Only last state change for multiple tool calls is being committed to state

[verdverm]

I have some custom functiontools and the LLM is calling it in batch via message parts. In those functiontools, I update the state, which is what the docs say to do: https://google.github.io/adk-docs/sessions/state/#a-warning-about-direct-state-modification

It looks like only the last tool call's state is being applied to the delta. I have confirmed the tools is called many times in one turn

Example state modifying functiontool

const CacheRemoveTool = "cache_remove"
const CacheRemoveDesc = `
adds or overwrites an entry in your working key/value cache and context
`

type CacheRemoveArgs struct {
	Key string `json:"key"` // path to a directory
}
type CacheRemoveResult struct {
	Key    string `json:"key"`    // path to a directory
	Status string `json:"status"` // "ok" or "error"
	Error  string `json:"error,omitempty"`
}

func NewCacheRemove() (tool.Tool, error) {
	handler := func(ctx tool.Context, input CacheRemoveArgs) (CacheRemoveResult, error) {
		k := fmt.Sprintf("cache:%s:%s", ctx.AgentName(), input.Key)
		fmt.Println("CACHE.REMOVE!", k)

		err := ctx.State().Set(k, nil)
		if err != nil {
			return CacheRemoveResult{Status: "error", Key: input.Key, Error: err.Error()}, err
		}
		return CacheRemoveResult{Status: "ok", Key: input.Key}, nil
	}
	return functiontool.New(functiontool.Config{
		Name:        CacheRemoveTool,
		Description: strings.TrimSpace(CacheRemoveDesc),
	}, handler)
}

Event with bulk functionCall

{
  "Content": {
    "parts": [
      {
        "text": "That is an insightful observation. Keeping the key with a marker like \"no relevant\" is a good way to maintain a memory of explored files without cluttering the context with content.\n\nI will now attempt to update a batch of keys using `cache_write` to see if it suffers from the same \"only the last command commits\" bug. I will use the marker value `\"no relevant\"` for the keys that currently show `<no value>` but have been examined.\n\nI will use the following keys for the test:\n1.  `package.json`\n2.  `src/main.tsx`\n3.  `src/App.tsx`\n4.  `src/vscodeApi.ts` (This will be the final command to see if it's the only one that persists)\n\n"
      },
      {
        "functionCall": {
          "id": "adk-3caaaeab-b818-4967-8a06-cf57908d0a72",
          "args": {
            "key": "package.json",
            "value": "no relevant"
          },
          "name": "cache_write"
        }
      },
      {
        "functionCall": {
          "id": "adk-7b44dc4e-205d-4882-b8c3-63dcec005eeb",
          "args": {
            "key": "src/main.tsx",
            "value": "no relevant"
          },
          "name": "cache_write"
        }
      },
      {
        "functionCall": {
          "id": "adk-60a05b82-cbfd-441f-a565-fd966db0d92f",
          "args": {
            "key": "src/App.tsx",
            "value": "no relevant"
          },
          "name": "cache_write"
        }
      },
      {
        "functionCall": {
          "id": "adk-dd68d7e9-08ee-4ba1-96d1-06784bdd4e89",
          "args": {
            "key": "src/vscodeApi.ts",
            "value": "no relevant"
          },
          "name": "cache_write"
        }
      }
    ],
    "role": "model"
  },
  "CitationMetadata": null,
  "GroundingMetadata": null,
  "UsageMetadata": {
    "cacheTokensDetails": [
      {
        "modality": "TEXT",
        "tokenCount": 6772
      }
    ],
    "cachedContentTokenCount": 6772,
    "candidatesTokenCount": 284,
    "promptTokenCount": 7262,
    "promptTokensDetails": [
      {
        "modality": "TEXT",
        "tokenCount": 7262
      }
    ],
    "totalTokenCount": 7546
  },
  "CustomMetadata": null,
  "LogprobsResult": null,
  "Partial": false,
  "TurnComplete": false,
  "Interrupted": false,
  "ErrorCode": "",
  "ErrorMessage": "",
  "FinishReason": "",
  "AvgLogprobs": 0,
  "ID": "2bf8cec9-5805-4209-832e-0eb085ef133a",
  "Timestamp": "2025-11-24T21:00:10.756461-08:00",
  "InvocationID": "e-f9ed648d-d584-49f5-9c5d-ebaa54c5ef07",
  "Branch": "",
  "Author": "hack",
  "Actions": {
    "StateDelta": {},
    "ArtifactDelta": null,
    "SkipSummarization": false,
    "TransferToAgent": "",
    "Escalate": false
  },
  "LongRunningToolIDs": null
}

Event with bulk functionResponse

{
  "Content": {
    "parts": [
      {
        "functionResponse": {
          "id": "adk-3caaaeab-b818-4967-8a06-cf57908d0a72",
          "name": "cache_write",
          "response": {
            "key": "package.json",
            "status": "ok"
          }
        }
      },
      {
        "functionResponse": {
          "id": "adk-7b44dc4e-205d-4882-b8c3-63dcec005eeb",
          "name": "cache_write",
          "response": {
            "key": "src/main.tsx",
            "status": "ok"
          }
        }
      },
      {
        "functionResponse": {
          "id": "adk-60a05b82-cbfd-441f-a565-fd966db0d92f",
          "name": "cache_write",
          "response": {
            "key": "src/App.tsx",
            "status": "ok"
          }
        }
      },
      {
        "functionResponse": {
          "id": "adk-dd68d7e9-08ee-4ba1-96d1-06784bdd4e89",
          "name": "cache_write",
          "response": {
            "key": "src/vscodeApi.ts",
            "status": "ok"
          }
        }
      }
    ],
    "role": "user"
  },
  "CitationMetadata": null,
  "GroundingMetadata": null,
  "UsageMetadata": null,
  "CustomMetadata": null,
  "LogprobsResult": null,
  "Partial": false,
  "TurnComplete": false,
  "Interrupted": false,
  "ErrorCode": "",
  "ErrorMessage": "",
  "FinishReason": "",
  "AvgLogprobs": 0,
  "ID": "d635feb2-4f3a-4353-9cb5-80b713e567b8",
  "Timestamp": "2025-11-24T21:00:10.76217-08:00",
  "InvocationID": "e-f9ed648d-d584-49f5-9c5d-ebaa54c5ef07",
  "Branch": "",
  "Author": "hack",
  "Actions": {
    "StateDelta": {
      "cache:hack:src/vscodeApi.ts": "no relevant"
    },
    "ArtifactDelta": null,
    "SkipSummarization": false,
    "TransferToAgent": "",
    "Escalate": false
  },
  "LongRunningToolIDs": null
}

Multiple tool calls in the logs

Chatting payload: handlers.ChatPayload{Text:"ok, we can see the right logs now. Try a bulk operation to write \"not relevant\" to all the cache keys", Sid:"065471a3-969a-4dc1-bc49-a3a4bcf9bcc3", Agent:"hack", Model:"gemini-2.5-flash"}
userMsg &{[0x140004ea230] user} map[env:map[clipboard:WRITE machineId:cf912dbf5f56042166332d2be6fdf1e21f8155680986088029fbcbf130665dda sid:065471a3-969a-4dc1-bc49-a3a4bcf9bcc3 user:verdverm vscodeSid:0ac5806d-099c-44ea-aa35-ea7c7005b2fc1764048527260 workspaceDir:/Users/tony/hof/hof/extensions/vscode/webviews/chat]]

BAC.hack
renderInstructions.Agent hack

BMC.hack

AMC.hack
model.LLMResponse{Content:(*genai.Content)(0x14000b5cde0), CitationMetadata:(*genai.CitationMetadata)(nil), GroundingMetadata:(*genai.GroundingMetadata)(nil), UsageMetadata:(*genai.GenerateContentResponseUsageMetadata)(0x14000858750), CustomMetadata:map[string]interface {}(nil), LogprobsResult:(*genai.LogprobsResult)(nil), Partial:false, TurnComplete:false, Interrupted:false, ErrorCode:"", ErrorMessage:"", FinishReason:"STOP", AvgLogprobs:0}
chat.event: &{{0x14000b5cde0 <nil> <nil> 0x14000858750 map[] <nil> false false false   STOP 0} d340e2dd-ed3b-4e8c-8376-5c89c45f46e0 2025-11-24 21:30:43.717199 -0800 PST m=+123.955414043 e-32938734-d387-4cfb-afe2-3de23b201181  hack {map[] map[] false  false} []}

BTC.hack.cache_write map[key:. value:not relevant]
CACHE.WRITE: cache:hack:.

ATC.hack.cache_write map[key:. value:not relevant] map[key:. status:ok]

BTC.hack.cache_write map[key:package.json value:not relevant]
CACHE.WRITE: cache:hack:package.json

ATC.hack.cache_write map[key:package.json value:not relevant] map[key:package.json status:ok]

BTC.hack.cache_write map[key:src value:not relevant]
CACHE.WRITE: cache:hack:src

ATC.hack.cache_write map[key:src value:not relevant] map[key:src status:ok]

BTC.hack.cache_write map[key:src/App.tsx value:not relevant]
CACHE.WRITE: cache:hack:src/App.tsx

ATC.hack.cache_write map[key:src/App.tsx value:not relevant] map[key:src/App.tsx status:ok]

BTC.hack.cache_write map[key:src/components value:not relevant]
CACHE.WRITE: cache:hack:src/components

ATC.hack.cache_write map[key:src/components value:not relevant] map[key:src/components status:ok]

BTC.hack.cache_write map[key:src/components/Header.tsx value:not relevant]
CACHE.WRITE: cache:hack:src/components/Header.tsx

ATC.hack.cache_write map[key:src/components/Header.tsx value:not relevant] map[key:src/components/Header.tsx status:ok]

BTC.hack.cache_write map[key:src/components/Messages.tsx value:not relevant]
CACHE.WRITE: cache:hack:src/components/Messages.tsx

ATC.hack.cache_write map[key:src/components/Messages.tsx value:not relevant] map[key:src/components/Messages.tsx status:ok]

BTC.hack.cache_write map[key:src/main.tsx value:not relevant]
CACHE.WRITE: cache:hack:src/main.tsx

ATC.hack.cache_write map[key:src/main.tsx value:not relevant] map[key:src/main.tsx status:ok]

BTC.hack.cache_write map[key:src/vscodeApi.ts value:not relevant]
CACHE.WRITE: cache:hack:src/vscodeApi.ts

ATC.hack.cache_write map[key:src/vscodeApi.ts value:not relevant] map[key:src/vscodeApi.ts status:ok]
chat.event: &{{0x14000a919e0 <nil> <nil> <nil> map[] <nil> false false false    0} 504e260a-8bc1-4b5a-91ae-4790f149eacf 2025-11-24 21:30:43.722247 -0800 PST m=+123.960462543 e-32938734-d387-4cfb-afe2-3de23b201181  hack {map[cache:hack:src/vscodeApi.ts:not relevant] map[] false  false} []}
renderInstructions.Agent hack

BMC.hack

AMC.hack
model.LLMResponse{Content:(*genai.Content)(0x14000c19470), CitationMetadata:(*genai.CitationMetadata)(nil), GroundingMetadata:(*genai.GroundingMetadata)(nil), UsageMetadata:(*genai.GenerateContentResponseUsageMetadata)(0x14000b0c240), CustomMetadata:map[string]interface {}(nil), LogprobsResult:(*genai.LogprobsResult)(nil), Partial:false, TurnComplete:false, Interrupted:false, ErrorCode:"", ErrorMessage:"", FinishReason:"STOP", AvgLogprobs:0}
chat.event: &{{0x14000c19470 <nil> <nil> 0x14000b0c240 map[] <nil> false false false   STOP 0} b8b6d228-20c2-4732-bd97-f2db9822ced8 2025-11-24 21:30:44.467564 -0800 PST m=+124.705775334 e-32938734-d387-4cfb-afe2-3de23b201181  hack {map[] map[] false  false} []}

AAC.hack

Legend:

BAC: BeforeAgentCallback
BMC: BeforeModelCallback
BTC: BeforeToolCallback

ATC: AfterToolCallback
AMC: AfterModelCallback
AAC: AfterAgentCallback

[verdverm]

Follow up, I looked into how the function call responses are merged

1. It doesn't look like State is at all
2. I changed my code to call ctx.Actions().EventState[key]=val, this exhibits the same behavior
3. I changed the merge code, and it works as expected, this seems like the correct code to fix what I see as a bug (without enough context to know how it might impact other parts of the system)

https://github.com/google/adk-go/blob/main/internal/llminternal/base_flow.go#L498

func mergeEventActions(base, other *session.EventActions) *session.EventActions {
	// flows/llm_flows/functions.py merge_parallel_function_response_events
	//
	// TODO: merge_parallel_function_response_events creates a "last one wins" scenario
	// except parts and requested_auth_configs. Check with the ADK team about
	// the intention.
	if other == nil {
		return base
	}
	if base == nil {
		return other
	}
	if other.SkipSummarization {
		base.SkipSummarization = true
	}
	if other.TransferToAgent != "" {
		base.TransferToAgent = other.TransferToAgent
	}
	if other.Escalate {
		base.Escalate = true
	}
	if other.StateDelta != nil {
		maps.Copy(base.StateDelta, other.StateDelta)    <==== Use maps.Copy to accumulate state changes
		// base.StateDelta = other.StateDelta                    <==== Overwrites so we only see the last entry
	}
	return base
}

lol, just noticed the TODO there as I was reviewing this reply

[verdverm]

fwiw, it looks like the adk-python project does a deep merge

https://github.com/google/adk-python/blob/main/src/google/adk/flows/llm_flows/functions.py#L850

---

Thinking about what the code does and what we might want to do on the UI side of things...

1. send a message
2. LLM goes, wants to call multiple tools
3. more than one of those tools wants to escalate or get confirmation or report an error

Would this require sending back multiple events instead of just one with all the function responses, or can we intermix, look at the results, and decide what to do (besides the escalation, I haven't gotten there yet or seen agents do it, but asking clarifying questions is one place I hope to!)