing only (same as Cowork) — no kiro plugin API or hooks required. Sessions captured as agent = "kiro" rows. Summaries and skillify triggered on idle sessions via the existing summarizeIdleSessions pattern.

Happy to implement this if the approach looks good to you.

[sumitvairagar]

Opened PR #365 with the full implementation.

The approach mirrors the Cowork harness exactly — transcript tailing, per-field secret redaction (same pattern as #363), lock file for concurrent MCP processes, idle-session summarization. The Kiro JSONL format maps cleanly to Hivemind rows with no ambiguity.

hivemind kiro install merges into ~/.kiro/settings/mcp.json non-destructively, preserving any existing servers. 67 new tests, TypeScript clean.