My AI session was compacted

What to do when your AI tool shortens a long session in the middle of an agent's work.

Your agent was working, the session got long, and your AI tool compacted it. This page says what that means, what to tell the agent, and what not to do. The short version: the same agent continues. Do not start a new one.

What happened

Compacted means your AI tool replaced the earlier part of the conversation with a summary, so the session could keep going. Claude Code calls this compaction or auto-compaction; Gemini CLI calls it compression. The agent keeps working, but it no longer has the full detail of what it read and did before the summary.

The Studio's files are not touched by a compaction. What the agent had saved there (its records, its memory, its working list) is still there. What the summary dropped is detail from the conversation itself, which is why the agent re-reads its records instead of relying on what it remembers.

What to do

Tell the agent to continue, with this exact sentence. It is written for the agent, so paste it as it is:

If this agent session was compacted or auto-compacted, or you no longer remember completing boot, do not activate and never run born. Run python3 vault/tools/tropo-compact-continue.py --agent <slug> before any other work.

<slug> is the agent's short name in lowercase, the name of its folder under agents/. For Darin it is darin.

The agent runs that command. The command gathers where the agent was from the Studio's own records (what it is working on, what it was about to do next, and anything that changed while it was busy) and prints a READ LIST: the specific files to read, in order. Then:

1. The agent opens and reads each file on the READ LIST, in order. The command can find the records; only the agent can read them.
2. The agent runs the attest command the tool prints, exactly as printed. That records that the reading was done.
3. The agent carries on as the same generation.

If the command asks which session it is

Usually the command can tell which session is calling. If it cannot, it says what it is missing, and the agent runs it again with two more values:

python3 vault/tools/tropo-compact-continue.py --agent <slug> --generation <issued-generation> --activation-run <this-run-directory>

• <issued-generation> is the generation this session was given when it started.
• <this-run-directory> is the folder that same start created, named like playbook-runs/agent-activation-<slug>-<generation>-<date>/. Its record file names the generation too.

The agent takes both values from this session's own start-up record, not from memory and not by picking the newest one it can find: a newer folder may belong to a different session. The command then prints its READ LIST and an attest command with those values filled in. The agent reads every file on the list first, then runs exactly that attest command. If neither the agent nor you can establish which start belongs to this session, stop there: the agent can still read, but it should not attest or retire until the question is settled.

Your tool may already have sent the sentence

The Studio ships two small reminder settings:

• Claude Code and Gemini CLI: the Studio's settings for these tools send the agent the sentence above as soon as a session is compacted.
• Codex, Cursor and Windsurf: the Studio ships no reminder setting for these. The sentence is in the instruction files those tools read, and some of these tools have hook features of their own, but Tropo does not configure them. If the agent seems unsure where it was, paste the sentence yourself.

What not to do

• Do not activate the agent again from its activation file in the same session. That begins a new generation and leaves the old one looking unfinished.
• Do not retire the agent because the session was compacted. Compaction is not the end of a session. You end sessions; the tool shortening one is not you ending it.

Terms used on this page

• Session: one continuous conversation between you and an agent in your AI tool.
• Boot: the reading an agent does when it starts: who it is, its memory and its current work. An agent that remembers finishing its boot is still in the session it started.
• Activate: start an agent from its activation file, which runs its boot as a new generation.
• born: the step inside an activation that gives the new generation its number. Running it for a session that is already going creates a second generation for the same work.
• Generation: one working life of an agent, from the moment it starts until you end its session. Each new start is the next generation. A compacted session is still the same generation.
• Activation file: the file you attach to start an agent, such as agents/darin/darin-activation.md.
• Attest: the agent's record that it has read every file on the READ LIST after the compaction. It comes after the reading, never instead of it.
• Retire: close an agent's generation on purpose, at the end of a session you chose to end.

More terms are in the Glossary.

Related

• Quickstart: getting a Studio running.
• Create an agent: how agents and their activation files are made.