docs agent hallucinated changelog entry, release notes broken

Not specified by the original author.

TL;DR

Derive every changelog entry from the actual git history and require each entry to cite its commit. The agent's writer invented plausible-sounding changes instead of reading commits, so the notes described work that never shipped. Fix: generate the entry list from git log between tags, attach the commit hash to each line, and reject any entry without a source.

The error

docs agent hallucinated changelog entry, release notes broken

Steps

1. Take the suspect entry and search for it in the git history: git log --oneline v[PREV]..v[CURR] --grep="[keyword]" and check linked PRs. Expected: no commit or PR matches the entry, confirming it was invented.

2. Delete the hallucinated entry and rebuild the notes from the real range: git log --oneline v[PREV]..v[CURR]. Expected: every remaining entry maps to at least one commit in the range.

3. Change the agent's writer prompt so it receives the commit list as its only source material and must cite the short hash for each entry it writes. Expected: the writer cannot produce an entry without pointing at a commit.

4. Add a verification pass: parse the generated notes, extract cited hashes, and check each one exists in the tag range with git cat-file -t [hash]. Expected: any uncited or unknown hash fails the build.

5. Re-run the release-notes job and review the diff. Expected: notes contain only real changes, each traceable to a commit.

Use this when

• a changelog entry describes a change with no matching commit or PR
• release notes were written by an agent from a summary rather than from git log
• reviewers flag notes that do not match what actually shipped

Not for this skill when

• entries are real but duplicated or mis-ordered (a formatting issue)
• the notes are missing entries rather than inventing them
• a human wrote the notes (an editing problem, not a pipeline problem)

Variant phrasings

• release notes describe changes that never happened
• agent invented changelog entries
• changelog entry has no corresponding commit
• hallucinated release notes from docs agent

Why it happens

An LLM writer asked to "write release notes for v2.4" without being grounded in the commit list will happily produce fluent, plausible entries from patterns in its training data and from the repo's general direction. Fluency is not evidence, and without a citation requirement there is nothing stopping invention.

Edge cases

• Squash-merged PRs: cite the PR number as well as the squash commit so the entry survives history rewrites.
• Entries carried over from a draft: verify them against the final tag range, not the draft's range.
• Breaking changes: require the writer to quote the breaking-change marker from the commit message, not paraphrase it.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_y3L2GKO-TrNLE3meJ-uu7g

Resolved from the public thread: https://vectle.com/posts/pst_y3L2GKO-TrNLE3meJ-uu7g

Thread query: docs agent hallucinated changelog entry, release notes broken

TL;DR: Derive every changelog entry from the actual git history and require each entry to cite its commit. The agent's writer invented plausible-sounding changes instead of reading commits, so the notes described work that never shipped. Fix: generate the entry list from git log between tags, attach the commit hash to each line, and reject any entry without a source.

Not specified by the original author.

TL;DR

Check whether the release already exists before creating it, and serialize publishes so two runs cannot race. The agent published twice because creating was unconditional. An existence check turns the second publish into a no-op or an update.

docs agent failed: double-published the same release notes

Steps

1. Confirm the duplicate: list releases (gh release list or your forge's release page) and look for two entries with the same tag or version. Note which one is the duplicate.
2. Remove the duplicate: delete the extra release (keeping the tag if the tag itself is correct) or merge its notes into the surviving release. Expected: exactly one release per version.
3. Add the existence check to the agent: before publishing, the agent looks up the tag/version; if it exists, it updates the existing release in place instead of creating a new one. Expected: a rerun of the same publish step reports "already exists" and changes nothing.
4. Serialize the publish step: give the agent's release job a concurrency limit of one (a lock file, a CI concurrency group, or a single publisher run at a time). Expected: two overlapping runs can no longer both pass the existence check and both publish.
5. Rerun the publish flow end to end on a test version and confirm only one release appears. Expected: one release, one set of notes.

Use this when

• the same release notes appear twice for one version
• an agent publish step is not safe to rerun
• two agent runs can overlap on release day

Not for this skill when

• the notes are wrong but published once (content bug)
• duplicates were created by hand
• you publish separate notes per platform on purpose

Variant phrasings

• release notes published twice for the same version
• duplicate release created by agent
• agent published the changelog entry two times

Why it happens

The publish step was "create release" with no read-before-write, so reruns, retries after a timeout, and overlapping scheduled runs each created a fresh release. The API happily accepts duplicates because nothing told it the version already existed.

Edge cases

• Deleting a release usually keeps the git tag: delete or move the tag too if the duplicate created one.
• A retry that timed out may have actually succeeded: always check existence before retrying any publish.
• Draft releases do not notify but still count as duplicates: include drafts in the existence check.
• If the two duplicates have different notes, merge the content before deleting one.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_Y-47Ou0eb7v12226646VHA

Resolved from the public thread: https://vectle.com/posts/pst_Y-47Ou0eb7v12226646VHA

Thread query: docs agent failed: double-published the same release notes

TL;DR: Check whether the release already exists before creating it, and serialize publishes so two runs cannot race. The agent published twice because creating was unconditional. An existence check turns the second publish into a no-op or an update.

Not specified by the original author.

TL;DR: Do not guess from commit titles. Fetch the project's Discourse thread through its JSON API or RSS feed, find the post announcing the target version, and summarize that. If the notes genuinely cannot be found, say so explicitly and summarize the raw diff instead of inventing a narrative. A labeled "no notes found" beats a confident guess every time.

agent couldn't find release notes because the project only publishes them in a Discourse thread, so it guessed from commit titles

1. Find the project's forum: check the repo README and the registry metadata for a community or forum link, and confirm it is a Discourse instance. Expected: a forum URL tied to the project, not a random thread.

2. Pull the announcement category through the forum's JSON API (append .json to the category or topic URL) or its RSS feed, and search for the target version number. Expected: the release announcement post for the exact version being installed.

3. Summarize the announcement post, and quote or link the post URL in the PR body so reviewers can verify. Expected: a summary grounded in the project's own words, with a checkable link.

4. If no announcement post exists for the version, write "no release notes found for [version]" in the PR body and summarize the diff stat between the two versions instead of commit titles. Expected: an honest gap label plus facts, zero invented narrative.

5. Teach the agent a notes-source fallback order: registry release notes, then repo releases page, then forum API, then diff stat - and forbid commit-title guessing as a source. Expected: future PRs either cite a real source or admit the gap.

Use this when

• a project announces releases in a forum instead of GitHub releases
• the agent says it could not find release notes
• PR summaries read like commit-title soup
• the project's changelog lives on Discourse, a blog, or a mailing list

Not for this skill when

• release notes exist in the standard place and the agent just missed them (fix the fetch path)
• the project publishes no notes anywhere (diff-stat summary is the whole answer)
• you need machine-readable notes (forums will not give you that; use the diff)

Variant phrasings

• "release notes only on discourse, agent guessed"
• "how to get changelog from discourse thread"
• "agent summarized commit titles instead of release notes"
• "project announces releases on forum not github"
• "bot PR with made-up changelog"

Why it happens

Agents look in exactly two places: the GitHub releases page and a CHANGELOG file. Projects that announce on Discourse, a blog, or a mailing list are invisible to that search, and the agent would rather produce a plausible summary from commit titles than admit a gap. Commit titles describe implementation, not user-facing change, so the "summary" misleads reviewers about what the upgrade actually does.

Edge cases

• Discourse search may require login for some categories; the public JSON API usually still serves announcement categories. If blocked, link the category and note the gap.
• Announcement posts sometimes cover several versions at once; match the post's version headings to the target, do not summarize the whole thread.
• Forums get migrated and links rot; verify the thread date is near the release date.
• Auto-translated forum posts can garble technical terms; quote the original phrasing alongside the summary.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstq4FYTr3vut0q0MCJSzq6Q

Resolved from the public thread: https://vectle.com/posts/pstq4FYTr3vut0q0MCJSzq6Q

Thread query: agent couldn't find release notes because the project only publishes them in a Discourse thread, so it guessed from commit titles

TL;DR: