#19329: feat: add per-agent compaction and context pruning overrides

# PR: Per-agent compaction and context pruning overrides ## Title feat: add per-agent compaction and context pruning overrides ## Body ### Summary Adds per-agent `compaction` and `contextPruning` configuration to `agents.list[]` entries, allowing different agents to use different compaction behavior. Currently, all agents share the same global `agents.defaults.compaction` settings — this is a significant gap when running agents on models with different context windows or when agents have fundamentally different session lifetimes. ### Motivation A multi-agent setup might include: - A **coding agent** on a large-context model that should compact aggressively to stay fast - A **companion agent** that should never compact (managed by a plugin or external process) - A **research agent** with a higher `reserveTokensFloor` to preserve more working context Today, all of these agents share identical compaction behavior. This PR lets each agent override: - `compaction.mode` — `"default"`, `"safeguard"`, or `"off"` (new) - `compaction.reserveTokensFloor` — per-agent token reserve - `compaction.maxHistoryShare` — per-agent history share limit - `compaction.memoryFlush` — per-agent flush prompt, system prompt, threshold, and enabled flag - `contextPruning` — per-agent pruning mode, TTL, and keepLastAssistants ### Example config ```json5 { agents: { defaults: { compaction: { mode: "safeguard", reserveTokensFloor: 20000, memoryFlush: { enabled: true } } }, list: [ { id: "researcher", compaction: { reserveTokensFloor: 80000 // higher floor, inherits mode + memoryFlush from defaults } }, { id: "plugin-managed", compaction: { mode: "off" // compaction handled externally } }, { id: "companion", compaction: { memoryFlush: { prompt: "Save emotional context, personal stories, and relationship details." } } } ] } } ``` ### What changed **Schema & types** (commit 1): - Extracted `CompactionSchema` from `zod-schema.agent-defaults.ts` into `zod-schema.agent-runtime.ts` for reuse - Added `compaction` and `contextPruning` fields to `AgentEntrySchema` - Added `"off"` to `AgentCompactionMode` type **Resolvers** (commit 2): - `resolveAgentCompaction(cfg, agentId?)` — deep merges per-agent overrides with global defaults; `memoryFlush` sub-object is shallow-merged one level deeper so partial overrides don't wipe sibling fields - `resolveAgentContextPruning(cfg, agentId?)` — merges per-agent context pruning with global defaults - `isCompactionEnabled(cfg, agentId?)` — guard helper that encapsulates the `mode !== "off"` check **Callsite wiring** (commit 3, 6 files): - `compact.ts` — skips `ensurePiCompactionReserveTokens` when mode is `"off"` - `attempt.ts` — same skip logic for run attempts - `extensions.ts` — uses resolved per-agent config for both compaction mode and context pruning - `pi-settings.ts` — `resolveCompactionReserveTokensFloor` now takes optional `agentId` - `memory-flush.ts` — `resolveMemoryFlushSettings` uses per-agent compaction config; returns `null` when mode is `"off"` - `agent-runner-memory.ts` — extracts `agentId` from session key and passes it through **Tests** (commit 4): - 38 tests covering all resolver functions, merge semantics, mode combinations, edge cases (missing agent, empty overrides, partial memoryFlush, undefined spread safety) **Review feedback** (commit 5): - Removed dead code in `buildContextPruningExtension` - Widened per-agent `contextPruning` type to include `ttl` and `keepLastAssistants` - Used named `AgentContextPruningConfig` type instead of inline return type ### Integration tested Tested on a live gateway with a dedicated test agent (`claude-haiku-4-5`, 200k context): | Test | Config | Result | |------|--------|--------| | Custom floor | `reserveTokensFloor: 80000` | Compacted at ~120k used (98k → 17k) ✅ | | Mode off | `mode: "off"` | No compaction at 222k+ tokens ✅ | ### Breaking changes None. All new fields are optional. Agents without per-agent overrides behave exactly as before (global defaults apply). ### Checklist - [x] Schema extraction (no duplication between defaults and per-agent) - [x] Deep merge with memoryFlush sub-object handling - [x] `mode: "off"` support with guard helper - [x] All callsites wired - [x] 38 unit tests passing - [x] Type check clean - [x] Integration tested on live gateway - [x] No unrelated changes in diff <!-- greptile_comment --> <h3>Greptile Summary</h3> Adds per-agent `compaction` and `contextPruning` configuration overrides to `agents.list[]` entries, enabling different agents to use different compaction behaviors. The implementation introduces a new `"off"` mode for compaction, along with resolver functions that deep-merge per-agent settings with global defaults. **Key changes:** - Extracted `CompactionSchema` for reuse between `AgentDefaultsSchema` and `AgentEntrySchema` - Added `resolveAgentCompaction()`, `resolveAgentContextPruning()`, and `isCompactionEnabled()` helpers with proper deep-merge semantics for `memoryFlush` sub-object - Wired per-agent config through all callsites (`compact.ts`, `attempt.ts`, `extensions.ts`, `pi-settings.ts`, `memory-flush.ts`, `agent-runner-memory.ts`) - Added comprehensive test coverage (38 tests covering merge semantics, edge cases, `mode: "off"` behavior) - Migrated existing e2e tests to unit tests in `agent-scope.test.ts` The implementation follows the existing pattern for per-agent model overrides and maintains backward compatibility - all new fields are optional and default to global settings when not specified. <h3>Confidence Score: 5/5</h3> - This PR is safe to merge with minimal risk. - The implementation is well-architected with proper deep-merge semantics for nested config objects, comprehensive test coverage (38 tests), clean extraction of shared schema components, and systematic wiring through all callsites. The `mode: "off"` feature is correctly guarded at every integration point. All changes maintain backward compatibility. - No files require special attention _{Last reviewed commit: 0a2dce7} <!-- greptile_other_comments_section --> _{(3/5) Reply to the agent's comments like "Can you suggest a fix for this @greptileai?" or ask follow-up questions!} <!-- /greptile_comment --> Closes #13736 Closes #17735 Relates to #16546 (this PR adds per-agent compaction; contextTokens override is a separate concern)