honeydo 0.1.1 → 0.1.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "honeydo",
-  "version": "0.1.1",
+  "version": "0.1.3",
   "description": "One CLI for every AI capability — chat, vision, image, video, audio. Built for AI agents.",
   "license": "MIT",
   "keywords": [

package/packages/gcli/dist/cli.d.ts

@@ -6,8 +6,9 @@
  * `gcli agy ...` routes to the agy CLI (explicit subcommand required); the
  * claude backend can switch cc-switch providers inline via
  * `claude -p ... --settings`. Adds value over calling the backends directly:
- * prompt via argv or stdin (`-p -`), 50k-char output truncation, a hard
- * timeout (spawn SIGTERM), explicit exit codes, and empty-output detection.
+ * prompt via argv or stdin (`-p -`), a hard timeout (spawn SIGTERM), explicit
+ * exit codes, and empty-output detection. Output is passed through unmodified
+ * — size limits are the endpoint's business, not ours.
  *
  * The claude backend resolves `--provider <name>` from the cc-switch SQLite
  * DB (read-only) and never rewrites ~/.claude/settings.json — the provider
@@ -25,7 +26,6 @@
  * backend's own timeout flag.
  */
 export declare const DEFAULT_TIMEOUT_MS = 300000;
-export declare const CHARACTER_LIMIT = 50000;
 export declare const API_DEFAULT_MAX_TOKENS = 80000;
 /**
  * Path to the cc-switch SQLite database. cc-switch ships provider configs
@@ -47,7 +47,28 @@ export declare const LAST_PROVIDER_PATH: string;
  * best-effort — the quota subtitle is an optimization, never an error.
  */
 export declare const QUOTA_CACHE_PATH: string;
-export type Subcommand = "agy" | "claude" | "api";
+/** hermes 主配置文件（model 段 + providers 段是本子命令的编辑目标）。 */
+export declare const HERMES_CONFIG_PATH: string;
+/** hermes env 文件（provider API key 的落点；全程保持 0o600）。 */
+export declare const HERMES_ENV_PATH: string;
+/** hermes cron 任务定义（重 pin 只读它；改动经 `hermes cron edit` 下发）。 */
+export declare const HERMES_CRON_JOBS_PATH: string;
+/** hermes 会话库（session_model_usage 表在此 —— 实测在顶层 state.db）。 */
+export declare const HERMES_STATE_DB_PATH: string;
+/** cc-switch name → hermes provider 字段映射（keyEnv/modelOverride）。 */
+export declare const HERMES_PROVIDERS_REGISTRY_PATH: string;
+/** 上一次切换的一跳记录（0o600，含旧 token 的 prevValue）。 */
+export declare const HERMES_STATE_PATH: string;
+/** hermes providers 条目的 transport 常量（cc-switch 只提供 anthropic 兼容端点）。 */
+export declare const HERMES_TRANSPORT = "anthropic_messages";
+/** 切换后验证 ping 的硬超时。 */
+export declare const HERMES_VERIFY_TIMEOUT_MS = 60000;
+/**
+ * 内置 seed（防推导造出与现有配置平行的条目）：键是 cc-switch provider 名，
+ * 值是 hermes 侧已手工建立好的 id/keyEnv。
+ */
+export declare const HERMES_PROVIDER_SEEDS: HermesProviderRegistry;
+export type Subcommand = "agy" | "claude" | "api" | "hermes";
 export type SubcommandResult = {
     subcommand: Subcommand | undefined;
     rest: string[];
@@ -95,6 +116,66 @@ export type RawProvider = {
     name: string;
     settingsConfig: unknown;
 };
+/** registry 单条：cc-switch name → hermes providers 条目字段。 */
+export type HermesProviderRegistryEntry = {
+    id: string;
+    keyEnv: string;
+    modelOverride?: string;
+};
+/** `~/.config/gcli/hermes-providers.json` 的形状（损坏/缺失视为空）。 */
+export type HermesProviderRegistry = Record<string, HermesProviderRegistryEntry>;
+/** 一跳切换的一端（from/to 同构）。 */
+export type HermesModelPoint = {
+    id: string;
+    model: string;
+    base_url: string;
+};
+/** 一条被重 pin 的 cron job 的旧 pin（rollback 回放用）。 */
+export type CronRepinTarget = {
+    jobId: string;
+    prevProvider: string;
+    prevModel: string | null;
+};
+/** `~/.config/gcli/hermes-state.json`（0o600，env.prevValue 含旧 token）。 */
+export type HermesStateFile = {
+    lastSwitch: {
+        ts: number;
+        ccName: string;
+        to: HermesModelPoint;
+        from: HermesModelPoint;
+        configBackup: string;
+        env: {
+            key: string;
+            prevValue: string | null;
+        };
+        cronRepinned: CronRepinTarget[];
+    };
+};
+/** editHermesConfig 的编辑指令。 */
+export type HermesConfigEdit = {
+    model: {
+        default: string;
+        provider: string;
+        base_url: string;
+    };
+    provider: {
+        id: string;
+        name: string;
+        base_url: string;
+        transport: string;
+        key_env: string;
+        default_model: string;
+    };
+};
+/** parseHermesConfig 的读取结果（status / 冲突检测共用）。 */
+export type HermesConfigInfo = {
+    model: {
+        default?: string;
+        provider?: string;
+        base_url?: string;
+    };
+    providerIds: string[];
+};
 export type ProviderLookup = {
     ok: true;
     providers: RawProvider[];
@@ -166,6 +247,29 @@ export type RunDeps = {
      * after an arrow-key picker confirm, before any backend spawn.
      */
     writeLastProvider: (name: string) => Promise<void>;
+    /** Spawn `hermes <args>`（cron edit / gateway / -z ping），照 runClaude 的 timeout/SIGTERM 契约。 */
+    runHermes: (args: string[], timeoutMs?: number) => Promise<SpawnResult>;
+    /** 读 UTF-8 文本文件；缺失/不可读 → undefined（不 throw）。 */
+    readTextFile: (path: string) => Promise<string | undefined>;
+    /** 原子写（tmp + rename）；mode 缺省时沿用目标文件现有权限（不存在则 0o600）。 */
+    writeTextFileAtomic: (path: string, text: string, mode?: number) => Promise<{
+        ok: true;
+    } | {
+        ok: false;
+        error: string;
+    }>;
+    /** 整文件复制（备份/整文件恢复）。 */
+    copyFile: (src: string, dest: string) => Promise<{
+        ok: true;
+    } | {
+        ok: false;
+        error: string;
+    }>;
+    /** 查 state.db session_model_usage 最新行（best-effort 验证）；任何失败 → undefined。 */
+    queryLastSessionModel: () => Promise<{
+        model: string;
+        provider: string;
+    } | undefined>;
 };
 export type RunOutcome = {
     exitCode: number;
@@ -185,7 +289,6 @@ export type RunOutcome = {
  * prompt="claude" — the subcommand must literally lead.
  */
 export declare function parseSubcommand(argv: string[]): SubcommandResult;
-export declare function truncate(text: string): string;
 export interface GcliOptions {
     /** Omit for interactive mode (no -p emitted). */
     prompt?: string;
@@ -331,6 +434,17 @@ export interface ApiRequest {
     prompt: string;
     stream: boolean;
     timeoutMs: number;
+    /**
+     * Explicit thinking control: "off" sends {type:"disabled"}, "on" sends
+     * {type:"enabled", budget_tokens: floor(maxTokens/2)}. Undefined (= CLI
+     * `--thinking auto`) omits the field entirely and keeps the endpoint's
+     * default — which on bigmodel GLM endpoints means thinking ON, where
+     * thinking shares the max_tokens budget with text (small budgets can end
+     * up thinking-only, no text block).
+     */
+    thinking?: "off" | "on";
+    /** Transient-failure retries (network error / 408/429/5xx / malformed JSON / empty body). 0 = single attempt. */
+    retries?: number;
 }
 /**
  * Build the HTTP body for an anthropic-compatible /v1/messages request.
@@ -340,9 +454,12 @@ export interface ApiRequest {
  * high-quality output in ~45s vs claude-agent's 53min). We rely on a sufficient
  * --max-tokens budget (default 80000) to cover both thinking and text, not on
  * disabling thinking. Disabling it would discard the very capability we chose
- * k3 for.
+ * k3 for. (`--thinking auto`, the CLI default, preserves this omitted-field
+ * behaviour. bigmodel GLM endpoints default thinking ON and share max_tokens
+ * between thinking and text — small explicit budgets can end up thinking-only;
+ * --thinking off/on send explicit disabled/enabled for that case.)
  */
-export declare function buildApiBody(req: Pick<ApiRequest, "model" | "maxTokens" | "prompt" | "stream">): Record<string, unknown>;
+export declare function buildApiBody(req: Pick<ApiRequest, "model" | "maxTokens" | "prompt" | "stream" | "thinking">): Record<string, unknown>;
 /**
  * Build the full URL for the messages endpoint. The cc-switch base_url is
  * stored with a trailing slash (e.g. `https://api.kimi.com/coding/`); we
@@ -368,6 +485,131 @@ export declare function extractTextDelta(line: string): string | null;
  * we concatenate every text block in order.
  */
 export declare function extractNonStreamText(body: unknown): string;
+/** What a single SSE `data:` line contributes to the stream aggregation. */
+export interface SseLineMeta {
+    /** Text fragment for content_block_delta/text_delta events, else null. */
+    text: string | null;
+    /** Characters contributed by a thinking_delta (thinking models spend the
+     * shared max_tokens budget here before any text is emitted). */
+    thinkingChars: number;
+    /** stop_reason carried by a message_delta event (e.g. "max_tokens"). */
+    stopReason?: string;
+    /** True when the line was a `data:` payload that JSON-parsed (any event type). */
+    parsed: boolean;
+}
+/**
+ * SSE line → structured meta (text delta + thinking volume + stop_reason).
+ *
+ * One JSON.parse per line feeding both the text aggregation and the "why did a
+ * stream produce no text" diagnosis (thinking-only streams that end in
+ * stop_reason=max_tokens are an endpoint budget issue, not a transport one).
+ * Non-`data:` lines / malformed JSON → empty meta.
+ */
+export declare function extractSseLineMeta(line: string): SseLineMeta;
+/** Non-streaming body → why extractNonStreamText found no text. */
+export interface NoTextBodyInfo {
+    /** content block type names in order, comma-joined ("" = no content at all). */
+    blocks: string;
+    sawThinking: boolean;
+    stopReason?: string;
+}
+/**
+ * Inspect a 200 non-streaming messages body that yielded no text, so the error
+ * can say WHY (thinking-only under a starved max_tokens budget vs an empty
+ * content array from a flaky gateway — the former is deterministic, the latter
+ * is worth retrying).
+ */
+export declare function describeNoTextBody(body: unknown): NoTextBodyInfo;
+/** HTTP statuses worth an automatic retry (transient server/gateway trouble). */
+export declare const API_RETRYABLE_STATUS: Set<number>;
+/** Parsed options for the hermes subcommand (strict: unknown flags error). */
+export interface ParsedHermesArgs {
+    /** Positional: provider name, or a reserved word (status/rollback). */
+    provider?: string;
+    model?: string;
+    dryRun: boolean;
+    verify: boolean;
+    keepOnFail: boolean;
+    help: boolean;
+}
+export type ParseHermesResult = ParsedHermesArgs | {
+    error: string;
+};
+/**
+ * Strip a trailing context-window marker like `[1M]` (a claude-agent
+ * convention cc-switch stores in model names; raw APIs reject it).
+ * Only a suffix anchored at the very end is removed.
+ */
+export declare function stripContextSuffix(model: string): string;
+/**
+ * Parse argv for `gcli hermes ...`. Strict like parseApiArgs (unknown flags
+ * are exit-2 errors — there is no backend to forward to). `--no-verify` is
+ * accepted via allowNegative. At most one positional.
+ */
+export declare function parseHermesArgs(argv: string[]): ParseHermesResult;
+/**
+ * Derive a hermes provider id from a cc-switch display name: lowercase,
+ * non-[a-z0-9] runs folded to `-`, leading/trailing dashes trimmed.
+ * May return "" for names with no alphanumerics (caller treats as error).
+ */
+export declare function deriveHermesId(ccName: string): string;
+/** Derive the .env key name from a hermes provider id: UPPER + `_API_KEY`. */
+export declare function deriveKeyEnv(id: string): string;
+/** Quote a scalar for our line-level YAML writer ('' escapes a single quote). */
+export declare function quoteYamlScalar(value: string): string;
+/**
+ * Targeted line-level YAML editor for hermes' config.yaml (zero-dep: no yaml
+ * lib guaranteed). Only the top-level `model:` (3 keys) and `providers:`
+ * (one entry) sections are touched; everything else is byte-preserved.
+ * Missing sections / unexpected nesting → {error}, never a guess (callers
+ * must not write on error). Idempotent by construction.
+ */
+export declare function editHermesConfig(text: string, edit: HermesConfigEdit): {
+    text: string;
+} | {
+    error: string;
+};
+/**
+ * Read (not edit) the parts of config.yaml the hermes backend needs: the
+ * current model section values and the list of providers entry ids.
+ * Missing model: section or unparseable model lines → {error}.
+ */
+export declare function parseHermesConfig(text: string): {
+    info: HermesConfigInfo;
+} | {
+    error: string;
+};
+/**
+ * Upsert one `KEY=value` line in a .env text (pure). Only active lines
+ * (`^KEY=`) match — commented-out occurrences are left alone; comments and
+ * blank lines are byte-preserved. value=null deletes the line(s); a missing
+ * key with a non-null value is appended at the end.
+ */
+export declare function upsertEnvLines(text: string, key: string, value: string | null): string;
+/**
+ * Plan cron re-pins: jobs that are enabled AND pinned to `oldProvider`
+ * (jobs.json shape `{jobs: [...]}`, a bare array tolerated). Malformed input
+ * yields an empty plan, never a throw.
+ */
+export declare function buildCronRepinPlan(jobsJson: unknown, oldProvider: string): CronRepinTarget[];
+/**
+ * Parse the provider registry (`hermes-providers.json`). Corrupt/missing
+ * content is best-effort → empty registry; entries lacking id/keyEnv are
+ * dropped.
+ */
+export declare function parseHermesRegistry(text: string): HermesProviderRegistry;
+export declare function serializeHermesRegistry(reg: HermesProviderRegistry): string;
+/**
+ * Parse `hermes-state.json` STRICTLY (畸形 → {error}，不猜): it carries the
+ * previous token and the rollback plan, so a half-broken state must not be
+ * acted on.
+ */
+export declare function parseHermesStateFile(text: string): {
+    state: HermesStateFile;
+} | {
+    error: string;
+};
+export declare function serializeHermesStateFile(s: HermesStateFile): string;
 export interface ParsedArgs {
     prompt?: string;
     model?: string;
@@ -399,6 +641,10 @@ export interface ParsedApiArgs {
     maxTokens: number;
     timeoutMs: number;
     stream: boolean;
+    /** `auto` (default) omits the thinking field; `off`/`on` send it explicitly. */
+    thinking: "auto" | "off" | "on";
+    /** Transient-failure retries (default 1; 0 disables). */
+    retries: number;
     version: boolean;
     help: boolean;
     /** --cwd was supplied (warned + ignored by the api backend). */
@@ -431,11 +677,20 @@ export declare function runClaudeInteractive(args: string[], cwd?: string): Prom
  *
  * - stream=true: reads the SSE body chunk-by-chunk, decodes UTF-8, splits on
  *   newlines, and aggregates `text_delta` payloads into stdout. Two clocks
- *   guard against hangs: an idle timer (reset on every text-bearing chunk)
- *   and an absolute timer (timeoutMs). Either firing aborts the fetch via
- *   AbortController → exit 1 with a timeout message.
+ *   guard against hangs: an idle timer (reset on EVERY received chunk —
+ *   thinking models can emit long non-text stretches) and an absolute timer
+ *   (timeoutMs). Either firing aborts the fetch via AbortController → exit 1
+ *   with a timeout message.
  * - stream=false: awaits the full JSON body and extracts `content[].text`,
- *   then applies the 50k-char truncation.
+ *   returned as-is (no size cap — endpoint limits are the endpoint's call).
+ *
+ * TRANSIENT failures are retried automatically (default 1 retry, `--retry N`):
+ * network-level fetch errors, HTTP 408/429/5xx, malformed JSON, empty-content
+ * 200 bodies, and streams that die or deliver no SSE events at all. Each retry
+ * note lands on stderr (stdout stays pipe-clean), so a recovered attempt still
+ * leaves a trace. Deterministic failures are NOT retried — timeouts, 4xx, and
+ * thinking-only responses (retrying cannot fix a max_tokens budget starved by
+ * endpoint-default thinking); those errors carry a diagnosis instead.
  *
  * HTTP errors (non-2xx, network failure, abort) → exit 1 with diagnostics on
  * stderr; stdout stays empty so the caller's empty-output guard still works.