Description

resolveAgentConfig() merges compaction.qualityGuard and compaction.memoryFlush using object spread on values that may be undefined (or null).

• AgentCompactionSchema makes both qualityGuard and memoryFlush optional; when omitted they are undefined.
• The merge condition uses typeof ... === "object" checks with ||, so it can enter the merge block when only one side is an object.
• It then spreads both sides:
  • { ...agentDefaults?.compaction?.qualityGuard, ...entry.compaction.qualityGuard }
  • { ...agentDefaults?.compaction?.memoryFlush, ...entry.compaction.memoryFlush }
• In JavaScript, object spread evaluates ToObject(source); spreading undefined/null throws TypeError: Cannot convert undefined or null to object.
• Additionally, typeof null === "object" means a null value (if it can enter config) would incorrectly pass the type check and then crash on spread.

Impact: a malformed or partial agent configuration (e.g., defaults include qualityGuard but an agent entry has compaction: {} without qualityGuard, or vice-versa) can crash config resolution, potentially causing a service startup failure or runtime denial-of-service if configs are reloadable.

Recommendation

Avoid spreading possibly undefined/null values. Guard both sides and/or default to {}.

Example fix:

const qualityGuardDefaults =
  agentDefaults?.compaction?.qualityGuard && typeof agentDefaults.compaction.qualityGuard === "object"
    ? agentDefaults.compaction.qualityGuard
    : undefined;
const qualityGuardEntry =
  entry.compaction.qualityGuard && typeof entry.compaction.qualityGuard === "object"
    ? entry.compaction.qualityGuard
    : undefined;

qualityGuard: qualityGuardDefaults || qualityGuardEntry
  ? { ...(qualityGuardDefaults ?? {}), ...(qualityGuardEntry ?? {}) }
  : undefined,

const memoryFlushDefaults =
  agentDefaults?.compaction?.memoryFlush && typeof agentDefaults.compaction.memoryFlush === "object"
    ? agentDefaults.compaction.memoryFlush
    : undefined;
const memoryFlushEntry =
  entry.compaction.memoryFlush && typeof entry.compaction.memoryFlush === "object"
    ? entry.compaction.memoryFlush
    : undefined;

memoryFlush: memoryFlushDefaults || memoryFlushEntry
  ? { ...(memoryFlushDefaults ?? {}), ...(memoryFlushEntry ?? {}) }
  : undefined,

Also consider rejecting null in schema (e.g., ensure .nullable() is not allowed) and/or use z.object(...).optional() which already avoids null unless upstream coercion introduces it.

Best possible solution:

Close this branch and converge the remaining per-agent compaction contract through #83637 and the canonical issue it closes.

Do we have a high-confidence way to reproduce the issue?

Yes. Source inspection shows current main rejects agents.list[].compaction in the strict agent schema, and PR-head inspection shows the branch's remaining schema/runtime gaps.

Is this the best way to solve the issue?

No. The feature direction is valid, but this branch is no longer the best merge vehicle because the active replacement PR owns the broader contract and proof path.

Security review:

Security review cleared: No concrete security or supply-chain issue was found in the final PR surface; the concerns are functional config compatibility and runtime scope.

What I checked:

• Live PR state: GitHub reports this PR open at head eadf56e with mergeStateStatus DIRTY; the body lists local checks but does not attach inspectable after-fix terminal output, logs, screenshots, recordings, or linked artifacts. (eadf56e7c978)
• Current main defaults contract: Current main accepts defaults compaction keys including midTurnPrecheck, memoryFlush.model, truncateAfterCompaction, and maxActiveTranscriptBytes under agents.defaults.compaction. (src/config/zod-schema.agent-defaults.ts:165, 29f8715f05c8)
• Current main still lacks public per-agent schema: Current main's strict AgentEntrySchema does not include a compaction property, so the feature request remains real but not implemented on main by this branch. (src/config/zod-schema.agent-runtime.ts:1004, 29f8715f05c8)
• PR schema is narrower than current defaults: The PR's extracted AgentCompactionSchema omits current defaults keys such as midTurnPrecheck, memoryFlush.model, truncateAfterCompaction, and maxActiveTranscriptBytes, so merging it as-is would narrow existing accepted config. (src/config/zod-schema.agent-compaction.ts:29, eadf56e7c978)
• PR accepts per-agent compaction before every consumer is effective: The PR adds compaction: AgentCompactionSchema to AgentEntrySchema, but the same head still leaves consumers such as compaction timeout and post-compaction sections reading defaults-only config. (src/config/zod-schema.agent-runtime.ts:833, eadf56e7c978)
• Defaults-only timeout consumer on PR head: resolveCompactionTimeoutMs() on the PR head still reads cfg?.agents?.defaults?.compaction?.timeoutSeconds, so accepted per-agent timeout config would not be honored by that path. (src/agents/pi-embedded-runner/compaction-safety-timeout.ts:18, eadf56e7c978)

Likely related people:

• vincentkoc: The central effective-agent resolver path traces to the agent-scope helper extraction that this feature would naturally extend. (role: agent-scope refactor author; confidence: high; commits: a372e4a15240; files: src/agents/agent-scope-config.ts, src/agents/agent-scope.ts)
• Peter Steinberger: Recent history touched defaults compaction, memory flush, and release/current-main config surfaces that this PR must preserve. (role: recent config and compaction area contributor; confidence: medium; commits: 3a7a67b21873, e0dfc776bba8, 50a2481652b6; files: src/config/zod-schema.agent-defaults.ts, src/config/types.agent-defaults.ts, src/auto-reply/reply/memory-flush.ts)
• asyncjason: The defaults-only timeout consumer relevant to this PR traces to the configurable compaction timeout change. (role: introduced related compaction timeout behavior; confidence: medium; commits: f77a6841317c; files: src/agents/pi-embedded-runner/compaction-safety-timeout.ts, src/config/zod-schema.agent-defaults.ts)
• Alex Knight: Current-main blame for several central files points through the latest managed gateway update snapshot, so Alex is a useful recent-routing signal but not necessarily the feature origin. (role: recent baseline contributor; confidence: low; commits: c81271ee6e3a; files: src/config/zod-schema.agent-defaults.ts, src/agents/agent-scope-config.ts, src/agents/pi-settings.ts)

Codex review notes: model gpt-5.5, reasoning high; reviewed against 29f8715f05c8.