Continue a long task
A long task can loop through reading, editing, testing, and fixing. Each Tool result may enter the next step; command output, web pages, and MCP results grow quickly.
Stage the work
- Explore read-only facts.
- Plan scope, risk, and acceptance.
- Implement only the agreed range.
- Validate and fix.
- Deliver evidence and limits.
Leave a short status at each stage.
Subagents and Team
If the task delegates work, check the returned files and results rather than relying only on the agent's summary. In Team, open the relevant work item and compare the delivery with its full request. Keep revision feedback in that item's discussion.
Continue after compaction
- Hover over Context usage beside the composer and check the updated usage.
- Ask Ghast to restate the goal, constraints, completed work and next step.
- Correct missing paths, decisions or acceptance criteria before it continues.
- If the next step might repeat an earlier command or external action, verify that action's current result first.
The ordinary timeline does not expose the compaction summary as a separate receipt. Lower usage is not proof that every important detail survived. See Manage context compaction for the command, settings and failure behavior.
Mechanisms with different lifetimes
| Mechanism | Lifetime | Purpose |
|---|---|---|
| Conversation history | Current conversation | Task process |
| Compaction summary | Current long task | Free context |
| Memory | Across conversations | Stable facts and preferences |
| Skill | Reusable method | How to perform a class of task |
| Team Memory | Team scope | Shared confirmed background |
Compaction does not automatically write Memory; Memory is not a permanent compaction copy.
What belongs in Memory
Store only confirmed information that is unlikely to change soon, project conventions reused across tasks, facts the user explicitly asked Ghast to remember, and durable background with a clear source and scope.
Do not store temporary paths, one-time codes, unconfirmed plans, production secrets, or long Tool output in Memory.
Handoff format
Before moving to a new conversation, fill in this short handoff from verified results. Keep exact file paths and object identifiers where needed, but exclude secrets and unrelated personal data.
Goal:
Allowed scope / prohibited actions:
Completed:
Current files / external objects:
Verified:
Unverified:
Blocked:
Next:
Open the new conversation, select the intended model and workspace, recheck command authorization, and send the handoff with a request to verify the current state before continuing. A new conversation or branch does not copy or isolate the actual workspace files.
Do not treat a summary as an audit log
A summary is useful for continuing work, but it does not prove exact command output, original file text, a successful external submission, the scope of a prior approval, or a third-party page's current state. Reread the original record when that evidence matters.