Add forkmind plugin — local-first LLM branching & debugging

.claude-plugin/marketplace.json

@@ -1670,6 +1670,24 @@
         "security",
         "compliance"
       ]
+    },
+    {
+      "name": "forkmind",
+      "source": "./plugins/forkmind",
+      "description": "Local-first LLM branching & debugging. Treat AI context windows like a Git repo — capture, branch, and regression-test LLM calls as a DAG. Free & local via Ollama, any OpenAI-compatible API, and MCP for agents.",
+      "version": "0.1.0",
+      "author": {
+        "name": "Medhovarsh Bayyapureddi",
+        "url": "https://github.com/Medhovarsh"
+      },
+      "category": "Development Engineering",
+      "homepage": "https://github.com/ccplugins/awesome-claude-code-plugins/tree/main/plugins/forkmind",
+      "keywords": [
+        "llm",
+        "debugging",
+        "observability",
+        "mcp"
+      ]
     }
   ]
 }

plugins/forkmind/.claude-plugin/plugin.json

@@ -0,0 +1,27 @@
+{
+  "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
+  "name": "forkmind",
+  "version": "0.1.0",
+  "description": "Local-first LLM branching & debugging. Treat AI context windows like a Git repo — capture, branch, and regression-test LLM calls as a DAG. Teaches Claude when and how to drive ForkMind.",
+  "author": {
+    "name": "Medhovarsh Bayyapureddi",
+    "url": "https://github.com/Medhovarsh"
+  },
+  "homepage": "https://medhovarsh.github.io/forkmind/",
+  "repository": "https://github.com/Medhovarsh/forkmind",
+  "license": "MIT",
+  "keywords": [
+    "llm",
+    "debugging",
+    "observability",
+    "ollama",
+    "mcp",
+    "local-first"
+  ],
+  "mcpServers": {
+    "forkmind": {
+      "command": "npx",
+      "args": ["-y", "github:medhovarsh/forkmind", "mcp"]
+    }
+  }
+}

plugins/forkmind/README.md

@@ -0,0 +1,13 @@
+# ForkMind 🧠
+
+Local-first LLM branching & debugging. Treat AI context windows like a Git repo —
+capture, branch, and regression-test LLM calls as a DAG. Free & local via Ollama,
+any OpenAI-compatible API, and an MCP server so agents query their own history.
+
+- **Skill** `forkmind` — Claude reaches for it on debug/compare/branch/regression.
+- **Command** `/forkmind` — start / branch / test / mcp.
+- **Agent** `forkmind-debugger` — isolated model/prompt compares, compact verdict.
+- **MCP** auto-wired — agents query their own `.forkmind/` history.
+
+Source & docs: https://github.com/Medhovarsh/forkmind
+Homepage: https://medhovarsh.github.io/forkmind/

plugins/forkmind/agents/forkmind-debugger.md

@@ -0,0 +1,43 @@
+---
+name: forkmind-debugger
+description: >
+  Drives ForkMind to debug, compare, or regression-test LLM / agent calls in an
+  isolated context. Spawn when the user wants to compare two prompts or models on
+  the same input, find why an LLM's answer changed, branch from a past turn, or
+  pin/verify a regression baseline. Returns a compact verdict (winner, diffs,
+  drift) — not raw transcripts — so main context stays small.
+tools: [Read, Bash, Glob, Grep]
+---
+
+You drive **ForkMind** — a local-first proxy that records LLM calls into
+`.forkmind/` as a branchable DAG. Your job: run the comparison/debug the caller
+asked for, then report a tight verdict.
+
+## Operating rules
+
+- ForkMind is local, free (Ollama default), no cloud. Never send data anywhere.
+- Start it if not running: `npx github:medhovarsh/forkmind start` (proxy +
+  dashboard on `:4500`). Clients hit `http://localhost:4500/v1`.
+- Inspect captured runs from `.forkmind/` (plain JSON on disk) — read nodes
+  directly rather than re-running when the data already exists.
+- If `forkmind mcp` history tools are available, use them to trace lineage.
+
+## Workflow
+
+1. Confirm what to compare: prompt A vs B, model X vs Y, or before/after a tweak.
+2. Ensure ForkMind is up; route each variant through the proxy so each becomes a
+   DAG node.
+3. Read the resulting nodes: request, response, tokens, provenance, stream flag.
+4. For regression: compare against the pinned baseline; flag drift.
+
+## Report format (return this, nothing more)
+
+```
+VERDICT: <one line — winner / root cause / drift yes-no>
+A (<model/prompt>): <key output trait> · <tokens>
+B (<model/prompt>): <key output trait> · <tokens>
+DIFF: <what actually changed and why it matters>
+NEXT: <recommended action — pin baseline / switch model / edit prompt>
+```
+
+Keep it short. Caller wants the conclusion, not the logs.

plugins/forkmind/commands/forkmind.md

@@ -0,0 +1,15 @@
+---
+description: Start ForkMind or branch/inspect/regression-test an LLM call
+---
+
+Invoke the `forkmind` skill. Based on `$ARGUMENTS`:
+
+- no args or `start` → tell the user how to launch ForkMind
+  (`npx github:medhovarsh/forkmind start`, proxy + dashboard on :4500) and how
+  to point their OpenAI-compatible client at `http://localhost:4500/v1`.
+- `branch` / `fork` → explain forking an alternative prompt or model from a
+  historical node in the dashboard.
+- `test` / `regression` → explain pinning baselines and catching drift in CI.
+- `mcp` → show the MCP server config so an agent can query its own history.
+
+Keep it actionable: give the exact command(s) for what the user asked.

plugins/forkmind/skills/forkmind/SKILL.md

@@ -0,0 +1,88 @@
+---
+name: forkmind
+description: >
+  Use when debugging, comparing, or regression-testing LLM / agent calls — when
+  the user wants to capture LLM traffic, see a conversation as a branchable DAG,
+  fork an alternative prompt or model from a past turn, or pin good outputs as
+  baselines to catch drift. ForkMind is local-first (no cloud, no account) and
+  proxies any OpenAI-compatible API, defaulting to free Ollama models.
+  Trigger: "debug this prompt", "compare two models", "branch from that turn",
+  "why did the LLM change its answer", "regression test my prompt", "capture LLM
+  calls", "forkmind".
+---
+
+# ForkMind
+
+ForkMind treats AI context windows like a Git repo. It runs a local proxy that
+records every LLM call into `.forkmind/` as a node in a DAG, serves a dashboard,
+and lets you branch alternative prompts/models from any historical turn.
+
+## When to reach for it
+
+| Situation | Why ForkMind |
+|---|---|
+| "This prompt sometimes regresses — catch it" | Pin baselines, re-run, diff drift |
+| "Compare llama3 vs gpt-4o on the same turn" | Fork the node, swap model, re-run |
+| "Why did the agent change its answer?" | Inspect the DAG: request/response, tokens, lineage |
+| "Capture all LLM traffic during this task" | Point client at the proxy; everything is logged |
+| Agent needs to recall its own past attempts | ForkMind MCP server exposes the history |
+
+Do NOT use for: production traffic logging at scale, hosted/cloud observability,
+or non-LLM HTTP debugging.
+
+## Run it
+
+ForkMind runs straight from the git link — no npm registry needed:
+
+```bash
+# starts proxy + dashboard on :4500
+npx github:medhovarsh/forkmind start
+```
+
+Then point any OpenAI-compatible client at the proxy:
+- `baseURL: http://localhost:4500/v1`
+- dashboard: `http://localhost:4500`
+
+Free local default: install [Ollama](https://ollama.com), `ollama pull llama3`.
+Any provider works (OpenAI, Anthropic, Groq, OpenRouter, Together, vLLM, LM Studio)
+by passing that provider's base URL + key through the proxy.
+
+## Core moves
+
+- **Capture** — route calls through `:4500/v1`; each call becomes a DAG node.
+- **Branch** — in the dashboard, "Fork from here" on any node → edit prompt /
+  swap model / change params → re-run only that subtree.
+- **Inspect** — node inspector shows request, response, tokens, provenance, and
+  whether the response was streamed.
+- **Regression test** — pin a known-good output as a baseline; re-run after a
+  prompt tweak; ForkMind flags drift (wire into CI).
+
+## MCP for agents
+
+Installed as the Claude Code plugin, ForkMind's MCP server is wired up
+automatically (`forkmind` server, launched via `npx … forkmind mcp`). An agent
+can query its own history mid-task — recall previous attempts, trace lineage,
+self-correct. No manual config needed.
+
+Outside the plugin, add it to any MCP client by hand:
+
+```json
+{
+  "mcpServers": {
+    "forkmind": { "command": "npx", "args": ["-y", "github:medhovarsh/forkmind", "mcp"] }
+  }
+}
+```
+
+## Delegating heavy compares
+
+For "compare model X vs Y" or "why did the answer change" tasks that would dump
+big transcripts into context, spawn the **`forkmind-debugger`** subagent. It runs
+the comparison in isolation and returns a compact verdict (winner, diff, drift),
+not raw logs.
+
+## Notes
+
+- Everything is plain JSON on disk under `.forkmind/`. No database, no telemetry.
+- Add `.forkmind/` to `.gitignore` (the repo already does for its own checkout).
+- Full docs: https://github.com/Medhovarsh/forkmind#readme