@gethmy/agent 1.7.1 → 1.7.2

package/dist/episode-writer.js

@@ -1,349 +0,0 @@
-import { log } from "./log.js";
-const TAG = "episode-writer";
-const MAX_APPROACH_SUMMARY_CHARS = 400;
-// Richer approach summary cap (#272). The single-turn trim used a 400-char cap;
-// the assembled multi-block summary is allowed to run longer but stays bounded
-// so it doesn't bloat the recall prompt.
-const MAX_RICH_APPROACH_CHARS = 1500;
-// Cap on the changed-file list persisted per episode (#272).
-const MAX_CHANGED_FILES = 30;
-/**
- * Rule-derived quality score (0..1) for an implement run. Failures default to 0.
- * Plan §"Quality score": +0.4 if build passed, +0.2 if lint passed, +0.2 if no
- * error thrown, +0.2 if run completed cleanly.
- */
-export function computeQualityScore(result, opts) {
-    if (!result.passed)
-        return 0;
-    let score = 0;
-    if (result.buildErrors.length === 0)
-        score += 0.4;
-    if (result.lintWarnings.length === 0)
-        score += 0.2;
-    if (!opts.errorThrown)
-        score += 0.2;
-    if (opts.runCompletedCleanly)
-        score += 0.2;
-    return Math.min(1, score);
-}
-/**
- * Clamp confidence into the documented [0.4, 1.0] band so failures retain a
- * minimum floor (plan §"Episode record shape").
- */
-function clampConfidence(qualityScore) {
-    return Math.max(0.4, Math.min(1.0, qualityScore));
-}
-/**
- * Trim a free-form summary down to the documented 400-char cap. v1 uses a
- * last-turn trim rather than an LLM rewrite (plan §"Write hook"). Empty or
- * whitespace-only input collapses to a marker so the episode still surfaces
- * as a recallable hit (rather than an empty bullet) in future prompts.
- */
-export function trimApproachSummary(text) {
-    const trimmed = text.trim();
-    if (trimmed.length === 0)
-        return "(no approach summary captured)";
-    if (trimmed.length <= MAX_APPROACH_SUMMARY_CHARS)
-        return trimmed;
-    return `${trimmed.slice(0, MAX_APPROACH_SUMMARY_CHARS - 1).trimEnd()}…`;
-}
-// Review rationale cap (#272 task 5). The reviewer's verdict reasoning is the
-// signal we want recallable, so it gets a far higher cap than the 400-char
-// implement trim (the upstream review summary is already sliced to ~2000).
-const MAX_REVIEW_RATIONALE_CHARS = 2000;
-/**
- * Cap the review rationale ("why approved / rejected") richly (#272 task 5)
- * rather than re-trimming to the 400-char implement bound. Empty input
- * collapses to a marker so the episode still surfaces as a recallable hit.
- */
-export function trimReviewRationale(text) {
-    const trimmed = text.trim();
-    if (trimmed.length === 0)
-        return "(no review rationale captured)";
-    if (trimmed.length <= MAX_REVIEW_RATIONALE_CHARS)
-        return trimmed;
-    return `${trimmed.slice(0, MAX_REVIEW_RATIONALE_CHARS - 1).trimEnd()}…`;
-}
-// Lines that read like a root-cause / gotcha / lesson — used to surface a cheap
-// deterministic `key_insight` without an LLM call (#272 task 3).
-const INSIGHT_RE = /\b(root cause|turned out|the (?:issue|problem|bug) (?:was|is)|the fix (?:was|is)|gotcha|caused by|because|the key (?:was|insight)|note that|caveat|the trick (?:was|is))\b/i;
-/**
- * Assemble a richer approach summary from the collected assistant text blocks
- * (#272). Joins the trailing blocks (most relevant context lives near the end
- * of a run) up to a longer bounded cap than the single-turn trim. Falls back to
- * the last-turn `fallback` text when no blocks were collected. No LLM call.
- */
-export function buildRichApproachSummary(blocks, fallback) {
-    const cleaned = (blocks ?? [])
-        .map((b) => b.trim())
-        .filter((b) => b.length > 0);
-    if (cleaned.length === 0)
-        return trimApproachSummary(fallback);
-    // Walk backwards accumulating blocks until we hit the cap, then restore order.
-    const picked = [];
-    let total = 0;
-    for (let i = cleaned.length - 1; i >= 0; i--) {
-        const block = cleaned[i];
-        const cost = block.length + (picked.length > 0 ? 2 : 0);
-        if (total + cost > MAX_RICH_APPROACH_CHARS && picked.length > 0)
-            break;
-        picked.unshift(block);
-        total += cost;
-        if (total >= MAX_RICH_APPROACH_CHARS)
-            break;
-    }
-    const joined = picked.join("\n\n").trim();
-    if (joined.length === 0)
-        return trimApproachSummary(fallback);
-    if (joined.length <= MAX_RICH_APPROACH_CHARS)
-        return joined;
-    return `${joined.slice(0, MAX_RICH_APPROACH_CHARS - 1).trimEnd()}…`;
-}
-/**
- * Extract a cheap deterministic "key insight" line from the run's assistant
- * text (#272 task 3). Scans for a sentence/line matching a root-cause / gotcha
- * pattern. Returns undefined when nothing matches — never fabricated, no LLM.
- */
-export function extractKeyInsight(blocks, fallback) {
-    const source = blocks && blocks.length > 0 ? blocks.join("\n") : fallback;
-    const lines = source
-        .split(/\n|(?<=[.!?])\s+/)
-        .map((l) => l.trim())
-        .filter((l) => l.length >= 20 && l.length <= 280);
-    for (const line of lines) {
-        if (INSIGHT_RE.test(line))
-            return line;
-    }
-    return undefined;
-}
-/**
- * Build the entity payload for one episode. Pure — returned object can be
- * snapshotted in tests without hitting the network.
- */
-export function buildEpisodePayload(input, projectId) {
-    if (input.kind === "implement") {
-        const qualityScore = computeQualityScore(input.result, {
-            errorThrown: input.errorMessage !== undefined,
-            runCompletedCleanly: input.result.passed,
-        });
-        const type = input.outcome === "success" ? "solution" : "error";
-        const importance = input.outcome === "success" ? 7 : 5;
-        // Richer approach summary (#272): assemble from all collected assistant
-        // blocks when available; fall back to the single last-turn trim otherwise.
-        const approachSummary = buildRichApproachSummary(input.approachBlocks, input.approachSummary);
-        const keyInsight = extractKeyInsight(input.approachBlocks, input.approachSummary);
-        const changedFiles = (input.changedFiles ?? []).slice(0, MAX_CHANGED_FILES);
-        const outcomeRationale = input.outcome === "success"
-            ? `Build ${input.result.buildErrors.length === 0 ? "passed" : "failed"}, lint ${input.result.lintWarnings.length === 0 ? "clean" : "issues"}.`
-            : `Verification failed: ${input.errorMessage ?? "see findings"}.`;
-        const metadata = {
-            episode_kind: "implement",
-            card_short_id: input.card.short_id,
-            card_title: input.card.title,
-            approach_summary: approachSummary,
-            outcome: input.outcome,
-            quality_score: qualityScore,
-            duration_ms: input.cost?.durationMs ?? 0,
-            token_cost: {
-                input: input.cost?.totalInputTokens ?? 0,
-                output: input.cost?.totalOutputTokens ?? 0,
-                usd: input.cost?.totalCostUsd ?? 0,
-            },
-            files_touched: input.filesEdited,
-            num_turns: input.cost?.numTurns ?? 0,
-            // Provenance (#273, task 3). Agent episodes are always agent-run; stamp
-            // the card + session ids we already carry so downstream hygiene + UI can
-            // tell auto-written episodes from human-curated patterns.
-            origin: {
-                source: "agent-run",
-                source_card_id: input.card.id,
-                author: "harmony-agent",
-                ...(input.agentSessionId
-                    ? { source_session_id: input.agentSessionId }
-                    : {}),
-            },
-        };
-        if (input.errorMessage)
-            metadata.error = input.errorMessage;
-        if (changedFiles.length > 0)
-            metadata.changed_files = changedFiles;
-        if (input.churn)
-            metadata.churn = input.churn;
-        if (keyInsight)
-            metadata.key_insight = keyInsight;
-        if (input.agentSessionId)
-            metadata.agent_session_id = input.agentSessionId;
-        const changedFilesLine = changedFiles.length > 0
-            ? `\n\nChanged files (${changedFiles.length}): ${changedFiles.join(", ")}`
-            : "";
-        const keyInsightLine = keyInsight ? `\n\nKey insight: ${keyInsight}` : "";
-        return {
-            workspace_id: input.workspaceId,
-            project_id: projectId,
-            type,
-            memory_tier: "episode",
-            scope: "project",
-            title: `Agent run implement — #${input.card.short_id}: ${input.card.title}`,
-            content: `${approachSummary}\n\nOutcome: ${outcomeRationale}${keyInsightLine}${changedFilesLine}`,
-            metadata,
-            importance,
-            confidence: clampConfidence(qualityScore),
-            tags: ["implement", input.outcome, `card:${input.card.short_id}`],
-            agent_identifier: "harmony-agent",
-        };
-    }
-    // Review episode
-    const qualityScore = input.verdict === "approved" ? 1 : 0.4;
-    const summary = trimReviewRationale(input.summary || "");
-    const metadata = {
-        episode_kind: "review",
-        card_short_id: input.card.short_id,
-        card_title: input.card.title,
-        approach_summary: summary,
-        outcome: input.verdict === "approved" ? "success" : "failure",
-        quality_score: qualityScore,
-        duration_ms: input.cost?.durationMs ?? 0,
-        token_cost: {
-            input: input.cost?.totalInputTokens ?? 0,
-            output: input.cost?.totalOutputTokens ?? 0,
-            usd: input.cost?.totalCostUsd ?? 0,
-        },
-        files_touched: 0,
-        num_turns: input.cost?.numTurns ?? 0,
-        // Provenance (#273, task 3). Review episodes are agent-run too. Prefer the
-        // review session id as the originating session when present.
-        origin: {
-            source: "agent-run",
-            source_card_id: input.card.id,
-            author: "harmony-agent",
-            ...((input.reviewSessionId ?? input.agentSessionId)
-                ? {
-                    source_session_id: (input.reviewSessionId ??
-                        input.agentSessionId),
-                }
-                : {}),
-        },
-    };
-    if (input.agentSessionId)
-        metadata.agent_session_id = input.agentSessionId;
-    if (input.reviewSessionId)
-        metadata.review_session_id = input.reviewSessionId;
-    if (input.originalEpisodeId)
-        metadata.original_episode_id = input.originalEpisodeId;
-    return {
-        workspace_id: input.workspaceId,
-        project_id: projectId,
-        type: "decision",
-        memory_tier: "episode",
-        scope: "project",
-        title: `Agent run review — #${input.card.short_id}: ${input.card.title}`,
-        content: `Review verdict: ${input.verdict}.\n\n${summary}`,
-        metadata,
-        importance: 8,
-        confidence: clampConfidence(qualityScore),
-        tags: ["review", input.verdict, `card:${input.card.short_id}`],
-        agent_identifier: "harmony-agent",
-    };
-}
-// DEFERRED (#272 task 6, on-plan): an optional best-effort Claude-CLI
-// distillation of the approach summary / key insight is explicitly out of scope
-// for this pass — no Anthropic SDK, no extra LLM spawn. The deterministic
-// assembly above is the v1 surface; revisit distillation in a later card.
-/**
- * Write one episode entity. Best-effort: any failure is logged and swallowed
- * so the calling pipeline can complete (plan D8: episode writes never block
- * run completion).
- *
- * Returns the entity id on success, or null on swallowed failure.
- */
-export async function writeEpisode(client, input) {
-    const payload = buildEpisodePayload(input, input.card.project_id);
-    try {
-        const { entity } = await client.createMemoryEntity({
-            ...payload,
-            metadata: payload.metadata,
-        });
-        const id = entity && typeof entity === "object" && "id" in entity
-            ? (entity.id ?? null)
-            : null;
-        log.info(TAG, `episode written for #${input.card.short_id}`, {
-            cardId: input.card.id,
-            event: "episode_write",
-            kind: input.kind,
-        });
-        return id;
-    }
-    catch (err) {
-        log.warn(TAG, `episode write failed for #${input.card.short_id}`, {
-            cardId: input.card.id,
-            event: "episode_write_failed",
-            kind: input.kind,
-            error: err instanceof Error ? err.message : String(err),
-        });
-        return null;
-    }
-}
-/**
- * Find the most recent implement episode for a given card so the review
- * pipeline can back-fill its verdict. Returns null when none exists or the
- * lookup throws — back-fill is best-effort.
- */
-export async function findLatestImplementEpisode(client, workspaceId, projectId, cardShortId) {
-    try {
-        const { entities } = await client.harmonyRecall({
-            workspaceId,
-            projectId,
-            type: ["solution", "error"],
-            memory_tier: "episode",
-            scope: "project",
-            tags: [`card:${cardShortId}`],
-            topK: 1,
-        });
-        const first = entities[0];
-        if (first &&
-            typeof first === "object" &&
-            "id" in first &&
-            typeof first.id === "string") {
-            return first.id;
-        }
-        return null;
-    }
-    catch (err) {
-        log.warn(TAG, "implement-episode lookup failed", {
-            event: "episode_lookup_failed",
-            cardShortId,
-            error: err instanceof Error ? err.message : String(err),
-        });
-        return null;
-    }
-}
-/**
- * Apply the review verdict to an earlier implement episode (plan §"Read hook"
- * back-fill block). Approved nudges the original episode's confidence up;
- * rejected tombstones it via superseded_by.
- */
-export async function backfillReviewVerdict(client, originalEpisodeId, verdict, reviewEpisodeId) {
-    try {
-        if (verdict === "approved") {
-            const { entity } = await client.getMemoryEntity(originalEpisodeId);
-            const current = entity?.confidence ?? 0.4;
-            const bumped = Math.min(1, current + 0.05);
-            await client.updateMemoryEntity(originalEpisodeId, {
-                confidence: bumped,
-            });
-        }
-        else {
-            await client.updateMemoryEntity(originalEpisodeId, {
-                superseded_by: reviewEpisodeId,
-            });
-        }
-    }
-    catch (err) {
-        log.warn(TAG, "review back-fill failed", {
-            event: "episode_backfill_failed",
-            originalEpisodeId,
-            verdict,
-            error: err instanceof Error ? err.message : String(err),
-        });
-    }
-}

package/dist/git-diff-stat.d.ts

@@ -1,24 +0,0 @@
-/** Default cap on the number of changed-file paths captured per episode. */
-export declare const MAX_CHANGED_FILES = 30;
-export interface DiffStat {
-    /** Changed file paths (authoritative — derived from the diff itself). */
-    files: string[];
-    /** Total lines added across the diff (best-effort, may be 0). */
-    insertions: number;
-    /** Total lines removed across the diff (best-effort, may be 0). */
-    deletions: number;
-}
-/**
- * Parse `git diff --numstat` output into a changed-file list + churn totals.
- * Pure — exported for testing. numstat lines look like:
- *   `12\t3\tpath/to/file.ts`
- * Binary files report `-\t-\t<path>`; those contribute 0 churn but still count
- * as a changed file. The file list is capped at `maxFiles`.
- */
-export declare function parseNumstat(raw: string, maxFiles?: number): DiffStat;
-/**
- * Capture a changed-file list + churn for the work on a branch versus its base,
- * via `git diff --numstat`. Best-effort and guarded: returns null on any
- * failure so callers can fall back to tracked paths (#272). Never throws.
- */
-export declare function captureDiffStat(worktreePath: string, baseBranch: string, maxFiles?: number): DiffStat | null;

package/dist/git-diff-stat.js

@@ -1,56 +0,0 @@
-import { execFileSync } from "node:child_process";
-import { log } from "./log.js";
-const TAG = "git-diff-stat";
-/** Default cap on the number of changed-file paths captured per episode. */
-export const MAX_CHANGED_FILES = 30;
-/**
- * Parse `git diff --numstat` output into a changed-file list + churn totals.
- * Pure — exported for testing. numstat lines look like:
- *   `12\t3\tpath/to/file.ts`
- * Binary files report `-\t-\t<path>`; those contribute 0 churn but still count
- * as a changed file. The file list is capped at `maxFiles`.
- */
-export function parseNumstat(raw, maxFiles = MAX_CHANGED_FILES) {
-    const files = [];
-    let insertions = 0;
-    let deletions = 0;
-    for (const line of raw.split("\n")) {
-        const trimmed = line.trim();
-        if (trimmed.length === 0)
-            continue;
-        const parts = trimmed.split("\t");
-        if (parts.length < 3)
-            continue;
-        const [add, del, ...pathParts] = parts;
-        const path = pathParts.join("\t");
-        if (!path)
-            continue;
-        const addN = Number.parseInt(add, 10);
-        const delN = Number.parseInt(del, 10);
-        if (Number.isFinite(addN))
-            insertions += addN;
-        if (Number.isFinite(delN))
-            deletions += delN;
-        if (files.length < maxFiles)
-            files.push(path);
-    }
-    return { files, insertions, deletions };
-}
-/**
- * Capture a changed-file list + churn for the work on a branch versus its base,
- * via `git diff --numstat`. Best-effort and guarded: returns null on any
- * failure so callers can fall back to tracked paths (#272). Never throws.
- */
-export function captureDiffStat(worktreePath, baseBranch, maxFiles = MAX_CHANGED_FILES) {
-    try {
-        const raw = execFileSync("git", ["diff", "--numstat", `${baseBranch}...HEAD`], { cwd: worktreePath, encoding: "utf-8", timeout: 30_000 });
-        return parseNumstat(raw, maxFiles);
-    }
-    catch (err) {
-        log.warn(TAG, "git diff --numstat failed", {
-            event: "diff_stat_failed",
-            error: err instanceof Error ? err.message : String(err),
-        });
-        return null;
-    }
-}

package/dist/git-pr.d.ts

@@ -1,38 +0,0 @@
-import type { Card } from "@harmony/shared";
-import type { AgentConfig } from "./types.js";
-export type GitProvider = "github" | "azure" | "gitlab" | "bitbucket" | "unknown";
-export declare function detectGitProvider(cwd?: string): GitProvider;
-/**
- * Validate that the CLI for the detected git provider is installed and authenticated.
- * Returns the provider name or throws with instructions.
- */
-export declare function validateGitProviderCli(provider: GitProvider, cwd?: string): void;
-export type PrState = "merged" | "open" | "closed" | "unknown";
-/**
- * Check whether a PR has been merged using the git provider CLI (async).
- */
-export declare function checkPrMergeStatus(prUrl: string, cwd: string, provider: GitProvider): Promise<PrState>;
-/**
- * Extract a PR URL from card description. Matches the format written by completion.ts:
- * `PR: https://...`
- */
-export declare function extractPrUrl(description: string | null): string | null;
-export declare function remoteBranchExists(branchName: string, cwd: string): boolean;
-export declare function pushBranch(branchName: string, cwd: string): void;
-/**
- * Push the current branch's tip to `newRef` on origin and delete `oldRef`.
- * Used when an approved attempt graduates from `agent-attempts/*` to
- * `agent/*` — keeps the commits durable across the rename and avoids any
- * window where the work is unreachable on origin.
- */
-export declare function renameRemoteBranch(oldRef: string, newRef: string, cwd: string): void;
-/**
- * Best-effort public branch URL for the recovery button on a failed session.
- * Returns null when we can't infer a tree URL — the daemon falls back to a
- * plain `git fetch && git checkout <ref>` instruction in that case.
- */
-export declare function getBranchWebUrl(branchName: string, cwd: string): string | null;
-export declare function buildPrBody(card: Card, commitLog: string): string;
-export declare function createPullRequest(card: Card, branchName: string, worktreePath: string, config: AgentConfig, provider: GitProvider): string | null;
-export declare function findExistingPr(branchName: string, worktreePath: string, provider: GitProvider): string | null;
-export declare function updateExistingPr(branchName: string, body: string, worktreePath: string, provider: GitProvider): void;