diff --git a/.changeset/auto-effort.md b/.changeset/auto-effort.md new file mode 100644 index 0000000..0ad8828 --- /dev/null +++ b/.changeset/auto-effort.md @@ -0,0 +1,5 @@ +--- +"opencode-adaptive-thinking": minor +--- + +Add an opt-in `autoSelect` mode that classifies the first user request in a session and carries the chosen reasoning effort until the agent, an explicit variant, or a model change overrides it. diff --git a/.changeset/effort-recommendation.md b/.changeset/effort-recommendation.md new file mode 100644 index 0000000..ddf8898 --- /dev/null +++ b/.changeset/effort-recommendation.md @@ -0,0 +1,5 @@ +--- +"opencode-adaptive-thinking": minor +--- + +Recommend a reasoning effort for each user turn and tell the agent when not to change it. The default guidance now relies on that recommendation and keeps a mid-turn reassessment clause. diff --git a/.changeset/quiet-resets.md b/.changeset/quiet-resets.md new file mode 100644 index 0000000..6175dfd --- /dev/null +++ b/.changeset/quiet-resets.md @@ -0,0 +1,5 @@ +--- +"opencode-adaptive-thinking": patch +--- + +Reset a temporary reasoning effort to the level in effect before the first override, so repeated changes no longer reset to an intermediate level or skip the reset. Read session history once per tool call with a bounded query, cache provider variants, merge guidance into an existing first system prompt, and remove the runtime Zod dependency. diff --git a/README.md b/README.md index c4c1b76..4391f1e 100644 --- a/README.md +++ b/README.md @@ -38,6 +38,7 @@ The plugin accepts these optional settings: - `enabled`: enable or disable the plugin without removing it from config. Defaults to `true`. - `quiet`: suppress toast notifications for configuration errors. Defaults to `false`. +- `autoSelect`: choose an initial effort level from the first user request in a session and carry it across turns. Defaults to `false`. - `toolName`: name of the reasoning-effort tool exposed to agents. Defaults to `set_reasoning_effort`. - `toolDescription`: description for the reasoning-effort tool. Defaults to `Set your reasoning effort`. - `systemPrompt`: custom adaptive-thinking guidance injected into the system prompt. @@ -89,6 +90,12 @@ Example persisted change: The system prompt lists the valid levels for the current session. Cached session state is only used when it is still valid for the current model, so switching models will not advertise or reuse an invalid effort level. +With `autoSelect` enabled, the plugin classifies the first user request in a session and sets the effort from the available variants for the current model, ignoring `none`. The chosen level is carried across turns until the agent changes it, the model changes, or a request carries an explicit variant. An explicit variant (UI picker, CLI flag, or API field) is honored and becomes the level carried forward. + +When `autoSelect` is off, the plugin still classifies each user turn and adds a recommendation to the guidance. If the recommended level differs from the current one, the agent is told to call the tool with that level; if it matches, the agent is told to keep the level unless the work grows beyond it. The default guidance relies on this recommendation instead of a blanket reassessment instruction. + +Adaptive-thinking guidance is merged into the first existing system prompt instead of a separate system message, so strict chat templates that allow only one system message keep working. + ## Troubleshooting If the tool returns `no valid reasoning effort levels are available for this session`, check that the active model exposes variants in OpenCode. Some providers or models may not support reasoning-effort levels. diff --git a/package.json b/package.json index 25c2132..c5e347e 100644 --- a/package.json +++ b/package.json @@ -37,9 +37,6 @@ "verify": "pnpm run format:check && pnpm run lint && pnpm run typecheck && pnpm run test && pnpm run build", "version-packages": "changeset version" }, - "dependencies": { - "zod": "^4.4.3" - }, "devDependencies": { "@changesets/changelog-github": "^0.7.0", "@changesets/cli": "^2.31.0", diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml index 54c6f2c..8698b75 100644 --- a/pnpm-lock.yaml +++ b/pnpm-lock.yaml @@ -14,9 +14,6 @@ importers: '@opencode-ai/sdk': specifier: '>=1' version: 1.17.13 - zod: - specifier: ^4.4.3 - version: 4.4.3 devDependencies: '@changesets/changelog-github': specifier: ^0.7.0 @@ -1318,9 +1315,6 @@ packages: zod@4.1.8: resolution: {integrity: sha512-5R1P+WwQqmmMIEACyzSvo4JXHY5WiAFHRMg+zBZKgKS+Q1viRa0C1hmUKtHltoIFKtIdki3pRxkmpP74jnNYHQ==} - zod@4.4.3: - resolution: {integrity: sha512-ytENFjIJFl2UwYglde2jchW2Hwm4GJFLDiSXWdTrJQBIN9Fcyp7n4DhxJEiWNAJMV1/BqWfW/kkg71UDcHJyTQ==} - snapshots: '@ai-sdk/provider@3.0.8': @@ -2379,5 +2373,3 @@ snapshots: yaml@2.9.0: {} zod@4.1.8: {} - - zod@4.4.3: {} diff --git a/src/complexity.test.ts b/src/complexity.test.ts new file mode 100644 index 0000000..7684d20 --- /dev/null +++ b/src/complexity.test.ts @@ -0,0 +1,21 @@ +import { expect, test } from "vitest"; +import { classifyPrompt, pickVariant } from "./complexity"; + +test("classifies prompt complexity", () => { + expect(classifyPrompt("fix typo in the README")).toBe("simple"); + expect(classifyPrompt("add a test for the parser")).toBe("normal"); + expect(classifyPrompt("debug the flaky test")).toBe("complex"); + expect(classifyPrompt("debug the flaky test and trace the root cause across modules")).toBe( + "deep", + ); +}); + +test("maps complexity to an available variant", () => { + const all = ["none", "low", "medium", "high", "xhigh"]; + expect(pickVariant(all, "simple")).toBe("low"); + expect(pickVariant(all, "normal")).toBe("medium"); + expect(pickVariant(all, "complex")).toBe("high"); + expect(pickVariant(all, "deep")).toBe("xhigh"); + expect(pickVariant(["low"], "deep")).toBe("low"); + expect(pickVariant([], "normal")).toBeUndefined(); +}); diff --git a/src/complexity.ts b/src/complexity.ts new file mode 100644 index 0000000..e67fa99 --- /dev/null +++ b/src/complexity.ts @@ -0,0 +1,54 @@ +export type Complexity = "simple" | "normal" | "complex" | "deep"; + +const signals: Array<[RegExp, number]> = [ + [/\b(debug|investigate|trace)\b/i, 2], + [/\b(root cause|flaky|race|deadlock|leak)\b/i, 2], + [ + /\b(architecture|architect|design|plan|refactor|migrat\w*|review|audit|security|performance|optimi[sz]e)\b/i, + 2, + ], + [/\b(error|exception|failing|fails?|failure|broken|bug|why)\b/i, 2], + [/\b(test|tests|coverage|benchmark)\b/i, 1], + [/\b(multi[- ]file|across|module|modules|system|end[- ]to[- ]end)\b/i, 1], + [/[?][\s\S]*[?]/, 1], + [/```/, 1], +]; + +const simpleSignals = /\b(typo|rename|format|lint|thanks|thank you|hello|hi|hey)\b/i; + +export const classifyPrompt = (text: string): Complexity => { + const input = text.trim(); + if (input.length === 0) return "normal"; + + let score = signals.reduce( + (total, [pattern, weight]) => total + (pattern.test(input) ? weight : 0), + 0, + ); + if (input.length > 400) score += 1; + if (input.length > 1200) score += 2; + if (input.length < 200 && simpleSignals.test(input)) score -= 3; + + if (score >= 6) return "deep"; + if (score >= 3) return "complex"; + if (score <= 0 && input.length < 160) return "simple"; + return "normal"; +}; + +export const pickVariant = ( + variants: Array, + complexity: Complexity, +): string | undefined => { + const usable = variants.filter((variant) => !/^(none|off|disabled)$/i.test(variant)); + if (usable.length === 0) return undefined; + if (usable.length === 1) return usable[0]; + + const last = usable.length - 1; + const index = { + simple: 0, + normal: Math.floor(last / 2), + complex: Math.round(last * 0.75), + deep: last, + }[complexity]; + + return usable[index]; +}; diff --git a/src/config.ts b/src/config.ts index bf83459..d4dd68c 100644 --- a/src/config.ts +++ b/src/config.ts @@ -1,26 +1,47 @@ -import * as z from "zod/v4"; - export const defaultSystemPrompt = - "You MUST manage reasoning effort actively. " + - "Lower it before trivial or routine turns; raise it for ambiguity, debugging, risky changes, or multi-step synthesis. " + - "Reassess at turn start, after meaningful new evidence, and when the task shifts. " + - "NEVER leave the current level unchanged by inertia, and NEVER reply to a trivial turn before considering a downshift."; + "Match reasoning effort to the work: lower it for routine or trivial turns, raise it for ambiguity, debugging, risky changes, or multi-step synthesis. " + + "Use the per-turn recommendation when one is provided, and reassess when new evidence changes the work; otherwise call the tool only when the current level no longer fits the task."; + +export type Config = { + enabled: boolean; + quiet: boolean; + autoSelect: boolean; + toolName: string; + toolDescription: string; + systemPrompt: string; +}; -export const configDefaults = { +const defaults: Config = { enabled: true, quiet: false, + autoSelect: false, toolName: "set_reasoning_effort", toolDescription: "Set your reasoning effort", systemPrompt: defaultSystemPrompt, }; -export const ConfigSchema = z - .object({ - enabled: z.boolean().default(configDefaults.enabled), - quiet: z.boolean().default(configDefaults.quiet), - toolName: z.string().default(configDefaults.toolName), - toolDescription: z.string().default(configDefaults.toolDescription), - systemPrompt: z.string().default(configDefaults.systemPrompt), - }) - .optional() - .default(configDefaults); +export const parseConfig = (options: unknown): { config: Config } | { error: string } => { + if (options === undefined) return { config: { ...defaults } }; + if (typeof options !== "object" || options === null || Array.isArray(options)) { + return { error: "options must be an object" }; + } + + const input = options as Record; + const config = { ...defaults }; + + for (const key of ["enabled", "quiet", "autoSelect"] as const) { + const value = input[key]; + if (value === undefined) continue; + if (typeof value !== "boolean") return { error: `${key} must be a boolean` }; + config[key] = value; + } + + for (const key of ["toolName", "toolDescription", "systemPrompt"] as const) { + const value = input[key]; + if (value === undefined) continue; + if (typeof value !== "string") return { error: `${key} must be a string` }; + config[key] = value; + } + + return { config }; +}; diff --git a/src/index.test.ts b/src/index.test.ts index 991f46a..ccf9a5d 100644 --- a/src/index.test.ts +++ b/src/index.test.ts @@ -1,418 +1,419 @@ import { describe, expect, test, vi } from "vitest"; import { AdaptiveThinkingPlugin } from "./index"; -const variants = { - none: {}, - low: {}, - medium: {}, - high: {}, - xhigh: {}, -}; - -const createClient = (sessionID: string, messages: Array = []) => { - const promptAsync = vi.fn(async () => ({ error: undefined })); - - return { - client: { - session: { - messages: vi.fn(async () => ({ data: messages, error: undefined })), - promptAsync, - }, - provider: { - list: vi.fn(async () => ({ - data: { - all: [ - { - id: "provider", - models: { - model: { variants }, - }, - }, - ], - }, - error: undefined, - })), - }, - app: { - agents: vi.fn(async () => ({ data: [], error: undefined })), - log: vi.fn(), - }, - tui: { - showToast: vi.fn(), - }, - }, - toolContext: { - sessionID, - messageID: "message", - agent: "agent", - directory: "/tmp", - worktree: "/tmp", - abort: new AbortController().signal, - metadata: vi.fn(), - ask: vi.fn(), - }, - }; -}; +const variants = { none: {}, low: {}, medium: {}, high: {}, xhigh: {} }; -const createMessage = (variant = "medium") => ({ +const message = (variant?: string, agent = "agent") => ({ info: { - id: "message", role: "user", - agent: "agent", - model: { - providerID: "provider", - modelID: "model", - variant, - }, + agent, + model: { providerID: "provider", modelID: "model", ...(variant ? { variant } : {}) }, }, parts: [], }); -const createAgentMessage = (agent: string) => ({ +const assistant = (variant?: string) => ({ info: { - id: `message-${agent}`, - role: "user", - agent, + role: "assistant", + agent: "agent", + providerID: "provider", + modelID: "model", + ...(variant ? { variant } : {}), }, parts: [], }); -const setReasoningEffort = async ( - plugin: Awaited>, - args: { level: string; persist: boolean }, - context: unknown, - toolName = "set_reasoning_effort", -) => plugin.tool![toolName]!.execute(args, context as never); +const createClient = ( + messages: Array = [], + modelVariants: Record = variants, + agents: Record = {}, +) => { + const prompts: Array = []; + const client = { + session: { + messages: vi.fn(async () => ({ data: messages, error: undefined })), + promptAsync: vi.fn(async ({ body }: any) => { + prompts.push(body); + messages.push(message(body.variant)); + return { error: undefined }; + }), + }, + provider: { + list: vi.fn(async () => ({ + data: { + all: [{ id: "provider", models: { model: { variants: modelVariants } } }], + }, + error: undefined, + })), + }, + app: { + agents: vi.fn(async () => ({ + data: Object.entries(agents).map(([name, variant]) => ({ name, variant })), + error: undefined, + })), + log: vi.fn(), + }, + tui: { showToast: vi.fn() }, + }; + return { client, prompts }; +}; -const touchTemporarySession = async (sessionID: string) => { - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); - const plugin = await AdaptiveThinkingPlugin({ client } as never); +const ctx = (sessionID: string, agent = "agent") => ({ sessionID, agent }); - await setReasoningEffort(plugin, { level: "low", persist: false }, toolContext); +const set = (plugin: any, level: string, persist: boolean, sessionID: string) => + plugin.tool.set_reasoning_effort.execute({ level, persist }, ctx(sessionID)); - return { client, plugin }; -}; +const idle = (plugin: any, sessionID: string) => + plugin.event({ event: { type: "session.idle", properties: { sessionID } } }); + +const transform = ( + plugin: any, + sessionID: string, + system: Array, + model: any = { variants }, +) => plugin["experimental.chat.system.transform"]({ sessionID, model }, { system }); describe("AdaptiveThinkingPlugin", () => { - test("uses defaults when no options are provided", async () => { - const sessionID = "default-options"; - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); - const plugin = await AdaptiveThinkingPlugin({ client } as never); - const system: string[] = []; - - await setReasoningEffort(plugin, { level: "high", persist: true }, toolContext); - await plugin["experimental.chat.system.transform"]!( - { - sessionID, - model: { variants }, - } as never, - { system }, - ); + test("applies defaults, persists a level and preserves agent and model", async () => { + const { client, prompts } = createClient([message("medium")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); - expect(plugin.tool?.set_reasoning_effort?.description).toBe("Set your reasoning effort"); - expect(client.session.promptAsync).toHaveBeenCalledWith( - expect.objectContaining({ body: expect.objectContaining({ variant: "high" }) }), + expect(plugin.tool.set_reasoning_effort.description).toBe("Set your reasoning effort"); + expect(await set(plugin, "high", true, "defaults")).toBe("Reasoning effort set to high"); + expect(prompts[0]).toMatchObject({ + agent: "agent", + variant: "high", + model: { providerID: "provider", modelID: "model" }, + }); + + const system: Array = []; + await transform(plugin, "defaults", system); + expect(system[0]).toContain("Match reasoning effort to the work"); + expect(system[0]).toContain("Current reasoning effort level: high."); + expect(system[0]).toContain( + "Valid reasoning effort levels for this session: none, low, medium, high, xhigh.", ); - expect(system[0]).toContain("You MUST manage reasoning effort actively"); expect(system[0]).toContain("set_reasoning_effort"); }); - test("preserves the calling agent and model when setting reasoning effort", async () => { - const sessionID = "preserve-agent-model"; - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); - toolContext.agent = "custom-worker"; - const plugin = await AdaptiveThinkingPlugin({ client } as never); - - await setReasoningEffort(plugin, { level: "high", persist: true }, toolContext); - - expect(client.session.promptAsync).toHaveBeenCalledWith( - expect.objectContaining({ - body: expect.objectContaining({ - agent: "custom-worker", - model: { providerID: "provider", modelID: "model" }, - variant: "high", - }), - }), - ); - }); + test("honors custom configuration and merges into an existing system prompt", async () => { + const { client } = createClient([message("medium")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never, { + toolName: "adjust_reasoning", + toolDescription: "Adjust it", + systemPrompt: "Prefer cheap.", + }); - test("returns no hooks when disabled", async () => { - const { client } = createClient("disabled-plugin"); - const plugin = await AdaptiveThinkingPlugin({ client } as never, { enabled: false }); + expect(plugin.tool.adjust_reasoning.description).toBe("Adjust it"); + expect(plugin.tool.set_reasoning_effort).toBeUndefined(); - expect(plugin).toEqual({}); + const system = ["Base prompt."]; + await transform(plugin, "custom", system); + expect(system).toHaveLength(1); + expect(system[0]?.startsWith("Base prompt.\n\nPrefer cheap.")).toBe(true); + expect(system[0]).not.toContain("Match reasoning effort"); + expect(system[0]).toContain("adjust_reasoning"); }); - test("logs configuration errors without creating hooks", async () => { - const { client } = createClient("invalid-config"); - const plugin = await AdaptiveThinkingPlugin({ client } as never, { enabled: "yes" }); - - expect(plugin).toEqual({}); - expect(client.tui.showToast).toHaveBeenCalledWith({ - body: { - variant: "error", - message: "Invalid Adaptive Thinking plugin configuration, see logs for details", - }, - }); - expect(client.app.log).toHaveBeenCalledWith({ - body: expect.objectContaining({ - level: "error", - service: "opencode-adaptive-thinking", - message: expect.stringContaining("Invalid Adaptive Thinking plugin configuration"), - }), - }); - }); + test("selects and carries an effort when autoSelect is enabled", async () => { + const { client } = createClient(); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never, { autoSelect: true }); + const input = { sessionID: "auto-select", model: { providerID: "provider", modelID: "model" } }; - test("suppresses configuration error toast when quiet is enabled", async () => { - const { client } = createClient("quiet-invalid-config"); - const plugin = await AdaptiveThinkingPlugin({ client } as never, { - enabled: "yes", - quiet: true, + const first = message().info; + await plugin["chat.message"](input, { + message: first, + parts: [{ type: "text", text: "fix a typo" }], }); + expect((first.model as { variant?: string }).variant).toBe("low"); - expect(plugin).toEqual({}); - expect(client.tui.showToast).not.toHaveBeenCalled(); - expect(client.app.log).toHaveBeenCalledWith({ - body: expect.objectContaining({ - level: "error", - service: "opencode-adaptive-thinking", - message: expect.stringContaining("Invalid Adaptive Thinking plugin configuration"), - }), + const second = message().info; + await plugin["chat.message"](input, { + message: second, + parts: [{ type: "text", text: "debug the flaky race" }], }); + expect((second.model as { variant?: string }).variant).toBe("low"); }); - test("keeps log service name internal", async () => { - const sessionID = "fixed-service-name"; - const { client, toolContext } = createClient(sessionID); - client.session.messages.mockResolvedValueOnce({ - error: { data: { message: "unavailable" } }, - } as never); - const plugin = await AdaptiveThinkingPlugin({ client } as never, { - serviceName: "custom-service", - }); + test("honors an explicit variant in autoSelect mode", async () => { + const { client } = createClient(); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never, { autoSelect: true }); + const input = { + sessionID: "auto-explicit", + model: { providerID: "provider", modelID: "model" }, + }; - await setReasoningEffort(plugin, { level: "low", persist: false }, toolContext); + const first = message().info; + await plugin["chat.message"]( + { ...input, variant: "high" }, + { message: first, parts: [{ type: "text", text: "fix a typo" }] }, + ); + expect((first.model as { variant?: string }).variant).toBe("high"); - expect(client.app.log).toHaveBeenCalledWith({ - body: expect.objectContaining({ - service: "opencode-adaptive-thinking", - }), + const second = message().info; + await plugin["chat.message"](input, { + message: second, + parts: [{ type: "text", text: "fix a typo" }], }); + expect((second.model as { variant?: string }).variant).toBe("high"); }); - test("uses configured tool name and description", async () => { - const sessionID = "custom-tool"; - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); - const plugin = await AdaptiveThinkingPlugin({ client } as never, { - toolName: "adjust_reasoning", - toolDescription: "Adjust reasoning depth for the next response", - }); + test("leaves messages alone when autoSelect is disabled or the prompt is ignored", async () => { + const { client } = createClient(); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); - expect(plugin.tool?.set_reasoning_effort).toBeUndefined(); - expect(plugin.tool?.adjust_reasoning?.description).toBe( - "Adjust reasoning depth for the next response", + const manual = message().info; + await plugin["chat.message"]( + { sessionID: "auto-off", model: { providerID: "provider", modelID: "model" } }, + { message: manual, parts: [{ type: "text", text: "debug the flaky race" }] }, ); + expect((manual.model as { variant?: string }).variant).toBeUndefined(); - await setReasoningEffort( - plugin, - { level: "high", persist: true }, - toolContext, - "adjust_reasoning", + const ignored = message().info; + await plugin["chat.message"]( + { sessionID: "auto-off", model: { providerID: "provider", modelID: "model" } }, + { message: ignored, parts: [{ type: "text", text: "debug the flaky race", ignored: true }] }, ); + expect((ignored.model as { variant?: string }).variant).toBeUndefined(); + }); - expect(client.session.promptAsync).toHaveBeenCalledWith( - expect.objectContaining({ body: expect.objectContaining({ variant: "high" }) }), - ); + test("recommends an effort change only when the turn needs one", async () => { + const { client } = createClient([message("low")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); + const input = { sessionID: "recommend", model: { providerID: "provider", modelID: "model" } }; + + await plugin["chat.message"](input, { + message: message("low").info, + parts: [{ type: "text", text: "debug the flaky race" }], + }); + const system: Array = []; + await transform(plugin, "recommend", system); + expect(system[0]).toContain("This turn suggests high effort"); + expect(system[0]).toContain("with `high`"); + + await plugin["chat.message"](input, { + message: message("low").info, + parts: [{ type: "text", text: "fix a typo" }], + }); + const next: Array = []; + await transform(plugin, "recommend", next); + expect(next[0]).toContain("only if the work grows beyond it"); }); - test("uses configured system prompt guidance", async () => { - const sessionID = "custom-system-prompt"; - const { client } = createClient(sessionID, [createMessage("medium")]); - const plugin = await AdaptiveThinkingPlugin({ client } as never, { - systemPrompt: "Prefer the cheapest reasoning level that can safely complete the task.", + test("clears the recommendation at the end of a turn", async () => { + const { client } = createClient([message("low")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); + const input = { + sessionID: "recommend-idle", + model: { providerID: "provider", modelID: "model" }, + }; + + await plugin["chat.message"](input, { + message: message("low").info, + parts: [{ type: "text", text: "debug the flaky race" }], }); - const system: string[] = []; - - await plugin["experimental.chat.system.transform"]!( - { - sessionID, - model: { variants }, - } as never, - { system }, - ); + await idle(plugin, "recommend-idle"); - expect(system).toHaveLength(1); + const system: Array = []; + await transform(plugin, "recommend-idle", system); expect(system[0]).toContain( - "Prefer the cheapest reasoning level that can safely complete the task.", + "Call `set_reasoning_effort` only when the current level no longer fits the task.", ); - expect(system[0]).not.toContain("You MUST manage reasoning effort actively"); }); - test("ignores cached reasoning effort when it is invalid for the current model", async () => { - const sessionID = "stale-cached-variant"; - const switchedModelVariants = { - low: {}, - medium: {}, - }; - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); - const plugin = await AdaptiveThinkingPlugin({ client } as never); - const system: string[] = []; - - await setReasoningEffort(plugin, { level: "high", persist: true }, toolContext); - await plugin["experimental.chat.system.transform"]!( - { - sessionID, - model: { variants: switchedModelVariants }, - } as never, - { system }, - ); + test("tells the agent that effort is managed when autoSelect is enabled", async () => { + const { client } = createClient(); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never, { autoSelect: true }); + const system: Array = []; - expect(system[0]).toContain("Current reasoning effort level: medium."); - expect(system[0]).toContain("Valid reasoning effort levels for this session: low, medium."); - expect(system[0]).not.toContain("Current reasoning effort level: high."); + await transform(plugin, "auto-instruction", system); + expect(system[0]).toContain("selected automatically"); }); - test("falls back to the newest agent variant when messages do not expose a model variant", async () => { - const sessionID = "newest-agent-fallback"; - const { client } = createClient(sessionID, [ - createAgentMessage("old-agent"), - createAgentMessage("new-agent"), - ]); - client.app.agents.mockResolvedValueOnce({ - data: [ - { name: "old-agent", variant: "high" }, - { name: "new-agent", variant: "low" }, - ], - error: undefined, - } as never); - const plugin = await AdaptiveThinkingPlugin({ client } as never); - const system: string[] = []; - - await plugin["experimental.chat.system.transform"]!( - { - sessionID, - model: { variants }, - } as never, - { system }, - ); - - expect(system[0]).toContain("Current reasoning effort level: low."); + test("returns no hooks when disabled", async () => { + const { client } = createClient(); + expect(await AdaptiveThinkingPlugin({ client } as never, { enabled: false })).toEqual({}); }); - test("resets a temporary reasoning effort once and keeps the reset prompt ignored", async () => { - const sessionID = "temporary-reset"; - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); - const plugin = await AdaptiveThinkingPlugin({ client } as never); - - await setReasoningEffort(plugin, { level: "low", persist: false }, toolContext); - await plugin.event!({ - event: { type: "session.idle", properties: { sessionID } }, - } as never); - await plugin.event!({ - event: { type: "session.idle", properties: { sessionID } }, - } as never); - - expect(client.session.promptAsync).toHaveBeenCalledTimes(2); - const promptCalls = client.session.promptAsync.mock.calls as Array>; - const secondPrompt = promptCalls[1]?.[0]; - expect(secondPrompt).toMatchObject({ + test("reports invalid configuration and suppresses the toast when quiet", async () => { + const invalid = createClient(); + expect( + await AdaptiveThinkingPlugin({ client: invalid.client } as never, { enabled: "yes" }), + ).toEqual({}); + expect(invalid.client.tui.showToast).toHaveBeenCalledWith({ body: { - agent: "agent", - model: { providerID: "provider", modelID: "model" }, - variant: "medium", - parts: [{ ignored: true, synthetic: true }], + variant: "error", + message: expect.stringContaining("Invalid Adaptive Thinking plugin configuration"), }, }); + expect(invalid.client.app.log).toHaveBeenCalledWith({ + body: expect.objectContaining({ + level: "error", + message: expect.stringContaining("enabled must be a boolean"), + }), + }); + + const quiet = createClient(); + await AdaptiveThinkingPlugin({ client: quiet.client } as never, { + enabled: "yes", + quiet: true, + }); + expect(quiet.client.tui.showToast).not.toHaveBeenCalled(); }); - test("evicts oldest session state when the cache is full", async () => { - const oldest = await touchTemporarySession("lru-oldest"); + test("rejects an invalid level without prompting", async () => { + const { client, prompts } = createClient([message("medium")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); - for (let i = 0; i < 500; i++) { - await touchTemporarySession(`lru-session-${i}`); - } + expect(await set(plugin, "extreme", true, "invalid-level")).toContain( + "Invalid reasoning effort level", + ); + expect(prompts).toHaveLength(0); + }); - await oldest.plugin.event!({ - event: { type: "session.idle", properties: { sessionID: "lru-oldest" } }, - } as never); + test("reports models without reasoning variants", async () => { + const { client, prompts } = createClient([message("medium")], {}); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); - expect(oldest.client.session.promptAsync).toHaveBeenCalledTimes(1); + expect(await set(plugin, "high", true, "no-variants")).toContain( + "no valid reasoning effort levels", + ); + expect(prompts).toHaveLength(0); }); - test("does not reset persisted reasoning effort on idle", async () => { - const sessionID = "persisted-effort"; - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); - const plugin = await AdaptiveThinkingPlugin({ client } as never); + test("resets a temporary level to the level from the session history", async () => { + const { client, prompts } = createClient([message("medium")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); - await setReasoningEffort(plugin, { level: "high", persist: true }, toolContext); - await plugin.event!({ - event: { type: "session.idle", properties: { sessionID } }, - } as never); + await set(plugin, "low", false, "temporary"); + await idle(plugin, "temporary"); - expect(client.session.promptAsync).toHaveBeenCalledTimes(1); + expect(prompts.map((prompt) => prompt.variant)).toEqual(["low", "medium"]); + expect(prompts[1]).toMatchObject({ + agent: "agent", + model: { providerID: "provider", modelID: "model" }, + parts: [{ synthetic: true, ignored: true }], + }); }); - test("rejects invalid reasoning effort levels before prompting", async () => { - const sessionID = "invalid-level"; - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); - const plugin = await AdaptiveThinkingPlugin({ client } as never); + test("reads the model and variant from assistant messages", async () => { + const { client, prompts } = createClient([message("medium"), assistant("high")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); - const result = await setReasoningEffort( - plugin, - { level: "extreme", persist: true }, - toolContext, - ); + await set(plugin, "low", false, "assistant-shape"); + await idle(plugin, "assistant-shape"); + + expect(prompts[0].model).toEqual({ providerID: "provider", modelID: "model" }); + expect(prompts.map((prompt) => prompt.variant)).toEqual(["low", "high"]); + }); + + test("keeps the original baseline across repeated temporary changes", async () => { + const changed = createClient([message("medium")]); + const changedPlugin: any = await AdaptiveThinkingPlugin({ client: changed.client } as never); + await set(changedPlugin, "low", false, "repeat-different"); + await set(changedPlugin, "high", false, "repeat-different"); + await idle(changedPlugin, "repeat-different"); + + const repeated = createClient([message("medium")]); + const repeatedPlugin: any = await AdaptiveThinkingPlugin({ client: repeated.client } as never); + await set(repeatedPlugin, "low", false, "repeat-same"); + await set(repeatedPlugin, "low", false, "repeat-same"); + await idle(repeatedPlugin, "repeat-same"); + + expect(changed.prompts.map((prompt) => prompt.variant)).toEqual(["low", "high", "medium"]); + expect(repeated.prompts.map((prompt) => prompt.variant)).toEqual(["low", "low", "medium"]); + }); - expect(result).toContain("Invalid reasoning effort level"); - expect(client.session.promptAsync).not.toHaveBeenCalled(); + test("keeps a temporary reset pending after a failed attempt", async () => { + const { client } = createClient([message("medium")]); + client.session.promptAsync + .mockResolvedValueOnce({ error: undefined }) + .mockResolvedValueOnce({ error: { data: { message: "unavailable" } } } as never) + .mockResolvedValueOnce({ error: undefined }); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); + + await set(plugin, "low", false, "reset-retry"); + await idle(plugin, "reset-retry"); + await idle(plugin, "reset-retry"); + + expect(client.session.promptAsync.mock.calls.map(([call]: any) => call.body.variant)).toEqual([ + "low", + "medium", + "medium", + ]); + expect(client.app.log).toHaveBeenCalled(); }); - test("handles provider lookup failure without prompting", async () => { - const sessionID = "provider-lookup-failure"; - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); + test("reports a provider lookup failure without prompting", async () => { + const { client, prompts } = createClient([message("medium")]); client.provider.list.mockResolvedValueOnce({ - error: { data: { message: "provider unavailable" } }, + data: undefined, + error: { message: "down" }, } as never); - const plugin = await AdaptiveThinkingPlugin({ client } as never); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); - const result = await setReasoningEffort(plugin, { level: "high", persist: true }, toolContext); + expect(await set(plugin, "high", true, "provider-failure")).toContain( + "no valid reasoning effort levels", + ); + expect(prompts).toHaveLength(0); + expect(client.app.log).toHaveBeenCalled(); + }); - expect(result).toContain("no valid reasoning effort levels"); - expect(client.session.promptAsync).not.toHaveBeenCalled(); - expect(client.app.log).toHaveBeenCalledWith({ - body: expect.objectContaining({ - level: "error", - message: expect.stringContaining("Failed to retrieve providers"), - }), - }); + test("does not reset a persisted level on idle", async () => { + const { client, prompts } = createClient([message("medium")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); + + await set(plugin, "high", true, "persisted"); + await idle(plugin, "persisted"); + expect(prompts).toHaveLength(1); }); - test("retries a failed temporary reset and clears it after a later success", async () => { - const sessionID = "recoverable-reset-failure"; - const { client, toolContext } = createClient(sessionID, [createMessage("medium")]); - client.session.promptAsync - .mockResolvedValueOnce({ error: undefined } as never) - .mockResolvedValueOnce({ error: { data: { message: "reset failed" } } } as never) - .mockResolvedValueOnce({ error: undefined } as never); - const plugin = await AdaptiveThinkingPlugin({ client } as never); - - await setReasoningEffort(plugin, { level: "low", persist: false }, toolContext); - await plugin.event!({ - event: { type: "session.idle", properties: { sessionID } }, - } as never); - await plugin.event!({ - event: { type: "session.idle", properties: { sessionID } }, - } as never); - await plugin.event!({ - event: { type: "session.idle", properties: { sessionID } }, + test("returns an error instead of throwing when messages cannot be read", async () => { + const { client } = createClient([message("medium")]); + client.session.messages.mockResolvedValueOnce({ + data: undefined, + error: { data: { message: "unavailable" } }, } as never); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); - expect(client.session.promptAsync).toHaveBeenCalledTimes(3); - expect(client.app.log).toHaveBeenCalledWith({ - body: expect.objectContaining({ - level: "error", - message: expect.stringContaining("Failed to reset reasoning effort"), - }), - }); + expect(await set(plugin, "high", true, "fetch-failure")).toContain( + "could not read the session", + ); + expect(client.app.log).toHaveBeenCalled(); + }); + + test("ignores a cached level that the current model does not support", async () => { + const { client } = createClient([message("medium")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); + await set(plugin, "high", true, "stale-cache"); + + const system: Array = []; + await transform(plugin, "stale-cache", system, { variants: { low: {}, medium: {} } }); + expect(system[0]).toContain("Valid reasoning effort levels for this session: low, medium."); + expect(system[0]).not.toContain("Current reasoning effort level: high."); + }); + + test("falls back to the agent variant when history has none", async () => { + const { client } = createClient([message(undefined, "worker")], variants, { worker: "low" }); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); + + const system: Array = []; + await transform(plugin, "agent-fallback", system, {}); + expect(system[0]).toContain("Current reasoning effort level: low."); + }); + + test("evicts the oldest session state when the cache is full", async () => { + const oldest = createClient([message("medium")]); + const oldestPlugin: any = await AdaptiveThinkingPlugin({ client: oldest.client } as never); + await set(oldestPlugin, "low", false, "lru-0"); + + for (let i = 1; i <= 500; i++) { + const { client } = createClient([message("medium")]); + const plugin: any = await AdaptiveThinkingPlugin({ client } as never); + await set(plugin, "low", false, `lru-${i}`); + } + + await idle(oldestPlugin, "lru-0"); + expect(oldest.prompts).toHaveLength(1); }); }); diff --git a/src/index.ts b/src/index.ts index 3b21c89..a6b26c8 100644 --- a/src/index.ts +++ b/src/index.ts @@ -1,18 +1,20 @@ import { tool, type Plugin } from "@opencode-ai/plugin"; -import type { Agent, Message, Model, Part, Provider } from "@opencode-ai/sdk/v2"; -import { ConfigSchema } from "./config.js"; +import type { Message, Part, Provider } from "@opencode-ai/sdk/v2"; import type { OpencodeClient } from "@opencode-ai/sdk"; +import { parseConfig } from "./config.js"; +import { classifyPrompt, pickVariant } from "./complexity.js"; const z = tool.schema; const serviceName = "opencode-adaptive-thinking"; const maxSessionStateSize = 500; +const historyLimit = 20; +type ModelInfo = { providerID: string; modelID: string }; +type SessionMessage = { info: Message; parts: Array }; type SessionState = { - currentVariant?: string; - persistedVariant?: string; - temporaryResetVariant?: string; - temporaryResetAgent?: string; - temporaryResetModel?: { providerID: string; modelID: string }; + variant?: string; + recommendation?: string; + reset?: { level: string; agent: string; model?: ModelInfo }; }; class SessionStateCache { @@ -34,28 +36,32 @@ class SessionStateCache { update(entry); this.entries.set(sessionID, entry); - while (this.entries.size > this.maxSize) { - const oldestSessionID = this.entries.keys().next().value; - if (!oldestSessionID) break; - this.entries.delete(oldestSessionID); + if (this.entries.size > this.maxSize) { + const oldest = this.entries.keys().next().value; + if (oldest) this.entries.delete(oldest); } } } const state = new SessionStateCache(maxSessionStateSize); -type PromptAsyncOptions = Parameters[0]; -type PromptAsyncBody = NonNullable & { variant: string }; -type ModelInfo = { providerID: string; modelID: string }; +type PromptAsyncBody = NonNullable< + Parameters[0]["body"] +> & { variant: string }; export const AdaptiveThinkingPlugin: Plugin = async ({ client }, options) => { - const configParseResult = ConfigSchema.safeParse(options); - if (!configParseResult.success) { - const quiet = - typeof options === "object" && options !== null && "quiet" in options - ? options.quiet === true - : false; - if (!quiet) { + const logError = (message: string, error: unknown) => + client.app.log({ + body: { + service: serviceName, + level: "error", + message: `${message}: ${JSON.stringify(error)}`, + }, + }); + + const parsed = parseConfig(options); + if ("error" in parsed) { + if (options?.quiet !== true) { client.tui.showToast({ body: { variant: "error", @@ -63,154 +69,104 @@ export const AdaptiveThinkingPlugin: Plugin = async ({ client }, options) => { }, }); } - client.app.log({ - body: { - service: serviceName, - level: "error", - message: `Invalid Adaptive Thinking plugin configuration: ${configParseResult.error.message}`, - }, - }); + logError("Invalid Adaptive Thinking plugin configuration", parsed.error); return {}; } - const config = configParseResult.data; + const config = parsed.config; if (!config.enabled) return {}; - const getModelVariants = (model: Model | undefined): string[] => { - const variants = model?.variants; - if (!variants) return []; - return Object.keys(variants); + const getModelVariants = (model: unknown) => { + const variants = (model as { variants?: Record } | undefined)?.variants; + return Object.keys(variants ?? {}); }; - const sendVariantPrompt = async ( - sessionID: string, - variant: string, - text: string, - agent: string, - model?: ModelInfo, - ignored = true, - ) => { - const body: PromptAsyncBody = { - noReply: true, - parts: [ - { - type: "text", - text, - synthetic: true, - ignored, - }, - ], - agent, - variant, - }; + let providers: Array | undefined; - if (model) { - body.model = model; + const variantsFor = async (model: ModelInfo) => { + if (!providers) { + const response = await client.provider.list(); + if (response.error) { + logError("Failed to retrieve providers", response.error); + return []; + } + providers = (response.data?.all ?? []) as unknown as Array; } - - return client.session.promptAsync({ - path: { id: sessionID }, - body, - }); + const provider = providers.find((candidate) => candidate.id === model.providerID); + return getModelVariants(provider?.models[model.modelID]); }; - const resolveValidVariants = async (sessionID: string, model?: Model): Promise => { - const inputModelVariants = getModelVariants(model); - if (inputModelVariants.length > 0) return inputModelVariants; - - const messagesResponse = await client.session.messages({ + const fetchMessages = async (sessionID: string) => { + const response = await client.session.messages({ path: { id: sessionID }, + query: { limit: historyLimit }, }); - if (messagesResponse.error) { - client.app.log({ - body: { - service: serviceName, - level: "error", - message: `Failed to retrieve messages for session ${sessionID}: ${JSON.stringify(messagesResponse.error.data)}`, - }, - }); - return []; + if (response.error) { + logError(`Failed to retrieve messages for session ${sessionID}`, response.error); + return; } - const modelInfo = resolveLatestModelInfo(messagesResponse.data); - if (!modelInfo) return []; + return (response.data ?? []) as Array; + }; - const providers = await client.provider.list(); - if (providers.error) { - client.app.log({ - body: { - service: serviceName, - level: "error", - message: `Failed to retrieve providers for session ${sessionID}: ${JSON.stringify(providers.error)}`, - }, - }); - return []; - } - const provider = providers.data?.all.find((p) => p.id === modelInfo.providerID) as - | Provider - | undefined; - if (!provider) return []; + const modelOf = (info: Message): ModelInfo => + "model" in info + ? { providerID: info.model.providerID, modelID: info.model.modelID } + : { providerID: info.providerID, modelID: info.modelID }; - return getModelVariants(provider.models[modelInfo.modelID] as Model | undefined); + const latestModel = (messages: Array): ModelInfo | undefined => { + const last = messages.at(-1)?.info; + return last ? modelOf(last) : undefined; }; - const resolveLatestModelInfo = ( - data: unknown, - ): { providerID: string; modelID: string } | undefined => { - const messages = data as Array<{ info: Message; parts: Array }>; + const latestVariant = async (messages: Array) => { + let agent: string | undefined; for (let i = messages.length - 1; i >= 0; i--) { - const message = messages[i]!; - if ("model" in message.info) { - const { providerID, modelID } = message.info.model; - if (providerID && modelID) { - return { providerID, modelID }; - } - } + const info = messages[i]!.info; + const variant = "model" in info ? info.model.variant : info.variant; + if (variant) return variant; + agent ??= info.agent; } - return; - }; + if (!agent) return; - const resolveCurrentVariant = async (sessionID: string): Promise => { - const messagesResponse = await client.session.messages({ - path: { id: sessionID }, - }); - if (messagesResponse.error) { - client.app.log({ - body: { - service: serviceName, - level: "error", - message: `Failed to retrieve messages for session ${sessionID}: ${JSON.stringify(messagesResponse.error.data)}`, - }, - }); + const response = await client.app.agents(); + if (response.error) { + logError("Failed to retrieve agents", response.error); return; } - let agentName: string | undefined; - const messages = messagesResponse.data as Array<{ info: Message; parts: Array }>; - for (let i = messages.length - 1; i >= 0; i--) { - const message = messages[i]!; - if ("model" in message.info && message.info.model.variant) { - return message.info.model.variant; - } - agentName ??= message.info.agent; - } - - if (!agentName) return; + const match = response.data?.find((candidate) => candidate.name === agent) as + | { variant?: string } + | undefined; + return match?.variant; + }; - const agentsResponse = await client.app.agents(); - if (agentsResponse.error) { - client.app.log({ - body: { - service: serviceName, - level: "error", - message: `Failed to retrieve agents for session ${sessionID}: ${JSON.stringify(agentsResponse.error)}`, - }, - }); - return; - } - const agents = agentsResponse.data; - const agent = agents?.find((a) => a.name === agentName) as Agent | undefined; - return agent?.variant; + const promptText = (parts: Array) => { + if (parts.some((part) => part.type === "text" && part.ignored)) return; + const text = parts + .filter((part) => part.type === "text") + .map((part) => part.text) + .join("\n") + .trim(); + return text.length > 0 ? text : undefined; }; + const sendVariantPrompt = ( + sessionID: string, + variant: string, + text: string, + agent: string, + model?: ModelInfo, + ) => + client.session.promptAsync({ + path: { id: sessionID }, + body: { + noReply: true, + parts: [{ type: "text", text, synthetic: true, ignored: true }], + agent, + variant, + ...(model ? { model } : {}), + } as PromptAsyncBody, + }); + return { tool: { [config.toolName]: tool({ @@ -230,52 +186,39 @@ export const AdaptiveThinkingPlugin: Plugin = async ({ client }, options) => { ), }, execute: async ({ level, persist }, { sessionID, agent }) => { - const validVariants = await resolveValidVariants(sessionID); - if (validVariants.length === 0) { + const messages = await fetchMessages(sessionID); + if (!messages) return "Failed to set reasoning effort: could not read the session"; + + const model = latestModel(messages); + const variants = model ? await variantsFor(model) : []; + if (variants.length === 0) { return "Failed to set reasoning effort: no valid reasoning effort levels are available for this session"; } - if (!validVariants.includes(level)) { - return `Invalid reasoning effort level: ${level}. Valid levels: ${validVariants.join(", ")}.`; + if (!variants.includes(level)) { + return `Invalid reasoning effort level: ${level}. Valid levels: ${variants.join(", ")}.`; } - const sessionState = state.get(sessionID); - const resetVariant = persist - ? undefined - : (sessionState?.persistedVariant ?? (await resolveCurrentVariant(sessionID))); - const currentModel = resolveLatestModelInfo( - (await client.session.messages({ path: { id: sessionID } })).data, - ); - - const promptResponse = await sendVariantPrompt( + const baseline = + state.get(sessionID)?.reset?.level ?? + state.get(sessionID)?.variant ?? + (await latestVariant(messages)); + const response = await sendVariantPrompt( sessionID, level, `Reasoning effort set to ${level}`, agent, - currentModel, + model, ); - if (promptResponse.error) { - return `Failed to set reasoning effort: ${JSON.stringify(promptResponse.error.data)}`; + if (response.error) { + return `Failed to set reasoning effort: ${JSON.stringify(response.error)}`; } state.update(sessionID, (entry) => { - entry.currentVariant = level; - if (persist) { - entry.persistedVariant = level; - delete entry.temporaryResetVariant; - delete entry.temporaryResetAgent; - delete entry.temporaryResetModel; - } else if (resetVariant && resetVariant !== level) { - entry.temporaryResetVariant = resetVariant; - entry.temporaryResetAgent = agent; - if (currentModel) { - entry.temporaryResetModel = currentModel; - } else { - delete entry.temporaryResetModel; - } + entry.variant = level; + if (persist || !baseline || baseline === level || !variants.includes(baseline)) { + delete entry.reset; } else { - delete entry.temporaryResetVariant; - delete entry.temporaryResetAgent; - delete entry.temporaryResetModel; + entry.reset ??= { level: baseline, agent, ...(model ? { model } : {}) }; } }); @@ -283,69 +226,131 @@ export const AdaptiveThinkingPlugin: Plugin = async ({ client }, options) => { }, }), }, + "chat.message": async ({ sessionID, variant }, { message, parts }) => { + const text = promptText(parts); + if (text === undefined) return; + + if (config.autoSelect && variant) { + (message.model as { variant?: string }).variant = variant; + state.update(sessionID, (entry) => { + entry.variant = variant; + delete entry.recommendation; + }); + return; + } + + const variants = await variantsFor(message.model); + const current = variant ?? (message.model as { variant?: string }).variant; + + if (config.autoSelect) { + const cached = state.get(sessionID)?.variant; + const level = + cached && variants.includes(cached) + ? cached + : pickVariant(variants, classifyPrompt(text)); + if (!level) return; + + (message.model as { variant?: string }).variant = level; + state.update(sessionID, (entry) => { + entry.variant = level; + delete entry.recommendation; + }); + return; + } + + if (current) { + state.update(sessionID, (entry) => { + entry.variant = current; + }); + } + + const recommended = pickVariant(variants, classifyPrompt(text)); + state.update(sessionID, (entry) => { + if (recommended) entry.recommendation = recommended; + else delete entry.recommendation; + }); + }, event: async ({ event }) => { - if (event.type === "session.idle") { - const sessionID = event.properties.sessionID; - const sessionState = state.get(sessionID); - const resetVariant = sessionState?.temporaryResetVariant; - const resetAgent = sessionState?.temporaryResetAgent; - const resetModel = sessionState?.temporaryResetModel; - if (!resetVariant || !resetAgent) return; - - const promptResponse = await sendVariantPrompt( - sessionID, - resetVariant, - `Reasoning effort reset to ${resetVariant}.`, - resetAgent, - resetModel, - ); - if (promptResponse.error) { - client.app.log({ - body: { - service: serviceName, - level: "error", - message: `Failed to reset reasoning effort on session idle: ${JSON.stringify(promptResponse.error.data)}`, - }, - }); - return; - } + if (event.type !== "session.idle") return; + + const sessionID = event.properties.sessionID; + const session = state.get(sessionID); + if (!session) return; + if (session.recommendation) { state.update(sessionID, (entry) => { - entry.currentVariant = resetVariant; - delete entry.temporaryResetVariant; - delete entry.temporaryResetAgent; - delete entry.temporaryResetModel; + delete entry.recommendation; }); + } + + const reset = session.reset; + if (!reset) return; + + const response = await sendVariantPrompt( + sessionID, + reset.level, + `Reasoning effort reset to ${reset.level}.`, + reset.agent, + reset.model, + ); + if (response.error) { + logError("Failed to reset reasoning effort on session idle", response.error); return; } + + state.update(sessionID, (entry) => { + entry.variant = reset.level; + delete entry.reset; + }); }, "experimental.chat.system.transform": async ({ sessionID, model }, { system }) => { if (!sessionID) return; - const variants = await resolveValidVariants(sessionID, model as Model); - if (variants.length === 0) return; + let variants = getModelVariants(model); + let variant = state.get(sessionID)?.variant; + if (variant && !variants.includes(variant)) variant = undefined; - const sessionState = state.get(sessionID); - const cachedVariant = sessionState?.currentVariant ?? sessionState?.persistedVariant; - let variant = cachedVariant && variants.includes(cachedVariant) ? cachedVariant : undefined; - if (!variant) { - const resolvedVariant = await resolveCurrentVariant(sessionID); - if (resolvedVariant && variants.includes(resolvedVariant)) { - variant = resolvedVariant; - state.update(sessionID, (entry) => { - entry.currentVariant = resolvedVariant; - }); + if (variants.length === 0 || !variant) { + const messages = await fetchMessages(sessionID); + if (!messages) return; + + if (variants.length === 0) { + const info = latestModel(messages); + variants = info ? await variantsFor(info) : []; + if (variants.length === 0) return; + } + + if (!variant) { + const resolved = await latestVariant(messages); + if (resolved && variants.includes(resolved)) { + variant = resolved; + state.update(sessionID, (entry) => { + entry.variant = resolved; + }); + } } } - system.push( - config.systemPrompt.trim() + - " " + - (variant ? `Current reasoning effort level: ${variant}. ` : "") + - `Valid reasoning effort levels for this session: ${variants.join(", ")}. ` + - `To change your reasoning effort, use the \`${config.toolName}\` tool with one of the valid levels. ` + - "Only call it when the task complexity justifies changing levels.", - ); + const recommendation = state.get(sessionID)?.recommendation; + const effortInstruction = config.autoSelect + ? `Reasoning effort is selected automatically; call \`${config.toolName}\` only to override it.` + : recommendation && recommendation !== variant + ? `This turn suggests ${recommendation} effort${variant ? ` (current: ${variant})` : ""}. Call \`${config.toolName}\` with \`${recommendation}\` before proceeding.` + : recommendation + ? `The current level fits this request; call \`${config.toolName}\` only if the work grows beyond it.` + : `Call \`${config.toolName}\` only when the current level no longer fits the task.`; + + const guidance = [ + config.systemPrompt.trim(), + variant ? `Current reasoning effort level: ${variant}.` : "", + `Valid reasoning effort levels for this session: ${variants.join(", ")}.`, + effortInstruction, + ] + .filter(Boolean) + .join(" "); + + const existing = system[0]; + system[0] = existing ? `${existing}\n\n${guidance}` : guidance; }, }; };