fix: populate cwd and assistant_response in Chronicle session store

[digitarald]

Two fields in the Chronicle local session store were always empty. This fixes both with a consumer-only approach (no OTel producer changes) and adds schema caveats for remaining data gaps.

Changes

cwd — set from vscode.workspace.workspaceFolders[0] during session init.

assistant_response — truncateForOTel produces invalid JSON for long responses (cuts mid-string, appends suffix). The new extractAssistantResponse() handles this:

1. Fast path: JSON.parse for non-truncated input
2. Fallback: substring extraction from the fixed JSON prefix + JSON unescape
3. Stored truncated to 1000 chars

Schema caveats — updated _getSchemaDescription() local branch to warn the LLM that:

• agent_name / agent_description may be empty for older sessions
• summary may contain raw JSON — prefer JOINing with turns.user_message
• assistant_response may be empty for older sessions
• session_files / session_refs may be empty for older sessions

What's NOT fixed (remaining from #312292)

• agent_name population — the code reads GenAiAttr.AGENT_NAME from spans but this attribute may not be set by all producers
• agent_description — no known source for this field
• summary containing raw JSON — requires changes to the summarization pipeline
• session_files / session_refs empty — file/ref tracking was added but may need producer-side work

Testing

• 9 new test cases in extractAssistantResponse.spec.ts (valid JSON, truncated JSON, JSON escape handling, edge cases)
• All chronicle tests pass, TypeScript compiles clean

Partially addresses #312292

[Copilot]

Pull request overview

This PR fixes missing data in the Chronicle local session store by (1) populating sessions.cwd at session initialization and (2) reliably extracting/storing turns.assistant_response from gen_ai.output.messages, including when that attribute has been truncated by truncateForOTel. It also updates the Chronicle intent’s local SQLite schema description so LLM-driven queries better match what’s actually stored.

Changes:

• Populate sessions.cwd from vscode.workspace.workspaceFolders?.[0] during session initialization.
• Replace JSON-parse-only assistant response extraction with extractAssistantResponse() that supports truncated OTel JSON payloads, and store a truncated (~1000 char) value in turns.assistant_response.
• Add Vitest coverage for extractAssistantResponse() across valid JSON, truncated JSON, and escaping edge cases.

Show a summary per file

File	Description
extensions/copilot/src/extension/intents/node/chronicleIntent.ts	Updates the local SQLite schema description for cwd and assistant_response.
extensions/copilot/src/extension/chronicle/vscode-node/sessionStoreTracker.ts	Sets cwd at session init; extracts and stores truncated assistant responses on turns.
extensions/copilot/src/extension/chronicle/common/test/extractAssistantResponse.spec.ts	Adds unit tests covering assistant response extraction, including truncated OTel JSON.
extensions/copilot/src/extension/chronicle/common/sessionStoreTracking.ts	Introduces extractAssistantResponse() with a fallback path for truncateForOTel-truncated JSON.

Copilot's findings

• Files reviewed: 4/4 changed files
• Comments generated: 2

[Copilot]

The truncated-JSON fallback doesn’t actually use the “fixed prefix” it documents (including the hard-coded 63-char note), and it searches for the first occurrence of "content": which can diverge from the fast-path behavior (fast path only returns parts where type === 'text'). Consider anchoring the fallback to the same structure as the expected OTel output (e.g. locate the "role":"assistant" + "type":"text" + "content":" sequence, then terminate at the truncation suffix), and drop/update the misleading fixed-length comment so future changes don’t accidentally break extraction.

Suggested change

	// Fallback: extract text from truncated JSON via substring.
	// The JSON prefix is fixed: [{"role":"assistant","parts":[{"type":"text","content":"
	// which is 63 chars. The content value follows, terminated by the truncation marker.
	if (!outputMessagesRaw.includes(OTEL_TRUNCATION_MARKER)) {
	return undefined;
	}
	const contentPrefix = '"content":"';
	const contentStart = outputMessagesRaw.indexOf(contentPrefix);
	if (contentStart === -1) {
	return undefined;
	}
	const textStart = contentStart + contentPrefix.length;
	// Fallback: extract text from truncated JSON by matching the serialized
	// assistant text-part prefix, then reading until the truncation marker.
	if (!outputMessagesRaw.includes(OTEL_TRUNCATION_MARKER)) {
	return undefined;
	}
	const assistantTextContentPrefix = '"role":"assistant","parts":[{"type":"text","content":"';
	const prefixStart = outputMessagesRaw.indexOf(assistantTextContentPrefix);
	if (prefixStart === -1) {
	return undefined;
	}
	const textStart = prefixStart + assistantTextContentPrefix.length;

[Copilot]

The local schema description says assistant_response is “a summary of the assistant’s reply”, but the implementation stores a truncated prefix (first ~1000 chars + optional ellipsis), which may cut mid-sentence and isn’t a true summary. To avoid misleading the model when it writes queries/interpretations, consider rewording this to something like “first ~1000 characters of the assistant reply” (and mention the ellipsis behavior if relevant).

Suggested change

	- **turns**: session_id, turn_index, user_message, assistant_response (truncated to ~1000 chars — a summary of the assistant's reply, not the full response), timestamp. The richest source of what actually happened — contains the user's prompts and the assistant's replies.
	- **turns**: session_id, turn_index, user_message, assistant_response (first ~1000 characters of the assistant reply, with an ellipsis if truncated — not the full response), timestamp. The richest source of what actually happened — contains the user's prompts and the assistant's replies.