diff --git a/apps/server/src/project/AgentSessionScanner.test.ts b/apps/server/src/project/AgentSessionScanner.test.ts index 30db62847e8e1..e1fbfc09b8f09 100644 --- a/apps/server/src/project/AgentSessionScanner.test.ts +++ b/apps/server/src/project/AgentSessionScanner.test.ts @@ -2327,7 +2327,7 @@ it.layer(NodeServices.layer)("AgentSessionScanner", (it) => { cwd: workspace, message: { role: "user", content: "First prompt" }, }), - // 600 x 64 KiB exceeds the 32 MiB selected-history budget; the shown window fits. + // 600 x 64 KiB exceeds the 32 MiB selected-history budget, so the oldest drop out. ...Array.from({ length: middleCount }, (_, index) => { const role = index % 2 === 0 ? "assistant" : "user"; return record({ type: role, message: { role, content: longText } }); @@ -2349,8 +2349,11 @@ it.layer(NodeServices.layer)("AgentSessionScanner", (it) => { const thread = outcomes[0]?._tag === "Importable" ? outcomes[0].thread : undefined; expect(thread?.messages[0]).toMatchObject({ role: "user", text: "First prompt" }); expect(thread?.messages.at(-1)).toMatchObject({ role: "user", text: "Latest prompt" }); - expect(thread?.messages).toHaveLength(200); - expect(thread?.omittedMessageCount).toBe(middleCount + 2 - 200); + expect(thread?.messages.length).toBeGreaterThan(200); + expect(thread?.omittedMessageCount).toBeGreaterThan(0); + expect((thread?.messages.length ?? 0) + (thread?.omittedMessageCount ?? 0)).toBe( + middleCount + 2, + ); }), ); @@ -2362,17 +2365,22 @@ it.layer(NodeServices.layer)("AgentSessionScanner", (it) => { const claudeHomePath = yield* makeTempDir("t3code-late-prompt-claude-"); const codexHomePath = yield* makeTempDir("t3code-late-prompt-codex-"); const workspace = yield* makeTempDir("t3code-late-prompt-workspace-"); + const record = (role: "user" | "assistant", content: string) => + encodeTranscriptRecord({ + type: role, + sessionId: "late-prompt", + message: { role, content }, + }); const contents = [ encodeTranscriptRecord({ type: "attachment", sessionId: "late-prompt", cwd: workspace }), - // The first prompt is still among the latest 200 messages after older replies drop out. - ...Array.from({ length: 250 }, (_, index) => { - const role = index < 60 || index % 2 === 1 ? "assistant" : "user"; - return encodeTranscriptRecord({ - type: role, - sessionId: "late-prompt", - message: { role, content: `${role} ${index}` }, - }); - }), + // 40 x 1 MiB replies before the first prompt exceed the 32 MiB budget. + ...Array.from({ length: 40 }, (_, index) => + record("assistant", `${index} ${"x".repeat(1024 * 1024)}`), + ), + record("user", "First prompt"), + ...Array.from({ length: 300 }, (_, index) => + record(index % 2 === 0 ? "assistant" : "user", `Message ${index}`), + ), ].join("\n"); const filePath = path.join(claudeHomePath, "projects", "-slug", "late-prompt.jsonl"); yield* writeTranscript({ filePath, contents, mtimeMs: nowMs - 60_000 }); @@ -2383,16 +2391,15 @@ it.layer(NodeServices.layer)("AgentSessionScanner", (it) => { workspaceRoot: workspace, }); - expect(thread?.messages[0]).toMatchObject({ role: "assistant", text: "assistant 50" }); - expect(thread).toEqual( - AgentSessionScanner.parseAgentSessionTranscript({ - source: "claudeAgent", - providerInstanceId: ProviderInstanceId.make("claudeAgent"), - fallbackSessionId: "late-prompt", - lastActiveAtMs: nowMs - 60_000, - contents, - }), - ); + const texts = thread?.messages.map((message) => message.text) ?? []; + expect(thread?.messages[0]?.role).toBe("assistant"); + expect(texts[0]?.startsWith("0 ")).toBe(false); + expect(texts.filter((text) => text === "First prompt")).toHaveLength(1); + expect(texts.slice(-301)).toEqual([ + "First prompt", + ...Array.from({ length: 300 }, (_, index) => `Message ${index}`), + ]); + expect(texts.length + (thread?.omittedMessageCount ?? 0)).toBe(341); }), ); @@ -2797,47 +2804,44 @@ describe("parseAgentSessionTranscript", () => { }, ); - it.each([199, 200, 201, 205])( - "counts only omitted conversation messages for a %s-message Claude transcript", - (count) => { - const thread = AgentSessionScanner.parseAgentSessionTranscript({ - contents: [ - ...Array.from({ length: count }, (_, index) => - encodeTranscriptRecord({ - type: index === 0 ? "user" : "assistant", - sessionId: "claude-session", - message: { content: `Message ${index}` }, - }), - ), - encodeTranscriptRecord({ type: "user", isMeta: true, message: { content: "Setup" } }), - encodeTranscriptRecord({ - type: "user", - isCompactSummary: true, - message: { content: "Summary" }, - }), - encodeTranscriptRecord({ - type: "assistant", - isSidechain: true, - message: { content: "Sidechain" }, - }), + it("keeps every visible message of a long Claude transcript", () => { + const count = 205; + const thread = AgentSessionScanner.parseAgentSessionTranscript({ + contents: [ + ...Array.from({ length: count }, (_, index) => encodeTranscriptRecord({ - type: "assistant", - message: { content: [{ type: "tool_use" }] }, + type: index === 0 ? "user" : "assistant", + sessionId: "claude-session", + message: { content: `Message ${index}` }, }), - "malformed", - ].join("\n"), - source: "claudeAgent", - providerInstanceId: ProviderInstanceId.make("claudeAgent"), - fallbackSessionId: "fallback", - lastActiveAtMs: Date.parse("2026-08-24T12:00:00.000Z"), - }); + ), + encodeTranscriptRecord({ type: "user", isMeta: true, message: { content: "Setup" } }), + encodeTranscriptRecord({ + type: "user", + isCompactSummary: true, + message: { content: "Summary" }, + }), + encodeTranscriptRecord({ + type: "assistant", + isSidechain: true, + message: { content: "Sidechain" }, + }), + encodeTranscriptRecord({ + type: "assistant", + message: { content: [{ type: "tool_use" }] }, + }), + "malformed", + ].join("\n"), + source: "claudeAgent", + providerInstanceId: ProviderInstanceId.make("claudeAgent"), + fallbackSessionId: "fallback", + lastActiveAtMs: Date.parse("2026-08-24T12:00:00.000Z"), + }); - expect(thread?.omittedMessageCount).toBe(Math.max(0, count - 200)); - expect(thread?.messages).toHaveLength(Math.min(200, count)); - expect(thread?.messages[0]?.text).toBe("Message 0"); - expect(thread?.messages.at(-1)?.text).toBe(`Message ${count - 1}`); - }, - ); + expect(thread?.messages).toHaveLength(count); + expect(thread?.messages[0]?.text).toBe("Message 0"); + expect(thread?.messages.at(-1)?.text).toBe(`Message ${count - 1}`); + }); it("keeps Claude text and titles while dropping malformed and tool records", () => { const thread = AgentSessionScanner.parseAgentSessionTranscript({ @@ -2938,7 +2942,8 @@ describe("parseAgentSessionTranscript", () => { expect(thread?.title).toBe(expected); }); - it.each([199, 200])("keeps the canonical first prompt with %s later Codex messages", (count) => { + it("keeps the canonical first prompt and every later Codex message", () => { + const count = 200; const canonicalPrompt = "\n Keep the canonical prompt \n"; const canonicalTimestamp = "2026-08-24T10:01:00.000Z"; const laterAssistantMessages = Array.from({ length: count }, (_, index) => @@ -2970,60 +2975,21 @@ describe("parseAgentSessionTranscript", () => { payload: { type: "user_message", message: canonicalPrompt }, }), ...laterAssistantMessages, - ].join("\n"), - source: "codex", - providerInstanceId: ProviderInstanceId.make("codex"), - fallbackSessionId: "fallback", - lastActiveAtMs: Date.parse("2026-08-24T12:00:00.000Z"), - }); - - expect(thread?.omittedMessageCount).toBe(count - 199); - expect(thread?.messages).toHaveLength(200); - expect(thread?.messages[0]).toMatchObject({ - role: "user", - text: canonicalPrompt, - createdAt: canonicalTimestamp, - }); - }); - - it("restores the canonical first prompt when a later user message remains", () => { - const canonicalPrompt = "\n Keep the canonical prompt \n"; - const canonicalTimestamp = "2026-08-24T10:01:00.000Z"; - const assistantMessages = Array.from({ length: 198 }, (_, index) => - encodeTranscriptRecord({ - type: "response_item", - timestamp: `2026-08-24T11:${String(index % 60).padStart(2, "0")}:00.000Z`, - payload: { - type: "message", - role: "assistant", - content: [{ type: "output_text", text: `Assistant message ${index}` }], - }, - }), - ); - const thread = AgentSessionScanner.parseAgentSessionTranscript({ - contents: [ - encodeTranscriptRecord({ type: "session_meta", payload: { id: "codex-session" } }), + encodeTranscriptRecord({ + type: "event_msg", + timestamp: "2026-08-24T11:58:30.000Z", + payload: { type: "user_message", message: "Keep this later prompt" }, + }), + // The response copy that follows its event is not imported twice. encodeTranscriptRecord({ type: "response_item", - timestamp: "2026-08-24T10:00:00.000Z", + timestamp: "2026-08-24T11:58:31.000Z", payload: { type: "message", role: "user", - internal_chat_message_metadata_passthrough: { turn_id: "turn-1" }, - content: [{ type: "input_text", text: "Keep the canonical prompt" }], + content: [{ type: "input_text", text: "Keep this later prompt" }], }, }), - encodeTranscriptRecord({ - type: "event_msg", - timestamp: canonicalTimestamp, - payload: { type: "user_message", message: canonicalPrompt }, - }), - ...assistantMessages, - encodeTranscriptRecord({ - type: "event_msg", - timestamp: "2026-08-24T11:58:30.000Z", - payload: { type: "user_message", message: "Keep this later prompt" }, - }), encodeTranscriptRecord({ type: "response_item", timestamp: "2026-08-24T11:59:00.000Z", @@ -3040,20 +3006,16 @@ describe("parseAgentSessionTranscript", () => { lastActiveAtMs: Date.parse("2026-08-24T12:00:00.000Z"), }); - expect(thread?.omittedMessageCount).toBe(1); - expect(thread?.messages).toHaveLength(200); + expect(thread?.messages).toHaveLength(count + 3); + expect(thread?.messages.slice(-2).map(({ role, text }) => ({ role, text }))).toEqual([ + { role: "user", text: "Keep this later prompt" }, + { role: "assistant", text: "Keep this latest response" }, + ]); expect(thread?.messages[0]).toMatchObject({ role: "user", text: canonicalPrompt, createdAt: canonicalTimestamp, }); - expect( - thread?.messages.filter((message) => message.text.trim() === canonicalPrompt.trim()), - ).toHaveLength(1); - expect(thread?.messages.some((message) => message.text === "Keep this later prompt")).toBe( - true, - ); - expect(thread?.messages.at(-1)?.text).toBe("Keep this latest response"); }); it("keeps mixed-format response users when turn IDs repeat after an assistant", () => { diff --git a/apps/server/src/project/AgentSessionScanner.ts b/apps/server/src/project/AgentSessionScanner.ts index d9ca41b8f26b1..511b15806fe81 100644 --- a/apps/server/src/project/AgentSessionScanner.ts +++ b/apps/server/src/project/AgentSessionScanner.ts @@ -81,7 +81,6 @@ const MAX_METADATA_RECORDS_PER_TRANSCRIPT = 1_000; const RECENT_THREAD_WINDOW_MS = 30 * 24 * 60 * 60 * 1000; // Screenshot-heavy transcripts can reach GiB; select fields while streaming to avoid allocating payloads. const MAX_IMPORTED_TRANSCRIPT_BYTES = 4 * 1024 * 1024 * 1024; -const MAX_IMPORTED_MESSAGES = 200; const MAX_IMPORT_HISTORY_BYTES = 32 * 1024 * 1024; const MAX_IMPORT_BYTES = 4 * 1024 * 1024 * 1024; const MAX_IMPORT_TRANSCRIPTS = 100; @@ -464,10 +463,13 @@ function parseAgentSessionRecords( let model = claudeFields.model ?? null; let derivedTitle: string | undefined; let hasCodexSessionId = false; - const messages: Array = []; - let firstUserMessage: - | (AgentSessionThreadMessage & { readonly codexResponseUser: boolean }) - | undefined; + type RetainedMessage = AgentSessionThreadMessage & { readonly codexResponseUser: boolean }; + const messages: Array = []; + // Codex prompt copies are matched within the turn since the last assistant message. Indexes + // keep that linear: scanning back per record is quadratic on a long run of user records. + let codexEventTextsInTurn = new Set(); + let codexResponseCopiesInTurn = new Map>(); + const removedCodexResponseCopies = new Set(); const canonicalCodexResponseUserIndices = new Set(); let canonicalUserTextsInTurn = new Set(); let responseUsersInTurn: Array<{ @@ -521,35 +523,23 @@ function parseAgentSessionRecords( finishCodexTurn(); } - let conversationMessageCount = 0; - const retainMessage = ( - message: AgentSessionThreadMessage & { readonly codexResponseUser: boolean }, - ) => { - if (firstUserMessage === undefined && message.role === "user") { - firstUserMessage = message; - } + const retainMessage = (message: RetainedMessage) => { if (derivedTitle === undefined && message.role === "user") { derivedTitle = deriveTitle(input.source, message.text); } - conversationMessageCount += 1; messages.push(message); - if (messages.length > MAX_IMPORTED_MESSAGES) messages.shift(); - }; - - const hasMatchingCodexEventInTurn = (text: string) => { - const comparisonText = text.trim(); - for (let index = messages.length - 1; index >= 0; index--) { - const message = messages[index]; - if (message?.role === "assistant") return false; - if ( - message?.role === "user" && - !message.codexResponseUser && - message.text.trim() === comparisonText - ) { - return true; - } + if (input.source !== "codex") return; + if (message.role === "assistant") { + codexEventTextsInTurn = new Set(); + codexResponseCopiesInTurn = new Map(); + } else if (message.codexResponseUser) { + const text = message.text.trim(); + const copies = codexResponseCopiesInTurn.get(text); + if (copies === undefined) codexResponseCopiesInTurn.set(text, [message]); + else copies.push(message); + } else { + codexEventTextsInTurn.add(message.text.trim()); } - return false; }; let recordIndex = -1; @@ -596,17 +586,9 @@ function parseAgentSessionRecords( if (record.type === "event_msg" && record.payload?.type === "user_message") { const text = record.payload.message ?? ""; if (text.trim().length === 0) continue; - // Codex can duplicate prompts across formats; remove only matching response copies. - for (let index = messages.length - 1; index >= 0; index--) { - const message = messages[index]; - if (message?.role === "assistant") break; - if (message?.codexResponseUser === true && message.text.trim() === text.trim()) { - if (firstUserMessage === message) firstUserMessage = undefined; - messages.splice(index, 1); - conversationMessageCount -= 1; - break; - } - } + // Codex can duplicate prompts across formats; remove only the latest matching response copy. + const copy = codexResponseCopiesInTurn.get(text.trim())?.pop(); + if (copy !== undefined) removedCodexResponseCopies.add(copy); retainMessage({ role: "user", text, @@ -628,7 +610,7 @@ function parseAgentSessionRecords( if (record.payload.role === "user" && canonicalCodexResponseUserIndices.has(recordIndex)) { continue; } - if (record.payload.role === "user" && hasMatchingCodexEventInTurn(extractedText)) { + if (record.payload.role === "user" && codexEventTextsInTurn.has(extractedText.trim())) { continue; } retainMessage({ @@ -639,15 +621,16 @@ function parseAgentSessionRecords( }); } - const visibleMessages = messages.map( + const visibleMessages = messages.filter((message) => !removedCodexResponseCopies.has(message)); + if ( + providerSessionId.trim().length === 0 || + !visibleMessages.some((message) => message.role === "user") + ) { + return null; + } + const retainedMessages = visibleMessages.map( ({ codexResponseUser: _codexResponseUser, ...message }) => message, ); - if (providerSessionId.trim().length === 0 || firstUserMessage === undefined) return null; - const firstUserMessageRetained = messages.includes(firstUserMessage); - const { codexResponseUser: _codexResponseUser, ...visibleFirstUserMessage } = firstUserMessage; - const retainedMessages = firstUserMessageRetained - ? visibleMessages - : [visibleFirstUserMessage, ...visibleMessages.slice(-(MAX_IMPORTED_MESSAGES - 1))]; return { source: input.source, @@ -658,7 +641,7 @@ function parseAgentSessionRecords( createdAt: retainedMessages[0]?.createdAt ?? fallbackTimestamp, updatedAt: fallbackTimestamp, messages: retainedMessages, - omittedMessageCount: conversationMessageCount + droppedMessageCount - retainedMessages.length, + omittedMessageCount: droppedMessageCount, }; } @@ -695,26 +678,39 @@ function shouldRetainDecodedRecord( ); } -// The import shows only the first user message and the latest messages, so a -// long Claude session holds just those and counts the rest as dropped. +// A Claude session longer than the memory budget keeps its first user message +// and drops its oldest other messages, counting them as dropped. function createClaudeHistoryWindow() { - const recent: Array<{ readonly record: DecodedTranscriptRecord; readonly bytes: number }> = []; + const held: Array<{ readonly record: DecodedTranscriptRecord; readonly bytes: number }> = []; + let start = 0; let firstUser: DecodedTranscriptRecord | undefined; let droppedMessageCount = 0; return { - // Takes a record shouldRetainDecodedRecord kept; returns the bytes it released. - push: (record: DecodedTranscriptRecord, bytes: number): number => { + // Takes a record shouldRetainDecodedRecord kept. + push: (record: DecodedTranscriptRecord, bytes: number) => { if (firstUser === undefined && record.type === "user") firstUser = record; - recent.push({ record, bytes }); - if (recent.length <= MAX_IMPORTED_MESSAGES) return 0; - const evicted = recent.shift(); - if (evicted === undefined || evicted.record === firstUser) return 0; - droppedMessageCount += 1; - return evicted.bytes; + held.push({ record, bytes }); + }, + // Drops the oldest message other than the first user message; null when none is left. + evict: (): number | null => { + while (start < held.length) { + const evicted = held[start++]; + if (evicted === undefined || evicted.record === firstUser) continue; + droppedMessageCount += 1; + // Compact once most of the array is dropped, so evicting stays linear. + if (start * 2 > held.length) { + held.splice(0, start); + start = 0; + } + return evicted.bytes; + } + return null; }, records: () => { - const held = recent.map((entry) => entry.record); - return firstUser === undefined || held.includes(firstUser) ? held : [firstUser, ...held]; + const records = held.slice(start).map((entry) => entry.record); + return firstUser === undefined || records.includes(firstUser) + ? records + : [firstUser, ...records]; }, droppedMessageCount: () => droppedMessageCount, }; @@ -1127,10 +1123,14 @@ export const make = Effect.gen(function* () { let bytesRead = 0; const reserve = (bytes: number) => { recordBytes += bytes; - if (historyBytes + recordBytes > MAX_IMPORT_HISTORY_BYTES) { - throw new TranscriptJsonLimitError({ - message: "Transcript selected history exceeds the 32 MiB memory budget", - }); + while (historyBytes + recordBytes > MAX_IMPORT_HISTORY_BYTES) { + const released = source === "claudeAgent" ? claudeHistory.evict() : null; + if (released === null) { + throw new TranscriptJsonLimitError({ + message: "Transcript selected history exceeds the 32 MiB memory budget", + }); + } + historyBytes -= released; } }; let reader = createTranscriptJsonReader(reserve, selectTranscriptPath); @@ -1154,7 +1154,8 @@ export const make = Effect.gen(function* () { } if (shouldRetainDecodedRecord(source, decoded.value)) { if (source === "claudeAgent") { - historyBytes += recordBytes - claudeHistory.push(decoded.value, recordBytes); + claudeHistory.push(decoded.value, recordBytes); + historyBytes += recordBytes; } else if ( source === "pi" && decoded.value.type === "message" && diff --git a/docs/user/welcome-page.md b/docs/user/welcome-page.md index 63fa86b52b98d..71cfe203f2659 100644 --- a/docs/user/welcome-page.md +++ b/docs/user/welcome-page.md @@ -90,17 +90,19 @@ are read from each enabled Pi provider's session directory, which follows `PI_CODING_AGENT_SESSION_DIR` and `PI_CODING_AGENT_DIR` in its environment. A Pi conversation imports its active branch, the one Pi reopens. -Conversation import is best effort. T3 Code keeps the first user prompt and the -newest remaining visible user and assistant messages, with 200 messages total. +Conversation import is best effort. T3 Code keeps every visible user and +assistant message. It omits tool activity and attachments. For Codex, it omits generated setup context only when a canonical user event and a valid shared turn ID identify the same user turn. Ambiguous legacy or response-only context stays in the imported conversation so T3 Code does not remove user text. It reads one conversation at -a time and skips files larger than 4 GiB. It also skips a conversation when the -message text and metadata it keeps need more than 32 MiB of memory. Tool output -does not count toward that limit. It skips a conversation with more than 100,000 -records every time, so opening the page again does not import it. It ignores -malformed records and skips unreadable or unparseable conversations. +a time and skips files larger than 4 GiB. The message text and metadata it keeps +may use up to 32 MiB of memory; tool output does not count. A longer Claude +conversation keeps its first prompt and the newest messages that fit, and the +thread says how many earlier messages it left out. A Codex or Pi conversation +that needs more, or has more than 100,000 records, is skipped every time, so +opening the page again does not import it. It ignores malformed records and +skips unreadable or unparseable conversations. Each import attempt reads up to 100 conversation files and 4 GiB per project. To continue a large batch, open the Welcome page again and pick the project in