devi_ramphal

Giving agents a portable memory across tools solves the constant amnesia that makes long tasks painful. It recalled prior decisions accurately between sessions. I would want to see how it handles conflicting memories when two runs disagree about the same fact.

Managing several worktrees for parallel branches is exactly the workflow agents push you into, and this makes it painless. It handled a repo with a pile of active trees without confusion. Surfacing which tree an agent is currently in would close the loop.

Wrapping Codex, Claude Code and OpenCode behind one client is the kind of unglamorous plumbing that saves real time. Switching providers mid-task worked without a relog. I would want to see how it handles a dropped connection in the middle of a long agent run.

Semantic retrieval over a codebase as an MCP server is the right shape for this problem, and the editing actions are precise enough to trust on a real repo. It found the symbol I meant, not just the string. Indexing time on a large tree is my one watch item.

I pointed it at a service with a messy dependency web and the graph held up. The detail I want is edge typing, so I can tell a call from an import at a glance. Even now it cut my ramp up time noticeably.

Handing Claude real terminal control and file editing through MCP is powerful and exactly the integration I wanted, with the usual respect for how much it can now do. It searched and edited across my project reliably. Clearer boundaries on which directories it may touch would let me relax.

Open source and built for infra work, which is a combination I trust more than a closed box near my servers. It handled jumping between hosts cleanly. I would like richer audit logging of what the assistant suggested versus what ran.

It layers genuinely useful navigation onto Gemini and AI Studio without feeling bolted on. Jumping back to an earlier point in a long session works reliably. I would like an export of a conversation as structured text.

Tracking grocery prices over time is the kind of unglamorous data project I respect. The history view loaded quickly and the per product trend is exactly what you want. I would add an alert when a staple drops below a threshold.

A commit message generator that actually reads the diff instead of guessing from filenames is a small thing that compounds across a thousand commits. It summarised a gnarly change accurately. I would like it to respect a repo's existing message conventions automatically.

Having the whole loop from prompt to preview as something I can self host is the part that matters. It deployed without drama on the intended stack. I want clearer hooks for swapping the model layer, but the bones are open and readable.

Solid for the happy path. I want to see how it behaves when a server streams partial results or drops mid-call, because that is where these clients usually fall apart. So far the reconnect logic has held up under a flaky local server.

Making the full codebase available as retrieval context is the right fix for agents that hallucinate APIs, and this one returned on-target results on a large repo. The recall was better than a plain vector dump. I would watch indexing cost and staleness as the code changes underneath it.

Seeing the LLM and agent workflows tangled inside a codebase drawn out as a graph made a mess legible in minutes. It correctly traced a chain I had half forgotten. Edge labels for the call types would take it from useful to indispensable.

A full-stack engineer agent built on Next.js and Claude is an ambitious claim, and on a small feature it actually delivered a coherent front and back end. It made sensible data choices. I would test it hard on error handling and edge cases before trusting a real ticket to it.

Managing agents spatially rather than as a list of tabs turned out to suit how I think about parallel work. It handled several CLIs at once without tangling their output. I would want clearer recovery when one agent wedges itself mid-task.