@oneuptime/common 12.0.34 → 13.0.1

package/Server/Utils/CodeRepository/GitHub/GitHubRunReply.ts

@@ -0,0 +1,421 @@
+import logger from "../../Logger";
+import ObjectID from "../../../../Types/ObjectID";
+import AIRunStatus from "../../../../Types/AI/AIRunStatus";
+import GitHubCommandType from "../../../../Types/CodeRepository/GitHubCommand";
+import CodeFixTaskContext, {
+  getGitHubConversationNumber,
+  getGitHubTaskContext,
+  GitHubTaskContext,
+} from "../../../../Types/AI/CodeFixTaskContext";
+import { DashboardClientUrl } from "../../../EnvironmentConfig";
+import AIRun from "../../../../Models/DatabaseModels/AIRun";
+import AIAgentTaskPullRequest from "../../../../Models/DatabaseModels/AIAgentTaskPullRequest";
+import AIRunService from "../../../Services/AIRunService";
+import AIAgentTaskPullRequestService from "../../../Services/AIAgentTaskPullRequestService";
+import GitHubConversation, { GitHubReaction } from "./GitHubConversation";
+import GitHubReplyComposer from "./GitHubReplyComposer";
+import CaptureSpan from "../../Telemetry/CaptureSpan";
+import LIMIT_MAX from "../../../../Types/Database/LimitMax";
+
+/*
+ * How a GitHub-triggered run talks back to the thread it came from.
+ *
+ * One command produces exactly ONE comment from this app, posted when the run
+ * is accepted and EDITED as it finishes. That constraint is the whole design:
+ * a long-running agent that appends a comment per state change turns a pull
+ * request into a status log, and a reviewer stops reading it.
+ *
+ * The outcome is delivered by a sweeper rather than by whoever happened to
+ * finalize the run. There are five places a run can reach a terminal
+ * state — the worker reporting in, the stale-heartbeat sweeper, the orphaned-
+ * queue sweeper, the claim-time executability guard, and an explicit cancel —
+ * and a comment posted from only some of them is worse than none: the user
+ * sees "on it" forever precisely in the cases where something went wrong.
+ * `reportedToGitHubAt` on the run's own context makes the sweep idempotent,
+ * so a GitHub outage costs a minute rather than the reply.
+ */
+export default class GitHubRunReply {
+  public static buildRunUrl(data: {
+    projectId: ObjectID;
+    aiRunId: ObjectID;
+  }): string {
+    return `${DashboardClientUrl.toString()}/${data.projectId.toString()}/ai/agents/${data.aiRunId.toString()}`;
+  }
+
+  public static buildProjectRunsUrl(projectId: ObjectID): string {
+    return `${DashboardClientUrl.toString()}/${projectId.toString()}/ai/agents`;
+  }
+
+  /*
+   * Acknowledge a command in the thread: a reaction on the comment that asked
+   * (instant, and unmistakably from this app) and one comment carrying the
+   * link to the run.
+   *
+   * Best-effort by design. The run is already queued and will do its work
+   * whether or not GitHub accepted the acknowledgement, and failing the
+   * command because a comment did not post would be strictly worse for the
+   * user than a quiet start.
+   */
+  @CaptureSpan()
+  public static async acknowledge(data: {
+    aiRunId: ObjectID;
+    projectId: ObjectID;
+    projectName: string | undefined;
+    github: GitHubTaskContext;
+  }): Promise<number | null> {
+    const conversationNumber: number = getGitHubConversationNumber(data.github);
+
+    if (data.github.triggerCommentId) {
+      await GitHubConversation.addReactionToComment({
+        installationId: data.github.installationId,
+        organizationName: data.github.organizationName,
+        repositoryName: data.github.repositoryName,
+        commentId: data.github.triggerCommentId,
+        reaction: GitHubReaction.Eyes,
+      });
+    } else {
+      await GitHubConversation.addReactionToIssue({
+        installationId: data.github.installationId,
+        organizationName: data.github.organizationName,
+        repositoryName: data.github.repositoryName,
+        issueNumber: conversationNumber,
+        reaction: GitHubReaction.Eyes,
+      });
+    }
+
+    try {
+      const comment: { commentId: number } =
+        await GitHubConversation.createIssueComment({
+          installationId: data.github.installationId,
+          organizationName: data.github.organizationName,
+          repositoryName: data.github.repositoryName,
+          issueNumber: conversationNumber,
+          body: GitHubReplyComposer.acknowledgement({
+            commandType: data.github.commandType,
+            runUrl: GitHubRunReply.buildRunUrl({
+              projectId: data.projectId,
+              aiRunId: data.aiRunId,
+            }),
+            instruction: data.github.instruction,
+            projectName: data.projectName,
+          }),
+        });
+
+      return comment.commentId;
+    } catch (error) {
+      logger.error(
+        `Could not post the GitHub acknowledgement for run ${data.aiRunId.toString()}: ${error}`,
+      );
+      return null;
+    }
+  }
+
+  /*
+   * Post the run's outcome, editing the acknowledgement when there is one.
+   *
+   * Returns true when GitHub accepted the write, so the caller can mark the
+   * run reported. A false answer leaves the marker unset and the next sweep
+   * tries again — which is what makes a GitHub outage recoverable instead of
+   * silently losing the reply.
+   */
+  @CaptureSpan()
+  public static async reportOutcome(data: {
+    run: AIRun;
+    github: GitHubTaskContext;
+  }): Promise<boolean> {
+    const aiRunId: ObjectID | null = data.run.id;
+    const projectId: ObjectID | undefined = data.run.projectId;
+
+    if (!aiRunId || !projectId) {
+      return false;
+    }
+
+    const runUrl: string = GitHubRunReply.buildRunUrl({
+      projectId: projectId,
+      aiRunId: aiRunId,
+    });
+
+    const body: string = await GitHubRunReply.composeOutcomeBody({
+      run: data.run,
+      github: data.github,
+      runUrl: runUrl,
+    });
+
+    const postFresh: () => Promise<void> = async (): Promise<void> => {
+      await GitHubConversation.createIssueComment({
+        installationId: data.github.installationId,
+        organizationName: data.github.organizationName,
+        repositoryName: data.github.repositoryName,
+        issueNumber: getGitHubConversationNumber(data.github),
+        body: body,
+      });
+    };
+
+    try {
+      if (data.github.acknowledgementCommentId) {
+        try {
+          await GitHubConversation.updateIssueComment({
+            installationId: data.github.installationId,
+            organizationName: data.github.organizationName,
+            repositoryName: data.github.repositoryName,
+            commentId: data.github.acknowledgementCommentId,
+            body: body,
+          });
+        } catch (updateError) {
+          /*
+           * The acknowledgement is gone — someone deleted it, which is an
+           * ordinary thing to do to a bot comment. Without this fallback the
+           * edit 404s forever: the outcome is never delivered, the run is
+           * never marked reported, and the sweep retries the same doomed write
+           * every minute until it ages out. A second comment is a far better
+           * outcome than silence.
+           */
+          logger.info(
+            `Could not edit the acknowledgement comment for run ${aiRunId.toString()} (${updateError}); posting the outcome as a new comment.`,
+          );
+
+          await postFresh();
+        }
+      } else {
+        await postFresh();
+      }
+    } catch (error) {
+      logger.warn(
+        `Could not post the GitHub outcome for run ${aiRunId.toString()}: ${error}`,
+      );
+      return false;
+    }
+
+    /*
+     * The reaction is a nicety and must never decide whether we reported —
+     * the comment has already landed by this point, and re-posting it on the
+     * next sweep because a reaction failed would be a duplicate. Wrapped as
+     * well as swallowed inside addReactionToComment, so this stays true even
+     * if that contract ever changes.
+     */
+    if (data.github.triggerCommentId) {
+      try {
+        await GitHubConversation.addReactionToComment({
+          installationId: data.github.installationId,
+          organizationName: data.github.organizationName,
+          repositoryName: data.github.repositoryName,
+          commentId: data.github.triggerCommentId,
+          reaction:
+            data.run.status === AIRunStatus.Completed
+              ? GitHubReaction.Rocket
+              : GitHubReaction.Confused,
+        });
+      } catch (reactionError) {
+        logger.debug(`Could not add the outcome reaction: ${reactionError}`);
+      }
+    }
+
+    return true;
+  }
+
+  private static async composeOutcomeBody(data: {
+    run: AIRun;
+    github: GitHubTaskContext;
+    runUrl: string;
+  }): Promise<string> {
+    if (data.run.status === AIRunStatus.Cancelled) {
+      /*
+       * Not GitHubReplyComposer.cancelled — that one answers the person who
+       * typed "cancel" and counts what it stopped across the thread. THIS
+       * comment is one run's own acknowledgement being closed out, and three
+       * cancelled runs must not each announce "cancelled 1 run on this
+       * thread". It also carries the run link, which is the one outcome where
+       * a reader most wants to see how far it got.
+       */
+      return GitHubReplyComposer.cancelledRun({ runUrl: data.runUrl });
+    }
+
+    if (data.run.status === AIRunStatus.NoFixFound) {
+      return GitHubReplyComposer.noChangeProposed({
+        reason: data.run.errorMessage,
+        runUrl: data.runUrl,
+      });
+    }
+
+    if (data.run.status !== AIRunStatus.Completed) {
+      return GitHubReplyComposer.failed({
+        reason: data.run.errorMessage,
+        runUrl: data.runUrl,
+      });
+    }
+
+    return GitHubReplyComposer.completed({
+      commandType: data.github.commandType,
+      runUrl: data.runUrl,
+      resultLines: await GitHubRunReply.buildResultLines(data),
+    });
+  }
+
+  /*
+   * What the run actually produced, in the reader's terms.
+   *
+   * A revision and a review deliberately do NOT list pull request links: the
+   * revision pushed to the pull request the comment is already on, and the
+   * review is right there in the thread. Linking a reader to the page they
+   * are reading is noise.
+   */
+  private static async buildResultLines(data: {
+    run: AIRun;
+    github: GitHubTaskContext;
+  }): Promise<Array<string>> {
+    if (data.github.commandType === GitHubCommandType.Revise) {
+      return [
+        "I pushed my changes to this pull request's branch. Please re-review the new commits.",
+      ];
+    }
+
+    if (data.github.commandType === GitHubCommandType.Review) {
+      return ["My review is posted on this pull request."];
+    }
+
+    const pullRequests: Array<AIAgentTaskPullRequest> =
+      await AIAgentTaskPullRequestService.findBy({
+        query: { aiRunId: data.run.id! },
+        select: { pullRequestUrl: true, pullRequestNumber: true },
+        limit: LIMIT_MAX,
+        skip: 0,
+        props: { isRoot: true },
+      });
+
+    if (pullRequests.length === 0) {
+      return ["The run finished, but it did not record a pull request."];
+    }
+
+    return pullRequests.map((pullRequest: AIAgentTaskPullRequest) => {
+      const url: string = pullRequest.pullRequestUrl?.toString() || "";
+
+      return pullRequest.pullRequestNumber
+        ? `Opened #${pullRequest.pullRequestNumber} — ${url}`
+        : `Opened a pull request — ${url}`;
+    });
+  }
+
+  /*
+   * Mark a run as having reported to GitHub.
+   *
+   * Written back into the run's own taskContext rather than a new column: the
+   * marker is meaningless without the conversation it belongs to, and a column
+   * would need a migration to say something the JSON already can.
+   *
+   * The whole taskContext is spread back in. TypeORM writes a jsonb column by
+   * REPLACING it, so writing `{ github: ... }` alone would silently drop every
+   * sibling key — traceId, serviceName, telemetryServiceId, the pinned
+   * investigation snapshot. A GitHub run happens to carry none of those today,
+   * which is exactly why this would go unnoticed until the first recipe that
+   * does.
+   */
+  @CaptureSpan()
+  public static async markReported(data: {
+    aiRunId: ObjectID;
+    taskContext: CodeFixTaskContext | undefined;
+    github: GitHubTaskContext;
+  }): Promise<void> {
+    await AIRunService.updateOneById({
+      id: data.aiRunId,
+      data: {
+        taskContext: {
+          ...(data.taskContext || {}),
+          github: {
+            ...data.github,
+            reportedToGitHubAt: new Date().toISOString(),
+          },
+        },
+      } as never,
+      props: { isRoot: true },
+    });
+  }
+
+  /*
+   * Mark a run whose GitHub context is unusable, so the sweep stops
+   * re-selecting it.
+   *
+   * These exist: the claim guard fails a run with an incomplete context, which
+   * makes it terminal with the same incomplete context still on it. Without a
+   * marker, the sweep picks it up, finds nothing it can report, and does that
+   * again every minute for twenty-four hours. Nothing reads a malformed
+   * context for anything else, so stamping the marker onto whatever is there
+   * is safe — and it is the only way to say "considered, cannot be reported".
+   */
+  @CaptureSpan()
+  public static async markUnreportable(data: {
+    aiRunId: ObjectID;
+    taskContext: CodeFixTaskContext | undefined;
+  }): Promise<void> {
+    await AIRunService.updateOneById({
+      id: data.aiRunId,
+      data: {
+        taskContext: {
+          ...(data.taskContext || {}),
+          github: {
+            ...((data.taskContext?.github || {}) as Partial<GitHubTaskContext>),
+            reportedToGitHubAt: new Date().toISOString(),
+          },
+        },
+      } as never,
+      props: { isRoot: true },
+    });
+  }
+
+  /*
+   * Whether a run's GitHub context is present but already reported. Separates
+   * "nothing to do" from "cannot be done", which the sweep has to treat
+   * differently — see markUnreportable.
+   */
+  public static hasReportableContext(run: AIRun): boolean {
+    return Boolean(getGitHubTaskContext(run.taskContext as CodeFixTaskContext));
+  }
+
+  // A run whose conversation is still waiting to hear how it went.
+  public static needsOutcomeReport(run: AIRun): GitHubTaskContext | null {
+    const github: GitHubTaskContext | null = getGitHubTaskContext(
+      run.taskContext as CodeFixTaskContext,
+    );
+
+    if (!github || github.reportedToGitHubAt) {
+      return null;
+    }
+
+    return github;
+  }
+
+  /*
+   * A one-line description of a live run, for "@mention status".
+   *
+   * Deliberately terse and free of internal vocabulary: someone reading a pull
+   * request wants "revising this pull request (started 4 minutes ago)", not a
+   * run id and an enum.
+   */
+  public static describeLiveRun(run: AIRun): string {
+    const github: GitHubTaskContext | null = getGitHubTaskContext(
+      run.taskContext as CodeFixTaskContext,
+    );
+
+    const what: string = github
+      ? GitHubRunReply.describeCommandForStatus(github.commandType)
+      : "working";
+
+    return run.status === AIRunStatus.Queued
+      ? `${what} — queued, waiting for an agent`
+      : `${what} — in progress`;
+  }
+
+  private static describeCommandForStatus(
+    commandType: GitHubCommandType,
+  ): string {
+    switch (commandType) {
+      case GitHubCommandType.Review:
+        return "Reviewing this pull request";
+      case GitHubCommandType.Revise:
+        return "Revising this pull request";
+      case GitHubCommandType.Implement:
+        return "Working on this issue";
+      default:
+        return "Working";
+    }
+  }
+}

package/Server/Utils/CodeRepository/GitHub/GitHubWebhookEvents.ts

@@ -0,0 +1,140 @@
+/*
+ * The GitHub webhook events the interactive app subscribes to, and the small
+ * amount of shape-reading needed to get from a raw payload to "who said what,
+ * where".
+ *
+ * Split out from the handler because these are the only places a raw,
+ * attacker-influenced JSON blob is turned into typed values, and they should
+ * be readable — and testable — without the surrounding orchestration.
+ */
+
+import { JSONArray, JSONObject } from "../../../../Types/JSON";
+
+// The label that hands an issue to the app when no other is configured.
+export const DEFAULT_GITHUB_TRIGGER_LABEL: string = "oneuptime";
+
+export enum GitHubWebhookEvent {
+  IssueComment = "issue_comment",
+  Issues = "issues",
+  PullRequest = "pull_request",
+  PullRequestReview = "pull_request_review",
+  PullRequestReviewComment = "pull_request_review_comment",
+  Installation = "installation",
+  InstallationRepositories = "installation_repositories",
+}
+
+export interface GitHubWebhookRepository {
+  organizationName: string;
+  repositoryName: string;
+}
+
+export interface GitHubWebhookSender {
+  login: string;
+  // "User" or "Bot". GitHub App comments are always "Bot".
+  type: string | undefined;
+}
+
+export default class GitHubWebhookEvents {
+  /*
+   * `full_name` is "owner/repo". Splitting on the FIRST slash is deliberate:
+   * an owner name cannot contain one, a repository name cannot either, and
+   * anything that does not split cleanly into two non-empty parts is not a
+   * shape we should act on.
+   */
+  public static getRepository(
+    payload: JSONObject,
+  ): GitHubWebhookRepository | null {
+    const fullName: string | undefined = (
+      payload["repository"] as JSONObject
+    )?.["full_name"]?.toString();
+
+    if (!fullName) {
+      return null;
+    }
+
+    const separatorIndex: number = fullName.indexOf("/");
+
+    if (separatorIndex <= 0 || separatorIndex === fullName.length - 1) {
+      return null;
+    }
+
+    return {
+      organizationName: fullName.substring(0, separatorIndex),
+      repositoryName: fullName.substring(separatorIndex + 1),
+    };
+  }
+
+  public static getInstallationId(payload: JSONObject): string | null {
+    const installationId: string | undefined = (
+      payload["installation"] as JSONObject
+    )?.["id"]?.toString();
+
+    return installationId || null;
+  }
+
+  public static getSender(payload: JSONObject): GitHubWebhookSender | null {
+    const sender: JSONObject | undefined = payload["sender"] as JSONObject;
+    const login: string | undefined = sender?.["login"]?.toString();
+
+    if (!login) {
+      return null;
+    }
+
+    return {
+      login: login,
+      type: sender?.["type"]?.toString(),
+    };
+  }
+
+  public static getAction(payload: JSONObject): string | null {
+    return payload["action"]?.toString() || null;
+  }
+
+  /*
+   * On `issue_comment`, GitHub sends the SAME event for issues and for pull
+   * requests — a pull request is an issue with a `pull_request` key. This is
+   * the only reliable way to tell them apart, and getting it wrong means
+   * trying to revise an issue or implement a pull request.
+   */
+  public static isPullRequestIssue(issue: JSONObject | undefined): boolean {
+    return Boolean(issue?.["pull_request"]);
+  }
+
+  public static getNumber(container: JSONObject | undefined): number | null {
+    const value: unknown = container?.["number"];
+
+    return typeof value === "number" ? value : null;
+  }
+
+  public static getLabelNames(issue: JSONObject | undefined): Array<string> {
+    const labels: JSONArray = (issue?.["labels"] as JSONArray) || [];
+
+    return labels
+      .map((label: JSONObject) => {
+        return ((label as JSONObject)?.["name"] as string) || "";
+      })
+      .filter((name: string) => {
+        return Boolean(name);
+      });
+  }
+
+  /*
+   * Case-insensitive because GitHub labels are case-preserving but
+   * case-insensitively unique, so "OneUptime" and "oneuptime" are the same
+   * label and a configuration that matched exactly would miss half of them.
+   */
+  public static hasLabel(data: {
+    labelNames: Array<string>;
+    labelName: string;
+  }): boolean {
+    const wanted: string = data.labelName.trim().toLowerCase();
+
+    if (!wanted) {
+      return false;
+    }
+
+    return data.labelNames.some((name: string) => {
+      return name.trim().toLowerCase() === wanted;
+    });
+  }
+}