feat(chat): display sub-agent model information in message info dialog

[larryro]

Summary

• Display all model calls (routing + sub-agents) in the message info dialog
• Track sub-agent model/provider/tokens through the completion handler pipeline
• Add context window viewer showing token distribution as markdown
• Add count operation to customer and product read tools

Closes #324

Test plan

• Send a message that triggers sub-agent calls (CRM, Web, etc.)
• Open message info dialog and verify sub-agent section shows
• Verify each sub-agent displays: agent type, model, provider, token usage
• Click context window link to view token distribution details
• Test customer/product read tools with count operation

🤖 Generated with Claude Code

Summary by CodeRabbit

Release Notes

• New Features

  • Added context window display with collapsible formatting in message info dialog
  • Added sub-agent call details showing agent name, model, provider, and token breakdown
  • Added context statistics display (token count, message count, approvals)
  • Added count operations for customers and products with pagination support

• Improvements

  • Enhanced message metadata tracking with model and provider information for sub-agents
  • Improved formatting and organization of agent response information

_{✏️ Tip: You can customize this high-level summary in your review settings.}

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR extends the message info dialog and related metadata infrastructure to display comprehensive model call information. It adds UI components to render context windows as expandable Markdown and displays sub-agent usage details (agent type, model, provider, tokens). The backend is refactored to collect and propagate model/provider metadata through sub-agent tool responses and to construct a structured context window with system instructions in collapsible HTML sections. Message metadata types are extended to include subAgentUsage, contextWindow, and contextStats. Agent instructions are wired into response generation, and context formatting is rewritten from XML-style tags to HTML

Details

collapsible sections. Product and customer read tools gain "count" operations with pagination support.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Possibly related PRs

• refactor: context management and sub-agents improvements #185 — Updates context-management and sub-agent response flows with refactored context window formatting and propagation of subAgentUsage metadata
• talecorp/poc2#207 — Modifies chat message metadata hooks, message info UI components, and metadata type definitions
• feat(platform): add human input tool and structured context window #258 — Extends agent response generation and message metadata to surface subAgentUsage, contextWindow, and contextStats in the UI

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 7

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

services/platform/convex/agent_tools/products/product_read_tool.ts (1)

41-46: ⚠️ Potential issue | 🟡 Minor

Update the fields description to mention count as well.
The fields help text still says “Ignored for 'list'.” With the new count op, it should say list/count.

📝 Suggested fix

-      "For 'get_by_id' only: Fields to return. Default: ['_id','name','description','price','currency','status','category','imageUrl','stock']. Ignored for 'list'.",
+      "For 'get_by_id' only: Fields to return. Default: ['_id','name','description','price','currency','status','category','imageUrl','stock']. Ignored for 'list' and 'count'.",

🤖 Fix all issues with AI agents

In `@services/platform/app/features/chat/components/message-info-dialog.tsx`:
- Around line 119-129: The token-usage block only renders when
metadata.totalTokens or metadata.contextWindow is present, so cases with only
metadata.inputTokens or metadata.outputTokens are skipped; update the gating
condition around the Field rendering (the JSX that currently checks
(metadata.totalTokens !== undefined || metadata.contextWindow)) to also consider
metadata.inputTokens and metadata.outputTokens (and any other token fields used)
so the Field and its children (including ContextWindowToken and any input/output
token displays) render whenever any token-related metadata exists; adjust the
condition and keep references to metadata, Field, and ContextWindowToken intact.
- Around line 57-67: The Markdown rendering in message-info-dialog.tsx uses
rehypeRaw to render untrusted HTML (contextWindow) and must be sanitized to
prevent XSS: import rehypeSanitize and add it to the Markdown rehypePlugins so
raw HTML is parsed then sanitized (rehypeRaw before rehypeSanitize), configuring
rehypeSanitize to allow <details> and <summary> (or extend the default schema)
so the existing disclosure UI still works; update the Markdown invocation in the
component that renders {contextWindow} to include rehypeSanitize and adjust
imports accordingly.

In `@services/platform/convex/agent_tools/customers/helpers/count_customers.ts`:
- Around line 32-33: Add a short explanatory comment next to the existing
`@ts-ignore` above the queryFn assignment to document why TS2589 is suppressed:
mention that Convex's API can produce excessively deep type instantiations
causing false-positive TS2589 errors, reference the Convex issue or internal
ticket/PR (e.g., "see Convex issue #... or internal RFC") and note this is a
recurring pattern so maintainers know the suppression is intentional for
internal.customers.internal_queries.query_customers.queryCustomersInternal; keep
it concise (one or two sentences) and include a TODO to remove the `@ts-ignore`
if/when the Convex types are fixed.

In `@services/platform/convex/lib/agent_response/generate_response.ts`:
- Around line 602-608: The selection of sub-agent data using directResult ??
outputDirect ?? outputValue can pick an empty or irrelevant object (e.g., {}
without model/usage). Update the logic in generate_response.ts to choose the
first candidate that actually contains sub-agent fields (e.g., has `model` or
`usage`) rather than just non-nullish: inspect `directResult`, then
`outputDirect`, then `outputValue` and pick the first where a predicate (e.g.,
hasSubAgentFields(obj) => obj && (obj.model || obj.usage)) returns true; if none
match, fall back to the original nullish coalescing. Apply this to how
`subAgentData` is derived so that `toolUsage` is taken from a validated
sub-agent object.

In `@services/platform/convex/lib/context_management/message_formatter.ts`:
- Around line 20-31: The wrapInDetails function inserts untrusted summary and
content directly into HTML; escape both inputs to prevent HTML structure
breakage and XSS. Update wrapInDetails to HTML-entity-escape summary and content
(at least &, <, >, ", and ') before interpolation, e.g. call a shared utility
like escapeHtml(summary) and escapeHtml(content) inside wrapInDetails while
keeping the existing open handling (openAttr and template). Ensure the escape
helper is used here so all callers get sanitized output.
- Around line 37-40: The summarize function can throw when JSON.stringify
encounters circular refs, BigInt, or other non-serializable values; update
summarize(output, max) to safely stringify inside a try-catch: attempt
JSON.stringify for non-strings, on failure fallback to a safe representation
(e.g. convert BigInt to string, use util.inspect or String(output), or return a
placeholder like '[unserializable]') and then truncate to max as before; mirror
the error-handling pattern used by estimateJsonTokens to ensure summarize never
throws and always returns a string.

In `@services/platform/lib/shared/schemas/pagination.ts`:
- Around line 3-7: Add a brief JSDoc comment above the id property in the
cursorPaginationOptsSchema to explain its purpose (e.g., request correlation or
deduplication) so future maintainers know when to provide it; locate the
z.object definition named cursorPaginationOptsSchema and document the id field
(z.number().optional()) with a one-line comment describing its semantics and
intended use.