Tips & troubleshooting

← Configure your ecosystem · Back to the guide

A short collection of the things that make Rosetta run smoothly, the mistakes that cost people time and tokens, and fixes for the problems you’re most likely to hit.

Getting the best results

• Use slash commands. Once your repo is set up, start every task with the right workflow command and describe your request naturally. See the scenarios.
• Be specific. Context up front means better output and fewer questions.
• Read plans before approving. The plan is your last checkpoint before work begins.
• Answer questions fully. When Rosetta asks, it’s targeting a specific gap. Short answers lead to incomplete work.
• Write requirements first. For anything non-trivial, /requirements-authoring-flow gives you a clean acceptance baseline and prevents scope creep.
• Keep your context files healthy. Accurate CONTEXT.md and ARCHITECTURE.md help every task. Don’t delete anything in docs/ — that’s Rosetta’s project memory.

Choosing a model

This is the single biggest lever on both quality and cost.

• Use a medium reasoning model: Sonnet 5 (medium), GPT-5.6-terra-medium, or gemini-3.7-flash-high. Set the reasoning effort to medium — higher effort can burn 30 minutes on unnecessary reasoning.
• Don’t run everything on a high-reasoning or Opus-class model. They spend heavily on reasoning and can drain a daily budget in one sitting. Rosetta already assigns an appropriate model to each step and switches automatically — you don’t need to force the most expensive one.
• Don’t use “Auto” model selection. It often quietly downgrades to a weaker model mid-task, hurting quality. Pick the model explicitly.

Lowering your token bill

• Prefer a flat-rate subscription (Claude Pro/Max, ChatGPT Plus/Pro) over pay-per-token API billing.

• To cut output tokens, add this line to your workspace AGENTS.md / CLAUDE.md:

  MUST ALWAYS think, reason, plan, chat, document in compressed/terse/unicode chars/terms/always/no hieroglyphs; Exclude final artifacts, any tool calls, all code, etc.
  

Managing sessions

When to start a fresh chat vs. continue:

• Same session for follow-ups on the work you just did — refining a change, fixing a missed edge case.
• New session when you move to a different feature or an unrelated change. Each top-level request should start fresh so context stays lean.

A very common mistake is running every task inside the chat that started with “what can you do?” — that bloats context and degrades results. Start a new session for the real task.

Watch your context usage. If it climbs above ~65%, wrap up and switch sessions. To hand off cleanly, ask the agent to save its state:

Please save execution state, workflow state, findings, original intent with clarifications, and tasks left to do as concise "agents/TEMP/execution-state.md" so that I can start a fresh new session and continue execution where you left it off.

Then start the new session with the same command you began with:

/<original-command> Please resume execution saved in "agents/TEMP/execution-state.md" according to flow instructions

If the agent gets stuck looping on the same problem, step in with specifics on how to solve it, or ask it to spin up a focused, more capable subagent for just that problem.

Common questions

Does Rosetta use more tokens? Yes, a bit up front — it loads guardrails and the relevant workflow. In return you get fewer wrong-path runs (which waste far more), earlier questions instead of endless back-and-forth, and more reliable results.

Why is the first message in a session slower? Rosetta does one-time prep — loading context and reading your project files. Later messages reuse that and are fast. You can’t skip prep, even for a one-liner; it’s lightweight and keeps trivial tasks from going wrong.

Does it work in plan mode, Auto mode, or full-access mode? Yes, in every mode. Permission modes only change what the agent may do without asking — they don’t turn off Rosetta’s prep, workflows, or approval gates.

How do I turn off the approval gates for one task? Include the literal phrase fully autonomous or no HITL in your request. That’s the only accepted opt-out (ambiguous phrasing won’t disable it, by design). Use it sparingly — those gates catch ambiguous intent and risky actions.

Fixing problems

The agent isn’t using Rosetta

If the agent responds like a generic assistant and ignores Rosetta:

• Re-run the check: ask What can you do, Rosetta? — it should list workflows.
• Confirm the plugin is actually installed for your IDE (see Install Rosetta) and that you didn’t install it in two places at once (e.g. both Claude Code and Cursor), which causes conflicts.
• Start a fresh session — especially right after initializing a repo, since new files only load at session start.
• Make sure you selected a real model, not “Auto.”

It stopped following Rosetta mid-session

This usually means a long session drifted or, in hosted setups, an expired connection. Start a fresh session and resume from a saved state file (see Managing sessions above).

Slow, empty, or low-quality responses

• Check your model choice — “Auto” or an over-powered reasoning model are the usual culprits (see Choosing a model).
• If your context is very full, switch to a fresh session.

Run a post-mortem on anything that disappointed you

Not just failures — whenever a run fought you or could have gone better, ask the agent to run /post-mortem. It root-causes the behavior across your prompt, workspace files, and Rosetta’s instructions, and recommends concrete fixes. With your explicit approval, it can also file sanitized feedback as a Rosetta issue — nothing leaves your repo without your say-so.

A plain-language glossary

You don’t need these to get started, but they show up in messages and docs.

More help

• Ask the agent directly: /help-flow ...
• Run /post-mortem after a rough run.
• Open an issue: https://github.com/griddynamics/rosetta/issues
• Email: rosetta-support@griddynamics.com
• Website: https://griddynamics.github.io/rosetta/
• Deeper technical detail lives in the repo’s TROUBLESHOOTING and FAQ.