@boyingliu01/xp-gate 0.27.6 → 0.28.1

package/plugins/dsh/lib/delphi-config.js

@@ -1,183 +0,0 @@
-/**
- * Delphi configuration loading for the DSH plugin.
- *
- * Reads the same `.delphi-config.json` the xp-gate skill already uses, so a
- * project configures its reviewers once and both the skill and the DSH tool
- * agree on who they are.
- *
- * @test REQ-DSH-016
- * @intent 复用项目既有的 .delphi-config.json，把 provider/api_key/专家席位解析成
- *         DSH 工具可用的结构，并对 ${ENV} 占位符与缺失密钥给出可操作的报错。
- * @covers AC-DSH-016-02
- */
-import { readFileSync } from "node:fs";
-import { homedir } from "node:os";
-import { join, resolve } from "node:path";
-import { validateSeats } from "./delphi.js";
-/** Raised when config on disk cannot be turned into a usable review setup. */
-export class DelphiSetupError extends Error {
-}
-/**
- * Expand `${VAR}` and `{env:VAR}` placeholders from the environment.
- * Supports the two spellings already present in xp-gate config files.
- */
-export function expandEnv(value, env = process.env) {
-    return value
-        .replace(/\$\{([A-Za-z_][A-Za-z0-9_]*)\}/g, (_m, name) => env[name] ?? "")
-        .replace(/\{env:([A-Za-z_][A-Za-z0-9_]*)\}/g, (_m, name) => env[name] ?? "");
-}
-/** Pick the active profile, falling back to the top level. */
-function selectProfile(cfg) {
-    const active = cfg.active_profile;
-    if (active && cfg.profiles?.[active])
-        return cfg.profiles[active];
-    if (cfg.profiles && Object.keys(cfg.profiles).length > 0) {
-        throw new DelphiSetupError(`No active_profile set. Available: ${Object.keys(cfg.profiles).join(", ")}`);
-    }
-    return cfg;
-}
-function seatFromRaw(raw, index, fallbackRole) {
-    const model = raw.model ?? raw.requested_model ?? "";
-    return {
-        id: raw.id ?? raw.name ?? `Expert ${String.fromCharCode(65 + index)}`,
-        role: raw.role ?? fallbackRole,
-        model,
-        maxTokens: raw.max_tokens ?? raw.maxTokens,
-    };
-}
-/** Normalize either experts shape into an ordered list with roles attached. */
-function normalizeExperts(experts) {
-    if (!experts)
-        return [];
-    if (Array.isArray(experts)) {
-        return experts.map((raw, i) => ({ raw, role: raw.role ?? `expert${i}` }));
-    }
-    // Object form: the KEY is the role, which is how xp-gate writes it.
-    return Object.entries(experts).map(([role, raw]) => ({ raw, role }));
-}
-/** Resolve every referenced provider's base_url + api_key, failing loudly. */
-function resolveProviders(profile, needed, env) {
-    if (needed.size === 0) {
-        throw new DelphiSetupError('No provider configured. Set "provider" on the profile or on each expert.');
-    }
-    const providerTable = profile.providers ?? {};
-    const resolved = new Map();
-    for (const name of needed) {
-        const entry = providerTable[name];
-        if (!entry) {
-            throw new DelphiSetupError(`Provider "${name}" is not defined under "providers". ` +
-                `Known: ${Object.keys(providerTable).join(", ") || "(none)"}`);
-        }
-        const baseUrl = entry.base_url ?? entry.baseUrl ?? "";
-        if (!baseUrl)
-            throw new DelphiSetupError(`Provider "${name}" has no base_url.`);
-        const rawKey = entry.api_key ?? entry.apiKey ?? "";
-        const apiKey = expandEnv(rawKey, env);
-        if (!apiKey) {
-            // Name the variable rather than echoing config, which may hold a literal key.
-            const varName = /\$\{([A-Za-z_][A-Za-z0-9_]*)\}/.exec(rawKey)?.[1] ??
-                /\{env:([A-Za-z_][A-Za-z0-9_]*)\}/.exec(rawKey)?.[1];
-            throw new DelphiSetupError(varName
-                ? `API key variable ${varName} is not set in the environment.`
-                : `Provider "${name}" has no api_key.`);
-        }
-        resolved.set(name, { baseUrl, apiKey });
-    }
-    return resolved;
-}
-/**
- * Turn raw config into a ready-to-run review setup.
- * Fails loudly and specifically: a missing key or an unexpanded placeholder is
- * the most common way a Delphi run silently degrades.
- */
-export function resolveConfig(cfg, env = process.env) {
-    const profile = selectProfile(cfg);
-    const normalized = normalizeExperts(profile.experts);
-    // An expert may name its own provider; otherwise there must be a profile-wide
-    // default. Resolve base_url/api_key per provider so mixed setups work.
-    const providerNames = new Set(normalized.map((e) => e.raw.provider ?? profile.provider).filter((p) => !!p));
-    const resolvedProviders = resolveProviders(profile, providerNames, env);
-    const seats = normalized.map((e, i) => seatFromRaw(e.raw, i, e.role));
-    // Surface an unexpanded placeholder as a config error, not a provider 404.
-    for (const s of seats) {
-        if (s.model.includes("${") || s.model.includes("{env:")) {
-            throw new DelphiSetupError(`Expert "${s.id}" model "${s.model}" is an unexpanded placeholder.`);
-        }
-    }
-    validateSeats(seats);
-    return buildResult(profile, normalized, resolvedProviders, seats, cfg);
-}
-/** Assemble the final result after seats and providers validated. */
-function buildResult(profile, normalized, resolvedProviders, seats, cfg) {
-    // All seats share one gateway in the supported setup; require agreement so a
-    // mixed list cannot silently send one seat's key to another host.
-    const chosen = resolvedProviders.get(normalized[0].raw.provider ?? profile.provider ?? "");
-    if (!chosen)
-        throw new DelphiSetupError("Unable to resolve a provider for the first expert.");
-    for (const e of normalized) {
-        const name = e.raw.provider ?? profile.provider ?? "";
-        const other = resolvedProviders.get(name);
-        if (!other || other.baseUrl !== chosen.baseUrl) {
-            throw new DelphiSetupError(`All experts must share one gateway base_url; "${name}" differs. ` +
-                `Split the review or align the providers.`);
-        }
-    }
-    // consensus may sit on the profile or at the document root; both occur.
-    const consensus = profile.consensus ?? cfg.consensus;
-    return {
-        seats,
-        baseUrl: chosen.baseUrl,
-        apiKey: chosen.apiKey,
-        thresholdPercent: consensus?.threshold_percent ?? 90,
-        maxRounds: consensus?.max_review_rounds ?? 5,
-        provider: normalized[0].raw.provider ?? profile.provider ?? "",
-    };
-}
-/** Read and resolve `.delphi-config.json` from a project root. */
-export function loadDelphiConfig(cwd, env = process.env, globalPath = join(homedir(), ".config", "xp-gate", ".delphi-config.json")) {
-    // Global fallback (AC-DSH-016-03): the machine-wide config at
-    // ~/.config/xp-gate/.delphi-config.json is the single source of truth for
-    // every project. Without it, each repo needed its own copy — and reading a
-    // stale or differently-priced config from another project's directory caused
-    // real money loss (2026-10-03: an expensive overseas model burned the daily
-    // quota). A project-local config still wins, so per-project overrides work.
-    const path = resolve(cwd, ".delphi-config.json");
-    let raw;
-    try {
-        raw = readFileSync(path, "utf8");
-    }
-    catch (projectError) {
-        // Fall back ONLY when the project file is genuinely absent (ENOENT); any
-        // other read error must surface rather than silently route this run to
-        // another config. Same semantics as delphi-external-review.cjs, which
-        // gates the fallback on existsSync (walkthrough R8 MA-01).
-        if (projectError?.code !== "ENOENT") {
-            throw new DelphiSetupError(`Cannot read ${path}: ${projectError.message}`);
-        }
-        try {
-            raw = readFileSync(globalPath, "utf8");
-        }
-        catch {
-            throw new DelphiSetupError(`No .delphi-config.json found at ${path} or the global location ${globalPath}.`);
-        }
-    }
-    let parsed;
-    try {
-        parsed = JSON.parse(raw);
-    }
-    catch (e) {
-        throw new DelphiSetupError(`.delphi-config.json is not valid JSON: ${e.message}`);
-    }
-    return resolveConfig(parsed, env);
-}
-/** Build the transport options a round needs from a resolved config. */
-export function toDelphiOptions(cfg) {
-    return {
-        baseUrl: cfg.baseUrl,
-        apiKey: cfg.apiKey,
-        seats: cfg.seats,
-        thresholdPercent: cfg.thresholdPercent,
-        maxRounds: cfg.maxRounds,
-    };
-}
-//# sourceMappingURL=delphi-config.js.map

package/plugins/dsh/lib/delphi-run.d.ts

@@ -1,24 +0,0 @@
-/**
- * Delphi review orchestration: iterate rounds until consensus or the cap.
- *
- * @test REQ-DSH-016
- * @intent 编排多轮交叉评审：首轮匿名独立评审，后续轮次反馈上一轮汇总，
- *         直到达到共识阈值（默认 90%）或触及轮次上限（默认 5），
- *         并把结果整理成可写成 .code-walkthrough-result.json 的证据。
- * @covers AC-DSH-016-03
- */
-import type { DelphiOptions, RoundResult } from "./delphi.js";
-export interface ReviewOutcome {
-    readonly rounds: readonly RoundResult[];
-    readonly consensus: boolean;
-    readonly finalRatio: number;
-    readonly totalRounds: number;
-}
-/**
- * Run a full review. Round 1 is anonymous; each later round sees the aggregate
- * plus whatever the caller says was addressed, then re-votes.
- */
-export declare function runReview(opts: DelphiOptions, subject: string, addressed?: string): Promise<ReviewOutcome>;
-/** Human-readable transcript for the tool result. */
-export declare function formatOutcome(outcome: ReviewOutcome): string;
-//# sourceMappingURL=delphi-run.d.ts.map

package/plugins/dsh/lib/delphi-run.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"delphi-run.d.ts","sourceRoot":"","sources":["../src/delphi-run.ts"],"names":[],"mappings":"AAAA;;;;;;;;GAQG;AAEH,OAAO,KAAK,EAAE,aAAa,EAAc,WAAW,EAAE,MAAM,aAAa,CAAA;AAOzE,MAAM,WAAW,aAAa;IAC5B,QAAQ,CAAC,MAAM,EAAE,SAAS,WAAW,EAAE,CAAA;IACvC,QAAQ,CAAC,SAAS,EAAE,OAAO,CAAA;IAC3B,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAA;IAC3B,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAA;CAC7B;AA0CD;;;GAGG;AACH,wBAAsB,SAAS,CAC7B,IAAI,EAAE,aAAa,EACnB,OAAO,EAAE,MAAM,EACf,SAAS,SAAK,GACb,OAAO,CAAC,aAAa,CAAC,CAoCxB;AAED,qDAAqD;AACrD,wBAAgB,aAAa,CAAC,OAAO,EAAE,aAAa,GAAG,MAAM,CAiB5D"}

package/plugins/dsh/lib/delphi-run.js

@@ -1,92 +0,0 @@
-/**
- * Delphi review orchestration: iterate rounds until consensus or the cap.
- *
- * @test REQ-DSH-016
- * @intent 编排多轮交叉评审：首轮匿名独立评审，后续轮次反馈上一轮汇总，
- *         直到达到共识阈值（默认 90%）或触及轮次上限（默认 5），
- *         并把结果整理成可写成 .code-walkthrough-result.json 的证据。
- * @covers AC-DSH-016-03
- */
-import { runRound } from "./delphi.js";
-/** Mirrors delphi.ts; declared here so orchestrators need not import internals. */
-const DEFAULT_THRESHOLD_PERCENT = 90;
-const DEFAULT_MAX_ROUNDS = 5;
-/** Anonymous first round: each seat answers the same question with no peers. */
-function roundOnePrompt(seat, subject) {
-    return (`You are the ${seat.role.toUpperCase()} expert. Review the material below INDEPENDENTLY. ` +
-        `Do not assume other reviewers exist or agree with you.\n\n${subject}\n\n` +
-        `Respond with <=300 words under the headings: "### Findings" (a markdown table of ` +
-        `ID | Severity | Description | Location | Fix, or a single NONE row), then "### VERDICT" ` +
-        `followed by exactly APPROVED or REQUEST_CHANGES, then "### CONFIDENCE" followed by an ` +
-        `integer 1-10. Approve only if you would personally ship this.`);
-}
-/**
- * Later rounds: feed back the aggregate so seats can revise their position.
- * Individual reports stay anonymous -- only the tally and the addressed items
- * are shared, which is what keeps the method honest.
- */
-function laterRoundPrompt(seat, subject, previous, addressed) {
-    const tally = previous.verdicts
-        .map((v) => `- ${v.role}: ${v.verdict}`)
-        .join("\n");
-    return (`You are the ${seat.role.toUpperCase()} expert. This is ROUND ${previous.round + 1} of a ` +
-        `Delphi consensus review. Individual peer reports are anonymous; only this aggregate is shared:\n\n` +
-        `${tally}\nApprovals: ${previous.approvals}/${previous.validCount} ` +
-        `(threshold ${DEFAULT_THRESHOLD_PERCENT}%).\n\n` +
-        `${addressed ? `What changed since the last round:\n${addressed}\n\n` : ""}` +
-        `Re-examine the material and revise or hold your position.\n\n${subject}\n\n` +
-        `Respond with <=300 words: "### Round assessment" (2-3 sentences), "### Remaining blocking ` +
-        `issues" (table, or a single NONE row), "### VERDICT" followed by exactly APPROVED or ` +
-        `REQUEST_CHANGES, then "### CONFIDENCE" and an integer 1-10.`);
-}
-/**
- * Run a full review. Round 1 is anonymous; each later round sees the aggregate
- * plus whatever the caller says was addressed, then re-votes.
- */
-export async function runReview(opts, subject, addressed = "") {
-    const threshold = opts.thresholdPercent ?? DEFAULT_THRESHOLD_PERCENT;
-    const maxRounds = opts.maxRounds ?? DEFAULT_MAX_ROUNDS;
-    const rounds = [];
-    for (let n = 1; n <= maxRounds; n++) {
-        const previous = rounds[rounds.length - 1];
-        const result = previous
-            ? await runRound(opts, n, "You are a rigorous senior code reviewer.", (seat) => laterRoundPrompt(seat, subject, previous, addressed))
-            : await runRound(opts, 1, "You are a rigorous senior code reviewer.", (seat) => roundOnePrompt(seat, subject));
-        rounds.push(result);
-        // Threshold is a percentage of the whole panel, so a round with a malformed
-        // report cannot reach consensus even if every valid seat approved.
-        if (result.approvals / opts.seats.length * 100 >= threshold) {
-            return {
-                rounds,
-                consensus: true,
-                finalRatio: result.approvals / opts.seats.length,
-                totalRounds: n,
-            };
-        }
-    }
-    const last = rounds[rounds.length - 1];
-    return {
-        rounds,
-        consensus: false,
-        finalRatio: last ? last.approvals / opts.seats.length : 0,
-        totalRounds: rounds.length,
-    };
-}
-/** Human-readable transcript for the tool result. */
-export function formatOutcome(outcome) {
-    const lines = [];
-    for (const r of outcome.rounds) {
-        lines.push(`## Round ${r.round} — approvals ${r.approvals}/${r.validCount} valid`);
-        for (const v of r.verdicts) {
-            const model = v.resolvedModel ? `${v.requestedModel} -> ${v.resolvedModel}` : v.requestedModel;
-            const note = v.error ? ` (${v.error})` : "";
-            lines.push(`- [${v.role}] ${v.id} (${model}): ${v.verdict} confidence=${v.confidence}${note}`);
-        }
-        lines.push("");
-    }
-    lines.push(outcome.consensus
-        ? `CONSENSUS REACHED: ${(outcome.finalRatio * 100).toFixed(0)}% after ${outcome.totalRounds} round(s).`
-        : `NO CONSENSUS: ${(outcome.finalRatio * 100).toFixed(0)}% after ${outcome.totalRounds} round(s) (cap reached).`);
-    return lines.join("\n");
-}
-//# sourceMappingURL=delphi-run.js.map

package/plugins/dsh/lib/delphi.d.ts

@@ -1,88 +0,0 @@
-/**
- * Delphi cross-review over an external OpenAI-compatible provider.
- *
- * This is the codified form of the manual run that produced PR #456: three
- * DISTINCT models, each asked the same question independently and anonymously,
- * iterated until the approval ratio clears the consensus threshold.
- *
- * Why a plugin tool rather than a shell one-liner: the DSH tool plane is where
- * a session actually lives, so the mechanism has to be callable from there.
- * DSH's own `tool-subagent` cannot serve this purpose -- subagents inherit one
- * host-configured model, so they cannot guarantee three distinct models, and
- * Delphi's whole value rests on that distinctness.
- *
- * @test REQ-DSH-016
- * @intent 用外部 provider 的多个不同模型做匿名交叉评审并判定共识，
- *         把「多模型交叉评审」固化成 DSH 可调用的机制而非一次性脚本。
- * @covers AC-DSH-016-01
- */
-/** One expert seat. `model` is the id sent to the provider. */
-export interface ExpertSeat {
-    readonly id: string;
-    readonly role: string;
-    readonly model: string;
-    /** Reasoning models burn hidden tokens, so seats need independent budgets. */
-    readonly maxTokens?: number;
-}
-/** A single expert's parsed verdict for one round. */
-export interface ExpertVerdict {
-    readonly id: string;
-    readonly role: string;
-    readonly requestedModel: string;
-    readonly resolvedModel: string | null;
-    readonly verdict: "APPROVED" | "REQUEST_CHANGES" | "INVALID_NO_VERDICT" | "ERROR";
-    readonly confidence: number;
-    /** Only well-formed verdicts may count toward consensus. */
-    readonly counted: boolean;
-    readonly report: string;
-    readonly error?: string;
-}
-/** Aggregated outcome of one round. */
-export interface RoundResult {
-    readonly round: number;
-    readonly verdicts: readonly ExpertVerdict[];
-    readonly validCount: number;
-    readonly approvals: number;
-    readonly consensusRatio: number;
-    readonly reachedConsensus: boolean;
-}
-export interface DelphiOptions {
-    /** OpenAI-compatible base URL, e.g. https://host/gpt-proxy */
-    readonly baseUrl: string;
-    /** Bearer token. Callers read this from the environment; never hardcode it. */
-    readonly apiKey: string;
-    readonly seats: readonly ExpertSeat[];
-    /** Percent, 0-100. Delphi default is 90. */
-    readonly thresholdPercent?: number;
-    readonly maxRounds?: number;
-    /** Per-request ceiling; the caller owns the overall timeout. */
-    readonly timeoutMs?: number;
-    readonly fetchImpl?: typeof fetch;
-}
-/** Raised when the configured seats cannot satisfy the Delphi invariants. */
-export declare class DelphiConfigError extends Error {
-}
-/**
- * Reject a seat configuration that cannot produce a valid Delphi review.
- * Checked BEFORE any network call so a misconfiguration never looks like a
- * provider failure.
- */
-export declare function validateSeats(seats: readonly ExpertSeat[]): void;
-/**
- * Parse an expert report into a verdict.
- *
- * The verdict token usually sits on the line AFTER a "### VERDICT" heading, so
- * a same-line `VERDICT: X` pattern misses every real report. A report with no
- * recognisable verdict is INVALID, never a guess: inferring approval from prose
- * would let the gate pass on an assumption.
- */
-export declare function parseVerdict(text: string): {
-    verdict: ExpertVerdict["verdict"];
-    confidence: number;
-};
-/**
- * Run ONE round. All seats are asked concurrently and never see each other's
- * output, which is what makes the round anonymous.
- */
-export declare function runRound(opts: DelphiOptions, round: number, system: string, buildPrompt: (seat: ExpertSeat) => string): Promise<RoundResult>;
-//# sourceMappingURL=delphi.d.ts.map

package/plugins/dsh/lib/delphi.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"delphi.d.ts","sourceRoot":"","sources":["../src/delphi.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,+DAA+D;AAC/D,MAAM,WAAW,UAAU;IACzB,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAA;IACnB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;IACrB,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAA;IACtB,8EAA8E;IAC9E,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAA;CAC5B;AAED,sDAAsD;AACtD,MAAM,WAAW,aAAa;IAC5B,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAA;IACnB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;IACrB,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAA;IAC/B,QAAQ,CAAC,aAAa,EAAE,MAAM,GAAG,IAAI,CAAA;IACrC,QAAQ,CAAC,OAAO,EAAE,UAAU,GAAG,iBAAiB,GAAG,oBAAoB,GAAG,OAAO,CAAA;IACjF,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAA;IAC3B,4DAA4D;IAC5D,QAAQ,CAAC,OAAO,EAAE,OAAO,CAAA;IACzB,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAA;IACvB,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAA;CACxB;AAED,uCAAuC;AACvC,MAAM,WAAW,WAAW;IAC1B,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAA;IACtB,QAAQ,CAAC,QAAQ,EAAE,SAAS,aAAa,EAAE,CAAA;IAC3C,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAA;IAC3B,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAA;IAC1B,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAA;IAC/B,QAAQ,CAAC,gBAAgB,EAAE,OAAO,CAAA;CACnC;AAED,MAAM,WAAW,aAAa;IAC5B,8DAA8D;IAC9D,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAA;IACxB,+EAA+E;IAC/E,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAA;IACvB,QAAQ,CAAC,KAAK,EAAE,SAAS,UAAU,EAAE,CAAA;IACrC,4CAA4C;IAC5C,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,CAAA;IAClC,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAA;IAC3B,gEAAgE;IAChE,QAAQ,CAAC,SAAS,CAAC,EAAE,MAAM,CAAA;IAC3B,QAAQ,CAAC,SAAS,CAAC,EAAE,OAAO,KAAK,CAAA;CAClC;AASD,6EAA6E;AAC7E,qBAAa,iBAAkB,SAAQ,KAAK;CAAG;AAE/C;;;;GAIG;AACH,wBAAgB,aAAa,CAAC,KAAK,EAAE,SAAS,UAAU,EAAE,GAAG,IAAI,CAoBhE;AAED;;;;;;;GAOG;AACH,wBAAgB,YAAY,CAAC,IAAI,EAAE,MAAM,GAAG;IAC1C,OAAO,EAAE,aAAa,CAAC,SAAS,CAAC,CAAA;IACjC,UAAU,EAAE,MAAM,CAAA;CACnB,CAWA;AA6DD;;;GAGG;AACH,wBAAsB,QAAQ,CAC5B,IAAI,EAAE,aAAa,EACnB,KAAK,EAAE,MAAM,EACb,MAAM,EAAE,MAAM,EACd,WAAW,EAAE,CAAC,IAAI,EAAE,UAAU,KAAK,MAAM,GACxC,OAAO,CAAC,WAAW,CAAC,CAmDtB"}

package/plugins/dsh/lib/delphi.js

@@ -1,161 +0,0 @@
-/**
- * Delphi cross-review over an external OpenAI-compatible provider.
- *
- * This is the codified form of the manual run that produced PR #456: three
- * DISTINCT models, each asked the same question independently and anonymously,
- * iterated until the approval ratio clears the consensus threshold.
- *
- * Why a plugin tool rather than a shell one-liner: the DSH tool plane is where
- * a session actually lives, so the mechanism has to be callable from there.
- * DSH's own `tool-subagent` cannot serve this purpose -- subagents inherit one
- * host-configured model, so they cannot guarantee three distinct models, and
- * Delphi's whole value rests on that distinctness.
- *
- * @test REQ-DSH-016
- * @intent 用外部 provider 的多个不同模型做匿名交叉评审并判定共识，
- *         把「多模型交叉评审」固化成 DSH 可调用的机制而非一次性脚本。
- * @covers AC-DSH-016-01
- */
-/** Delphi's defining invariant: exactly three seats and a 90% approval bar. */
-const REQUIRED_EXPERTS = 3;
-const DEFAULT_THRESHOLD_PERCENT = 90;
-const DEFAULT_MAX_TOKENS = 8000;
-const DEFAULT_TIMEOUT_MS = 180_000;
-/** Raised when the configured seats cannot satisfy the Delphi invariants. */
-export class DelphiConfigError extends Error {
-}
-/**
- * Reject a seat configuration that cannot produce a valid Delphi review.
- * Checked BEFORE any network call so a misconfiguration never looks like a
- * provider failure.
- */
-export function validateSeats(seats) {
-    if (seats.length !== REQUIRED_EXPERTS) {
-        throw new DelphiConfigError(`Delphi requires exactly ${REQUIRED_EXPERTS} experts, got ${seats.length}.`);
-    }
-    for (const s of seats) {
-        if (!s.model || s.model.trim() === "") {
-            throw new DelphiConfigError(`Expert "${s.id}" has no model (a local fallback cannot count).`);
-        }
-        if (s.model.includes("${")) {
-            throw new DelphiConfigError(`Expert "${s.id}" model "${s.model}" looks like an unexpanded env var.`);
-        }
-    }
-    const distinct = new Set(seats.map((s) => s.model.trim()));
-    if (distinct.size !== REQUIRED_EXPERTS) {
-        throw new DelphiConfigError(`All ${REQUIRED_EXPERTS} experts must use distinct models; got: ${[...distinct].join(", ")}.`);
-    }
-}
-/**
- * Parse an expert report into a verdict.
- *
- * The verdict token usually sits on the line AFTER a "### VERDICT" heading, so
- * a same-line `VERDICT: X` pattern misses every real report. A report with no
- * recognisable verdict is INVALID, never a guess: inferring approval from prose
- * would let the gate pass on an assumption.
- */
-export function parseVerdict(text) {
-    const section = /VERDICT[\s\S]{0,120}?\b(APPROVED|REQUEST_CHANGES|PASS_WITH_CAVEATS|BLOCKED)\b/i.exec(text);
-    if (!section)
-        return { verdict: "INVALID_NO_VERDICT", confidence: 0 };
-    const raw = section[1].toUpperCase();
-    // PASS_WITH_CAVEATS / BLOCKED are not approvals under the 90% rule.
-    const verdict = raw === "APPROVED" ? "APPROVED" : "REQUEST_CHANGES";
-    const c = /CONFIDENCE[\s\S]{0,40}?(\d{1,2})/i.exec(text);
-    const confidence = c ? Math.min(10, Math.max(0, Number(c[1]))) : 0;
-    return { verdict, confidence };
-}
-/** One provider call. Throws on transport failure or empty content. */
-async function chatOnce(opts, seat, system, user) {
-    const doFetch = opts.fetchImpl ?? fetch;
-    const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(), opts.timeoutMs ?? DEFAULT_TIMEOUT_MS);
-    try {
-        const res = await doFetch(`${opts.baseUrl.replace(/\/+$/, "")}/v1/chat/completions`, {
-            method: "POST",
-            headers: {
-                Authorization: `Bearer ${opts.apiKey}`,
-                "Content-Type": "application/json",
-            },
-            body: JSON.stringify({
-                model: seat.model,
-                messages: [
-                    { role: "system", content: system },
-                    { role: "user", content: user },
-                ],
-                max_tokens: seat.maxTokens ?? DEFAULT_MAX_TOKENS,
-                temperature: 0.2,
-            }),
-            signal: controller.signal,
-        });
-        if (!res.ok) {
-            const body = await res.text().catch(() => "");
-            throw new Error(`HTTP ${res.status}: ${body.slice(0, 200)}`);
-        }
-        const json = (await res.json());
-        const content = json.choices?.[0]?.message?.content ?? "";
-        const finishReason = json.choices?.[0]?.finish_reason ?? "unknown";
-        // Empty content is a real failure -- typically a reasoning model that spent
-        // its whole budget on hidden tokens. Never let it pass as an empty review.
-        if (content.trim() === "") {
-            throw new Error(`empty content (finish_reason=${finishReason}, max_tokens=${seat.maxTokens ?? DEFAULT_MAX_TOKENS}); ` +
-                `raise max_tokens if this is a reasoning model`);
-        }
-        return { content, resolvedModel: json.model ?? null, finishReason };
-    }
-    finally {
-        clearTimeout(timer);
-    }
-}
-/**
- * Run ONE round. All seats are asked concurrently and never see each other's
- * output, which is what makes the round anonymous.
- */
-export async function runRound(opts, round, system, buildPrompt) {
-    const settled = await Promise.allSettled(opts.seats.map(async (seat) => {
-        const r = await chatOnce(opts, seat, system, buildPrompt(seat));
-        return { seat, ...r };
-    }));
-    const verdicts = settled.map((s, i) => {
-        const seat = opts.seats[i];
-        if (s.status !== "fulfilled") {
-            return {
-                id: seat.id,
-                role: seat.role,
-                requestedModel: seat.model,
-                resolvedModel: null,
-                verdict: "ERROR",
-                confidence: 0,
-                counted: false,
-                report: "",
-                error: s.reason instanceof Error ? s.reason.message : String(s.reason),
-            };
-        }
-        const { verdict, confidence } = parseVerdict(s.value.content);
-        return {
-            id: seat.id,
-            role: seat.role,
-            requestedModel: seat.model,
-            resolvedModel: s.value.resolvedModel,
-            verdict,
-            confidence,
-            counted: verdict !== "INVALID_NO_VERDICT",
-            report: s.value.content,
-        };
-    });
-    const counted = verdicts.filter((v) => v.counted);
-    const approvals = counted.filter((v) => v.verdict === "APPROVED").length;
-    // Consensus is measured over VALID responses: a malformed report is not a
-    // dissent, but it cannot be counted as agreement either.
-    const ratio = counted.length === 0 ? 0 : approvals / counted.length;
-    const threshold = opts.thresholdPercent ?? DEFAULT_THRESHOLD_PERCENT;
-    return {
-        round,
-        verdicts,
-        validCount: counted.length,
-        approvals,
-        consensusRatio: ratio,
-        reachedConsensus: counted.length === REQUIRED_EXPERTS && ratio * 100 >= threshold,
-    };
-}
-//# sourceMappingURL=delphi.js.map

package/plugins/dsh/lib/dialect.d.ts

@@ -1,24 +0,0 @@
-/**
- * Shell dialect selection and quoting.
- *
- * DSH chooses its shell backend per platform: `dsh-base` disables
- * `bash-sandbox`/`tool-bash` and enables `pwsh-sandbox`/`tool-pwsh` on win32.
- * A command string written for the wrong dialect fails to parse, so callers
- * must pick the dialect to match the executor rather than assume POSIX.
- */
-export type ShellDialect = "posix" | "powershell";
-/** Quote one argument for the target dialect. */
-export declare function shellQuote(value: string, dialect: ShellDialect): string;
-/**
- * POSIX single-quote shell escaping: `'` becomes `'\''`.
- * @deprecated Use {@link shellQuote} with an explicit dialect.
- */
-export declare function shq(value: string): string;
-/**
- * Pick the dialect for the current platform.
- *
- * `process.platform` is the reliable signal available inside a tool body; it
- * mirrors exactly where DSH enables each sandbox/executor pair.
- */
-export declare function detectDialect(platform?: NodeJS.Platform): ShellDialect;
-//# sourceMappingURL=dialect.d.ts.map

package/plugins/dsh/lib/dialect.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"dialect.d.ts","sourceRoot":"","sources":["../src/dialect.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAEH,MAAM,MAAM,YAAY,GAAG,OAAO,GAAG,YAAY,CAAA;AAEjD,iDAAiD;AACjD,wBAAgB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,OAAO,EAAE,YAAY,GAAG,MAAM,CASvE;AAED;;;GAGG;AACH,wBAAgB,GAAG,CAAC,KAAK,EAAE,MAAM,GAAG,MAAM,CAEzC;AAED;;;;;GAKG;AACH,wBAAgB,aAAa,CAAC,QAAQ,GAAE,MAAM,CAAC,QAA2B,GAAG,YAAY,CAExF"}

package/plugins/dsh/lib/dialect.js

@@ -1,36 +0,0 @@
-/**
- * Shell dialect selection and quoting.
- *
- * DSH chooses its shell backend per platform: `dsh-base` disables
- * `bash-sandbox`/`tool-bash` and enables `pwsh-sandbox`/`tool-pwsh` on win32.
- * A command string written for the wrong dialect fails to parse, so callers
- * must pick the dialect to match the executor rather than assume POSIX.
- */
-/** Quote one argument for the target dialect. */
-export function shellQuote(value, dialect) {
-    if (dialect === "posix") {
-        // POSIX single-quote escaping: `'` becomes `'\''`.
-        return `'${value.replace(/'/g, "'\\''")}'`;
-    }
-    // PowerShell single-quoted strings escape an embedded quote by doubling it.
-    // Backticks and `$(…)` are literal inside single quotes, so no further
-    // escaping is needed to keep them inert.
-    return `'${value.replace(/'/g, "''")}'`;
-}
-/**
- * POSIX single-quote shell escaping: `'` becomes `'\''`.
- * @deprecated Use {@link shellQuote} with an explicit dialect.
- */
-export function shq(value) {
-    return shellQuote(value, "posix");
-}
-/**
- * Pick the dialect for the current platform.
- *
- * `process.platform` is the reliable signal available inside a tool body; it
- * mirrors exactly where DSH enables each sandbox/executor pair.
- */
-export function detectDialect(platform = process.platform) {
-    return platform === "win32" ? "powershell" : "posix";
-}
-//# sourceMappingURL=dialect.js.map

package/plugins/dsh/lib/gate-runner.d.ts

@@ -1,18 +0,0 @@
-import type { ShellExecRequest, ShellExecSpec, ShellRunResult } from "@deepseek-ai/dsh-shell";
-/**
- * Structural view of `ctx.shell` so the runner is testable with a fake shell
- * without importing the abstract `ShellExecutor` service.
- */
-export interface XpGateShell {
-    resolve(request: ShellExecRequest): ShellExecSpec;
-    run(spec: ShellExecSpec): Promise<ShellRunResult>;
-}
-/** Render a completed shell run as a single model-facing text block. */
-export declare function renderGateResult(result: ShellRunResult): string;
-/**
- * Run an xp-gate command through the harness shell seam. Nonzero exits, timeouts,
- * and sandbox denials resolve (and render) rather than reject; only cancellation
- * (the caller's AbortSignal fired first) throws an AbortError with code ABORTED.
- */
-export declare function runXpGate(shell: XpGateShell, signal: AbortSignal, command: string, workdir: string, timeoutMs: number): Promise<string>;
-//# sourceMappingURL=gate-runner.d.ts.map

package/plugins/dsh/lib/gate-runner.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"gate-runner.d.ts","sourceRoot":"","sources":["../src/gate-runner.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,gBAAgB,EAAE,aAAa,EAAE,cAAc,EAAE,MAAM,wBAAwB,CAAA;AAE7F;;;GAGG;AACH,MAAM,WAAW,WAAW;IAC1B,OAAO,CAAC,OAAO,EAAE,gBAAgB,GAAG,aAAa,CAAA;IACjD,GAAG,CAAC,IAAI,EAAE,aAAa,GAAG,OAAO,CAAC,cAAc,CAAC,CAAA;CAClD;AAED,wEAAwE;AACxE,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,cAAc,GAAG,MAAM,CAqB/D;AAED;;;;GAIG;AACH,wBAAsB,SAAS,CAC7B,KAAK,EAAE,WAAW,EAClB,MAAM,EAAE,WAAW,EACnB,OAAO,EAAE,MAAM,EACf,OAAO,EAAE,MAAM,EACf,SAAS,EAAE,MAAM,GAChB,OAAO,CAAC,MAAM,CAAC,CASjB"}