Accept wrapped memory agent selections

Author: SqlRush

docs/cc-100-roadmap.md

@@ -172,6 +172,8 @@ M6 补充：relevant memory prefetch 现在有可取消 runtime，并在 `Memory
 
 M6 补充：model-backed relevant memory selector prompt 现在包含 recent successful tools 和 already-surfaced memory paths 的有界上下文，模型侧选择与 deterministic prefilter 的 tool/surfaced 约束更一致。
 
+M6 补充：session-memory recall agent 和 relevant-memory selector 现在递归解包 `data`/`payload`/`body`、JSON:API `resource`/`attributes`/`properties`/`attrs`，以及 GraphQL `viewer`/`edge`/`node`/`nodes`/`edges` selection wrapper，API-shaped model response 中的 session IDs 和 memory paths 会按模型顺序保留。
+
 M6/M7 补充：Runner 会把 `RelevantMemoryDir` 透传到 tool metadata 的 internal auto-memory path context，让 Read tool freshness prefix 和 permission internal-path policy 在同一 memory dir 配置下生效。
 
 M6 补充：transcript resume 在 fallback 转换 attachment message 时会保留 raw attachment payload，确保恢复出的 `relevant_memories` attachment 仍可被 request 构建路径展开为 system-reminder。
@@ -502,6 +504,8 @@ M7 补充：prompt history `LogEntry` 读取现在接受 `sessionID`/`session`/`
 
 本轮补充：conversation runner 现在会在用户消息入队后计算 compact token warning state，并在触发 warning/error/auto-compact/blocking 阈值时发出 `token_warning` event；warning state 接入 blocking-limit override，auto-compact threshold 判断接入 `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE`，使 runtime warning 和自动压缩使用同一套 window 输入。
 
+本轮补充：session-memory recall agent 和 relevant-memory selector 现在递归解包 `data`/`payload`/`body`、JSON:API `resource`/`attributes`/`properties`/`attrs`，以及 GraphQL `viewer`/`edge`/`node`/`nodes`/`edges` selection wrapper，API-shaped model response 中的 session IDs 和 memory paths 会按模型顺序保留。
+
 本轮补充：microcompact disk cache loader 现在读取 Go 默认、camelCase、snake_case 和相邻实现常见字段别名/包装形态，包括 `result`/`data`/`cache`/`value` wrapper、`content`/`text` summary、`cacheKey`/`key`/`hash` digest、cache-hit 别名、计数字段别名/数字字符串、RFC3339/Unix 秒/Unix 毫秒时间字段，以及 `createdAt` + `ttlSeconds`/`ttlMs`/`expiresIn`/`maxAge` 等相对 TTL 推导，提升 cached microcompact 文件在不同实现/版本间的恢复率。
 
 本轮补充：microcompact disk cache loader 现在也接受 `cacheEntry`/`cache_entry`、`micro_compact`/`micro_compact_result` wrapper，以及 `summaryMarkdown`/`resultSummary`/`compressedText`、`cacheDigest`/`digestHash`/`fingerprint`、`summarizedCount`/`retainedCount`、`formatVersion` 和 `ttlMilliseconds`/`expiresInMilliseconds`/`maxAgeMilliseconds` 等相邻实现字段别名。

docs/claude-code-go-rewrite-plan.md

@@ -247,6 +247,7 @@ test/parity/                 # golden tests against TS/official behavior
 - 本轮补充：memory 层增加可取消 `PrefetchRelevantMemories` runtime，复用现有 gating/selector/surfacing 逻辑返回 plan、selection 和 attachments；conversation `RunTurn` 会在用户消息进入后启动 relevant memory prefetch，并在第一轮 model request 消费结果，预取文件系统错误 fail-open 且不阻断主请求。
 - 本轮补充：relevant memory prefetch 接入 model-backed sideQuery selector：当 `MemoryAgentClient` 可用时，先向模型提供候选 memory manifest，请模型返回 `memory_paths`/`memoryPaths`/`filePath`/`matches`/嵌套 selection 等路径别名，按模型顺序读取附件；模型错误或无效路径会 fail-open 回落 deterministic selector。完整官方 prompt/telemetry parity 仍继续推进。
 - 本轮补充：model-backed relevant memory selector prompt 现在包含 recent successful tools 和 already-surfaced memory paths 的有界上下文，模型侧选择与 deterministic prefilter 的 tool/surfaced 约束更一致。
+- 本轮补充：session-memory recall agent 和 relevant-memory selector 现在递归解包 `data`/`payload`/`body`、JSON:API `resource`/`attributes`/`properties`/`attrs`，以及 GraphQL `viewer`/`edge`/`node`/`nodes`/`edges` selection wrapper，API-shaped model response 中的 session IDs 和 memory paths 会按模型顺序保留。
 - 本轮补充：conversation runner 会在用户消息入队后基于 compact window 计算 token warning state，并在达到 warning/error/auto-compact/blocking 阈值时发出 `token_warning` event；warning state 接入 blocking-limit env override，auto-compact 判断接入既有 `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE`，避免 warning 和 compact 阈值来源分叉。
 - 本轮补充：microcompact disk cache loader 现在读取 Go 默认、camelCase 和 snake_case 字段别名，并容忍计数字段的数字字符串，以及 RFC3339、Unix 秒、Unix 毫秒时间字段，提升 cached microcompact 文件在不同实现/版本间的恢复率。
 - 本轮补充：microcompact disk cache loader 现在接受 `cacheEntry`/`cache_entry`、`micro_compact`/`micro_compact_result` wrapper，以及 `summaryMarkdown`/`resultSummary`/`compressedText`、`cacheDigest`/`digestHash`/`fingerprint`、`summarizedCount`/`retainedCount`、`formatVersion` 和 `ttlMilliseconds`/`expiresInMilliseconds`/`maxAgeMilliseconds` 等相邻实现字段别名。

docs/first-second-parity-audit.md

@@ -137,6 +137,7 @@ M6 progress now includes:
 - `internal/compact`: microcompact disk cache loading now treats a `value` text content-block object as a direct summary payload while still recovering digest, version, timestamp, and related cache metadata from that same `value` sidecar object when present.
 - `internal/conversation`: optional auto-compact can now run before the main request, fail open with a consecutive-failure circuit breaker, emit runtime `token_warning` events from compact warning state, persist compact boundary metadata to transcript, write a session-memory summary, inject recalled session-memory snippets into API request context when enabled, and optionally extract turn-end memory facts into session memory.
 - `internal/memory`: session-memory summaries can now be loaded with frontmatter field aliases, rolled up/pruned into an archive summary with prior rollup archives excluded from candidate compaction and merged across archive IDs, rune-safe rollup truncation, recalled by query with deterministic scoring and recency ordering, ranked through a model-assisted candidate/session-id selection path with fallback when selected IDs are invalid or excluded, alternate/camel response-key parsing, fenced/prose JSON extraction, scalar session-id parsing, nested/wrapped/collection-alias selection parsing, and nested selected-memory item parsing, and injected into resume context through the same optional recall agent path; deterministic and model-backed memory fact extraction can summarize user preferences, requests, decisions, and tool-use facts, including fenced/prose JSON, wrapped facts responses, alternate and structured fact field names, nested source objects, nested fact response shapes, and fact kind aliases.
+- `internal/memory`: session-memory recall and relevant-memory model selectors now recurse through GraphQL/JSON:API selection wrappers such as `data`, `payload`, `body`, `resource`, `attributes`, `properties`, `attrs`, `viewer`, `edge`, `node`, `nodes`, and `edges`, preserving model-ranked session IDs and memory paths from API-shaped responses.
 - `internal/session`: project session listing and pagination, prompt-history lock/buffered flush/field aliases, lightweight transcript index/title/text-preview inference, line-offset transcript indexing/window/byte-budget-window/parent-chain/resume/tail/byte-budget-tail loading, AI-title/last-prompt/task/agent/PR/worktree transcript metadata loading and type/field aliases, transcript message/session UUID field aliases, tombstone metadata loading with target/session/reason aliases and delete/relink replay, agent-scoped content replacement metadata/record field-alias loading, session-scoped metadata reappend including AI-title/last-prompt/task-summary, streaming transcript search snippets for resume/search UI, official `subagents/agent-*.jsonl` transcript layout with legacy sidechain listing, agent metadata sidecar read/write/field aliases, sidechain runtime start/append/finish/cancel/fail summary bridging plus parent-chain append/finish, sidechain manager orchestration for spawn/append/finish/cancel/fail/list/resume, sidechain state/list/resume support with content-field aliases, sidechain resume context construction, sidechain conversation and agent-scoped content-replacement reconstruction, transcript tail/window loading and byte-budget tail loading for bounded-memory resume/UI paths, lightweight transcript metadata loading, remote-history token refresh retry, page-field/event-list/records/entries/last-id/cursor/event-id/has-next aliases plus wrapped-data/links/paging/bare-array/keyed-event-map responses, before_id resume, fallback field fill, and duplicate-aware parent linking during transcript materialization.
 - `internal/session`: transcript full loader and line-offset index now accept top-level record UUID aliases such as `messageUuid`, `messageUUID`, `message_id`, `messageId`, and `id`, plus parent message UUID/ID aliases for progress bridge and indexed resume paths.
 - `internal/session`: indexed resume chains now distinguish budget-truncated parents from genuinely missing parents, exposing both markers while preserving bounded transcript reads.

internal/memory/agent.go

@@ -525,8 +525,22 @@ func parseRecallAgentJSON(raw string) (string, []contracts.ID, bool) {
 		CandidateMemoriesCamel  []json.RawMessage `json:"candidateMemories"`
 		Candidates              []json.RawMessage `json:"candidates"`
 		Results                 []json.RawMessage `json:"results"`
+		Nodes                   []json.RawMessage `json:"nodes"`
+		Edges                   []json.RawMessage `json:"edges"`
+		Items                   []json.RawMessage `json:"items"`
+		Resources               []json.RawMessage `json:"resources"`
 		Selection               json.RawMessage   `json:"selection"`
 		Selected                json.RawMessage   `json:"selected"`
+		Data                    json.RawMessage   `json:"data"`
+		Payload                 json.RawMessage   `json:"payload"`
+		Body                    json.RawMessage   `json:"body"`
+		Resource                json.RawMessage   `json:"resource"`
+		Attributes              json.RawMessage   `json:"attributes"`
+		Properties              json.RawMessage   `json:"properties"`
+		Attrs                   json.RawMessage   `json:"attrs"`
+		Viewer                  json.RawMessage   `json:"viewer"`
+		Edge                    json.RawMessage   `json:"edge"`
+		Node                    json.RawMessage   `json:"node"`
 		Result                  json.RawMessage   `json:"result"`
 		Response                json.RawMessage   `json:"response"`
 		Recall                  json.RawMessage   `json:"recall"`
@@ -584,6 +598,10 @@ func parseRecallAgentJSON(raw string) (string, []contracts.ID, bool) {
 				object.CandidateMemoriesCamel,
 				object.Candidates,
 				object.Results,
+				object.Nodes,
+				object.Edges,
+				object.Items,
+				object.Resources,
 			)
 		}
 		query := strings.TrimSpace(object.Query)
@@ -606,7 +624,25 @@ func parseRecallAgentJSON(raw string) (string, []contracts.ID, bool) {
 			query = strings.TrimSpace(object.ExpandedQueryCamel)
 		}
 		if len(ids) == 0 {
-			query, ids = recallSelectionFromNestedPayloads(query, object.Selection, object.Selected, object.Result, object.Response, object.Recall, object.MemoryRecall, object.MemoryRecallCamel)
+			query, ids = recallSelectionFromNestedPayloads(query,
+				object.Selection,
+				object.Selected,
+				object.Data,
+				object.Payload,
+				object.Body,
+				object.Resource,
+				object.Attributes,
+				object.Properties,
+				object.Attrs,
+				object.Viewer,
+				object.Edge,
+				object.Node,
+				object.Result,
+				object.Response,
+				object.Recall,
+				object.MemoryRecall,
+				object.MemoryRecallCamel,
+			)
 		}
 		return query, ids, query != "" || len(ids) > 0
 	}
@@ -646,6 +682,20 @@ func parseRelevantMemoryAgentJSON(raw string) (string, []string, bool) {
 			query, ids = relevantMemorySelectionFromNestedPayloads(query, rawObjectValues(rawObject,
 				"selected",
 				"selection",
+				"data",
+				"payload",
+				"body",
+				"resource",
+				"attributes",
+				"properties",
+				"attrs",
+				"viewer",
+				"edge",
+				"node",
+				"nodes",
+				"edges",
+				"items",
+				"resources",
 				"result",
 				"response",
 				"memory_selection",
@@ -702,8 +752,22 @@ func parseRelevantMemoryAgentJSON(raw string) (string, []string, bool) {
 		Memories                     []json.RawMessage `json:"memories"`
 		FilesList                    []json.RawMessage `json:"files_list"`
 		FilesListCamel               []json.RawMessage `json:"filesList"`
+		Nodes                        []json.RawMessage `json:"nodes"`
+		Edges                        []json.RawMessage `json:"edges"`
+		Items                        []json.RawMessage `json:"items"`
+		Resources                    []json.RawMessage `json:"resources"`
 		Selected                     json.RawMessage   `json:"selected"`
 		Selection                    json.RawMessage   `json:"selection"`
+		Data                         json.RawMessage   `json:"data"`
+		Payload                      json.RawMessage   `json:"payload"`
+		Body                         json.RawMessage   `json:"body"`
+		Resource                     json.RawMessage   `json:"resource"`
+		Attributes                   json.RawMessage   `json:"attributes"`
+		Properties                   json.RawMessage   `json:"properties"`
+		Attrs                        json.RawMessage   `json:"attrs"`
+		Viewer                       json.RawMessage   `json:"viewer"`
+		Edge                         json.RawMessage   `json:"edge"`
+		Node                         json.RawMessage   `json:"node"`
 		Result                       json.RawMessage   `json:"result"`
 		Response                     json.RawMessage   `json:"response"`
 		MemorySelection              json.RawMessage   `json:"memory_selection"`
@@ -753,6 +817,10 @@ func parseRelevantMemoryAgentJSON(raw string) (string, []string, bool) {
 				object.Memories,
 				object.FilesList,
 				object.FilesListCamel,
+				object.Nodes,
+				object.Edges,
+				object.Items,
+				object.Resources,
 			)
 		}
 		query := strings.TrimSpace(object.Query)
@@ -778,6 +846,16 @@ func parseRelevantMemoryAgentJSON(raw string) (string, []string, bool) {
 			query, ids = relevantMemorySelectionFromNestedPayloads(query,
 				object.Selected,
 				object.Selection,
+				object.Data,
+				object.Payload,
+				object.Body,
+				object.Resource,
+				object.Attributes,
+				object.Properties,
+				object.Attrs,
+				object.Viewer,
+				object.Edge,
+				object.Node,
 				object.Result,
 				object.Response,
 				object.MemorySelection,
@@ -1092,6 +1170,26 @@ var recallNestedItemKeys = []string{
 	"candidateSession",
 	"candidate_memory",
 	"candidateMemory",
+	"data",
+	"payload",
+	"body",
+	"result",
+	"response",
+	"resource",
+	"attributes",
+	"properties",
+	"attrs",
+	"viewer",
+	"edge",
+	"node",
+	"nodes",
+	"edges",
+	"items",
+	"resources",
+	"record",
+	"entry",
+	"item",
+	"value",
 }
 
 var relevantMemoryNestedItemKeys = []string{
@@ -1106,6 +1204,26 @@ var relevantMemoryNestedItemKeys = []string{
 	"relevantMemory",
 	"candidate_memory",
 	"candidateMemory",
+	"data",
+	"payload",
+	"body",
+	"result",
+	"response",
+	"resource",
+	"attributes",
+	"properties",
+	"attrs",
+	"viewer",
+	"edge",
+	"node",
+	"nodes",
+	"edges",
+	"items",
+	"resources",
+	"record",
+	"entry",
+	"item",
+	"value",
 }
 
 func stripMarkdownFence(raw string) string {

internal/memory/memory_test.go

@@ -442,6 +442,44 @@ func TestMemoryAgentSelectRelevantMemoriesParsesNestedAliasesAndFallsBack(t *tes
 	}
 }
 
+func TestMemoryAgentSelectRelevantMemoriesParsesGraphQLResourceSelections(t *testing.T) {
+	dir := t.TempDir()
+	dbPath := filepath.Join(dir, "db.md")
+	opsPath := filepath.Join(dir, "ops.md")
+	writeFile(t, dbPath, "---\ndescription: database permissions migration\n---\ndb rules\n")
+	writeFile(t, opsPath, "---\ndescription: deployment runbook\n---\nops rules\n")
+	client := &fakeMemoryClient{response: &anthropic.Response{
+		ID:    "msg_memory_graphql",
+		Type:  "message",
+		Role:  "assistant",
+		Model: "sonnet",
+		Content: []contracts.ContentBlock{contracts.NewTextBlock(`{
+			"data":{
+				"viewer":{
+					"memorySelection":{
+						"query":"database access",
+						"edges":[
+							{"node":{"attrs":{"filePath":"ops.md"}}},
+							{"edge":{"node":{"properties":{"memoryPath":"db.md"}}}}
+						]
+					}
+				}
+			}
+		}`)},
+	}}
+
+	result, err := (Agent{Client: client}).SelectRelevantMemories(context.Background(), dir, "database permissions", RelevantMemorySelectorOptions{Limit: 2})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Fallback || result.Query != "database access" || strings.Join(result.SelectedIDs, ",") != "ops.md,db.md" {
+		t.Fatalf("result = %#v", result)
+	}
+	if len(result.Selected) != 2 || result.Selected[0].Path != opsPath || result.Selected[1].Path != dbPath {
+		t.Fatalf("selected = %#v", result.Selected)
+	}
+}
+
 func TestPrefetchRelevantMemoriesCanUseMemoryAgentSelector(t *testing.T) {
 	dir := t.TempDir()
 	dbPath := filepath.Join(dir, "db.md")
@@ -1291,6 +1329,56 @@ func TestMemoryAgentRecallParsesWrappedSelectionsAndScalarID(t *testing.T) {
 	}
 }
 
+func TestMemoryAgentRecallParsesGraphQLResourceSelections(t *testing.T) {
+	root := filepath.Join(t.TempDir(), "session-memory")
+	for _, item := range []struct {
+		id      contracts.ID
+		summary string
+		updated int64
+	}{
+		{id: "prior", summary: "database access policy notes", updated: 200},
+		{id: "other", summary: "credential rotation notes", updated: 100},
+	} {
+		if _, err := WriteSessionSummary(SessionSummaryOptions{
+			Root:      root,
+			SessionID: item.id,
+			Summary:   item.summary,
+			UpdatedAt: time.Unix(item.updated, 0).UTC(),
+		}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	client := &fakeMemoryClient{response: &anthropic.Response{
+		ID:    "msg_recall_graphql",
+		Type:  "message",
+		Role:  "assistant",
+		Model: "sonnet",
+		Content: []contracts.ContentBlock{contracts.NewTextBlock(`{
+			"data":{
+				"viewer":{
+					"memoryRecall":{
+						"searchQuery":"database access",
+						"edges":[
+							{"node":{"attributes":{"sessionId":"prior"}}},
+							{"edge":{"node":{"properties":{"summaryId":"other"}}}}
+						]
+					}
+				}
+			}
+		}`)},
+	}}
+	result, err := (Agent{Client: client}).Recall(context.Background(), root, "what did we decide about db access?", RecallOptions{Limit: 2})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.Fallback || result.Query != "database access" || strings.Join(contractIDStrings(result.SelectedIDs), ",") != "prior,other" {
+		t.Fatalf("result = %#v", result)
+	}
+	if len(result.Matches) != 2 || result.Matches[0].Summary.SessionID != "prior" || result.Matches[1].Summary.SessionID != "other" {
+		t.Fatalf("matches = %#v", result.Matches)
+	}
+}
+
 func TestMemoryAgentRecallExtractsFencedJSONFromModelProse(t *testing.T) {
 	root := filepath.Join(t.TempDir(), "session-memory")
 	if _, err := WriteSessionSummary(SessionSummaryOptions{