Skip to main content
Severitycritical
Statusfixed
Found2026-04-18
Fixed2026-04-18
Areagateway/session
Commitbbb1474

​
Symptom

Long multi-agent sessions progressively forgot early context. After the first compaction pass, facts captured in that summary (user’s original goal, early agent results, pinned translations, entity names) were preserved correctly. After the second pass, those facts silently disappeared — the planner could no longer reference them, gave “I don’t know” answers, or hallucinated replacements. Observable only on sessions long enough to trigger two or more compactions. Pattern: session starts coherent, drifts over time, ends disconnected from its own early history.

​
Root cause

gateway/src/session/compaction.ts:115-124 (pre-fix): 
const { error: sessErr } = await client
  .from("gateway_sessions")
  .update({
    compaction_summary: summary,          // bare UPDATE, not a merge
    compaction_at: new Date().toISOString(),
  })
  .eq("id", sessionID)
compaction_summary is a single text column. Each pass did a bare UPDATE; the second pass clobbered the first pass’s paragraph. Worse, the summarizer on pass #2 was never given the prior summary as input. summarize() received only the new head messages. The resulting paragraph captured only what happened since the last compaction, not the accumulated history. A subtler secondary bug lived in the same path: session.history() prepends the prior summary as a synthetic user message with a freshly-minted UUID. On pass #2 that synthetic landed in head and would be “re-summarized” — but then the subsequent UPDATE ... WHERE id IN (head_ids) was a silent no-op for the synthetic id (no matching row), so only real head rows got marked compacted. Meanwhile the summarizer had already paraphrased a paraphrase, losing fidelity on every iteration. Mental model: “compaction is a rolling summary, so each pass summarizes the latest chunk.” True at pass #1. At pass #N the “latest chunk” excludes everything from passes #1 through #N-1 — which are now gone from gateway_messages (compacted) AND gone from the summary column (overwritten).

​
Fix

  • summarize() grows an optional priorSummary?: string | null. When non-empty, it’s injected as a leading user message tagged [PRIOR SUMMARY — preserve every fact below].
  • System prompt gains explicit fact-preservation language: “If a PRIOR SUMMARY is provided, treat its facts as authoritative and carry them forward. The new summary must be a SUPERSET of the prior summary, not a replacement.”
  • runCompaction reads compaction_summary from the session row before summarizing and passes it as priorSummary. It also filters synthetic messages out of history before splitting, so the no-op UPDATE path and the paraphrase-of-paraphrase path are both gone.
  • The column overwrite is now safe because the new value is a superset of the old one by construction.
See commit bbb1474 and tests at gateway/tests/session/summary.test.ts — three cases covering marker injection, closing-prompt variant selection, and whitespace handling.

​
Why the tests didn’t catch it

No test ran a second compaction pass. The session test suite was a thin sanity check around message persistence; compaction was untested end-to-end. A single-pass test would pass, because pass #1 correctly produces a summary of the head. The bug only manifests on pass #2 or later — a case no test exercised. Compounding: the unit I’d need to test is the accumulation of facts across multiple compactions, which requires an LLM. Stubbing the LLM is possible (as the fix’s regression test does), but nobody had written such a harness before this bug was found.

​
Class of bug — where else to watch

“Overwriting a lossy-compressed store” is the core pattern. Any time a column or row is written as a summarization / paraphrase / hash-with-collisions of its inputs, replacing it wholesale on each update compounds loss. The only safe update is one that takes the prior value as input. Other places this shape could hide in the codebase:
  • The planner’s agent_catalog column in gateway_sessions is wholesale-overwritten on every turn (gateway_sessions schema). External may omit an agent from the catalog on a subsequent turn — we silently drop it even though the session’s history references its tool calls. Same shape: overwrite of a lossy-valid store. Tracked in known-issues.md under agent-catalog-overwrite.
  • Any future “session summary” features on top of compaction (e.g. per-turn summaries, cross-session rollups) must preserve-on-merge, not overwrite.
  • The gateway_tasks.usage JSON column aggregates polling counts and signature stats. Today it’s only written once per task, but if anything ever updates it in place, the same pattern applies.
Rule of thumb: before writing to a column that replaces a previous value, ask: is this column a pure projection of its source data (safe to overwrite) or a lossy compression (must merge)? If lossy, the update must read-then-write, not write-only.