create-epic-recap

SkillDocs & knowledge

Lets your agent write a progress or wrap-up summary of a Jira Epic from its child issues, PRs, and release notes.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the create-epic-recap skill

About this capability

Use when an engineer or manager asks to recap, summarize, or post an update on a Jira Epic — a progress update for an in-progress Epic (how far along it is, what's shipped so far, what's next) or a resolution recap for a finished one. Gathers child-issue progress, merged GitHub PRs, release notes, a

What this skill tells your AI

The instructions your AI receives, as published by datadog/datadog-agent in .agents/skills/create-epic-recap/SKILL.md and read by ahel’s review.

Generate a recap for the Jira Epic $ARGUMENTS, aggregating child-issue progress, merged GitHub PRs, and release notes. The recap adapts to the Epic's state: a progress update while it's in flight (how far along it is, what's shipped so far, what's next) or a resolution recap once it's done — see Determine the recap mode in Step 2. Show a preview and post it as a comment on the Epic only after explicit user approval. This lets an engineer communicate progress or resolution to PMs and stakeholders without losing flow (motivation: OTAGENT-1038).

Owning team: team/opentelemetry-agent (@DataDog/opentelemetry-agent)

Reference files (load as needed)

  • references/runtime-tooling.md — runtime detection, the Cursor vs Claude Code tool mapping, cloudId, large responses, JQL. Read this before Step 2.
  • references/pr-discovery.md — the full Step 4 algorithm: Phase A/B, tier classification + regex, drop rules, throttling, the {{pr_discovery_note}} variants, and the Claude Code capability gap. Read this before Step 4.

This skill runs in two runtimes with different Atlassian MCP servers (Cursor's mcp-atlassian and Claude Code's Atlassian Rovo). The key gap: Rovo has no dev-status endpoint, so Phase A1/Tier 0 is Cursor-only and cloudId is required on every Rovo call. Whenever a step says "call the Fetch issue / Search children / Post comment tool", look up the exact tool and params in references/runtime-tooling.md.

Prerequisites

If any check fails, stop and tell the user what to fix.

  1. Atlassian MCP server — connected and authenticated (Cursor: user-atlassian; Claude Code: Atlassian Rovo). Probe with a known issue fetch; if it fails, ask the user to authenticate/connect.
  2. GitHub CLI (gh) — installed and authenticated for the DataDog org. Run gh auth status; if no active account, ask the user to run gh auth login.

Example

Input: /create-epic-recap OTAGENT-304 --dry-run

Fetches Epic OTAGENT-304, finds its completed child issues, discovers merged PRs across all child keys (Jira Development panel on Cursor + GitHub search), reads release notes from the PR file lists, renders the recap, prints a preview, and — because of --dry-run — saves a draft without posting:

Saved draft to /tmp/OTAGENT-304-recap.md

Step 1: Parse arguments

  • EPIC-KEY (required, first positional): matches ^[A-Z][A-Z0-9_]+-\d+$, e.g. OTAGENT-820. If missing or malformed, stop and ask the user.
  • --dry-run (optional flag): render and preview only, never post.
  • --mode <resolution|progress> (optional): override the recap mode. When omitted, Step 2 auto-detects it from the Epic status. Use progress for an in-progress Epic (a status update on how far along it is) and resolution for a finished Epic.

Step 2: Fetch the Epic

Call the Fetch issue tool (see references/runtime-tooling.md) requesting fields summary, description, status, issuetype, labels, assignee, reporter and the Epic's comments (Cursor: comment_limit: 20; Claude Code: add "comment" to the fields array with responseContentFormat: "markdown"). See Reading comments in references/runtime-tooling.md.

Validate:

  • If the issue cannot be found, stop and inform the user.
  • Read the issue type from whichever shape the runtime returns — accept both issuetype.name (Rovo) and issue_type.name (some mcp-atlassian versions). If the resolved name is not Epic, stop and tell the user this skill only works on Epics (suggest /run-jira for non-Epics). Do not reject just because one of the two shapes is absent.

Determine the recap mode (epic_mode), used from here on to shape wording and sections:

  • If --mode was passed in Step 1, use it verbatim.
  • Otherwise auto-detect from the Epic's status.category (accept status.statusCategory.key too): category Doneresolution; anything else (indeterminate/In Progress, new/To Do) → progress.
  • resolution = the Epic is finished, produce a "Resolution recap". progress = the Epic is still in flight, produce a "Progress update" (how far along it is, what's shipped so far, what's next).

Read the Epic comments you fetched: skim the most recent ones for context that is not in the description or PRs — decisions, scope changes, blockers, and (especially in progress mode) status updates on how far along the work is. Capture this as epic_comment_context for {{summary}} and the progress narrative.

Keep summary, description, status, labels, epic_mode, and epic_comment_context in memory for rendering.

Step 3: Fetch child issues

Call the Search children tool with the Epic-children JQL (see references/runtime-tooling.md), fields: summary, status, issuetype, assignee, labels, limit 50. On Claude Code, also request customfield_10000 in this call so Step 4 Phase A2 counts come back for free. If the response spills to a file, parse with jq.

Collect each child's key, summary, status.name, and status.category (accept status.statusCategory.key too).

Classify children by status category into three buckets:

  • done_children — category Done (statuses like Done, Closed, Resolved).
  • in_progress_children — category indeterminate (In Progress, In Review, etc.).
  • todo_children — category new (To Do, Backlog, etc.).

Compute progress counts for rendering: done = len(done_children), total = <count of all children>, percent = round(100 * done / total) (guard against total == 0).

  • In resolution mode, PR discovery and the recap body are driven by done_children; unfinished items go into skipped_children and are only mentioned if asked (as before).
  • In progress mode, done_children still drive PR discovery (merged PRs), while in_progress_children and todo_children are surfaced in the Progress section as remaining work.

Read comments on the relevant child issues too — useful context often lives only in task comments, so don't skip them. For each relevant child (in resolution mode: done_children; in progress mode: prioritise in_progress_children, then done_children), call the Fetch issue tool individually with comment_limit / fields:["comment"] and skim the latest comments. Do not request the comment field in the bulk Search children call (it blows up the response — see Large responses / Reading comments in references/runtime-tooling.md). Bound the work: cap at ~10 issues and the latest ~10 comments each; capture anything material as child_comment_context.

An empty list of completed children is fine — some Epics are resolved by PRs that reference the Epic key directly. Continue with just <EPIC-KEY> as the search term.

Step 4: Find merged PRs

Read references/pr-discovery.md and follow it. In short:

  • Build the key list [EPIC-KEY, <completed child keys>]. PR discovery is merged-only: even in progress mode only done_children contribute keys — in_progress_children/todo_children are represented as remaining work in the Progress section, not searched for PRs.
  • Phase A (Cursor only): A1 reads Tier 0 PRs from the Jira Development panel; A2 reads merged-PR counts from customfield_10000 into jira_pr_counts for cross-validation. On Claude Code, skip A1 (no dev-status) and use A2 + Phase B only.
  • Phase B (both runtimes): gh search prs once per key, then classify each hit into Tier 1 (include) / Tier 2 (include) / Tier 3 (opt-in, surfaced in preview) / Tier 4 (cross-ref, drop).
  • Apply the revert/bot drop rules, dedup across phases, and record tier3_candidates and pr_shortfall.
  • If zero PRs are found, use the manual-URL / empty-section / cancel fallback from the reference.

Step 5: Fetch PR details

For each merged PR, fetch details (run in parallel when possible):

gh pr view <number> \
  --repo <owner>/<repo> \
  --json title,body,files,labels,mergedAt,baseRefName,author,mergeCommit

Collect:

  • title, body, mergedAt, baseRefName.
  • mergeCommit.oidstore as mergeSha; Step 6 needs it to read release-note files added by the PR that aren't on the base branch. If mergeCommit is null (rebase/squash merge), fall back to the last commit's oid: gh pr view <number> --repo <owner>/<repo> --json commits --jq '.commits[-1].oid'.
  • files[].path — used in Steps 6 and 7.
  • labels[].name — note team/opentelemetry, component/*, changelog/*, qa/*.

Step 6: Read release notes from the PRs

For each PR, filter files[].path for entries starting with releasenotes/notes/ (main Agent), releasenotes-dca/notes/ (Cluster Agent), or releasenotes-installscript/notes/ (Install script). PRs may live in datadog-agent or other Datadog repos using the same convention.

Fetch each matching path from the PR's baseRefName:

gh api "repos/<owner>/<repo>/contents/<path>?ref=<baseRefName>" --jq '.content' | base64 -d

If the file was added by the PR (not yet on base) or has since been removed, fall back to the merge commit via mergeSha:

gh api "repos/<owner>/<repo>/contents/<path>?ref=<mergeSha>" --jq '.content' | base64 -d

If mergeSha is unavailable, skip the file and note its release note could not be read — do not fail; Step 9's PR-body fallback covers it.

Parse each YAML note and collect the section name (features, enhancements, fixes, upgrade, deprecations, security, other, issues) and its prose. Keep the original wording — release notes are already customer-facing.

Empty release notes are common, not an error. Several teams (notably team/opentelemetry-agent, which routinely labels DDOT PRs changelog/no-changelog) ship user-visible behaviour without reno entries. If none are found, do not stop or warn — Step 9 derives What's new from PR titles/bodies. Record this so the preview can note _None of the linked PRs included release notes_.

Step 7: Classify the change

Build a signals object from PR file paths and release-note prose. Each field can have multiple values; omit it from the recap when no signal matches.

Signal path (file-path prefixes):

  • comp/otelcol/, comp/core/configsync/, cmd/otel-agent/, pkg/config/otel/agent-otel-ingest and/or ddot
  • pkg/opentelemetry-mapping-go/dd-exporter-contrib
  • Helm charts, chart/, Dockerfile.otel, images/otel-agent/standalone-ddot

Signal type (file-path prefixes; a change can hit several):

  • pkg/logs/, comp/logs/logs
  • pkg/metrics/, pkg/opentelemetry-mapping-go/otlp/metrics/, comp/metrics/metrics
  • pkg/trace/, cmd/trace-agent/traces
  • pkg/collector/corechecks/ebpf/, pkg/gpu/, pkg/security/, pkg/profiler/profiles/system

API & config changes — scan PR diffs and release-note content for paths like pkg/config/setup/config.go, pkg/config/**/*.yaml, comp/core/config/, cmd/*/subcommands/*/command.go, or prose with config/option/setting/API/endpoint/flag. If found, list the concrete config keys / API surfaces (from release notes when available, else the diff). Otherwise mark "None".

Repositories touched — distinct repository.nameWithOwner from Step 4, sorted alphabetically.

Step 8: Ask the user for the remaining sections

Use a single multi-question AskUserQuestion for the pieces that cannot be derived from code, each with a free-text option plus the canned answer:

  1. Performance impact — text; Not measured valid. Encourage benchmark numbers / load-test / regression-detector links.
  2. Agent footprint — text; No change valid. Encourage RSS / CPU / binary-size deltas with quality-gates dashboard links.
  3. Customer utilisation tracking — how PMs track adoption: dashboard URL, metric name, log query, telemetry event, or Not tracked yet.

Step 9: Render the recap

Read recap-template.md and substitute each {{placeholder}}:

PlaceholderSource
{{epic_key}}Step 1
{{epic_summary}}Step 2
{{recap_title}}Step 2 epic_mode: Resolution recap (resolution) or Progress update (progress).
{{summary}}Synthesised 1-2 sentences for PMs, informed by epic_comment_context. resolution: what shipped and that the Epic is done — prefer Epic summary + release-note headlines; if no release notes, combine the Epic summary with the most user-relevant PR titles. progress: where the work stands — what's shipped so far and what's next, leading with the progress count.
{{progress}}progress mode only (omit the section otherwise). From Step 3: a bold **<done> of <total> issues complete (<percent>%).** line, then a Remaining: bullet list of in_progress_children (label In progress) and todo_children (label To do) as [<KEY>](<url>) — <summary>. Fold in status notes from epic_comment_context / child_comment_context when they explain where things stand.
{{whats_new}}Bullet list of user-facing wins, in order of preference: (1) features/enhancements release-note prose; (2) fixes/upgrade/deprecations if user-visible; (3) fallback when release notes are empty: one bullet per PR from the title (strip the [OTAGENT-XXX] prefix, rewrite in user-facing language) + a one-sentence summary of the PR body's ### What does this PR do?. The fallback is the normal path for changelog/no-changelog teams. Drop internal refactors, behaviourless dep bumps, and test-only PRs.
{{signal_path}}Step 7 bullet list, or omit the section if empty
{{signal_type}}Step 7 bullet list, or omit the section if empty
{{api_config_changes}}Step 7 content, or omit if "None" and no relevant release notes
{{performance_impact}}Step 8 answer, or omit if Not measured AND no perf-related release notes
{{agent_footprint}}Step 8 answer, or omit if No change AND no footprint-relevant release notes
{{repositories_touched}}Step 7 list, bullet form
{{customer_tracking}}Step 8 answer, or omit if Not tracked yet
{{linked_prs}}Bullet list - [<repo>#<number>](<url>) — <title>, then indented release-note bullets ( - <section>: <one-line excerpt>). Group Tier 0 PRs first with a _(linked via Jira)_ annotation, then Tier 1/2 from GitHub search. On Claude Code there are no Tier 0 PRs — start with Tier 1/2.
{{pr_discovery_note}}One of the quiet/loud/empty variants — see PR discovery note in references/pr-discovery.md. Quiet whenever Tier 0 was unavailable (always on Rovo) and no shortfall; loud whenever pr_shortfall is non-empty; empty only when Tier 0 was available (Cursor) and pr_shortfall is empty.

Drop the HTML rendering-rules comment from the template before producing the final markdown. When omitting an optional section, remove its ## heading too — no empty headings.

Step 10: Preview and approval

Print the rendered markdown under ### Preview — <EPIC-KEY> recap.

Before the recap, print a one-line PR discovery summary (on Claude Code the Tier 0 count is always 0 — make clear discovery was GitHub-only):

> Found N PRs: X via Jira Development panel (Tier 0), Y via GitHub search (Tier 1/2). Z Tier 3 candidates skipped (see below).

If pr_shortfall is non-empty, print a warning block after the summary (the same shortfall is also rendered into the posted report via {{pr_discovery_note}}):

> ⚠️ OTAGENT-307: Jira says 4 linked PRs, found 2. Check Tier 3 candidates or the Jira Development panel.

After the recap, if tier3_candidates is non-empty, print a separate ### Skipped (Tier 3 — opt-in) block (shown to the user only, not posted to Jira):

### Skipped (Tier 3 — opt-in)

The following PRs mention the searched Jira keys in their body but without a closing keyword (`Resolves`/`Closes`/`Fixes`/`JIRA:`). They are excluded by default. Pick `Edit` and say "include #N, #M" to add them.

- [<repo>#<number>](<url>) — <title>
  - Searched key: <KEY>
  - Body context: «…<the 1-2 lines around the key match>…»

Then call AskUserQuestion with options:

  • Post — proceed to Step 11.
  • Edit — ask for free-text instructions (e.g. "shorten the summary", "drop the perf section", "include #N" / "include all Tier 3" to promote candidates, "add a note about backport"), apply, and loop back to the preview.
  • Cancel — go to Step 12.

If --dry-run was set, skip the question and jump to Step 12 with cancel semantics, printing a notice that the recap was not posted. The Tier 3 block is still printed in dry-run.

Step 11: Post the comment

Call the Post comment tool (see references/runtime-tooling.md) with the rendered markdown from Step 9 (including the attribution footer). Do not set visibility/commentVisibility — this is a regular comment.

POST-action verification: re-fetch the Epic (Cursor: comment_limit=5; Claude Code: fields: ["comment"]) and confirm the new comment is present (match the footer string Generated by create-epic-recap). If verification fails, surface the error and do not retry automatically.

On success, print:

Recap posted: https://datadoghq.atlassian.net/browse/<EPIC-KEY>

Step 12: Save and exit (when not posting)

When the user picks Cancel or --dry-run was specified:

  1. Write the final markdown to /tmp/<EPIC-KEY>-recap.md.
  2. Print the path, e.g. Saved draft to /tmp/OTAGENT-820-recap.md.
  3. Exit cleanly.

Errors and edge cases

  • Atlassian MCP auth failure — stop, do not try a different transport, ask the user to authenticate.
  • Claude Code cloudId errors — see references/runtime-tooling.md (use getAccessibleAtlassianResources to resolve the UUID).
  • jira_get_issue_development_info 500/empty — the dev-status endpoint is fragile (no parallel calls, occasional downtime, exact CamelCase "GitHub"). On failure, A2 + Phase B carry the pipeline; on Rovo this tool doesn't exist at all. Full handling in references/pr-discovery.md.
  • gh not authenticatedgh auth status; if it fails, ask the user to gh auth login.
  • gh search prs rate-limited — back off 60 s, retry once; if still failing, ask for manual PR URLs (Step 4 fallback).
  • Very large PR set (> 25 PRs) — present a numbered list and ask via AskUserQuestion whether to include all or narrow by date / label / repo.
  • Posting fails — keep the markdown on disk (Step 12 path), report the error verbatim, do not silently retry.

Important constraints

  • Never post without explicit user approval--dry-run and Cancel must result in no Jira write.
  • Never modify the Epic description or other fields — comments only.
  • Always include the attribution footer so future readers know the recap was AI-generated.
  • Never include secrets, internal-only URLs, or sensitive customer data. Mask customer names as <customer> and flag during preview.
  • Comment body is markdown — pass it directly; do not pre-render to ADF/Wiki. On Cursor jira_add_comment takes Markdown in body; on Claude Code pass commentBody with contentFormat: "markdown".

Signals

GitHub stars
4k
Forks
1k
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
create-epic-recap
Source
github.com/datadog/datadog-agent