SeaShell plans before it edits by running 2 separate agents. The plan agent is read-only: it can read your repository and draft an approach, but it cannot write a file. When the plan is complete SeaShell asks whether to switch to the build agent, and only then can anything be written. In 2026 that separation is what lets you argue with an approach before it costs you a diff.
Why planning is a separate agent, not a prompt
Asking a model to describe its plan first is 1 instruction it can ignore. SeaShell instead runs planning in a read-only agent that has no write tool at all, so a plan cannot quietly become an edit halfway through a response.
The difference matters most on the changes you would most want to review. A large refactor described in prose can look reasonable and still touch the wrong files. When the planning agent is structurally unable to write, the review step is guaranteed rather than requested, and the model's compliance is not part of your safety story.
The handoff you approve
When a plan is finished SeaShell asks 1 question: whether to switch to the build agent and start implementing. Answering no leaves your working tree exactly as it was, and the plan stays available to revise or discard.
This is the point where most of the value lands. Rejecting a plan costs nothing, because no file has been touched, and revising a plan is far cheaper than reviewing and reverting a set of edits that already exist. Teams that plan first tend to spend their review attention on the approach rather than on the diff.
Planning and permissions reinforce each other
Approving a plan does not approve its edits. Each file change still computes a unified diff and passes through the allow, ask, or deny permission check, so there are 2 independent gates between an idea and a written file.
Planning catches the wrong approach; diff review catches the wrong implementation of a right approach. They fail differently, which is why SeaShell keeps both rather than treating an approved plan as blanket consent for whatever the build agent decides to do next.
Frequently asked questions
- Does SeaShell plan before editing files?
- Yes. SeaShell drafts a plan in a read-only plan agent that cannot write files, then asks before switching to a build agent to implement it.
- Can the plan agent change my code by accident?
- No. The plan agent has no write capability, so planning cannot modify your working tree regardless of what the model decides to do.
- What happens if I reject a plan?
- Nothing is written. Your working tree is unchanged and the plan can be revised or discarded.
- Does approving a plan approve all its edits?
- No. Every file edit still computes a unified diff and passes the allow, ask, or deny permission check before it is written.
Try SeaShell in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install SeaShellRelated guides
Atlas for Pandas: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Pandas. Vectorize df.apply, fix chained assignment under Copy-on-Write, and pin DataFrames with assert_frame_equal.
Atlas with Gemma 3 27B Instruct in 2026
Explore Atlas with Gemma 3 27B Instruct, Google's 27B open-weight model. Leverage its 202,752 token context and cost-effective $0.12/$0.2 per Mtok pricing for coding tasks in 2026.
Atlas with OpenAI o3-mini in 2026
Explore OpenAI o3-mini's role with Atlas in 2026. This December 2024 model offers a 200K context window and affordable $1.10/$4.40 per Mtok pricing, ideal for cost-effective reasoning tasks and subagent work.
Atlas with DeepSeek-R1 (local via Ollama) in 2026
Explore DeepSeek-R1 (local via Ollama) for Atlas in 2026. Leverage its visible chain-of-thought for debugging complex agent behavior, with variants from 1.5B to 671B and flexible pricing.
Atlas with Ollama Cloud in 2026: direct Scaling for Your AI Coding Agent
In 2026, drive Atlas with Ollama Cloud for direct scaling of open-weights models. Access massive context windows up to 1,048,576 tokens and powerful models like qwen3-coder:480b, maintaining local Ollama continuity.
Atlas with Gemma 4 31B (Ollama) in 2026
Drive Atlas with Gemma 4 31B (Ollama) locally. This 20GB model offers a 256K token context window for free, ideal for code-tuned AI development in 2026.
Atlas for Phoenix in 2026
Atlas is a terminal-native AI coding agent for Phoenix in 2026. It reads contexts, LiveView modules, and Ecto changesets, then runs mix test behind a prompt.
Atlas with Phi-4 (local via Ollama) in 2026
Explore Atlas with Phi-4 (local via Ollama) in 2026. This 14B parameter model offers strong reasoning for Atlas's planning agent, is free (self-hosted), and fits on a 12GB GPU, though its 16K token context window limits