dsh-continual-evolve 0.11.0 → 0.12.0

package/lib/validate.js

@@ -1,181 +0,0 @@
-import { MEMORY_TYPE_KEY, isMemoryType } from "./types.js";
-import { validateSkillEntryContent } from "./skillquality.js";
-const ACTIONS = new Set(["create", "update", "delete", "archive"]);
-const KINDS = new Set(["prompt", "memory", "skill", "subagent"]);
-export const BASE_SYSTEM_PROMPT_ID = "base_system_prompt";
-/**
- * Gap C2: mechanical check that an edit's declared blast radius is coherent
- * with the scope it targets. A session-scoped edit claiming "general" would
- * silently read like a cross-project tactical rule; a global edit claiming
- * "session" would contradict its persistence. Absent blastRadius is NOT
- * rejected (pre-C2 data and manual edits stay compatible) — the planner is
- * instructed to always declare it, and this rule catches what it declares
- * incoherently.
- */
-export function validateBlastRadiusScope(scope, blastRadius) {
-    if (scope === "local" && blastRadius === "general") {
-        return "local-scope edit must declare blastRadius \"session\" or \"project\" (\"general\" would claim a cross-project rule)";
-    }
-    if (scope === "global" && blastRadius === "session") {
-        return "global-scope edit must declare blastRadius \"general\" or \"project\" (\"session\" contradicts cross-session persistence)";
-    }
-    if (scope === "project" && blastRadius === "session") {
-        return "project-scope edit must declare blastRadius \"project\" or \"general\" (\"session\" contradicts cross-session persistence)";
-    }
-    return undefined;
-}
-/**
- * Returns a human-readable failure reason, or undefined when the edit passes.
- *
- * `before` (the entry the edit targets, absent for creates/unknown ids) lets
- * update rules distinguish "carrying the persisted value" from "changing it"
- * — e.g. a rollback inverse re-carries the stored skill_kind, which must
- * pass, while a genuine executable↔guidance switch must not.
- */
-export function validateEdit(edit, computedId, scope, before) {
-    if (!ACTIONS.has(edit.action)) {
-        return `unsupported action ${String(edit.action)}`;
-    }
-    if (!KINDS.has(edit.kind)) {
-        return `unsupported kind ${String(edit.kind)}`;
-    }
-    if (edit.kind === "prompt" && (edit.id === BASE_SYSTEM_PROMPT_ID || computedId === BASE_SYSTEM_PROMPT_ID)) {
-        return "base system prompt is not editable";
-    }
-    if (edit.action !== "create" && !edit.id) {
-        return `${edit.action} requires id`;
-    }
-    // Gap C2: blast-radius/scope coherence is a mechanical property of the
-    // edit payload itself — checked for every action, not only create/update.
-    if (scope && edit.blastRadius !== undefined) {
-        const blastError = validateBlastRadiusScope(scope, edit.blastRadius);
-        if (blastError)
-            return blastError;
-    }
-    // Archive only names an existing entry: no title/content payload, and the
-    // base system prompt stays immutable under every action.
-    if (edit.action === "archive") {
-        return undefined;
-    }
-    // Create needs the full payload (nothing to fall back to). Update may
-    // carry any subset — apply merges with `?? before` — but must carry at
-    // least one change (review audit 2026-08-28 B3/B5: the old blanket
-    // "update requires title and content" broke every partial-update path:
-    // skill archive/demote, consolidate, and the evolve_update tool contract
-    // all construct payload-subset updates).
-    if (edit.action === "create" && (!edit.title || !edit.content)) {
-        return "create requires title and content";
-    }
-    if (edit.action === "update" &&
-        edit.title === undefined &&
-        edit.content === undefined &&
-        edit.path === undefined &&
-        edit.metadata === undefined &&
-        edit.reference === undefined &&
-        edit.arguments === undefined) {
-        return "update carries no changes";
-    }
-    if (edit.kind === "memory" && (edit.action === "create" || edit.action === "update")) {
-        // One-fact typed memories: every memory create declares its recall
-        // type; feedback/project (pitfalls, conventions) additionally carry
-        // the Why + How-to-apply structure so the fact is reusable, not just
-        // recorded. Updates are partial — only carried fields are checked,
-        // resolved against the persisted entry.
-        const carriedType = edit.metadata?.[MEMORY_TYPE_KEY];
-        if (edit.action === "create" && !isMemoryType(carriedType)) {
-            return `create memory requires metadata.${MEMORY_TYPE_KEY} one of user|feedback|project|reference (one fact per entry; update the existing entry instead of adding a duplicate)`;
-        }
-        if (carriedType !== undefined && !isMemoryType(carriedType)) {
-            return `invalid metadata.${MEMORY_TYPE_KEY} ${JSON.stringify(carriedType)} (one of user|feedback|project|reference)`;
-        }
-        const effectiveType = isMemoryType(carriedType)
-            ? carriedType
-            : edit.action === "update"
-                ? before?.metadata?.[MEMORY_TYPE_KEY]
-                : undefined;
-        if ((effectiveType === "feedback" || effectiveType === "project") &&
-            (edit.action === "create" || edit.content !== undefined)) {
-            const content = edit.content ?? "";
-            if (!/why/i.test(content) || !/how/i.test(content)) {
-                return `${effectiveType} memory content must carry a Why and a How to apply section (the fact plus why it matters and how to apply it next time)`;
-            }
-        }
-    }
-    if (edit.action === "create" && edit.kind === "skill") {
-        // Guidance skills are SKILL.md documents: no python reference (a
-        // reference on a guidance skill would be an invented contract) and
-        // no arguments contract. Executable skills keep the full contract.
-        if (edit.skill_kind === "guidance") {
-            if (edit.reference !== undefined && Object.keys(edit.reference).length > 0) {
-                return "guidance skill must not carry a python reference (it is a SKILL.md document, not an executable)";
-            }
-            if (edit.arguments !== undefined && Object.keys(edit.arguments).length > 0) {
-                return "guidance skill must not carry an arguments contract (only executable skills declare inputs)";
-            }
-        }
-        else {
-            const contractError = validateSkillContract(edit);
-            if (contractError)
-                return contractError;
-        }
-        // The entry body materializes as a SKILL.md under generated
-        // frontmatter; content-level mechanics (no shadowing `---`, no
-        // escaping resource refs) are code-enforced so a bad body never
-        // reaches the store (mirrors skill-creator's validate-frontmatter).
-        const contentProblems = validateSkillEntryContent(edit.content ?? "");
-        if (contentProblems.length > 0) {
-            return contentProblems.join("; ");
-        }
-    }
-    if (edit.action === "update" && edit.kind === "skill") {
-        // skill_kind is immutable on update: a switch without the matching
-        // contract pair would leave an invalid entry (guidance carrying a
-        // python reference, or an executable with none). Re-carrying the
-        // persisted value (rollback inverses do) passes. Recreate the entry
-        // to change the kind.
-        if (edit.skill_kind !== undefined && before !== undefined && before.skill_kind !== edit.skill_kind) {
-            return "update cannot change skill_kind — delete and recreate the entry instead";
-        }
-        // A carried non-empty reference replaces the contract wholesale, so it
-        // must be a complete executable contract on its own. Empty references
-        // (guidance skills) and absent payloads pass through — apply keeps
-        // the persisted values.
-        if (edit.reference !== undefined && Object.keys(edit.reference).length > 0) {
-            const contractError = validateSkillContract(edit);
-            if (contractError)
-                return contractError;
-        }
-        if (edit.content !== undefined) {
-            const contentProblems = validateSkillEntryContent(edit.content);
-            if (contentProblems.length > 0) {
-                return contentProblems.join("; ");
-            }
-        }
-    }
-    return undefined;
-}
-function validateSkillContract(edit) {
-    if (edit.arguments === undefined) {
-        return "create/update skill requires arguments";
-    }
-    const reference = edit.reference;
-    if (!reference || typeof reference !== "object") {
-        return "create/update skill requires python reference";
-    }
-    const ref = reference;
-    if (ref.type !== "python") {
-        return "create/update skill reference.type must be python";
-    }
-    const hasImport = (typeof ref.import === "string" && ref.import.length > 0) ||
-        (typeof ref.python_import === "string" && ref.python_import.length > 0);
-    const hasCallable = (typeof ref.callable === "string" && ref.callable.length > 0) ||
-        (typeof ref.call_pattern === "string" && ref.call_pattern.length > 0);
-    if (!hasImport) {
-        return "create/update skill requires python import";
-    }
-    if (!hasCallable) {
-        return "create/update skill requires callable or call_pattern";
-    }
-    return undefined;
-}
-//# sourceMappingURL=validate.js.map

package/lib/wrapup-candidates.d.ts

@@ -1,127 +0,0 @@
-/**
- * Wrap-up candidate domain: the mechanical audit of a session's local
- * entries (candidates, global coverage detection, usage hints), the LLM
- * assessment parser, and the shared wrap-up types. Pure logic — the LLM
- * call lives in `wrapup.ts`, the proposal builders in
- * `wrapup-proposals.ts`.
- */
-import type { HarnessEntry, HarnessState, RefinementKind } from "./types.js";
-export type WrapupVerdict = "promote" | "archive" | "keep";
-/** A classified local entry: `key` matches one audited candidate exactly. */
-export interface WrapupItem {
-    /** `kind:id` of the candidate this verdict refers to. */
-    key: string;
-    verdict: WrapupVerdict;
-    reason: string;
-    /**
-     * Optional split-promotion payload (verdict "archive" only): the entry is
-     * archived as a whole, but a CLEANED cross-session-reusable part is
-     * offered for promotion — the durable fact distilled out of the mixed
-     * entry, with the ephemeral snapshot left behind in the archive.
-     */
-    promote?: {
-        title: string;
-        content: string;
-    };
-    /**
-     * Valence signal (P1 效价反馈): the assessor found trajectory or global
-     * evidence contradicting this entry's content (user correction, overridden
-     * fact). Strictly parsed — only an explicit JSON `true` sets it. Keep
-     * verdicts with this flag get a negative-valence stamp; the assessor is
-     * instructed to prefer "archive" for contradicted entries.
-     */
-    contradicted?: boolean;
-}
-/** A real global entry worth showing the assessor for the same topic. */
-export interface GlobalHint {
-    id: string;
-    title: string;
-}
-/** The model's full classification of a session's local entries. */
-export interface WrapupAssessment {
-    items: WrapupItem[];
-    rationale: string;
-}
-/** A local entry offered for assessment, plus its deterministic audit flags. */
-export interface WrapupCandidate {
-    kind: RefinementKind;
-    id: string;
-    title: string;
-    content: string;
-    path: string;
-    version: number;
-    metadata: Record<string, unknown>;
-    /**
-     * True when the global store already covers this topic by a STRONG
-     * signal: a title that normalizes equal to, or (beyond a length floor)
-     * contains, the candidate's title. Collisions on id alone with a wildly
-     * different title are deliberately NOT coverage — see {@link globalHintsFor}.
-     */
-    coveredGlobally: boolean;
-    /**
-     * Actual global entries that touch the same topic (same id, equal
-     * normalized title, or title overlap). Shown to the assessor so it judges
-     * against real titles instead of a bare boolean; a bare same-id collision
-     * shows up here precisely so the model can tell whether the global copy
-     * really covers the local content.
-     */
-    globalHints: GlobalHint[];
-    /**
-     * Injection usage count (gap B1): how many times this entry was included
-     * in a system-prompt assembly. Zero means the entry was never used — a
-     * strong staleness signal the assessor can weigh.
-     */
-    injectionCount: number;
-    /**
-     * Staleness flag (gap B2): true when the entry has both zero injection
-     * usage AND a recency score below the staleness threshold (old + unused).
-     * The assessor is instructed to prefer "archive" for stale entries.
-     */
-    stale: boolean;
-    /**
-     * Negative-valence counter (P1 效价反馈): how many prior assessments marked
-     * this entry contradicted (metadata {@link VALENCE_NEGATIVE_KEY}). Shown to
-     * the assessor so it can prefer "archive"; also demotes the entry in
-     * injection ranking (inject.ts rankEntries).
-     */
-    negativeCount: number;
-}
-export declare function candidateKey(kind: RefinementKind, id: string): string;
-/**
- * Deterministic global-coverage check (STRONG signal): the global store
- * already covers a topic when it holds a title that normalizes equal to, or
- * (beyond a length floor) contains, the candidate's normalized title.
- * Archived global entries count too — the topic was already judged
- * cross-session; a local duplicate would only re-sediment it.
- *
- * The bare same-id case is deliberately NOT coverage: ids are slugs derived
- * from titles, so a real collision is usually caught by the title check
- * below. A same-id entry with a wildly different title is a weak signal — the
- * caller routes it through {@link globalHintsFor} for the assessor to judge
- * against the actual global title (real case: local `memory` "用户产品愿景与
- * 收入需求（本会话）" vs global `memory` "用户画像（持续更新）").
- */
-export declare function globalCoverageDetected(globalState: HarnessState, kind: RefinementKind, entry: Pick<HarnessEntry, "id" | "title">): boolean;
-/**
- * The actual global entries that touch the same topic as a local candidate:
- * same id (regardless of title — the weak collision signal that is NOT
- * coverage on its own), equal normalized title, or title overlap. The raw ids
- * and titles let the assessor judge enrichment against real global content
- * (does the global copy already hold what the local one adds?) rather than a
- * bare boolean. Bounded: a handful of best matches, never the whole store.
- */
-export declare function globalHintsFor(globalState: HarnessState, kind: RefinementKind, entry: Pick<HarnessEntry, "id" | "title">): GlobalHint[];
-export declare function listLocalCandidates(state: HarnessState, globalState: HarnessState, baseDir?: string): WrapupCandidate[];
-/**
- * Parse and validate the model's assessment JSON. Defense is mechanical:
- * keys outside the candidate list are dropped, verdicts outside the enum
- * collapse to "keep", and candidates the model omitted default to "keep" —
- * a malformed reply can never change an entry's fate by itself.
- *
- * Split promotion (verdict "archive" with a `promote` sub-object): the
- * sub-object is accepted ONLY on archive verdicts and ONLY when both cleaned
- * title and content are non-empty strings — a dropped/malformed sub-object
- * silently degrades to a plain archive (the entry is never half-promoted).
- */
-export declare function parseWrapupAssessment(text: string, candidates: readonly WrapupCandidate[]): WrapupAssessment;
-//# sourceMappingURL=wrapup-candidates.d.ts.map

package/lib/wrapup-candidates.js

@@ -1,161 +0,0 @@
-import { PROMOTED_TO_KEY, VALENCE_NEGATIVE_KEY, isArchived } from "./types.js";
-import { extractJsonObject } from "./plan.js";
-import { getUsageCount, loadUsage } from "./usage.js";
-import { recencyScore } from "./injection-rank.js";
-export function candidateKey(kind, id) {
-    return `${kind}:${id}`;
-}
-/** Lowercase, punctuation-stripped title used for cheap coverage matching. */
-function normalizeKey(value) {
-    return value.toLowerCase().replace(/[^a-z0-9\u3400-\u9fff]+/g, "").trim();
-}
-/**
- * Deterministic global-coverage check (STRONG signal): the global store
- * already covers a topic when it holds a title that normalizes equal to, or
- * (beyond a length floor) contains, the candidate's normalized title.
- * Archived global entries count too — the topic was already judged
- * cross-session; a local duplicate would only re-sediment it.
- *
- * The bare same-id case is deliberately NOT coverage: ids are slugs derived
- * from titles, so a real collision is usually caught by the title check
- * below. A same-id entry with a wildly different title is a weak signal — the
- * caller routes it through {@link globalHintsFor} for the assessor to judge
- * against the actual global title (real case: local `memory` "用户产品愿景与
- * 收入需求（本会话）" vs global `memory` "用户画像（持续更新）").
- */
-export function globalCoverageDetected(globalState, kind, entry) {
-    const records = globalState.entries[kind];
-    const title = normalizeKey(entry.title);
-    if (title.length === 0)
-        return false;
-    for (const other of Object.values(records)) {
-        const otherTitle = normalizeKey(other.title);
-        if (otherTitle.length === 0)
-            continue;
-        if (otherTitle === title)
-            return true;
-        if (title.length >= 4 && otherTitle.length >= 4 && (title.includes(otherTitle) || otherTitle.includes(title))) {
-            return true;
-        }
-    }
-    return false;
-}
-/**
- * The actual global entries that touch the same topic as a local candidate:
- * same id (regardless of title — the weak collision signal that is NOT
- * coverage on its own), equal normalized title, or title overlap. The raw ids
- * and titles let the assessor judge enrichment against real global content
- * (does the global copy already hold what the local one adds?) rather than a
- * bare boolean. Bounded: a handful of best matches, never the whole store.
- */
-export function globalHintsFor(globalState, kind, entry) {
-    const records = globalState.entries[kind];
-    const title = normalizeKey(entry.title);
-    const hints = [];
-    for (const other of Object.values(records)) {
-        const otherTitle = normalizeKey(other.title);
-        if (otherTitle.length === 0 && other.id !== entry.id)
-            continue;
-        const matches = other.id === entry.id ||
-            (otherTitle.length > 0 && (otherTitle === title || (title.length >= 4 && otherTitle.length >= 4 && (title.includes(otherTitle) || otherTitle.includes(title)))));
-        if (matches) {
-            hints.push({ id: other.id, title: other.title });
-        }
-    }
-    return hints;
-}
-/**
- * The auditable local candidates of a session: every non-archived local
- * entry that has not already been promoted (a promoted entry's lifecycle is
- * finished — the global copy is the live one). Each carries its
- * `coveredGlobally` flag so the assessor never wastes a promote on a topic
- * the global store already owns.
- */
-/** Staleness threshold: entries with recency below this AND zero usage are stale. */
-const STALE_RECENCY_THRESHOLD = 0.1;
-export function listLocalCandidates(state, globalState, baseDir) {
-    const usage = baseDir ? loadUsage(baseDir) : undefined;
-    const now = Date.now();
-    const candidates = [];
-    for (const kind of Object.keys(state.entries)) {
-        for (const entry of Object.values(state.entries[kind])) {
-            if (entry.scope !== "local")
-                continue;
-            if (isArchived(entry))
-                continue;
-            if (typeof entry.metadata[PROMOTED_TO_KEY] === "string")
-                continue;
-            const injectionCount = usage ? getUsageCount(usage, kind, entry.id) : 0;
-            const stale = injectionCount === 0 && recencyScore(entry, now) < STALE_RECENCY_THRESHOLD;
-            const rawNegative = entry.metadata[VALENCE_NEGATIVE_KEY];
-            const negativeCount = typeof rawNegative === "number" && Number.isFinite(rawNegative) && rawNegative > 0 ? rawNegative : 0;
-            candidates.push({
-                kind,
-                id: entry.id,
-                title: entry.title,
-                content: entry.content,
-                path: entry.path,
-                version: entry.version,
-                metadata: entry.metadata,
-                coveredGlobally: globalCoverageDetected(globalState, kind, entry),
-                globalHints: globalHintsFor(globalState, kind, entry),
-                injectionCount,
-                stale,
-                negativeCount,
-            });
-        }
-    }
-    return candidates;
-}
-/**
- * Parse and validate the model's assessment JSON. Defense is mechanical:
- * keys outside the candidate list are dropped, verdicts outside the enum
- * collapse to "keep", and candidates the model omitted default to "keep" —
- * a malformed reply can never change an entry's fate by itself.
- *
- * Split promotion (verdict "archive" with a `promote` sub-object): the
- * sub-object is accepted ONLY on archive verdicts and ONLY when both cleaned
- * title and content are non-empty strings — a dropped/malformed sub-object
- * silently degrades to a plain archive (the entry is never half-promoted).
- */
-export function parseWrapupAssessment(text, candidates) {
-    const value = extractJsonObject(text);
-    if (typeof value !== "object" || value === null || Array.isArray(value)) {
-        throw new Error("wrap-up assessment JSON must be an object");
-    }
-    const record = value;
-    const allowed = new Set(candidates.map((candidate) => candidateKey(candidate.kind, candidate.id)));
-    const items = [];
-    if (Array.isArray(record["items"])) {
-        for (const raw of record["items"]) {
-            if (typeof raw !== "object" || raw === null)
-                continue;
-            const item = raw;
-            const key = typeof item["key"] === "string" ? item["key"] : "";
-            if (!allowed.has(key))
-                continue;
-            const verdict = item["verdict"] === "promote" || item["verdict"] === "archive" ? item["verdict"] : "keep";
-            const built = { key, verdict, reason: typeof item["reason"] === "string" ? item["reason"] : "" };
-            if (item["contradicted"] === true) {
-                built.contradicted = true;
-            }
-            if (verdict === "archive" && typeof item["promote"] === "object" && item["promote"] !== null) {
-                const sub = item["promote"];
-                const subTitle = typeof sub["title"] === "string" ? sub["title"].trim() : "";
-                const subContent = typeof sub["content"] === "string" ? sub["content"].trim() : "";
-                if (subTitle.length > 0 && subContent.length > 0) {
-                    built.promote = { title: subTitle, content: subContent };
-                }
-            }
-            items.push(built);
-        }
-    }
-    for (const candidate of candidates) {
-        const key = candidateKey(candidate.kind, candidate.id);
-        if (!items.some((item) => item.key === key)) {
-            items.push({ key, verdict: "keep", reason: "not mentioned by the assessor" });
-        }
-    }
-    return { items, rationale: typeof record["rationale"] === "string" ? record["rationale"] : "" };
-}
-//# sourceMappingURL=wrapup-candidates.js.map

package/lib/wrapup-command.d.ts

@@ -1,10 +0,0 @@
-/**
- * The `/evolve wrapup` subcommand handler. Extracted from command.ts (P2-2).
- */
-import type { Context } from "@deepseek-ai/cordis";
-import type { CommandInvocation, CommandResult } from "@deepseek-ai/dsh-commands";
-import type { EvolutionEngine } from "./service.js";
-import { type PromotionPolicy } from "./promotion.js";
-import { type RecordLanguagePreference } from "./record-language.js";
-export declare function executeWrapupCommand(ctx: Context, engine: EvolutionEngine, invocation: CommandInvocation, policy?: PromotionPolicy, recordLanguage?: RecordLanguagePreference): Promise<CommandResult>;
-//# sourceMappingURL=wrapup-command.d.ts.map