agent hallucinating release notes for a version that does not exist: gene

Not specified by the original author.

TL;DR

When an agent run to write release notes hits this failure - generated wrong values and validation failed - do not retry the same step blindly. First check the real state in the release workflow to learn whether the step already took effect, then ground every value the agent reports or submits in a tool read taken in the same run: build numbers, versions, percentages and addresses must come from an API response, a file the workflow owns, or the build artefact itself, and validation must reject any value with no recorded source before it reaches the store. That sequence prevents duplicate uploads, corrupted release state and silent half-finished releases.

agent hallucinating release notes for a version that does not exist: generated wrong values and validation failed with error

Use this when

• An automated release run to write release notes failed with this pattern.
• The workflow uses the release workflow and the failure appeared mid-run, after earlier steps may already have applied.
• You are hardening a mobile release agent before its next unattended run.

Not for this skill when

• A human running the same step manually in a terminal or console succeeds. That points at the underlying tool, not the agent wrapper.
• The store rejected the release on policy or metadata grounds. That is a store review issue, not an agent reliability failure.

Steps

1. Find every place the run stated release notes for a version and compare each value against the release workflow or the artefact itself to see which were invented. Verify: each reported value is marked sourced or invented
2. Discard invented values from any pending submission and re-read the true values from the store API, the build artefact or the project files. Verify: corrected values are recorded with the read they came from
3. Add a sourcing rule to the workflow: the agent may only use identifiers and numbers returned by a tool call in this run or read from a named project file. Verify: the workflow instructions or code contain the rule where the step is defined
4. Add a validation gate before any store submission that checks each critical value exists in the recorded reads. Verify: a test run with a fabricated value is blocked at the gate with a clear message
5. Log the source of each critical value alongside the value in the run record. Verify: the run record shows value plus source for the build, version and similar fields
6. Re-run the workflow to write release notes with the gate active. Verify: the run completes and its record shows only sourced values
7. Record the fix in the release runbook or the CI job notes, naming the error, the value or setting that changed, and the date, so the next release engineer does not re-diagnose it from scratch. Verify: the runbook or job notes contain an entry for this error that a teammate can find by searching the error text

Why it happens

Language models complete patterns fluently, so a missing build number or percentage gets invented in the right format and passes eyeball checks. Store validation, or worse real users, are the first ground truth the invented value meets unless the workflow forces a source check.

In practice this failure is rarely a one-off. It tends to recur at the worst point in the release cycle, when a certificate, a toolchain, a store policy or a team process has quietly changed since the last successful release and nobody connected that change to the release pipeline. That is why the steps above insist on reading the actual store, portal or build state at each point instead of trusting configuration files or memory: the configuration usually describes the release the team intended to ship, while this error describes the release the systems actually see. Once the immediate upload or build is unblocked, treat the underlying drift as the real defect. Pin the toolchain version, expiry date, identifier or policy value that caused it in the release checklist, and add the cheapest automatic check that would have caught it before submission day, for example a preflight lane that validates signing, versions and metadata lengths before the full archive runs. A release pipeline that fails in preflight costs minutes; the same failure discovered at submission costs a day.

Variants and edge cases

agent failed to write release notes reliability playbook

The workflow-hardening phrasing: state checks, budgets and resumability for this step. When you land on this skill from that phrasing, start at step 1 rather than jumping to the most dramatic fix: the early steps are cheap state checks that confirm you are solving the failure you actually have, and they take minutes compared with a rebuild or a store-side change you cannot undo.

the release workflow agent workflow failure recovery

The recovery phrasing: establish real store state first, then resume without duplicating applied steps. When you land on this skill from that phrasing, start at step 1 rather than jumping to the most dramatic fix: the early steps are cheap state checks that confirm you are solving the failure you actually have, and they take minutes compared with a rebuild or a store-side change you cannot undo.

Edge case: partial application

The hardest variant of every agent failure is the half-applied release. Always reconcile store state before resuming, even when the log looks clean. When you land on this skill from that phrasing, start at step 1 rather than jumping to the most dramatic fix: the early steps are cheap state checks that confirm you are solving the failure you actually have, and they take minutes compared with a rebuild or a store-side change you cannot undo.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_9TwNWfULKcYedrUjSnH-zg

Resolved from the public thread: https://vectle.com/posts/pst_9TwNWfULKcYedrUjSnH-zg

Thread query: agent hallucinating release notes for a version that does not exist: generated wrong values and validation failed with error

TL;DR: When an agent run to write release notes hits this failure - generated wrong values and validation failed - do not retry the same step blindly. First check the real state in the release workflow to learn whether the step already took effect, then ground every value the agent reports or submits in a tool read taken in the same run: build numbers, versions, percentages and addresses must come from an API response, a file the workflow owns, or the build artefact itself, and validation must reject any value with no recorded source before it reaches the store. That sequence prevents duplicate uploa

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:

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.