Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@packages/core/src/llm-core/agent/types.ts`:
- Around line 197-200: AgentObservation now includes DirectToolOutput but
consumers assume string/array; add type guards to handle DirectToolOutput
explicitly: where code reads the observation variable (e.g., the length/null
checks and string interpolation using observation and observationPrefix, and
places that feed ToolMessage content), first call
isDirectToolOutput(observation) (or inline check for the DirectToolOutput shape)
and either convert it to a string/appropriate ToolMessage content via the same
serialization used by coerceToAgentObservation or branch to handle its fields,
ensuring you don't access observation.length or interpolate the object directly;
update both the observation length/null check site and the buffer +
`\n${observationPrefix}${observation}\n` concatenation to guard/transform
DirectToolOutput before use.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@packages/core/src/middlewares/chat/stop_chat.ts`:
- Around line 84-87: The code is incorrectly re-checking request ownership by
calling ctx.chatluna.conversationRuntime.getRequestId(session, conversation.id),
which narrows stop permissions back to the original requester; instead remove
the session-based lookup and call stopConversationRequest with the conversation
id directly (use conversation.id as the identifier passed to
ctx.chatluna.conversationRuntime.stopConversationRequest), delete the requestId
variable usage, and keep the earlier permission: 'manage' semantics so
admins/shared managers can stop others' active requests.

In `@packages/core/src/services/conversation_runtime.ts`:
- Around line 449-478: dispose() aborts active requests then immediately removes
entries from activeByConversation, which leaves appendPendingMessage() Promises
waiting on roundDecisionResolvers unresolved; before deleting an active you must
either resolve its pending roundDecisionResolvers (resolve(false)) or call
completeRequest(active.conversationId, active.requestId) so the pending promise
can drain. Update dispose(platform?) to, for each matching active, first
clear/resolve the roundDecisionResolvers for that active (or invoke
completeRequest(active.conversationId, active.requestId)), then call
active.abortController.abort(...) and only after that delete from
activeByConversation and interfaces to avoid permanently hanging
appendPendingMessage() callers.

In `@packages/core/src/services/conversation.ts`:
- Around line 1828-1834: The current logic always defaults
constraint.users/constraint.excludeUsers to '[]' so an explicit null (meaning
"no restriction") is treated as an empty list and incorrectly blocks everyone;
update the checks so you first test whether constraint.users === null (in which
case skip the inclusion test), otherwise JSON.parse(constraint.users) into users
and then run the existing users != null && !users.includes(session.userId)
check; do the analogous change for constraint.excludeUsers (treat null as "no
excludes", only JSON.parse when not null) so the variables users and
excludeUsers reflect the intent of a null value rather than being coerced to [].

In `@packages/shared-adapter/src/client.ts`:
- Line 152: The supportImageInput list currently includes 'gemma' but
adapter-gemini's processImageParts only handles models containing 'gemma2',
causing images to be silently dropped; either remove 'gemma' from the
supportImageInput set or update processImageParts in
packages/adapter-gemini/src/utils.ts to also handle 'gemma' (or log a warning
when a model matches 'gemma' but not 'gemma2'); locate the supportImageInput
declaration in client.ts and the processImageParts function in adapter-gemini's
utils.ts and make the model-name checks consistent (or add a clear warning log
in processImageParts when an unsupported gemma variant is passed).
- Line 132: The fallback hardcoded return value in getModelMaxContextSize
(currently "return 200_000") increases the default context window for unmapped
models and can break token/counting logic downstream; change it to avoid a large
hardcoded value—either restore the previous fallback to use
getModelContextSize('o1-mini') or derive the fallback from
modelMaxContextSizeTable/configuration (or expose a configurable
DEFAULT_MODEL_MAX_CONTEXT_SIZE) so unknown models use the conservative prior
value; update getModelMaxContextSize and any callers that assume the smaller
default and add a short comment documenting the chosen fallback behavior.

---

Outside diff comments:
In `@packages/adapter-gemini/src/utils.ts`:
- Around line 211-220: The model-image support check in the condition using
model.includes('gemma2') is inconsistent with
imageModelMatchers/supportImageInput which include 'gemma' and with requester.ts
which filters by 'gemma', causing gemma models' images to be dropped; update the
condition in the function containing this snippet to check for 'gemma' as well
(e.g., include model.includes('gemma') alongside model.includes('gemma2') or
replace 'gemma2' with 'gemma' to match the rest of the adapter), keep the
existing exclusion for 'gemini-1.0', and run relevant tests or verify
supportImageInput and requester.ts behaviors remain consistent.