The cross-harness, self-improving memory layer for AI agents. Join our community: https://discord.com/invite/zdNChS2fBu
Go agent-ai-cliagent-harnessesagent-securitybrowser-extensionciclaude-codecloudcodexcursordetection-engineeringjevknowledge-basememorymobile-device-managementobservabilitysecuritysecurity-information-and-event-managementsecurity-toolstelemetrytraces
fix: never use completion/message id as the conversation id (Bugbot)
shukieshahf2df69929e1c
diff --git a/browser-extension/README.md b/browser-extension/README.mdindex 52e77b25..0793c8af 100644--- a/browser-extension/README.md+++ b/browser-extension/README.md@@ -105,6 +105,12 @@ Beacon collector (prompt → response → tool calls → `runtime.jsonl` → S3service worker for streaming would need additional interception (deferred).- **Single capture path.** No DOM-scraping fallback yet; if a site changes its SSE wire format, theadapter needs updating (the live-smoke drift alarm is designed to catch this).+- **Telemetry integrity on the injected page.** The MAIN-world interceptor and any other page script+ share the same JS world, so the ISOLATED content script cannot cryptographically distinguish the+ extension's own `postMessage` events from forged ones. A malicious script already running on+ claude.ai could therefore inject fabricated chat records into the local collector. Impact is+ telemetry *integrity* only (no credential/code exposure). A proper fix is service-worker-driven+ MAIN injection (`chrome.scripting`) with a per-tab nonce the page can't read; deferred for V0.> **Fixtures.** `fixtures/claude/simple-turn.sse` is a **real, sanitized** claude.ai capture> (via `npm run record:fixtures`); the adapter is verified against it. `with-tool-call.sse` isdiff --git a/browser-extension/src/adapters/claude.ts b/browser-extension/src/adapters/claude.tsindex dbd05e49..a6adf410 100644--- a/browser-extension/src/adapters/claude.ts+++ b/browser-extension/src/adapters/claude.ts@@ -86,9 +86,11 @@ class ClaudeParser implements TurnParser {case 'message_start': {const m = obj.message ?? {};if (typeof m.model === 'string') this.model = m.model;+ // The completion/message id is a per-turn id — capture it for turnId,+ // but never use it as the conversation/session id (that comes only from+ // the URL/request body; using message.id would mis-correlate turns).if (typeof m.uuid === 'string') this.messageId = m.uuid;else if (typeof m.id === 'string') this.messageId = m.id;- if (typeof m.id === 'string' && !this.sessionId) this.sessionId = m.id;if (typeof m.usage?.input_tokens === 'number') this.inputTokens = m.usage.input_tokens;break;}@@ -140,9 +142,11 @@ class ClaudeParser implements TurnParser {toolCalls.push({ ...entry.tool, arguments: args });}}- const sessionId = this.sessionId || `claude-${this.messageId || this.reqId}`;- // Prefer the stable assistant message id; fall back to the interceptor's- // per-page reqId only when message_start was never seen (degenerate case).+ // Session id comes only from the conversation id (URL/body). If truly+ // absent, use a clearly-synthetic per-stream id — never the message id.+ const sessionId = this.sessionId || `claude-unknown-${this.reqId}`;+ // Turn key prefers the stable assistant message id; falls back to the+ // interceptor's per-page reqId only when message_start was never seen.const turnKey = this.messageId || `r${this.reqId}`;return {site: 'claude_web',