| name | compact-session |
| description | Use when asked to manually compact a session's memory, trigger session summarization, or reduce a long session's context footprint. |
| disable-model-invocation | true |
| allowed-tools | ["Bash(ll-compact-session:*)","Read"] |
| metadata | {"short-description":"Use when asked to manually compact a session's memory or trigger session summ"} |
Compact Session Skill
Manually triggers session-memory compaction for one session — the same LCM
compaction path (session_store.compact_session) that fires automatically in
the background once a session crosses the soft token threshold (FEAT-2598).
Produces leaf/condensed summary_nodes rows and prints the resulting
CompactResult (summary text, covered message count, token estimate).
This is the LCM/semantic-summarization axis — distinct from ll-session compact, which sweeps the separate retention axis (kind='retention'
raw_events summarization, ENH-1906/ENH-2581). Do not confuse the two.
When to Activate
Proactively offer or invoke this skill when the user:
- Asks to compact, summarize, or condense a session manually
- Wants to reduce a long-running session's context footprint before continuing
- Asks whether a session has been compacted yet, or wants to see its summary
- Mentions the session is approaching a context limit and wants to trigger
summarization early rather than waiting for the automatic soft-threshold trigger
Arguments
$ARGUMENTS
- session_id (required) — the session to compact. If the user doesn't name
one, use the current session's ID (available from the host CLI's session
context) or ask which session they mean.
--db PATH (optional) — session database path (default: .ll/history.db)--json (optional) — machine-readable output
How to Use
Run the ll-compact-session CLI command:
ll-compact-session SESSION_ID
For programmatic access (e.g. from another skill or loop evaluator):
ll-compact-session SESSION_ID --json
The JSON payload includes new_leaves (leaf nodes created this run),
summary_text (the condensed summary, null if the session has fewer than
two leaf blocks), compacted_messages (covered message_events ids), and
context_token_estimate.
Compaction is idempotent — repeated calls do not create duplicate nodes.
Note that history.compaction.enabled (default false in
.ll/ll-config.json) gates whether an LLM summary is produced at all; when
disabled, ll-compact-session still runs but the leaf/condensed summaries
fall back to deterministic truncation (no LLM cost) per the existing LCM
three-level escalation in session_store._summarize_block.
Examples
| User says | Action |
|---|
| "Compact this session before we keep going" | ll-compact-session <current-session-id> |
| "Has session abc123 been summarized?" | ll-compact-session abc123 --json, report summary_text |
| "This loop's session is getting huge, trigger compaction now" | ll-compact-session <session-id> |
