aiden-runtime 4.1.1 → 4.1.2

package/dist/cli/v4/promotionPrompt.js

@@ -0,0 +1,202 @@
+"use strict";
+/**
+ * Copyright (c) 2026 Shiva Deore (Taracod).
+ * Licensed under AGPL-3.0. See LICENSE for details.
+ *
+ * Aiden — local-first agent.
+ */
+/**
+ * cli/v4/promotionPrompt.ts — Phase v4.1.2-memory-D.
+ *
+ * REPL-side glue for the durable-facts promotion flow:
+ *   - `parsePromotionInput(raw, count)`  — pure: parse user reply into
+ *                                          a 0-indexed array of approved
+ *                                          candidate indices.
+ *   - `formatCandidateList(candidates)`  — pure: render the prompt body
+ *                                          the user sees.
+ *   - `promptForApproval(api, ...)`      — drives the prompt loop;
+ *                                          re-prompts ONCE on garbage,
+ *                                          then defaults to skip.
+ *   - `writeApprovedDurableFacts(...)`   — append approved candidates
+ *                                          to MEMORY.md `## Durable facts`
+ *                                          via MemoryGuard.replaceSection.
+ *
+ * Input grammar (per Phase D's Q3):
+ *   - "all"                  → every shown candidate
+ *   - "none" / "skip" / ""   → none
+ *   - "1,3"                  → 0-indexed 0 and 2
+ *   - "1-3"                  → 0-indexed 0, 1, 2 (inclusive range)
+ *   - "1, 3-5"               → mixed; whitespace tolerated
+ *   - Anything unparseable   → re-prompt once, then default skip
+ *
+ * The function intentionally keeps the parser pure so unit tests
+ * don't have to drive a prompt API. The prompt-loop function wires
+ * the parser to the existing `ChatPromptApi.readLine`.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.parsePromotionInput = parsePromotionInput;
+exports.formatCandidateList = formatCandidateList;
+exports.promptForApproval = promptForApproval;
+exports.buildDurableFactsBody = buildDurableFactsBody;
+exports.readExistingDurableFactsBody = readExistingDurableFactsBody;
+exports.writeApprovedDurableFacts = writeApprovedDurableFacts;
+// ── Parser ────────────────────────────────────────────────────────────────
+/**
+ * Parse a user reply into the set of approved candidate indices
+ * (0-indexed). Returns `null` to signal "unparseable input — re-prompt
+ * once" so callers can distinguish "explicit skip" (empty array) from
+ * "garbage typed".
+ *
+ * Pure, deterministic; safe for unit tests.
+ */
+function parsePromotionInput(raw, count) {
+    const trimmed = raw.trim().toLowerCase();
+    if (trimmed === '' || trimmed === 'none' || trimmed === 'skip')
+        return [];
+    if (trimmed === 'all') {
+        return Array.from({ length: count }, (_, i) => i);
+    }
+    const out = new Set();
+    let sawAnyValid = false;
+    // Tolerate "1, 3-5 ,7"  with mixed whitespace.
+    for (const token of trimmed.split(',')) {
+        const piece = token.trim();
+        if (!piece)
+            continue;
+        const range = piece.match(/^(\d+)\s*-\s*(\d+)$/);
+        if (range) {
+            const start = Number.parseInt(range[1], 10);
+            const end = Number.parseInt(range[2], 10);
+            if (!Number.isFinite(start) || !Number.isFinite(end))
+                continue;
+            const [lo, hi] = start <= end ? [start, end] : [end, start];
+            for (let n = lo; n <= hi; n += 1) {
+                if (n >= 1 && n <= count) {
+                    out.add(n - 1);
+                    sawAnyValid = true;
+                }
+            }
+            continue;
+        }
+        const single = piece.match(/^\d+$/);
+        if (single) {
+            const n = Number.parseInt(piece, 10);
+            if (n >= 1 && n <= count) {
+                out.add(n - 1);
+                sawAnyValid = true;
+            }
+            continue;
+        }
+        // Non-numeric token alongside others — treat the WHOLE input as
+        // unparseable so the user gets one re-prompt instead of a silent
+        // partial selection.
+        return null;
+    }
+    if (!sawAnyValid)
+        return []; // numbers given but all out of range
+    return [...out].sort((a, b) => a - b);
+}
+// ── Renderer ──────────────────────────────────────────────────────────────
+/**
+ * Build the text the user sees. Pure — caller writes this to display.
+ */
+function formatCandidateList(candidates) {
+    const lines = [];
+    lines.push(`${candidates.length} thing${candidates.length === 1 ? '' : 's'} worth remembering this session. Promote which?`);
+    lines.push('');
+    for (let i = 0; i < candidates.length; i += 1) {
+        const c = candidates[i];
+        const sourceTag = c.source === 'explicit' ? '[user said]'
+            : c.source === 'decision' ? '[decision]'
+                : '[open item]';
+        lines.push(`  [${i + 1}] ${sourceTag} ${c.text}`);
+    }
+    lines.push('');
+    lines.push('Reply: numbers to approve (e.g. "1,3" or "1-3"), "all", or skip.');
+    return lines.join('\n');
+}
+/**
+ * Drive the approval prompt. Renders the candidate list, reads ONE
+ * line, parses, returns approved Candidate[]. On unparseable input
+ * re-prompts ONCE; second failure defaults to skip with a dim line
+ * explaining why nothing was promoted.
+ *
+ * No mid-session state leakage — purely a session-end interaction.
+ */
+async function promptForApproval(api, display, candidates) {
+    if (candidates.length === 0)
+        return [];
+    display.write('\n' + formatCandidateList(candidates) + '\n');
+    for (let attempt = 0; attempt < 2; attempt += 1) {
+        const raw = await api.readLine('Promote > ');
+        const parsed = parsePromotionInput(raw, candidates.length);
+        if (parsed !== null) {
+            if (parsed.length === 0) {
+                display.dim('Nothing promoted to durable facts.');
+                return [];
+            }
+            return parsed.map((i) => candidates[i]);
+        }
+        if (attempt === 0) {
+            display.warn('Could not parse input. Use numbers ("1,3"), ranges ("1-3"), "all", or "skip".');
+        }
+    }
+    display.dim('Skipped: input still unparseable. Nothing promoted to durable facts.');
+    return [];
+}
+// ── Persistence ───────────────────────────────────────────────────────────
+const DURABLE_FACTS_HEADER = '## Durable facts';
+/**
+ * Render the section body for `## Durable facts` by combining existing
+ * entries with newly-approved candidates. Newest at the BOTTOM so
+ * read order reflects when each fact landed — matches how users scan
+ * MEMORY.md.
+ *
+ * Pure — caller passes existing body (extracted via the same regex
+ * pattern MemoryGuard uses in replaceSection).
+ */
+function buildDurableFactsBody(existingBody, approved) {
+    const existingLines = existingBody
+        .split('\n')
+        .map((l) => l.trim())
+        .filter((l) => l.length > 0);
+    const newLines = approved.map((c) => `- ${c.text}`);
+    return [...existingLines, ...newLines].join('\n');
+}
+/**
+ * Read the current `## Durable facts` body from MEMORY.md (returns
+ * empty string when the section doesn't yet exist). Mirrors the
+ * regex pattern MemoryGuard.replaceSection uses.
+ */
+async function readExistingDurableFactsBody(memoryManager) {
+    const snap = await memoryManager.loadSnapshot();
+    const md = snap.memoryMd ?? '';
+    const headerEscaped = DURABLE_FACTS_HEADER.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+    const sectionRe = new RegExp(`${headerEscaped}[^\\n]*\\n([\\s\\S]*?)(?=\\n## |$)`);
+    const m = md.match(sectionRe);
+    return m ? (m[1] ?? '').trim() : '';
+}
+/**
+ * Persist the approved candidates. Reads existing body (so a second
+ * session-end appends rather than overwrites), folds in new lines,
+ * and writes via MemoryGuard.replaceSection — which handles
+ * verify-on-disk + section auto-creation.
+ *
+ * Returns the GuardedResult so the caller can dim-log success or
+ * warn on a failed verify.
+ */
+async function writeApprovedDurableFacts(memoryManager, memoryGuard, approved) {
+    if (approved.length === 0) {
+        return { ok: true, verified: true, entryCount: 0 };
+    }
+    const existingBody = await readExistingDurableFactsBody(memoryManager);
+    const newBody = buildDurableFactsBody(existingBody, approved);
+    const entryCount = newBody.split('\n').filter((l) => l.trim().length > 0).length;
+    const result = await memoryGuard.replaceSection('memory', DURABLE_FACTS_HEADER, newBody);
+    return {
+        ok: result.ok,
+        verified: result.verified,
+        reason: result.reason,
+        entryCount,
+    };
+}

package/dist/cli/v4/providerBootSelector.js

@@ -0,0 +1,144 @@
+"use strict";
+/**
+ * Copyright (c) 2026 Shiva Deore (Taracod).
+ * Licensed under AGPL-3.0. See LICENSE for details.
+ *
+ * Aiden — local-first agent.
+ */
+/**
+ * cli/v4/providerBootSelector.ts — Phase v4.1.2-bug1.
+ *
+ * Boot-time provider/model picker. Replaces the hardcoded
+ * `groq + llama-3.3-70b-versatile` fallback that bit new users:
+ * users authenticated with ChatGPT Plus OAuth (the post-v4.1.1
+ * onboarding default) booted into Groq anyway, and llama-3.3-70b's
+ * tool emission triggers Groq's first-party 400
+ * ("Failed to call a function. Please adjust your prompt.").
+ *
+ * Resolution precedence (caller in `aidenCLI.ts` enforces order):
+ *   1. Both CLI flags `--provider` + `--model` set      → use as-is
+ *   2. One CLI flag set                                  → use it, resolve other
+ *   3. Persisted config (model-selection.json) complete  → use as-is
+ *   4. Persisted config partial                          → use it, resolve other
+ *   5. Neither → priority-list auto-pick                 → THIS MODULE
+ *   6. Nothing authed                                    → hardcoded fallback
+ *
+ * `resolveBootProvider()` covers cases 1–5; the caller composes the
+ * input shape and falls back when this returns `null` (case 6).
+ *
+ * Test surface: the enumerator is injected so unit tests mock it.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.BOOT_PRIORITY = void 0;
+exports.findProviderForModel = findProviderForModel;
+exports.resolveBootProvider = resolveBootProvider;
+const doctorLiveness_1 = require("./doctorLiveness");
+const registry_1 = require("../../providers/v4/registry");
+/**
+ * Provider id ordering for auto-pick. Higher in the list = preferred.
+ * OAuth subscription flows lead because they have no API-key
+ * onboarding friction and are the default v4.1.1 install path.
+ */
+exports.BOOT_PRIORITY = [
+    'chatgpt-plus', // OAuth — primary onboarding flow
+    'claude-pro', // OAuth — Anthropic equivalent
+    'anthropic', // API key — power-user tier
+    'openai', // API key — power-user tier
+    // Phase v4.1.2-deepseek: paid tier, strong tool-caller, ranked
+    // above groq for the same first-run-UX reason — groq's free-tier
+    // tool emission was the original bug1 (llama-3.3-70b 400s on
+    // tool calls).
+    'deepseek', // API key — paid, strong tool-caller (V4 Pro)
+    'groq', // free-tier API key — common but tool-emission flaky
+    'ollama', // local — only if daemon up
+];
+/**
+ * Walk every provider's `modelIds` and return the entry that lists
+ * `modelId`. Used by the `--model`-only path to validate the model
+ * is at least known to one provider before we accept it.
+ */
+function findProviderForModel(modelId) {
+    for (const entry of Object.values(registry_1.PROVIDER_REGISTRY)) {
+        if (entry.modelIds.includes(modelId))
+            return entry;
+    }
+    return null;
+}
+/**
+ * Resolve the boot provider + model. Returns `null` when no choice
+ * could be inferred AND nothing is authed (the caller falls back to
+ * the hardcoded `groq + llama-3.3-70b-versatile` default in that
+ * one case).
+ *
+ * Throws (`Error`) when input is internally inconsistent — e.g. the
+ * user passed `--model foo` for a model no provider declares. Caller
+ * surfaces the message via the standard error path.
+ */
+async function resolveBootProvider(input, enumerate) {
+    const { cliProviderId, cliModelId, cfgProviderId, cfgModelId } = input;
+    // Case 1: both CLI flags set.
+    if (cliProviderId && cliModelId) {
+        return { providerId: cliProviderId, modelId: cliModelId, source: 'cli-flag' };
+    }
+    // Case 2a: `--provider` only. Use that provider + its first
+    // non-codex model. The registry might not know this provider; we
+    // accept it as-is (the runtime resolver will throw a clearer error
+    // later if it's bogus).
+    if (cliProviderId && !cliModelId) {
+        const entry = registry_1.PROVIDER_REGISTRY[cliProviderId];
+        const modelId = entry ? (0, doctorLiveness_1.pickProbeModel)(entry) : '';
+        if (!modelId) {
+            throw new Error(`--provider '${cliProviderId}' set but no model could be inferred. ` +
+                `Pass --model explicitly.`);
+        }
+        return { providerId: cliProviderId, modelId, source: 'cli-flag-partial' };
+    }
+    // Case 2b: `--model` only. Verify the model is known to some
+    // provider; pick the matching provider.
+    if (!cliProviderId && cliModelId) {
+        const entry = findProviderForModel(cliModelId);
+        if (!entry) {
+            throw new Error(`--model '${cliModelId}' is not declared by any provider in the ` +
+                `registry. Run \`aiden model\` to see available models, or pass ` +
+                `--provider explicitly.`);
+        }
+        return { providerId: entry.id, modelId: cliModelId, source: 'cli-flag-partial' };
+    }
+    // Case 3: persisted config — both fields set.
+    if (cfgProviderId && cfgModelId) {
+        return { providerId: cfgProviderId, modelId: cfgModelId, source: 'persisted-config' };
+    }
+    // Case 4a: config has provider, no model — resolve via pickProbeModel.
+    if (cfgProviderId && !cfgModelId) {
+        const entry = registry_1.PROVIDER_REGISTRY[cfgProviderId];
+        const modelId = entry ? (0, doctorLiveness_1.pickProbeModel)(entry) : '';
+        if (modelId) {
+            return { providerId: cfgProviderId, modelId, source: 'config-partial' };
+        }
+        // Unknown provider in config — fall through to auto-pick rather
+        // than refusing to boot.
+    }
+    // Case 4b: config has model only — same shape as Case 2b but
+    // softer: fall through to auto-pick if the model isn't found in
+    // the registry (config can lag the codebase).
+    if (!cfgProviderId && cfgModelId) {
+        const entry = findProviderForModel(cfgModelId);
+        if (entry) {
+            return { providerId: entry.id, modelId: cfgModelId, source: 'config-partial' };
+        }
+    }
+    // Case 5: auto-pick from priority list.
+    const configured = await enumerate();
+    for (const id of exports.BOOT_PRIORITY) {
+        const hit = configured.find((c) => c.entry.id === id && c.configured);
+        if (hit) {
+            return {
+                providerId: id,
+                modelId: (0, doctorLiveness_1.pickProbeModel)(hit.entry),
+                source: 'auto-priority',
+            };
+        }
+    }
+    // Case 6: nothing authed — caller falls back to hardcoded default.
+    return null;
+}

package/dist/cli/v4/sessionSummaryGate.js

@@ -0,0 +1,66 @@
+"use strict";
+/**
+ * Copyright (c) 2026 Shiva Deore (Taracod).
+ * Licensed under AGPL-3.0. See LICENSE for details.
+ *
+ * Aiden — local-first agent.
+ */
+/**
+ * cli/v4/sessionSummaryGate.ts — Phase v4.1.2-followup-2.
+ *
+ * Pure decision helpers extracted from `ChatSession.maybeAutoSummarize`
+ * so the threshold + mtime/size-grew logic is unit-testable without
+ * standing up a full ChatSession + mocked agent loop.
+ *
+ *   shouldAutoSummarize → returns {fire: true} or
+ *                         {fire: false, reason: 'short'|'unconfigured'|'no-paths'}.
+ *                         ChatSession uses the reason tag to log the right
+ *                         user-visible message.
+ *
+ *   memoryGrewBetween → strict size-or-mtime comparison so the caller can
+ *                       detect "the model actually fired session_summary"
+ *                       even when the tool wrote without growing the file
+ *                       length (e.g. replaced a previous same-length entry).
+ *
+ * No I/O here. ChatSession owns the fs.stat + display.warn / display.dim.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.SESSION_SUMMARY_MIN_TURNS = void 0;
+exports.shouldAutoSummarize = shouldAutoSummarize;
+exports.memoryGrewBetween = memoryGrewBetween;
+/** Minimum user-message turns required before auto-summary triggers. */
+exports.SESSION_SUMMARY_MIN_TURNS = 3;
+/**
+ * Decide whether the /quit auto-summary should fire. Threshold lives
+ * here as the single source of truth; ChatSession imports the constant
+ * so the user-facing log message ("need 3+") cites the exact value.
+ */
+function shouldAutoSummarize(input) {
+    if (input.userTurns < exports.SESSION_SUMMARY_MIN_TURNS) {
+        return { fire: false, reason: 'short' };
+    }
+    if (input.unconfigured) {
+        return { fire: false, reason: 'unconfigured' };
+    }
+    if (!input.memoryPath) {
+        return { fire: false, reason: 'no-paths' };
+    }
+    return { fire: true };
+}
+/**
+ * True iff MEMORY.md grew (longer) or was touched (newer mtime) between
+ * the two snapshots. Used to detect whether the agent actually fired
+ * the session_summary tool inside the synthetic turn — if not, the
+ * user sees a warning instead of a misleading "saved" message.
+ *
+ * The size-OR-mtime disjunction (not just size>before) covers the case
+ * where session_summary replaces an existing same-length entry: file
+ * size stays the same but mtime advances.
+ */
+function memoryGrewBetween(before, after) {
+    if (after.size > before.size)
+        return true;
+    if (after.mtime > before.mtime)
+        return true;
+    return false;
+}

package/dist/cli/v4/toolPreview.js

@@ -0,0 +1,139 @@
+"use strict";
+/**
+ * Copyright (c) 2026 Shiva Deore (Taracod).
+ * Licensed under AGPL-3.0. See LICENSE for details.
+ *
+ * Aiden — local-first agent.
+ */
+/**
+ * cli/v4/toolPreview.ts — Phase v4.1.2 alive-core.
+ *
+ * Clean per-tool argument previews. Replaces the old
+ * `JSON.stringify(args)` blob in `display.toolPreview` with a
+ * tool-aware lookup that extracts the primary argument (the one
+ * actually useful at a glance — `command` for terminal, `path` for
+ * file ops, `query` for search, etc.).
+ *
+ * Falls back to the original full-JSON stringification when the tool
+ * isn't in the map or the primary arg is absent. This keeps unknown
+ * tools rendering exactly as before — additive only.
+ *
+ * Adding a new tool with a non-obvious primary arg? Add it here.
+ * Tools whose `args` shape is "the arg is meaningful at-a-glance"
+ * (a path, a query, a command, a URL, an id, a name) belong in this map.
+ * Tools whose args are a small flag bag (e.g. system_info has no args
+ * worth showing) can be omitted — the renderer hides the args block
+ * entirely when the map returns `null` and the arg object is empty.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.TOOL_PRIMARY_ARG = void 0;
+exports.buildToolPreview = buildToolPreview;
+/**
+ * Map of tool-name → name of the property in `args` that should render
+ * as the at-a-glance preview. Stable contract; tests assert specific
+ * entries.
+ */
+exports.TOOL_PRIMARY_ARG = {
+    // ── terminal / execution ─────────────────────────────────────────────
+    shell_exec: 'command',
+    execute_code: 'code',
+    // ── file ops ─────────────────────────────────────────────────────────
+    file_read: 'path',
+    file_write: 'path',
+    file_patch: 'path',
+    file_list: 'path',
+    file_copy: 'source',
+    file_move: 'source',
+    file_delete: 'path',
+    // ── web ──────────────────────────────────────────────────────────────
+    web_search: 'query',
+    deep_research: 'query',
+    youtube_search: 'query',
+    fetch_url: 'url',
+    fetch_page: 'url',
+    open_url: 'url',
+    // ── browser ──────────────────────────────────────────────────────────
+    browser_navigate: 'url',
+    browser_click: 'selector',
+    browser_fill: 'selector',
+    browser_type: 'selector',
+    browser_scroll: 'selector',
+    browser_extract: 'selector',
+    browser_get_url: '', // no args — present so map lookup hits
+    browser_screenshot: 'path',
+    browser_close: '',
+    // ── memory ───────────────────────────────────────────────────────────
+    memory_add: 'content',
+    memory_remove: 'content',
+    memory_replace: 'old',
+    // ── skills ───────────────────────────────────────────────────────────
+    skill_view: 'name',
+    skill_manage: 'action',
+    skills_list: '',
+    // ── sessions ─────────────────────────────────────────────────────────
+    session_search: 'query',
+    session_list: '',
+    session_summary: 'trigger',
+    // ── process ──────────────────────────────────────────────────────────
+    process_spawn: 'command',
+    process_kill: 'pid',
+    process_list: '',
+    process_wait: 'pid',
+    process_log_read: 'pid',
+    // ── subagent ─────────────────────────────────────────────────────────
+    subagent_fanout: 'mode',
+    // ── system / misc ────────────────────────────────────────────────────
+    system_info: '',
+    now_playing: '',
+    get_natural_events: '',
+};
+/**
+ * Maximum visible characters for the preview value. Long commands /
+ * full file contents get truncated with an ellipsis so a single tool
+ * row stays on one line at typical terminal widths.
+ */
+const PREVIEW_MAX_CHARS = 120;
+/**
+ * Build the per-tool preview string for `args`. Returns:
+ *   - `null` when the tool isn't in the map (caller falls back to the
+ *     legacy JSON.stringify path),
+ *   - `''` (empty string) when the tool is in the map but has no
+ *     meaningful primary arg (caller renders just the tool name),
+ *   - the truncated string value of the primary arg otherwise.
+ *
+ * Exposed for unit tests. Pure function, no side effects.
+ */
+function buildToolPreview(toolName, args) {
+    if (!Object.prototype.hasOwnProperty.call(exports.TOOL_PRIMARY_ARG, toolName)) {
+        return null;
+    }
+    const argKey = exports.TOOL_PRIMARY_ARG[toolName];
+    if (argKey === '')
+        return '';
+    if (!args || typeof args !== 'object')
+        return '';
+    const raw = args[argKey];
+    if (raw === undefined || raw === null)
+        return '';
+    let str;
+    if (typeof raw === 'string') {
+        str = raw;
+    }
+    else if (typeof raw === 'number' || typeof raw === 'boolean') {
+        str = String(raw);
+    }
+    else {
+        try {
+            str = JSON.stringify(raw);
+        }
+        catch {
+            str = String(raw);
+        }
+    }
+    // Collapse whitespace so multi-line commands stay on one preview row.
+    str = str.replace(/\s+/g, ' ').trim();
+    if (str.length > PREVIEW_MAX_CHARS) {
+        str = `${str.slice(0, PREVIEW_MAX_CHARS - 1)}…`;
+    }
+    return str;
+}