howtofixcli 0.7.14 → 0.7.15

package/dist/model-catalog.js

@@ -1,217 +0,0 @@
-/**
- * Real context windows and reasoning support, from models.dev.
- *
- * `GET /v1/models` on the gateway returns ids and nothing else — verified against
- * the live service, whose response carries only `id`, `object`, `created`,
- * `owned_by` and `supported_endpoint_types`. So a synced model has no declared
- * window and no declared capability, and `/connect` filled the gap with two
- * guesses that were both wrong in the same direction:
- *
- *   - `contextWindow: 128_000` for every model. The gateway serves models with
- *     1,000,000-token windows (deepseek-v4.1-flash, kimi-k3, mimo-v2.5), so the
- *     context meter read ~8x too full and the CLI compacted a conversation that
- *     had barely started.
- *   - no `reasoning` field at all. The SDK gates thinking on it — the schema is
- *     `reasoning: boolean`, and `getSupportedThinkingLevels` returns `["off"]`
- *     the moment it is falsy — so every synced model lost its thinking levels
- *     even though the model reasons. That is the bug a user reports as "thinking
- *     is missing on a model that has it".
- *
- * models.dev publishes exactly these facts per model (`limit.context`,
- * `limit.output`, `reasoning`). It is a public catalogue, so the CLI can read it
- * instead of guessing. Nothing here is bundled: the catalogue is fetched once,
- * slimmed to the three fields that are used, and cached next to the agent state,
- * so a later start can repair models without a network call.
- *
- * The pure half (index, lookup) is separated from the network half so the
- * matching rules are testable without a fetch.
- */
-/** Where the catalogue lives. One document, every provider. */
-export const MODELS_DEV_URL = "https://models.dev/api.json";
-/** How long to wait for the catalogue. It is ~5 MB, so this is generous. */
-export const CATALOG_TIMEOUT_MS = 20_000;
-/**
- * Providers whose entry wins when the same model is listed by several.
- *
- * The catalogue is community-maintained and mirrors every reseller, so one model
- * id resolves to a dozen near-identical entries. The first-party listing is the
- * one whose numbers describe the model rather than a middleman's limits, so the
- * search prefers it. Order matters only among the preferred set; everything else
- * is ordered alphabetically so the result never depends on object iteration.
- */
-const PREFERRED_PROVIDERS = ["anthropic", "openai", "google", "deepseek", "zai", "xai", "moonshotai", "minimax", "alibaba", "xiaomi", "meta", "tencent", "nvidia", "venice", "openrouter"];
-/** The comparison form of a model id: case and punctuation are not meaningful. */
-export function normalizeModelId(id) {
-    return id.toLowerCase().replace(/[^a-z0-9]+/g, "");
-}
-/** The last path segment of an id, normalized: `z-ai/glm-5.3-fast` -> `glm53fast`. */
-function lastSegment(id) {
-    const parts = id.split("/");
-    return normalizeModelId(parts[parts.length - 1] ?? id);
-}
-/**
- * Index every model in the catalogue document.
- *
- * One entry per normalized id; when several providers list the same model, the
- * preferred provider wins (see {@link PREFERRED_PROVIDERS}). A model whose entry
- * carries no `limit.context` is skipped: without a window there is nothing to
- * correct, and indexing it would let a bad entry shadow a good one.
- */
-export function buildCatalogIndex(raw) {
-    const index = new Map();
-    const owners = new Map();
-    if (!raw || typeof raw !== "object")
-        return index;
-    const providers = raw;
-    const priority = (provider) => {
-        const at = PREFERRED_PROVIDERS.indexOf(provider);
-        return at === -1 ? PREFERRED_PROVIDERS.length : at;
-    };
-    // Sorted so the winner among equally-preferred providers is decided by name,
-    // not by the insertion order of a parsed JSON object — this number feeds a
-    // token budget, so it must not depend on how the document happened to be built.
-    for (const provider of Object.keys(providers).sort()) {
-        const models = providers[provider]?.models;
-        if (!models || typeof models !== "object")
-            continue;
-        for (const id of Object.keys(models)) {
-            const model = models[id];
-            const context = Number(model?.limit?.context);
-            if (!Number.isFinite(context) || context <= 0)
-                continue;
-            const key = normalizeModelId(id);
-            const rank = priority(provider);
-            const held = owners.get(key);
-            // A preferred provider always wins; otherwise the first (alphabetical) one
-            // keeps the slot. Never let a later, less-preferred listing overwrite it.
-            if (held !== undefined && held <= rank)
-                continue;
-            const output = Number(model?.limit?.output);
-            index.set(key, {
-                contextWindow: Math.floor(context),
-                maxTokens: Number.isFinite(output) && output > 0 ? Math.floor(output) : undefined,
-                reasoning: model?.reasoning === true,
-                vision: Array.isArray(model?.modalities?.input) && model.modalities.input.includes("image"),
-            });
-            owners.set(key, rank);
-        }
-    }
-    return index;
-}
-/**
- * The catalogue entry for a gateway model id, or undefined when it is unknown.
- *
- * Three passes, in order of how literal the match is: the whole id, the last path
- * segment, then a suffix. The gateway renames models (`deepseek-v4.1-flash` is
- * `DeepSeek-V4.1-Flash` in the catalogue, `glm-5.3-flash` is `zai/glm-5.3-flash`)
- * and drops vendor prefixes (`anthropic/claude-sonnet-5` keeps its own), so a
- * single exact match finds less than half. A miss is not a failure: the caller
- * keeps its conservative default, which is the behaviour before this module.
- */
-export function lookupModel(index, id) {
-    const wanted = normalizeModelId(id);
-    if (!wanted)
-        return undefined;
-    const exact = index.get(wanted);
-    if (exact)
-        return exact;
-    const wantedTail = lastSegment(id);
-    for (const [key, entry] of index) {
-        if (lastSegment(key) === wantedTail)
-            return entry;
-    }
-    for (const [key, entry] of index) {
-        if (key.endsWith(wanted) || wanted.endsWith(key))
-            return entry;
-    }
-    return undefined;
-}
-/** Where the slimmed catalogue is cached, so a later start needs no network. */
-export function catalogCachePath(agentDir) {
-    return `${agentDir}/models-dev-cache.json`;
-}
-/**
- * The serialized form of an index: `{ id: [context, output, reasoning, vision] }`.
- *
- * An array rather than an object per entry because the catalogue is 8,000+ rows
- * and repeating the keys eight thousand times triples the file for no reader's
- * benefit. The cache is internal — nothing outside this module parses it.
- *
- * `vision` was added after the first release, so a cache written by the older
- * build is a 3-tuple. {@link deserializeCatalog} reads the fourth slot as
- * optional and defaults it to `false`, which is the behaviour that build had;
- * the next fetch fills it in. Bumping a format number would instead throw the
- * whole cache away and cost every user a 5 MB download on the next start.
- */
-export function serializeCatalog(index) {
-    const out = {};
-    for (const [key, entry] of index) {
-        out[key] = [entry.contextWindow, entry.maxTokens ?? 0, entry.reasoning ? 1 : 0, entry.vision ? 1 : 0];
-    }
-    return JSON.stringify(out);
-}
-/** Read back {@link serializeCatalog}. Returns an empty index for anything unusable. */
-export function deserializeCatalog(text) {
-    const index = new Map();
-    let parsed;
-    try {
-        parsed = JSON.parse(text);
-    }
-    catch {
-        return index;
-    }
-    if (!parsed || typeof parsed !== "object")
-        return index;
-    for (const [key, value] of Object.entries(parsed)) {
-        if (!Array.isArray(value) || value.length < 3)
-            continue;
-        const context = Number(value[0]);
-        const output = Number(value[1]);
-        if (!Number.isFinite(context) || context <= 0)
-            continue;
-        index.set(key, {
-            contextWindow: Math.floor(context),
-            maxTokens: Number.isFinite(output) && output > 0 ? Math.floor(output) : undefined,
-            reasoning: value[2] === 1 || value[2] === true,
-            // Absent in a cache written before this field existed: read it as "no
-            // vision", which is what that build assumed, and let the next fetch correct it.
-            vision: value[3] === 1 || value[3] === true,
-        });
-    }
-    return index;
-}
-/**
- * Fetch the catalogue and index it.
- *
- * Never throws: a catalogue that cannot be fetched is a reason to keep the
- * conservative defaults, not a reason for `/connect` to fail. The caller reports
- * which of the two happened.
- */
-export async function fetchCatalog(options) {
-    const url = options?.url ?? MODELS_DEV_URL;
-    const doFetch = options?.fetchImpl ?? fetch;
-    const timeoutMs = options?.timeoutMs ?? CATALOG_TIMEOUT_MS;
-    const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(), timeoutMs);
-    try {
-        const response = await doFetch(url, { signal: controller.signal, headers: { Accept: "application/json" } });
-        if (!response.ok) {
-            return { ok: false, message: `the model catalogue answered ${response.status}` };
-        }
-        const index = buildCatalogIndex(await response.json());
-        if (index.size === 0) {
-            return { ok: false, message: "the model catalogue was empty or in an unexpected shape" };
-        }
-        return { ok: true, index };
-    }
-    catch (cause) {
-        const aborted = cause instanceof Error && cause.name === "AbortError";
-        return {
-            ok: false,
-            message: aborted ? `the model catalogue did not answer within ${Math.round(timeoutMs / 1000)}s` : "could not reach the model catalogue",
-        };
-    }
-    finally {
-        clearTimeout(timer);
-    }
-}

package/dist/model-facts.js

@@ -1,35 +0,0 @@
-/**
- * The two facts a model picker is useless without: how much context it takes and
- * what it costs.
- *
- * Every model entry already carries `contextWindow` and `cost`, but the picker
- * showed only the provider, so a free model and an expensive one looked the
- * same. These helpers turn those numbers into one short line for the list.
- */
-/** 200000 -> "200K", 1000000 -> "1M", 1500000 -> "1.5M". Empty when unknown. */
-export function formatContextWindow(tokens) {
-    if (!tokens || !Number.isFinite(tokens) || tokens <= 0)
-        return "";
-    if (tokens >= 1_000_000) {
-        // Rounded to one decimal, then trimmed: 1_048_576 is "1M", not "1.0M".
-        return `${Number((tokens / 1_000_000).toFixed(1))}M`;
-    }
-    if (tokens >= 1_000)
-        return `${Math.round(tokens / 1_000)}K`;
-    return String(tokens);
-}
-/** Price per million tokens for input and output, or "free" when both are zero. */
-export function formatModelPrice(cost) {
-    if (!cost)
-        return "";
-    if (cost.input <= 0 && cost.output <= 0)
-        return "free";
-    const money = (value) => (value === 0 ? "0" : String(Number(value >= 1 ? value.toFixed(2) : value.toFixed(3))));
-    return `$${money(cost.input)}/$${money(cost.output)} per M`;
-}
-/** "1M ctx · free" — the summary shown beside a model in the picker. */
-export function describeModelFacts(model) {
-    const context = formatContextWindow(model.contextWindow);
-    const price = formatModelPrice(model.cost);
-    return [context ? `${context} ctx` : "", price].filter(Boolean).join(" · ");
-}

package/dist/model-resolve.js

@@ -1,65 +0,0 @@
-/**
- * Resolve a model spec the way the CLI's own `--model` flag does.
- *
- * Exact `provider/id` or a bare `id` first, then a case-insensitive substring
- * match, so both the name `/model` prints and a shorthand work. Lifted out of
- * `session.ts` when subagents gained a model of their own: two resolvers would
- * drift, and a subagent picking a different model than the parent asked for
- * would read as the option being broken.
- */
-export function resolveModel(models, spec) {
-    const exact = models.find((model) => `${model.provider}/${model.id}` === spec || model.id === spec);
-    if (exact)
-        return exact;
-    const needle = spec.toLowerCase();
-    return models.find((model) => model.id.toLowerCase().includes(needle) || `${model.provider}/${model.id}`.toLowerCase().includes(needle));
-}
-/**
- * Order a catalogue for choosing a default model, best first.
- *
- * A gateway lists its catalogue in no useful order — `anthropic/claude-opus-4.8`
- * can come first while the fast general-purpose models are far down the list — so
- * "use whatever is first" hands a new user the most expensive thing on the
- * account, or an image model that cannot hold a conversation.
- *
- * Image, embedding, audio and video models are pushed to the end: they cannot
- * answer a coding prompt, so selecting one would look like a broken install. The
- * caller then probes the first few and takes the first that answers.
- */
-const PREFERRED_FAMILIES = ["deepseek", "qwen", "claude", "gpt-5", "gpt-4", "gemini", "glm", "mimo", "kimi", "llama"];
-const NOT_CHAT = /(image|embedding|embed|whisper|tts|audio|video|rerank|moderation|dall-e|stable-diffusion|sora)/i;
-export function orderModelsForDefault(ids) {
-    const rank = (id) => {
-        if (NOT_CHAT.test(id))
-            return 100;
-        const family = PREFERRED_FAMILIES.findIndex((name) => id.toLowerCase().includes(name));
-        return family === -1 ? 50 : family;
-    };
-    // Stable within a rank, so the gateway's own order decides ties.
-    return ids
-        .map((id, index) => ({ id, index }))
-        .sort((a, b) => rank(a.id) - rank(b.id) || a.index - b.index)
-        .map((entry) => entry.id);
-}
-/**
- * The model a session should start on when nothing was asked for.
- *
- * The model runtime knows every model in the catalog it ships, including
- * providers this account holds no credential for, and its own default is the
- * catalog's — `opencode/kimi-k2.6` — not the account's. Starting there is not a
- * cosmetic problem: the first prompt goes to a provider that cannot answer and
- * comes back as an empty assistant message, with nothing in the transcript to
- * explain why it looks hung.
- *
- * So the default comes from the models that are actually available, and a
- * HowToFix provider wins over any other reachable one: an install that signed in
- * to HowToFix should run HowToFix's models.
- */
-export function defaultModelFor(available, ownProviders = []) {
-    if (available.length === 0)
-        return undefined;
-    const own = available.filter((model) => ownProviders.includes(model.provider));
-    const pool = own.length > 0 ? own : available;
-    const first = orderModelsForDefault(pool.map((model) => model.id))[0];
-    return pool.find((model) => model.id === first) ?? pool[0];
-}

package/dist/notes.js

@@ -1,78 +0,0 @@
-/**
- * A shared scratchpad for facts that must survive context compaction and reach
- * every agent working on the task.
- *
- * A shared notes store for the session.
- * Why it matters: the `task` tool spawns subagents with their own context
- * window, so anything the main agent learned while scouting is invisible to them,
- * and anything they discover dies with their run. A note written once is visible
- * to the parent and to every sibling, and it survives a compaction that would
- * otherwise drop the discovery from the transcript.
- *
- * The store is process-wide on purpose: `agents.ts` opens subagent sessions in
- * this same process, so they share it. It is in-memory only — notes are working
- * state for one session, not project documentation (that is `AGENTS.md`).
- */
-import { StringEnum } from "@netrodev/hxf-ai";
-import { defineTool } from "@netrodev/hxf-coding-agent";
-import { Type } from "typebox";
-const notes = new Map();
-/** Longest note accepted, so one call cannot flood every later request. */
-const MAX_VALUE = 2000;
-/** Longest key accepted; keys are the lookup handle, so they stay short. */
-const MAX_KEY = 80;
-export function clearNotes() {
-    notes.clear();
-}
-export function noteCount() {
-    return notes.size;
-}
-/** The block injected into the prompt, or an empty string when there is nothing. */
-export function renderNotes() {
-    if (notes.size === 0)
-        return "";
-    const lines = [...notes].map(([key, value]) => `- ${key}: ${value}`);
-    return ["<shared-notes>", "Working notes recorded with the `context` tool. They survive compaction and are visible to every agent on this task.", ...lines, "</shared-notes>"].join("\n");
-}
-export function applyNote(op, key, value) {
-    if (op === "list") {
-        if (notes.size === 0)
-            return { text: "No notes." };
-        return { text: [...notes].map(([k, v]) => `${k} = ${v}`).join("\n") };
-    }
-    if (!key)
-        return { text: "Error: key is required", error: "key required" };
-    if (key.length > MAX_KEY)
-        return { text: `Error: key longer than ${MAX_KEY} characters`, error: "key too long" };
-    if (op === "forget") {
-        const existed = notes.delete(key);
-        return { text: existed ? `Removed ${key}` : `No note named ${key}` };
-    }
-    if (op === "set") {
-        if (!value)
-            return { text: "Error: value is required for set", error: "value required" };
-        const clipped = value.length > MAX_VALUE ? `${value.slice(0, MAX_VALUE)}…` : value;
-        notes.set(key, clipped);
-        return { text: `Recorded ${key} (${notes.size} note${notes.size === 1 ? "" : "s"} held)` };
-    }
-    return { text: `Error: unknown op ${JSON.stringify(op)}`, error: "unknown op" };
-}
-export const contextTool = defineTool({
-    name: "context",
-    label: "context",
-    description: "Record a durable working note (a decision, a file's role, a fact you would otherwise re-derive). Notes survive context compaction and are visible to every agent on this task, including subagents. Use `set` to write, `list` to read, `forget` to remove.",
-    promptSnippet: "context: durable notes shared with every agent on this task",
-    parameters: Type.Object({
-        op: StringEnum(["set", "list", "forget"], { description: "set writes a note, list reads all, forget removes one" }),
-        key: Type.Optional(Type.String({ description: "Short lookup key, e.g. 'auth-entrypoint'" })),
-        value: Type.Optional(Type.String({ description: "The fact to keep (for set)" })),
-    }),
-    async execute(_toolCallId, params) {
-        const { text, error } = applyNote(params.op, params.key, params.value);
-        return {
-            content: [{ type: "text", text }],
-            details: { op: params.op, notes: Object.fromEntries(notes) },
-            ...(error ? { isError: true } : {}),
-        };
-    },
-});