@tarquinen/opencode-dcp 3.2.5-beta0 → 3.2.6-beta0

package/index.ts

@@ -0,0 +1,141 @@
+import type { Plugin } from "@opencode-ai/plugin"
+import { getConfig } from "./lib/config"
+import { createCompressMessageTool, createCompressRangeTool } from "./lib/compress"
+import {
+    compressDisabledByOpencode,
+    hasExplicitToolPermission,
+    type HostPermissionSnapshot,
+} from "./lib/host-permissions"
+import { Logger } from "./lib/logger"
+import { createSessionState } from "./lib/state"
+import { PromptStore } from "./lib/prompts/store"
+import {
+    createChatMessageHandler,
+    createChatMessageTransformHandler,
+    createCommandExecuteHandler,
+    createEventHandler,
+    createSystemPromptHandler,
+    createTextCompleteHandler,
+} from "./lib/hooks"
+import { configureClientAuth, isSecureMode } from "./lib/auth"
+
+const id = "opencode-dynamic-context-pruning"
+
+const server: Plugin = (async (ctx) => {
+    const config = getConfig(ctx)
+
+    if (!config.enabled) {
+        return {}
+    }
+
+    const logger = new Logger(config.debug)
+    const state = createSessionState()
+    const prompts = new PromptStore(logger, ctx.directory, config.experimental.customPrompts)
+    const hostPermissions: HostPermissionSnapshot = {
+        global: undefined,
+        agents: {},
+    }
+
+    if (isSecureMode()) {
+        configureClientAuth(ctx.client)
+        // logger.info("Secure mode detected, configured client authentication")
+    }
+
+    logger.info("DCP initialized", {
+        strategies: config.strategies,
+    })
+
+    const compressToolContext = {
+        client: ctx.client,
+        state,
+        logger,
+        config,
+        prompts,
+    }
+
+    return {
+        "experimental.chat.system.transform": createSystemPromptHandler(
+            state,
+            logger,
+            config,
+            prompts,
+        ),
+        "experimental.chat.messages.transform": createChatMessageTransformHandler(
+            ctx.client,
+            state,
+            logger,
+            config,
+            prompts,
+            hostPermissions,
+        ) as any,
+        "chat.message": createChatMessageHandler(state, logger, config, hostPermissions),
+        "experimental.text.complete": createTextCompleteHandler(),
+        "command.execute.before": createCommandExecuteHandler(
+            ctx.client,
+            state,
+            logger,
+            config,
+            ctx.directory,
+            hostPermissions,
+        ),
+        event: createEventHandler(state, logger),
+        tool: {
+            ...(config.compress.permission !== "deny" && {
+                compress:
+                    config.compress.mode === "message"
+                        ? createCompressMessageTool(compressToolContext)
+                        : createCompressRangeTool(compressToolContext),
+            }),
+        },
+        config: async (opencodeConfig) => {
+            if (
+                config.compress.permission !== "deny" &&
+                compressDisabledByOpencode(opencodeConfig.permission)
+            ) {
+                config.compress.permission = "deny"
+            }
+
+            if (config.commands.enabled && config.compress.permission !== "deny") {
+                opencodeConfig.command ??= {}
+                opencodeConfig.command["dcp"] = {
+                    template: "",
+                    description: "Show available DCP commands",
+                }
+            }
+
+            const toolsToAdd: string[] = []
+            if (config.compress.permission !== "deny" && !config.experimental.allowSubAgents) {
+                toolsToAdd.push("compress")
+            }
+
+            if (toolsToAdd.length > 0) {
+                const existingPrimaryTools = opencodeConfig.experimental?.primary_tools ?? []
+                opencodeConfig.experimental = {
+                    ...opencodeConfig.experimental,
+                    primary_tools: [...existingPrimaryTools, ...toolsToAdd],
+                }
+            }
+
+            if (!hasExplicitToolPermission(opencodeConfig.permission, "compress")) {
+                const permission = opencodeConfig.permission ?? {}
+                opencodeConfig.permission = {
+                    ...permission,
+                    compress: config.compress.permission,
+                } as typeof permission
+            }
+
+            hostPermissions.global = opencodeConfig.permission
+            hostPermissions.agents = Object.fromEntries(
+                Object.entries(opencodeConfig.agent ?? {}).map(([name, agent]) => [
+                    name,
+                    agent?.permission,
+                ]),
+            )
+        },
+    }
+}) satisfies Plugin
+
+export default {
+    id,
+    server,
+}

package/lib/analysis/tokens.ts

@@ -0,0 +1,225 @@
+/**
+ * Shared Token Analysis
+ * Computes a breakdown of token usage across categories for a session.
+ *
+ * TOKEN CALCULATION STRATEGY
+ * ==========================
+ * We minimize tokenizer estimation by leveraging API-reported values wherever possible.
+ *
+ * WHAT WE GET FROM THE API (exact):
+ *   - tokens.input    : Input tokens for each assistant response
+ *   - tokens.output   : Output tokens generated (includes text + tool calls)
+ *   - tokens.reasoning: Reasoning tokens used
+ *   - tokens.cache    : Cache read/write tokens
+ *
+ * HOW WE CALCULATE EACH CATEGORY:
+ *
+ *   SYSTEM = firstAssistant.input + cache.read + cache.write - tokenizer(firstUserMessage)
+ *            The first response's total input (input + cache.read + cache.write)
+ *            contains system + first user message. On the first request of a
+ *            session, the system prompt appears in cache.write (cache creation),
+ *            not cache.read.
+ *
+ *   TOOLS  = tokenizer(toolInputs + toolOutputs) - prunedTokens
+ *            We must tokenize tools anyway for pruning decisions.
+ *
+ *   USER   = tokenizer(all user messages)
+ *            User messages are typically small, so estimation is acceptable.
+ *
+ *   ASSISTANT = total - system - user - tools
+ *               Calculated as residual. This absorbs:
+ *               - Assistant text output tokens
+ *               - Reasoning tokens (if persisted by the model)
+ *               - Any estimation errors
+ *
+ *   TOTAL  = input + output + reasoning + cache.read + cache.write
+ *            Matches opencode's UI display.
+ *
+ * WHY ASSISTANT IS THE RESIDUAL:
+ *   If reasoning tokens persist in context (model-dependent), they semantically
+ *   belong with "Assistant" since reasoning IS assistant-generated content.
+ */
+
+import type { AssistantMessage, TextPart, ToolPart } from "@opencode-ai/sdk/v2"
+import type { SessionState, WithParts } from "../state"
+import { isIgnoredUserMessage } from "../messages/query"
+import { isMessageCompacted } from "../state/utils"
+import { countTokens, extractCompletedToolOutput } from "../token-utils"
+
+export type MessageStatus = "active" | "pruned"
+
+export interface TokenBreakdown {
+    system: number
+    user: number
+    assistant: number
+    tools: number
+    toolCount: number
+    toolsInContextCount: number
+    prunedTokens: number
+    prunedToolCount: number
+    prunedMessageCount: number
+    total: number
+    messageCount: number
+}
+
+export interface TokenAnalysis {
+    breakdown: TokenBreakdown
+    messageStatuses: MessageStatus[]
+}
+
+export function emptyBreakdown(): TokenBreakdown {
+    return {
+        system: 0,
+        user: 0,
+        assistant: 0,
+        tools: 0,
+        toolCount: 0,
+        toolsInContextCount: 0,
+        prunedTokens: 0,
+        prunedToolCount: 0,
+        prunedMessageCount: 0,
+        total: 0,
+        messageCount: 0,
+    }
+}
+
+export function analyzeTokens(state: SessionState, messages: WithParts[]): TokenAnalysis {
+    const breakdown = emptyBreakdown()
+    const messageStatuses: MessageStatus[] = []
+    breakdown.prunedTokens = state.stats.totalPruneTokens
+
+    let firstAssistant: AssistantMessage | undefined
+    for (const msg of messages) {
+        if (msg.info.role !== "assistant") continue
+        const assistantInfo = msg.info as AssistantMessage
+        if (
+            assistantInfo.tokens?.input > 0 ||
+            assistantInfo.tokens?.cache?.read > 0 ||
+            assistantInfo.tokens?.cache?.write > 0
+        ) {
+            firstAssistant = assistantInfo
+            break
+        }
+    }
+
+    let lastAssistant: AssistantMessage | undefined
+    for (let i = messages.length - 1; i >= 0; i--) {
+        const msg = messages[i]
+        if (msg.info.role !== "assistant") continue
+        const assistantInfo = msg.info as AssistantMessage
+        if (assistantInfo.tokens?.output > 0) {
+            lastAssistant = assistantInfo
+            break
+        }
+    }
+
+    const apiInput = lastAssistant?.tokens?.input || 0
+    const apiOutput = lastAssistant?.tokens?.output || 0
+    const apiReasoning = lastAssistant?.tokens?.reasoning || 0
+    const apiCacheRead = lastAssistant?.tokens?.cache?.read || 0
+    const apiCacheWrite = lastAssistant?.tokens?.cache?.write || 0
+    breakdown.total = apiInput + apiOutput + apiReasoning + apiCacheRead + apiCacheWrite
+
+    const userTextParts: string[] = []
+    const toolInputParts: string[] = []
+    const toolOutputParts: string[] = []
+    const allToolIds = new Set<string>()
+    const activeToolIds = new Set<string>()
+    const prunedByMessageToolIds = new Set<string>()
+    const allMessageIds = new Set<string>()
+
+    let firstUserText = ""
+    let foundFirstUser = false
+
+    for (const msg of messages) {
+        const ignoredUser = msg.info.role === "user" && isIgnoredUserMessage(msg)
+        if (ignoredUser) continue
+
+        allMessageIds.add(msg.info.id)
+        const parts = Array.isArray(msg.parts) ? msg.parts : []
+        const compacted = isMessageCompacted(state, msg)
+        const pruneEntry = state.prune.messages.byMessageId.get(msg.info.id)
+        const messagePruned = !!pruneEntry && pruneEntry.activeBlockIds.length > 0
+        const messageActive = !compacted && !messagePruned
+
+        breakdown.messageCount += 1
+        messageStatuses.push(messageActive ? "active" : "pruned")
+
+        for (const part of parts) {
+            if (part.type === "tool") {
+                const toolPart = part as ToolPart
+                if (toolPart.callID) {
+                    allToolIds.add(toolPart.callID)
+                    if (!compacted) activeToolIds.add(toolPart.callID)
+                    if (messagePruned) prunedByMessageToolIds.add(toolPart.callID)
+                }
+
+                const toolPruned = toolPart.callID && state.prune.tools.has(toolPart.callID)
+                if (!compacted && !toolPruned) {
+                    if (toolPart.state?.input) {
+                        const inputText =
+                            typeof toolPart.state.input === "string"
+                                ? toolPart.state.input
+                                : JSON.stringify(toolPart.state.input)
+                        toolInputParts.push(inputText)
+                    }
+                    const outputText = extractCompletedToolOutput(toolPart)
+                    if (outputText !== undefined) {
+                        toolOutputParts.push(outputText)
+                    }
+                }
+                continue
+            }
+
+            if (part.type === "text" && msg.info.role === "user" && !compacted) {
+                const textPart = part as TextPart
+                const text = textPart.text || ""
+                userTextParts.push(text)
+                if (!foundFirstUser) firstUserText += text
+            }
+        }
+
+        if (msg.info.role === "user" && !foundFirstUser) {
+            foundFirstUser = true
+        }
+    }
+
+    const prunedByToolIds = new Set<string>()
+    for (const toolID of allToolIds) {
+        if (state.prune.tools.has(toolID)) prunedByToolIds.add(toolID)
+    }
+
+    const prunedToolIds = new Set<string>([...prunedByToolIds, ...prunedByMessageToolIds])
+    breakdown.toolCount = allToolIds.size
+    breakdown.toolsInContextCount = [...activeToolIds].filter(
+        (id) => !prunedByToolIds.has(id),
+    ).length
+    breakdown.prunedToolCount = prunedToolIds.size
+
+    for (const [messageID, entry] of state.prune.messages.byMessageId) {
+        if (allMessageIds.has(messageID) && entry.activeBlockIds.length > 0) {
+            breakdown.prunedMessageCount += 1
+        }
+    }
+
+    const firstUserTokens = countTokens(firstUserText)
+    breakdown.user = countTokens(userTextParts.join("\n"))
+    const toolInputTokens = countTokens(toolInputParts.join("\n"))
+    const toolOutputTokens = countTokens(toolOutputParts.join("\n"))
+
+    if (firstAssistant) {
+        const firstInput =
+            (firstAssistant.tokens?.input || 0) +
+            (firstAssistant.tokens?.cache?.read || 0) +
+            (firstAssistant.tokens?.cache?.write || 0)
+        breakdown.system = Math.max(0, firstInput - firstUserTokens)
+    }
+
+    breakdown.tools = toolInputTokens + toolOutputTokens
+    breakdown.assistant = Math.max(
+        0,
+        breakdown.total - breakdown.system - breakdown.user - breakdown.tools,
+    )
+
+    return { breakdown, messageStatuses }
+}

package/lib/auth.ts

@@ -0,0 +1,37 @@
+export function isSecureMode(): boolean {
+    return !!process.env.OPENCODE_SERVER_PASSWORD
+}
+
+export function getAuthorizationHeader(): string | undefined {
+    const password = process.env.OPENCODE_SERVER_PASSWORD
+    if (!password) return undefined
+
+    const username = process.env.OPENCODE_SERVER_USERNAME ?? "opencode"
+    // Use Buffer for Node.js base64 encoding (btoa may not be available in all Node versions)
+    const credentials = Buffer.from(`${username}:${password}`).toString("base64")
+    return `Basic ${credentials}`
+}
+
+export function configureClientAuth(client: any): any {
+    const authHeader = getAuthorizationHeader()
+
+    if (!authHeader) {
+        return client
+    }
+
+    // The SDK client has an internal client with request interceptors
+    // Access the underlying client to add the interceptor
+    const innerClient = client._client || client.client
+
+    if (innerClient?.interceptors?.request) {
+        innerClient.interceptors.request.use((request: Request) => {
+            // Only add auth header if not already present
+            if (!request.headers.has("Authorization")) {
+                request.headers.set("Authorization", authHeader)
+            }
+            return request
+        })
+    }
+
+    return client
+}

package/lib/commands/compression-targets.ts

@@ -0,0 +1,137 @@
+import type { CompressionBlock, PruneMessagesState } from "../state"
+
+export interface CompressionTarget {
+    displayId: number
+    runId: number
+    topic: string
+    compressedTokens: number
+    durationMs: number
+    grouped: boolean
+    blocks: CompressionBlock[]
+}
+
+function byBlockId(a: CompressionBlock, b: CompressionBlock): number {
+    return a.blockId - b.blockId
+}
+
+function buildTarget(blocks: CompressionBlock[]): CompressionTarget {
+    const ordered = [...blocks].sort(byBlockId)
+    const first = ordered[0]
+    if (!first) {
+        throw new Error("Cannot build compression target from empty block list.")
+    }
+
+    const grouped = first.mode === "message"
+    return {
+        displayId: first.blockId,
+        runId: first.runId,
+        topic: grouped ? first.batchTopic || first.topic : first.topic,
+        compressedTokens: ordered.reduce((total, block) => total + block.compressedTokens, 0),
+        durationMs: ordered.reduce((total, block) => Math.max(total, block.durationMs), 0),
+        grouped,
+        blocks: ordered,
+    }
+}
+
+function groupMessageBlocks(blocks: CompressionBlock[]): CompressionTarget[] {
+    const grouped = new Map<number, CompressionBlock[]>()
+
+    for (const block of blocks) {
+        const existing = grouped.get(block.runId)
+        if (existing) {
+            existing.push(block)
+            continue
+        }
+        grouped.set(block.runId, [block])
+    }
+
+    return Array.from(grouped.values()).map(buildTarget)
+}
+
+function splitTargets(blocks: CompressionBlock[]): CompressionTarget[] {
+    const messageBlocks: CompressionBlock[] = []
+    const singleBlocks: CompressionBlock[] = []
+
+    for (const block of blocks) {
+        if (block.mode === "message") {
+            messageBlocks.push(block)
+        } else {
+            singleBlocks.push(block)
+        }
+    }
+
+    const targets = [
+        ...singleBlocks.map((block) => buildTarget([block])),
+        ...groupMessageBlocks(messageBlocks),
+    ]
+    return targets.sort((a, b) => a.displayId - b.displayId)
+}
+
+export function getActiveCompressionTargets(
+    messagesState: PruneMessagesState,
+): CompressionTarget[] {
+    const activeBlocks = Array.from(messagesState.activeBlockIds)
+        .map((blockId) => messagesState.blocksById.get(blockId))
+        .filter((block): block is CompressionBlock => !!block && block.active)
+
+    return splitTargets(activeBlocks)
+}
+
+export function getRecompressibleCompressionTargets(
+    messagesState: PruneMessagesState,
+    availableMessageIds: Set<string>,
+): CompressionTarget[] {
+    const allBlocks = Array.from(messagesState.blocksById.values()).filter((block) => {
+        return availableMessageIds.has(block.compressMessageId)
+    })
+
+    const messageGroups = new Map<number, CompressionBlock[]>()
+    const singleTargets: CompressionTarget[] = []
+
+    for (const block of allBlocks) {
+        if (block.mode === "message") {
+            const existing = messageGroups.get(block.runId)
+            if (existing) {
+                existing.push(block)
+            } else {
+                messageGroups.set(block.runId, [block])
+            }
+            continue
+        }
+
+        if (block.deactivatedByUser && !block.active) {
+            singleTargets.push(buildTarget([block]))
+        }
+    }
+
+    for (const blocks of messageGroups.values()) {
+        if (blocks.some((block) => block.deactivatedByUser && !block.active)) {
+            singleTargets.push(buildTarget(blocks))
+        }
+    }
+
+    return singleTargets.sort((a, b) => a.displayId - b.displayId)
+}
+
+export function resolveCompressionTarget(
+    messagesState: PruneMessagesState,
+    blockId: number,
+): CompressionTarget | null {
+    const block = messagesState.blocksById.get(blockId)
+    if (!block) {
+        return null
+    }
+
+    if (block.mode !== "message") {
+        return buildTarget([block])
+    }
+
+    const blocks = Array.from(messagesState.blocksById.values()).filter(
+        (candidate) => candidate.mode === "message" && candidate.runId === block.runId,
+    )
+    if (blocks.length === 0) {
+        return null
+    }
+
+    return buildTarget(blocks)
+}

package/lib/commands/context.ts

@@ -0,0 +1,132 @@
+/**
+ * DCP Context Command
+ * Shows a visual breakdown of token usage in the current session.
+ * Token calculation logic lives in ../analysis/tokens.ts
+ *
+ * TOKEN CALCULATION STRATEGY
+ * ==========================
+ * We minimize tokenizer estimation by leveraging API-reported values wherever possible.
+ *
+ * WHAT WE GET FROM THE API (exact):
+ *   - tokens.input    : Input tokens for each assistant response
+ *   - tokens.output   : Output tokens generated (includes text + tool calls)
+ *   - tokens.reasoning: Reasoning tokens used
+ *   - tokens.cache    : Cache read/write tokens
+ *
+ * HOW WE CALCULATE EACH CATEGORY:
+ *
+ *   SYSTEM = firstAssistant.input + cache.read + cache.write - tokenizer(firstUserMessage)
+ *            The first response's total input (input + cache.read + cache.write)
+ *            contains system + first user message. On the first request of a
+ *            session, the system prompt appears in cache.write (cache creation),
+ *            not cache.read.
+ *
+ *   TOOLS  = tokenizer(toolInputs + toolOutputs) - prunedTokens
+ *            We must tokenize tools anyway for pruning decisions.
+ *
+ *   USER   = tokenizer(all user messages)
+ *            User messages are typically small, so estimation is acceptable.
+ *
+ *   ASSISTANT = total - system - user - tools
+ *               Calculated as residual. This absorbs:
+ *               - Assistant text output tokens
+ *               - Reasoning tokens (if persisted by the model)
+ *               - Any estimation errors
+ *
+ *   TOTAL  = input + output + reasoning + cache.read + cache.write
+ *            Matches opencode's UI display.
+ *
+ * WHY ASSISTANT IS THE RESIDUAL:
+ *   If reasoning tokens persist in context (model-dependent), they semantically
+ *   belong with "Assistant" since reasoning IS assistant-generated content.
+ */
+
+import type { Logger } from "../logger"
+import type { SessionState, WithParts } from "../state"
+import { sendIgnoredMessage } from "../ui/notification"
+import { formatTokenCount } from "../ui/utils"
+import { getCurrentParams } from "../token-utils"
+import { analyzeTokens, type TokenBreakdown } from "../analysis/tokens"
+
+export interface ContextCommandContext {
+    client: any
+    state: SessionState
+    logger: Logger
+    sessionId: string
+    messages: WithParts[]
+}
+function createBar(value: number, maxValue: number, width: number, char: string = "█"): string {
+    if (maxValue === 0) return ""
+    const filled = Math.round((value / maxValue) * width)
+    const bar = char.repeat(Math.max(0, filled))
+    return bar
+}
+
+function formatContextMessage(breakdown: TokenBreakdown): string {
+    const lines: string[] = []
+    const barWidth = 30
+
+    const toolsLabel = `Tools (${breakdown.toolsInContextCount})`
+
+    const categories = [
+        { label: "System", value: breakdown.system, char: "█" },
+        { label: "User", value: breakdown.user, char: "▓" },
+        { label: "Assistant", value: breakdown.assistant, char: "▒" },
+        { label: toolsLabel, value: breakdown.tools, char: "░" },
+    ] as const
+
+    const maxLabelLen = Math.max(...categories.map((c) => c.label.length))
+
+    lines.push("╭───────────────────────────────────────────────────────────╮")
+    lines.push("│                  DCP Context Analysis                     │")
+    lines.push("╰───────────────────────────────────────────────────────────╯")
+    lines.push("")
+    lines.push("Session Context Breakdown:")
+    lines.push("─".repeat(60))
+    lines.push("")
+
+    for (const cat of categories) {
+        const bar = createBar(cat.value, breakdown.total, barWidth, cat.char)
+        const percentage =
+            breakdown.total > 0 ? ((cat.value / breakdown.total) * 100).toFixed(1) : "0.0"
+        const labelWithPct = `${cat.label.padEnd(maxLabelLen)} ${percentage.padStart(5)}% `
+        const valueStr = formatTokenCount(cat.value).padStart(13)
+        lines.push(`${labelWithPct}│${bar.padEnd(barWidth)}│${valueStr}`)
+    }
+
+    lines.push("")
+    lines.push("─".repeat(60))
+    lines.push("")
+
+    lines.push("Summary:")
+
+    if (breakdown.prunedTokens > 0) {
+        const withoutPruning = breakdown.total + breakdown.prunedTokens
+        const pruned = []
+        if (breakdown.prunedToolCount > 0) pruned.push(`${breakdown.prunedToolCount} tools`)
+        if (breakdown.prunedMessageCount > 0)
+            pruned.push(`${breakdown.prunedMessageCount} messages`)
+        lines.push(
+            `  Pruned:          ${pruned.join(", ")} (~${formatTokenCount(breakdown.prunedTokens)})`,
+        )
+        lines.push(`  Current context: ~${formatTokenCount(breakdown.total)}`)
+        lines.push(`  Without DCP:     ~${formatTokenCount(withoutPruning)}`)
+    } else {
+        lines.push(`  Current context: ~${formatTokenCount(breakdown.total)}`)
+    }
+
+    lines.push("")
+
+    return lines.join("\n")
+}
+
+export async function handleContextCommand(ctx: ContextCommandContext): Promise<void> {
+    const { client, state, logger, sessionId, messages } = ctx
+
+    const { breakdown } = analyzeTokens(state, messages)
+
+    const message = formatContextMessage(breakdown)
+
+    const params = getCurrentParams(state, messages, logger)
+    await sendIgnoredMessage(client, sessionId, message, params, logger)
+}