Add memory_compactions tool for browsing compaction checkpoints

Queries compaction-type parts in the DB to find session compaction
events, then retrieves the summary text from the adjacent assistant
message. Presents compactions as navigable checkpoints with list and
read modes.
This commit is contained in:
2026-04-21 09:20:53 +00:00
parent f7bb7f94cf
commit 38b84b8163
3 changed files with 128 additions and 1 deletions

View File

@@ -59,12 +59,13 @@ src/
| `experimental.chat.system.transform` | Inject context % used + advisory into system prompt | | `experimental.chat.system.transform` | Inject context % used + advisory into system prompt |
| `event` | Feed SSE events to ContextTracker | | `event` | Feed SSE events to ContextTracker |
### Tools (7) ### Tools (8)
| Tool | Purpose | | Tool | Purpose |
|------|---------| |------|---------|
| `memory_context` | Current context window usage (% , tokens, model, status) | | `memory_context` | Current context window usage (% , tokens, model, status) |
| `memory_compact` | Trigger compaction via `ctx.client.session.summarize()` | | `memory_compact` | Trigger compaction via `ctx.client.session.summarize()` |
| `memory_compactions` | List/read compaction checkpoints (summaries) for a session |
| `memory_summary` | Quick counts: projects, sessions, messages, todos | | `memory_summary` | Quick counts: projects, sessions, messages, todos |
| `memory_sessions` | List recent sessions, optionally filtered by project path | | `memory_sessions` | List recent sessions, optionally filtered by project path |
| `memory_messages` | Read messages from a specific session | | `memory_messages` | Read messages from a specific session |
@@ -99,6 +100,15 @@ usable = model.limit.input ? model.limit.input - reserved
The `tokens.input` on the last assistant message approximates context size. We track against model context limit from config, falling back to 200k. The `tokens.input` on the last assistant message approximates context size. We track against model context limit from config, falling back to 200k.
### Compaction Data in DB
When compaction occurs, OpenCode creates:
1. A synthetic `user` message with a `compaction`-type part (`part.data = {type: "compaction", auto: true/false, overflow: true/false}`)
2. `message.data.summary = {diffs: [...]}` on the compaction message
3. The assistant message immediately after contains the actual summary text in a `text`-type part
The `memory_compactions` tool queries for `compaction`-type parts and retrieves the adjacent summary text, presenting them as navigable checkpoints.
### Write Operations ### Write Operations
All write operations (compaction triggering) go through the OpenCode client SDK (`ctx.client.session.summarize`). The plugin never writes to the database or any OpenCode files. All write operations (compaction triggering) go through the OpenCode client SDK (`ctx.client.session.summarize`). The plugin never writes to the database or any OpenCode files.

View File

@@ -82,6 +82,7 @@ All backed by read-only `bun:sqlite` queries to `${XDG_DATA_HOME:-$HOME/.local/s
| Tool | Purpose | | Tool | Purpose |
|------|---------| |------|---------|
| `memory_compactions` | List/read compaction checkpoints for a session |
| `memory_summary` | Quick counts: projects, sessions, messages, todos | | `memory_summary` | Quick counts: projects, sessions, messages, todos |
| `memory_sessions` | List recent sessions with metadata, sorted by update time | | `memory_sessions` | List recent sessions with metadata, sorted by update time |
| `memory_messages` | Read messages from a specific session as markdown | | `memory_messages` | Read messages from a specific session as markdown |

View File

@@ -282,6 +282,122 @@ export const createTools = (
}, },
}), }),
memory_compactions: tool({
description:
"List compaction events (checkpoints) for a session. Compactions are moments where the conversation was summarized to free context space. Use 'read' to get the full summary text of a specific compaction — these act as checkpoints showing what the agent considered important at that point in the session.",
args: {
sessionId: z.string().describe("Session ID to check compactions for."),
read: z
.number()
.optional()
.describe("Index (1-based) of a specific compaction to read the full summary text for."),
},
async execute(args) {
try {
type CompactionMeta = {
compaction_msg_id: string;
compaction_time: number;
time: string;
is_auto: number;
overflow: number;
};
type SummaryRow = {
summary_text: string;
};
const compactions = runQuery<CompactionMeta>(
`
SELECT
cp_msg.id AS compaction_msg_id,
cp_msg.time_created AS compaction_time,
datetime(cp_msg.time_created/1000, 'unixepoch', 'localtime') AS time,
COALESCE(json_extract(cp_part.data, '$.auto'), 0) AS is_auto,
COALESCE(json_extract(cp_part.data, '$.overflow'), 0) AS overflow
FROM part cp_part
JOIN message cp_msg ON cp_msg.id = cp_part.message_id
WHERE cp_msg.session_id = $sessionId
AND json_extract(cp_part.data, '$.type') = 'compaction'
ORDER BY cp_msg.time_created ASC
`,
{ $sessionId: args.sessionId },
);
if (!compactions || compactions.length === 0) {
return "No compactions found for this session.";
}
if (args.read !== undefined) {
const idx = args.read - 1;
if (idx < 0 || idx >= compactions.length) {
return `Invalid compaction index. Session has ${compactions.length} compaction(s). Use 1-${compactions.length}.`;
}
const comp = compactions[idx];
const summaryRows = runQuery<SummaryRow>(
`
SELECT json_extract(p.data, '$.text') AS summary_text
FROM message m
JOIN part p ON p.message_id = m.id
WHERE m.session_id = $sessionId
AND json_extract(m.data, '$.role') = 'assistant'
AND json_extract(p.data, '$.type') = 'text'
AND m.time_created > $compactionTime
ORDER BY m.time_created ASC
LIMIT 1
`,
{ $sessionId: args.sessionId, $compactionTime: comp.compaction_time },
);
const summaryText =
summaryRows && summaryRows.length > 0 && summaryRows[0].summary_text
? summaryRows[0].summary_text
: "(no summary text found)";
const header = [
`# Compaction ${args.read}`,
`Time: ${comp.time}`,
`Auto: ${comp.is_auto ? "yes" : "no"}`,
`Overflow: ${comp.overflow ? "yes" : "no"}`,
"",
].join("\n");
return `${header}${summaryText}`;
}
const lines = [`# Compactions (${compactions.length})\n`];
lines.push("| # | Time | Auto | Summary |");
lines.push("|---|------|------|---------|");
for (let i = 0; i < compactions.length; i++) {
const comp = compactions[i];
const summaryRows = runQuery<SummaryRow>(
`
SELECT substr(json_extract(p.data, '$.text'), 1, 150) AS summary_text
FROM message m
JOIN part p ON p.message_id = m.id
WHERE m.session_id = $sessionId
AND json_extract(m.data, '$.role') = 'assistant'
AND json_extract(p.data, '$.type') = 'text'
AND m.time_created > $compactionTime
ORDER BY m.time_created ASC
LIMIT 1
`,
{ $sessionId: args.sessionId, $compactionTime: comp.compaction_time },
);
const preview = summaryRows?.[0]?.summary_text
? `${summaryRows[0].summary_text.replace(/\n/g, " ").substring(0, 60)}...`
: "(no summary)";
lines.push(`| ${i + 1} | ${comp.time} | ${comp.is_auto ? "yes" : "no"} | ${preview} |`);
}
lines.push(
"",
`Use memory_compactions with a "read" argument (1-${compactions.length}) to view the full summary for a specific compaction.`,
);
return lines.join("\n");
} catch (err) {
return `Failed to query compactions: ${err instanceof Error ? err.message : String(err)}`;
}
},
}),
memory_plans: tool({ memory_plans: tool({
description: "List saved plan files from OpenCode's plans directory.", description: "List saved plan files from OpenCode's plans directory.",
args: { args: {