Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

fix(release): enrich AI release notes + render changelog in release flow (#4035)

LearningCircuit074285a26d7f
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 16b691358a..f3eafefece 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -379,14 +379,52 @@ jobs:
with:
egress-policy: audit
- - name: Sparse checkout release notes
+ - name: Sparse checkout release sources
+ # changelog.d/ + pyproject.toml are needed by the "Render changelog"
+ # step below (towncrier reads its [tool.towncrier] config from
+ # pyproject and consumes fragments from changelog.d/). The --version
+ # CLI flag bypasses towncrier's package-import path, so we don't need
+ # src/ or pdm.lock here.
uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
with:
- sparse-checkout: docs/release_notes
+ sparse-checkout: |
+ docs/release_notes/
+ changelog.d/
+ pyproject.toml
sparse-checkout-cone-mode: false
fetch-depth: 1
persist-credentials: false
+ - name: Set up Python
+ uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+ with:
+ python-version: '3.12'
+
+ - name: Install towncrier
+ run: pip install 'towncrier~=24.8'
+
+ - name: Render changelog from fragments
+ # Produces docs/release_notes/${RELEASE_VERSION}.md from changelog.d/
+ # fragments using the towncrier config in pyproject.toml. Towncrier
+ # errors on an empty fragment directory, so guard with a glob check —
+ # a maintenance release with no fragments should not fail the job.
+ # README.md is excluded because changelog.d/README.md is the
+ # contributor guide, not a fragment.
+ # Fragments are consumed (deleted) locally on the runner; the
+ # cleanup-changelog job re-runs this and opens a PR to persist the
+ # changes back to main.
+ env:
+ RELEASE_VERSION: ${{ needs.build.outputs.version }}
+ run: |
+ set -euo pipefail
+ fragment_count=$(find changelog.d -maxdepth 1 -name '*.md' ! -name 'README.md' | wc -l)
+ if [[ "$fragment_count" -gt 0 ]]; then
+ towncrier build --yes --version "$RELEASE_VERSION"
+ echo "Rendered docs/release_notes/${RELEASE_VERSION}.md from ${fragment_count} fragment(s)"
+ else
+ echo "No changelog fragments found — skipping towncrier build"
+ fi
+
- name: Download build artifacts
uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
with:
@@ -409,10 +447,12 @@ jobs:
# workflows could see partial notes).
#
# Body layout (each section is omitted if its source is empty):
- # 1. AI-generated TL;DR (from OpenRouter, model = vars.AI_MODEL)
+ # 1. AI-generated narrative (model = vars.AI_MODEL). Fed the
+ # hand-written notes, the auto PR list, every PR's title+body,
+ # and the actual diff vs the previous release.
# 2. Hand-written narrative — docs/release_notes/<version>.md,
- # rendered from changelog.d/ fragments by
- # `pdm run towncrier build` at release prep time.
+ # rendered earlier in this job by towncrier from
+ # changelog.d/ fragments.
# 3. Horizontal rule
# 4. Auto-generated, label-categorized PR list (already starts
# with "## What's Changed", emitted by GitHub's generate-notes
@@ -438,7 +478,12 @@ jobs:
NOTES_FILE="$(mktemp)"
PROMPT_FILE="$(mktemp)"
AI_RESPONSE_FILE="$(mktemp)"
- trap 'rm -f "$NOTES_FILE" "$PROMPT_FILE" "$AI_RESPONSE_FILE"' EXIT
+ PR_BODIES_FILE="$(mktemp)"
+ DIFF_FILE="$(mktemp)"
+ COMPARE_JSON="$(mktemp)"
+ # ${DIFF_FILE}.clean is the staging path for the NUL-strip step;
+ # rm -f tolerates the case where it doesn't exist.
+ trap 'rm -f "$NOTES_FILE" "$PROMPT_FILE" "$AI_RESPONSE_FILE" "$PR_BODIES_FILE" "$DIFF_FILE" "${DIFF_FILE}.clean" "$COMPARE_JSON"' EXIT
# `set -e` does NOT exit on a failing command substitution in an
# assignment, so wrap with explicit failure handling.
@@ -451,25 +496,147 @@ jobs:
fi
# The hand-written narrative is the per-version file produced by
- # `pdm run towncrier build` from changelog.d/ fragments. Empty if
- # the maintainer forgot to render before bumping __version__.py —
- # release proceeds with auto-notes only.
+ # the "Render changelog from fragments" step above. Empty if there
+ # were no fragments in changelog.d/ for this release (towncrier is
+ # skipped in that case) — release then proceeds with auto-notes
+ # plus whatever the AI produces from PR bodies + diff.
HAND_NOTES=""
if [[ -f "$RELEASE_NOTES_FILE" ]]; then
HAND_NOTES=$(cat "$RELEASE_NOTES_FILE")
else
- echo "WARNING: $RELEASE_NOTES_FILE not found — release proceeds with auto-notes only"
+ echo "WARNING: $RELEASE_NOTES_FILE not found — release proceeds without hand-written notes"
fi
- # ---- AI TL;DR (best-effort) -----------------------------------------
+ # ---- PR descriptions (batch via GraphQL, one round-trip) ----
+ # Extract PR numbers from the auto-body's pull/<num> URLs. Dedupe
+ # in case the same PR is linked twice. GraphQL aliases let us fetch
+ # all PR titles + bodies in a single request, sidestepping the
+ # rate-limit cost of N sequential REST calls.
+ PR_NUMBERS=$(printf '%s' "$AUTO_BODY" | grep -oE 'pull/[0-9]+' | grep -oE '[0-9]+' | sort -u || true)
+ : > "$PR_BODIES_FILE"
+ if [[ -n "$PR_NUMBERS" ]]; then
+ # GraphQL has a node/complexity ceiling. One aliased
+ # pullRequest(number:) lookup per PR with three shallow fields
+ # is cheap, but unbounded fan-out can still trip the limit on a
+ # huge release. Cap at 200 PRs per query — sized for LDR
+ # releases that bundle dependency-bump traffic (v1.6.10 had
+ # 144 unique PRs). The 750k char overall prompt cap protects
+ # against the per-PR body fan-out blowing the context window.
+ pr_count=$(printf '%s\n' "$PR_NUMBERS" | grep -c . || true)
+ if [[ "$pr_count" -gt 200 ]]; then
+ echo "WARNING: $pr_count PRs in release — capping to first 200 for GraphQL fetch"
+ PR_NUMBERS=$(printf '%s\n' "$PR_NUMBERS" | head -200)
+ fi
+ OWNER="${GITHUB_REPOSITORY%/*}"
+ REPO="${GITHUB_REPOSITORY#*/}"
+ QUERY="query { repository(owner: \"$OWNER\", name: \"$REPO\") {"
+ i=0
+ while IFS= read -r num; do
+ [[ -z "$num" ]] && continue
+ QUERY+=" pr${i}: pullRequest(number: ${num}) { number title body }"
+ i=$((i + 1))
+ done <<< "$PR_NUMBERS"
+ QUERY+=" } }"
+
+ if gh api graphql -f query="$QUERY" \
+ --jq '.data.repository | to_entries[] | .value | select(. != null) | "### PR #\(.number): \(.title)\n\((.body // "") | .[0:2000])\n"' \
+ > "$PR_BODIES_FILE" 2>/dev/null; then
+ echo "Fetched $(printf '%s\n' "$PR_NUMBERS" | wc -l) PR descriptions ($(wc -c < "$PR_BODIES_FILE") chars)"
+ else
+ echo "WARNING: GraphQL PR fetch failed — proceeding without PR descriptions"
+ : > "$PR_BODIES_FILE"
+ fi
+ fi
+
+ # ---- Diff between previous release and current ----
+ # /releases/latest excludes drafts and prereleases. Returns 404 if
+ # there is no prior stable release — the AI then runs without a
+ # diff input. Filter noise (lockfiles, generated docs, SBOM, static
+ # assets, snapshot fixtures, binary patches) so the LLM sees signal.
+ # Hard cap diff size at 700,000 chars (~175k tokens) to leave
+ # headroom in Kimi K2 Thinking's 262k context window for the rest
+ # of the prompt + 64k output budget.
+ : > "$DIFF_FILE"
+ PREV_TAG=""
+ if PREV_TAG=$(gh api "/repos/${GITHUB_REPOSITORY}/releases/latest" --jq '.tag_name' 2>/dev/null) && [[ -n "$PREV_TAG" && "$PREV_TAG" != "null" ]]; then
+ # On a workflow re-run after the release already exists,
+ # /releases/latest returns the just-created tag. Fall back to
+ # the second-most-recent stable release so the diff is
+ # meaningful instead of empty.
+ if [[ "$PREV_TAG" == "$RELEASE_TAG" ]]; then
+ echo "NOTE: /releases/latest is the current tag (re-run scenario) — looking up previous"
+ # Pass RELEASE_TAG as a jq variable rather than splicing it
+ # into the program text. RELEASE_TAG is already validated as
+ # bare semver above, but --arg keeps shell quoting and jq
+ # quoting fully separated regardless.
+ RELEASES_JSON=$(gh api "/repos/${GITHUB_REPOSITORY}/releases?per_page=10" 2>/dev/null || printf '[]')
+ PREV_TAG=$(printf '%s' "$RELEASES_JSON" | jq -r --arg rel "$RELEASE_TAG" \
+ '[.[] | select(.draft == false and .prerelease == false and .tag_name != $rel)][0].tag_name // empty' \
+ 2>/dev/null || true)
+ if [[ -z "$PREV_TAG" ]]; then
+ echo "WARNING: could not find a prior stable release — skipping diff"
+ fi
+ fi
+ fi
+ if [[ -n "$PREV_TAG" ]]; then
+ echo "Fetching diff $PREV_TAG...$RELEASE_TAG"
+ if gh api "/repos/${GITHUB_REPOSITORY}/compare/${PREV_TAG}...${RELEASE_TAG}" > "$COMPARE_JSON" 2>/dev/null; then
+ # GitHub's compare API caps the .files[] array at 300 entries
+ # per response (it paginates commits, not files). Log when we
+ # hit that boundary so a future release's diff truncation can
+ # be diagnosed quickly.
+ compare_file_count=$(jq -r '.files | length' "$COMPARE_JSON" 2>/dev/null || echo 0)
+ if [[ "$compare_file_count" -ge 300 ]]; then
+ echo "WARNING: compare API returned $compare_file_count files — at the 300-file limit; some files may be omitted from the AI diff input"
+ fi
+ # Don't swallow jq stderr — if the filter fails (e.g. the
+ # compare API returned malformed JSON), we want the actual
+ # error in the runner log, not a silent empty diff.
+ if ! jq -r '.files[]
+ | select(.patch != null)
+ | select((.filename | test("(package-lock\\.json$|pdm\\.lock$|\\.min\\.(js|css)$|^docs/CONFIGURATION\\.md$|^docs/ci/workflow-status\\.md$|^sbom-|/__pycache__/|\\.pyc$|^docs/images/|^src/local_deep_research/web/static/|/(golden_master|snapshot)|\\.snap$)")) | not)
+ | "diff --git a/\(.filename) b/\(.filename)\n\(.patch)\n"' \
+ "$COMPARE_JSON" > "$DIFF_FILE"; then
+ echo "WARNING: jq failed processing compare API output — diff input will be empty"
+ : > "$DIFF_FILE"
+ fi
+ # Character-aware truncation. `head -c` is byte-oriented and
+ # can split multi-byte UTF-8 sequences (towncrier emits emoji
+ # headers 💥/🔒/✨/🐛 that appear in diffs of
+ # docs/release_notes/*.md). Splitting mid-emoji produces
+ # invalid UTF-8, which jq --rawfile then refuses to encode.
+ python3 - "$DIFF_FILE" 700000 'diff truncated at 700,000 chars' <<'PYEOF'
+ import sys
+ path, max_chars, marker = sys.argv[1], int(sys.argv[2]), sys.argv[3]
+ with open(path, encoding='utf-8', errors='replace') as f:
+ content = f.read()
+ if len(content) > max_chars:
+ content = content[:max_chars] + f"\n[... {marker} ...]\n"
+ with open(path, 'w', encoding='utf-8') as f:
+ f.write(content)
+ print(f"Truncated to {len(content)} chars", file=sys.stderr)
+ PYEOF
+ # NUL bytes break jq --rawfile JSON encoding; strip them.
+ tr -d '\000' < "$DIFF_FILE" > "${DIFF_FILE}.clean" && mv "${DIFF_FILE}.clean" "$DIFF_FILE"
+ echo "Diff size: $(wc -c < "$DIFF_FILE") bytes"
+ else
+ echo "WARNING: compare API call failed — proceeding without diff"
+ : > "$DIFF_FILE"
+ fi
+ else
+ echo "No previous stable release found — skipping diff"
+ fi
+
+ # ---- AI narrative (best-effort) -------------------------------------
AI_SUMMARY=""
if [[ -z "${OPENROUTER_API_KEY:-}" ]]; then
echo "OPENROUTER_API_KEY not set — skipping AI summary"
else
# Compose the prompt: instructions (from $SUMMARY_PROMPT env)
- # followed by the two inputs, separated by clearly-labelled
- # markers (avoiding `---` which is a Markdown HR and could
- # confuse the model).
+ # followed by labelled input sections. Order matters for
+ # truncation safety — diff is last because it's the largest
+ # variable input and gets trimmed first if the overall prompt
+ # exceeds the 750k char safety cap.
{
printf '%s\n' "$SUMMARY_PROMPT"
printf '\n----- HAND-WRITTEN NOTES -----\n'
@@ -479,14 +646,46 @@ jobs:
printf '(none)\n'
fi
printf '\n----- AUTO-GENERATED PR LIST -----\n%s\n' "$AUTO_BODY"
+ if [[ -s "$PR_BODIES_FILE" ]]; then
+ printf '\n----- PR DESCRIPTIONS -----\n'
+ cat "$PR_BODIES_FILE"
+ fi
+ if [[ -s "$DIFF_FILE" ]]; then
+ printf '\n----- DIFF (%s...%s) -----\n' "$PREV_TAG" "$RELEASE_TAG"
+ cat "$DIFF_FILE"
+ fi
} > "$PROMPT_FILE"
+ # Overall prompt safety cap. Kimi K2 Thinking has a 262k token
+ # context; leaving 64k for max_tokens output means ~198k for input,
+ # which at ~3 chars/token for code-heavy diff is ~600k chars. We
+ # cap at 750k chars to absorb the per-char/token variance — if
+ # PR descriptions are unusually long alongside a near-cap diff,
+ # the tail (diff) gets trimmed first since it sits last.
+ # Character-aware truncation (Python) — see diff section above.
+ python3 - "$PROMPT_FILE" 750000 'prompt truncated at 750,000 chars' <<'PYEOF'
+ import sys
+ path, max_chars, marker = sys.argv[1], int(sys.argv[2]), sys.argv[3]
+ with open(path, encoding='utf-8', errors='replace') as f:
+ content = f.read()
+ if len(content) > max_chars:
+ content = content[:max_chars] + f"\n[... {marker} ...]\n"
+ with open(path, 'w', encoding='utf-8') as f:
+ f.write(content)
+ print(f"Truncated to {len(content)} chars", file=sys.stderr)
+ PYEOF
+
+ # reasoning.max_tokens caps the thinking trace so the model
+ # cannot burn the whole max_tokens budget on reasoning and emit
+ # empty content. This was the root cause of the v1.6.10 release
+ # AI step failing with finish_reason: length and empty .content.
JSON_PAYLOAD=$(jq -n \
--arg model "$AI_MODEL" \
--rawfile content "$PROMPT_FILE" \
--argjson temperature "$AI_TEMPERATURE" \
--argjson max_tokens "$AI_MAX_TOKENS" \
- '{model:$model, messages:[{role:"user", content:$content}], temperature:$temperature, max_tokens:$max_tokens}')
+ --argjson reasoning_max_tokens "$AI_REASONING_MAX_TOKENS" \
+ '{model:$model, messages:[{role:"user", content:$content}], temperature:$temperature, max_tokens:$max_tokens, reasoning:{max_tokens:$reasoning_max_tokens}}')
# Up to 2 attempts (1 retry) with a short backoff. The model
# endpoint can hiccup on cold starts or transient rate-limits;
@@ -517,20 +716,25 @@ jobs:
done
if [[ "$HTTP_CODE" =~ ^2 ]]; then
- # Some thinking models on some providers route the final answer
- # into `reasoning_content` rather than `content`. Try `content`
- # first; fall back to `reasoning_content` if it's empty.
+ # OpenRouter normalizes Moonshot's thinking trace to .reasoning
+ # (not .reasoning_content — that key is Moonshot's native API
+ # shape, exposed only on direct calls). Try .content first;
+ # fall back to .reasoning_content then .reasoning. The triple
+ # fallback is defense-in-depth: provider normalization is in
+ # flux and either variant may surface in the future.
AI_SUMMARY=$(jq -r '.choices[0].message.content // empty' "$AI_RESPONSE_FILE" 2>/dev/null || true)
if [[ -z "$AI_SUMMARY" ]]; then
AI_SUMMARY=$(jq -r '.choices[0].message.reasoning_content // empty' "$AI_RESPONSE_FILE" 2>/dev/null || true)
- if [[ -n "$AI_SUMMARY" ]]; then
- echo "NOTE: empty .content; using .reasoning_content from $AI_MODEL"
- fi
+ [[ -n "$AI_SUMMARY" ]] && echo "NOTE: empty .content; using .reasoning_content from $AI_MODEL"
+ fi
+ if [[ -z "$AI_SUMMARY" ]]; then
+ AI_SUMMARY=$(jq -r '.choices[0].message.reasoning // empty' "$AI_RESPONSE_FILE" 2>/dev/null || true)
+ [[ -n "$AI_SUMMARY" ]] && echo "NOTE: empty .content/.reasoning_content; using .reasoning from $AI_MODEL"
fi
if [[ -n "$AI_SUMMARY" ]]; then
echo "AI summary generated (model=$AI_MODEL, ${#AI_SUMMARY} chars)"
else
- echo "WARNING: AI response 2xx but neither .content nor .reasoning_content — skipping summary"
+ echo "WARNING: AI response 2xx but no usable content field — skipping summary"
# Diagnostic: dump the response shape so the next failure
# can be debugged without rerunning. Bounded to keep noise
# out of logs.
@@ -543,7 +747,7 @@ jobs:
echo "WARNING: AI summary call failed (HTTP $HTTP_CODE) — skipping summary"
fi
fi
- # ---- end AI TL;DR ----------------------------------------------------
+ # ---- end AI narrative -----------------------------------------------
{
if [[ -n "$AI_SUMMARY" ]]; then
@@ -558,6 +762,23 @@ jobs:
printf '%s\n' "$AUTO_BODY"
} > "$NOTES_FILE"
+ # GitHub release body limit is 125,000 chars — the API rejects
+ # longer bodies with HTTP 422 and gh CLI does not pre-validate.
+ # Truncate from the tail (the auto PR list is the largest variable
+ # section in big releases) with a visible marker. Character-aware
+ # truncation (Python) — see diff section above.
+ python3 - "$NOTES_FILE" 124400 "[notes truncated to fit GitHub's 125,000-char body limit]" <<'PYEOF'
+ import sys
+ path, max_chars, marker = sys.argv[1], int(sys.argv[2]), sys.argv[3]
+ with open(path, encoding='utf-8', errors='replace') as f:
+ content = f.read()
+ if len(content) > max_chars:
+ content = content[:max_chars] + f"\n\n_{marker}_\n"
+ with open(path, 'w', encoding='utf-8') as f:
+ f.write(content)
+ print(f"Release body truncated to {len(content)} chars", file=sys.stderr)
+ PYEOF
+
# Includes SBOM files, Sigstore signatures, and SLSA provenance
gh release create "$RELEASE_TAG" \
--repo "$GITHUB_REPOSITORY" \
@@ -580,7 +801,18 @@ jobs:
OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
AI_MODEL: ${{ vars.AI_MODEL || 'moonshotai/kimi-k2-thinking' }}
AI_TEMPERATURE: ${{ vars.AI_RELEASE_SUMMARY_TEMPERATURE || '0.3' }}
- AI_MAX_TOKENS: ${{ vars.AI_RELEASE_SUMMARY_MAX_TOKENS || '4000' }}
+ # 64000 matches the AI code-reviewer's working budget. The previous
+ # 4000-token cap was the root cause of the v1.6.10 release AI failure:
+ # kimi-k2-thinking burned the whole budget on reasoning tokens (which
+ # count against max_tokens) and emitted finish_reason=length with an
+ # empty .content. Reasoning tokens are pricey ($2.50/M output rate),
+ # so reasoning.max_tokens is capped separately below.
+ AI_MAX_TOKENS: ${{ vars.AI_RELEASE_SUMMARY_MAX_TOKENS || '64000' }}
+ # Caps the thinking trace so the model cannot exhaust max_tokens
+ # before emitting visible content. 16000 is well above Kimi's typical
+ # 5k-15k reasoning footprint for summarization without giving it
+ # room to spiral on a large diff.
+ AI_REASONING_MAX_TOKENS: ${{ vars.AI_RELEASE_SUMMARY_REASONING_MAX_TOKENS || '16000' }}
# Curl --max-time for the OpenRouter call. Default 900s (15 min) —
# thinking models need substantial headroom on a multi-PR release
# prompt (the previous 120s default timed out on v1.6.8 with
@@ -588,32 +820,153 @@ jobs:
# if you want the AI step to fail faster on stuck calls.
AI_MAX_TIME: ${{ vars.AI_RELEASE_SUMMARY_MAX_TIME || '900' }}
SUMMARY_PROMPT: >
- You are summarizing a software release for the project's GitHub release page.
- Two inputs follow: the hand-written release notes from the repository (may be empty)
- and the auto-generated, label-categorized list of merged pull requests.
+ You are writing the top section of a GitHub release page for an open-source
+ project. The audience is end users of the software who want to know what
+ changed and whether this release matters to them.
+
+
+ Several inputs follow, each introduced by a header line of the form
+ `----- HAND-WRITTEN NOTES -----`, `----- AUTO-GENERATED PR LIST -----`,
+ `----- PR DESCRIPTIONS -----`, `----- DIFF (oldtag...newtag) -----`:
+
+ - HAND-WRITTEN NOTES: a maintainer-curated changelog rendered from per-PR
+ fragments. May be empty for maintenance releases. When present, these are
+ the most authoritative description of intent. They appear verbatim below
+ your text in the published release page — do not repeat them line-by-line.
+
+ - AUTO-GENERATED PR LIST: GitHub's label-categorized list of merged PRs for
+ this release. Useful for completeness; weak signal for intent. This list
+ also appears verbatim below your text — do not repeat it.
+
+ - PR DESCRIPTIONS: the body text of each merged PR. Often explains the why
+ behind a change better than the title.
+ - DIFF: the actual code change between the previous release and this one.
+ May be truncated. Use to disambiguate what a PR actually did or to catch
+ user-visible changes not well described in PR text.
- Write the section a user would skim first to decide whether this release matters to them.
- Lead with the most user-facing change. If there are breaking changes, surface them next
- with a one-line "what to do" for each. Then notable new features and meaningful fixes.
- Skip dependency bumps, internal CI, refactors, and docs unless they're load-bearing for users.
+ Write a helpful narrative that goes at the top of the release page. Lead
+ with the most user-visible changes. Flag breaking changes early with a brief
+ "what to do" if applicable. Group related changes when natural. Skip
+ dependency bumps, internal CI, refactors, and docs-only changes unless
+ they're load-bearing for users.
- Output plain GitHub-flavored markdown — no preamble, no closing meta-commentary.
- Open with the literal heading `## TL;DR`. Use `###` for any subsections inside it.
- Do not emit `#` (the release title above the body provides the H1) or anything
- deeper than `###`. Reference PRs only when they exist in the inputs (`#1234`
- style); never invent numbers, features, or behavior. If the hand-written notes
- contradict the PR list, trust the hand-written notes — they are the authoritative
- human narrative.
+ Size the writeup to the material. A small maintenance release might get one
+ short paragraph; a substantial release with new features and breaking
+ changes might warrant several paragraphs with subsections. Don't pad. Don't
+ open with a meta-preamble like "this release contains X PRs across Y
+ categories" — just describe what happened.
- If both inputs are sparse, keep the summary to one or two sentences acknowledging
- this is a small/maintenance release.
+ Output GitHub-flavored markdown. Open with a `##` heading naming the theme
+ of the release (e.g. `## Friendlier OpenAI-compatible error messages and
+ multi-migration upgrade fixes`, or for small releases simply `## What's
+ new`). Use `###` for any subsections. Do not emit `#` (the release title
+ above is the H1). Reference PRs as `#1234` only when their numbers appear
+ in the inputs — never invent PR numbers, features, or behavior. If the
+ hand-written notes contradict the PR list or diff, trust the hand-written
+ notes; they are the authoritative human narrative.
- Keep it dense: a short opening paragraph plus bullets, or 2–3 short paragraphs —
- whichever fits the material better. Aim for something a user can read in 30 seconds.
+
+ No preamble, no closing meta-commentary. Just the section.
+
+ # Persists the towncrier render back to main. The create-release job runs
+ # `towncrier build` in a throwaway sparse checkout to produce the rendered
+ # release notes for the GitHub release body, but those changes (consumed
+ # fragments + new docs/release_notes/<version>.md) never reach the repo
+ # from there. Without this job, the next release would re-render all the
+ # same fragments. Branch protection on main has 0 required reviews and
+ # 0 required status checks, so the maintainer can squash-merge this PR
+ # without ceremony.
+ cleanup-changelog:
+ needs: [build, create-release]
+ if: ${{ !cancelled() && needs.create-release.result == 'success' && needs.build.outputs.release_exists == 'false' }}
+ runs-on: ubuntu-latest
+ permissions:
+ contents: write
+ pull-requests: write
+
+ steps:
+ - name: Harden the runner (Audit all outbound calls)
+ uses: step-security/harden-runner@a5ad31d6a139d249332a2605b85202e8c0b78450 # v2.19.1
+ with:
+ egress-policy: audit
+
+ - name: Checkout the release commit
+ # Pinning to ${{ github.sha }} (not `ref: main`) avoids a race:
+ # if a new PR with new fragments merges into main between
+ # create-release and cleanup-changelog, `ref: main` would consume
+ # those new fragments into THIS release's docs/release_notes file
+ # and delete them prematurely — misattributing them and stealing
+ # them from the next release. github.sha is the commit that
+ # triggered the workflow run (same commit create-release rendered
+ # against), so the set of fragments is identical.
+ uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+ with:
+ ref: ${{ github.sha }}
+ fetch-depth: 1
+
+ - name: Set up Python
+ uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
+ with:
+ python-version: '3.12'
+
+ - name: Install towncrier
+ run: pip install 'towncrier~=24.8'
+
+ - name: Render and consume fragments
+ env:
+ RELEASE_VERSION: ${{ needs.build.outputs.version }}
+ run: |
+ set -euo pipefail
+ if [[ ! "$RELEASE_VERSION" =~ ^[0-9]+\.[0-9]+\.[0-9]+([-+][a-zA-Z0-9.+_-]+)?$ ]]; then
+ echo "ERROR: RELEASE_VERSION '$RELEASE_VERSION' is not a valid bare semver string" >&2
+ exit 1
+ fi
+ # Skip cleanly when there are no fragments — nothing to clean up,
+ # no PR needed. README.md is preserved per the contributor guide
+ # in changelog.d/README.md.
+ if [[ "$(find changelog.d -maxdepth 1 -name '*.md' ! -name 'README.md' | wc -l)" -eq 0 ]]; then
+ echo "No changelog fragments to clean up — skipping PR"
+ echo "skip=true" >> "$GITHUB_ENV"
+ exit 0
+ fi
+ towncrier build --yes --version "$RELEASE_VERSION"
+ echo "skip=false" >> "$GITHUB_ENV"
+
+ - name: Open cleanup PR
+ if: env.skip != 'true'
+ uses: peter-evans/create-pull-request@5f6978faf089d4d20b00c7766989d076bb2fc7f1 # v8.1.1
+ with:
+ token: ${{ secrets.GITHUB_TOKEN }}
+ # Pin base explicitly: when this workflow is triggered by a tag
+ # push, github.sha may not sit on main HEAD, and the action's
+ # default-branch resolution could pick a non-main base. We always
+ # want this PR to target main.
+ base: main
+ commit-message: "chore: clear changelog fragments for ${{ needs.build.outputs.version }}"
+ branch: "chore/post-release-cleanup-${{ needs.build.outputs.version }}"
+ delete-branch: true
+ title: "chore: clear changelog fragments for ${{ needs.build.outputs.version }}"
+ body: |
+ ## Summary
+ Post-release housekeeping for **${{ needs.build.outputs.version }}**:
+
+ - Renders `changelog.d/` fragments into `docs/release_notes/${{ needs.build.outputs.version }}.md`
+ - Removes the consumed fragments
+
+ The same render was already used to compose the GitHub release body for
+ [${{ needs.build.outputs.tag }}](https://github.com/${{ github.repository }}/releases/tag/${{ needs.build.outputs.tag }}).
+ This PR exists only to persist the cleanup so the next release does not
+ re-render the same fragments.
+
+ Auto-merge-friendly: no review value beyond a sanity check that the
+ rendered notes look right.
+ labels: |
+ automation
+ maintenance
trigger-prerelease-docker:
needs: [build, provenance]
diff --git a/changelog.d/+3878-openai-compat-friendly-errors.feature.md b/changelog.d/+3878-openai-compat-friendly-errors.feature.md
deleted file mode 100644
index 9beb1a2f7d..0000000000
--- a/changelog.d/+3878-openai-compat-friendly-errors.feature.md
+++ /dev/null
@@ -1 +0,0 @@
-OpenAI-compatible runtime errors (LM Studio / vLLM / llama.cpp server / OpenRouter / custom endpoint) now surface a message that names the provider, configured base URL, and model. The existing `Error type: <code>` token convention is extended with seven `openai_*` codes (`openai_connection_refused`, `openai_timeout`, `openai_auth`, `openai_permission_denied`, `openai_model_not_found`, `openai_bad_request`, `openai_unknown`), and `ErrorReporter` maps each to the right `ErrorCategory`. The original exception text is preserved in a `Details:` suffix; userinfo embedded in a base URL (`https://u:key@host/v1`) is stripped before display.
diff --git a/changelog.d/+3905-ldr-disable-rate-limiting-alias.misc.md b/changelog.d/+3905-ldr-disable-rate-limiting-alias.misc.md
deleted file mode 100644
index 5de7fc2a9b..0000000000
--- a/changelog.d/+3905-ldr-disable-rate-limiting-alias.misc.md
+++ /dev/null
@@ -1 +0,0 @@
-Added `LDR_DISABLE_RATE_LIMITING` as the canonical name for disabling HTTP rate limiting, matching the `LDR_` env-var convention. The legacy `DISABLE_RATE_LIMITING` still works but emits a deprecation warning. Note: this is distinct from `LDR_RATE_LIMITING_ENABLED`, which controls the adaptive search-engine rate limiter — a different subsystem (#3905).
diff --git a/changelog.d/+3905-upload-rate-limits.feature.md b/changelog.d/+3905-upload-rate-limits.feature.md
deleted file mode 100644
index 97f6c3aa0a..0000000000
--- a/changelog.d/+3905-upload-rate-limits.feature.md
+++ /dev/null
@@ -1 +0,0 @@
-Upload rate limits are now configurable via `LDR_SECURITY_RATE_LIMIT_UPLOAD_USER` and `LDR_SECURITY_RATE_LIMIT_UPLOAD_IP`. The default cap is raised from `10 per minute;100 per hour` to `60 per minute;1000 per hour`, unblocking bulk RAG library uploads.
diff --git a/changelog.d/+3936-ldr-disable-rate-limiting-workflow-migration.misc.md b/changelog.d/+3936-ldr-disable-rate-limiting-workflow-migration.misc.md
deleted file mode 100644
index 01f155b97e..0000000000
--- a/changelog.d/+3936-ldr-disable-rate-limiting-workflow-migration.misc.md
+++ /dev/null
@@ -1 +0,0 @@
-Migrated CI workflows to the canonical `LDR_DISABLE_RATE_LIMITING` env var (added in #3936) and closed a test-isolation gap in `test_enabled_by_default` that did not clear the canonical var.
diff --git a/changelog.d/+context-overflow-drop-dead-charts.misc.md b/changelog.d/+context-overflow-drop-dead-charts.misc.md
deleted file mode 100644
index 5707dd9bcc..0000000000
--- a/changelog.d/+context-overflow-drop-dead-charts.misc.md
+++ /dev/null
@@ -1 +0,0 @@
-Removed unused chart-rendering JavaScript on the context-overflow analytics page (token-usage-over-time stacked bar, model-token-stats table, context-limits-distribution doughnut, and the configured-limits list). These functions stopped being called when the page was reorganised to focus on truncation diagnosis (#3792); deleting them now drops ~200 lines of dead code without any user-visible change.
diff --git a/changelog.d/+context-overflow-link-wires.feature.md b/changelog.d/+context-overflow-link-wires.feature.md
deleted file mode 100644
index e010efc286..0000000000
--- a/changelog.d/+context-overflow-link-wires.feature.md
+++ /dev/null
@@ -1 +0,0 @@
-Context-overflow warnings (front-page banner, result-page banner, completion toast) now show a clickable action that takes you straight to the relevant diagnostics — historical warnings link to the context-overflow metrics page, while per-research warnings link to that research's overflow section on its details page.
diff --git a/changelog.d/+context-overflow-log-fields.feature.md b/changelog.d/+context-overflow-log-fields.feature.md
deleted file mode 100644
index 23ad101548..0000000000
--- a/changelog.d/+context-overflow-log-fields.feature.md
+++ /dev/null
@@ -1 +0,0 @@
-Context-overflow warnings now include structured fields (`research_id`, `model`, `provider`) and detect both input-only and total-context (input+output) overflow. Four distinct tags (`[provider-confirmed]`, `[total-context]`, `[estimated]`, `[estimated-total-context]`) make different overflow causes distinguishable in logs and alerts. Hosted providers (OpenAI, Anthropic, OpenRouter) get estimation-based detection so their truncation events are no longer invisible. The high-context VRAM warning now links directly to the context-overflow metrics page.
diff --git a/changelog.d/+context-overflow-page-rework.feature.md b/changelog.d/+context-overflow-page-rework.feature.md
deleted file mode 100644
index e638c4db6a..0000000000
--- a/changelog.d/+context-overflow-page-rework.feature.md
+++ /dev/null
@@ -1 +0,0 @@
-The `/metrics/context-overflow` page is reorganised around per-research truncation diagnosis: redundant aggregate sections (Token Usage Overview, Token Usage Over Time, Model Usage Breakdown, Context Limit Distribution doughnut, Token Usage by Phase) are dropped in favour of empty-state messages and a recoloured and re-shaped scatter chart that buckets points by context-utilisation ratio (green circle / amber triangle / red diamond — shape gives a redundant signal for colourblind users) rather than per-limit dashed reference lines. Tooltips now explicitly label requests with unknown context limits or unknown token counts. The main `/metrics` dashboard gains a small Context Overflow summary panel linking to the full diagnostics page; on aggregation failure the panel shows `—` rather than a misleading `0%`.
diff --git a/changelog.d/+context-overflow-research-mode-filter.bugfix.md b/changelog.d/+context-overflow-research-mode-filter.bugfix.md
deleted file mode 100644
index 3dab84e592..0000000000
--- a/changelog.d/+context-overflow-research-mode-filter.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-The Context Overflow summary panel on `/metrics` now honors the dashboard's research-mode filter (Quick / Detailed / All) the same way the rest of the panels do.
diff --git a/changelog.d/+ipv6-transition-prefixes.security.md b/changelog.d/+ipv6-transition-prefixes.security.md
deleted file mode 100644
index 4222a19145..0000000000
--- a/changelog.d/+ipv6-transition-prefixes.security.md
+++ /dev/null
@@ -1,25 +0,0 @@
-SSRF defense-in-depth: block IPv6 transition prefixes that can wrap
-private IPv4 destinations on hosts with kernel sit0/NAT64 routes.
-
-- `2002::/16` (6to4, RFC 3056 — deprecated by RFC 7526)
-- `64:ff9b::/96` (NAT64 well-known, RFC 6052)
-- `64:ff9b:1::/48` (NAT64 local-use, RFC 8215 — same SSRF threat class
- as the WKP; missing it earned a HackerOne bounty against
- ssrf_filter)
-- `2001::/32` (Teredo, RFC 4380)
-- `100::/64` (IPv6 discard, RFC 6666)
-- `::/96` (IPv4-Compatible IPv6, RFC 4291 §2.5.5.1 — DEPRECATED 2006;
- same SSRF threat class as the transition prefixes)
-
-The metadata-IP block is hardened against IPv6-wrapped IMDS access:
-when an IPv6 destination falls in a NAT64 prefix, the embedded IPv4 is
-extracted and matched against `ALWAYS_BLOCKED_METADATA_IPS`, so
-`[64:ff9b::a9fe:a9fe]` cannot reach 169.254.169.254 even on a NAT64
-host.
-
-Operators on IPv6-only deployments using DNS64+NAT64 (where outbound
-IPv4 traffic is synthesized through `64:ff9b::/96`) can opt back in via
-the env-only setting `security.allow_nat64`
-(`LDR_SECURITY_ALLOW_NAT64=true`). The opt-in is scoped strictly to
-the two NAT64 prefixes — 6to4, Teredo, and discard remain blocked
-unconditionally, and the IMDS embedded-IPv4 carve-out still applies.
diff --git a/changelog.d/+prerelease-floating-tag.misc.md b/changelog.d/+prerelease-floating-tag.misc.md
deleted file mode 100644
index 3cbd336acd..0000000000
--- a/changelog.d/+prerelease-floating-tag.misc.md
+++ /dev/null
@@ -1 +0,0 @@
-The prerelease Docker workflow now also re-points the floating `prerelease` tag at each new RC manifest. Testers can pin `image: localdeepresearch/local-deep-research:prerelease` in their compose file and run `docker compose pull && docker compose up -d` to fetch the latest RC without editing the tag each cycle. Versioned tags (`prerelease-vX.Y.Z-<sha>`) are still published for reproducible testing.
diff --git a/changelog.d/+ssrf-metadata-block-log-hygiene.security.md b/changelog.d/+ssrf-metadata-block-log-hygiene.security.md
deleted file mode 100644
index 0f4abc64dd..0000000000
--- a/changelog.d/+ssrf-metadata-block-log-hygiene.security.md
+++ /dev/null
@@ -1,8 +0,0 @@
-Hardened SSRF defenses against AWS ECS task metadata
-(`169.254.170.2`, `169.254.170.23`), Tencent Cloud (`169.254.0.23`),
-and AlibabaCloud (`100.100.100.200`) metadata endpoints — these are
-now always blocked alongside the existing AWS IMDS / Azure / OCI /
-DigitalOcean entry (`169.254.169.254`). Redacted credentials, path,
-and query from URL-rejection logs (operators with grep/regex tooling
-on the rejection log lines will see authority-only `scheme://host:port`
-instead of full URLs going forward).
diff --git a/changelog.d/+ssrf-parser-differential.security.md b/changelog.d/+ssrf-parser-differential.security.md
deleted file mode 100644
index f54eb7025d..0000000000
--- a/changelog.d/+ssrf-parser-differential.security.md
+++ /dev/null
@@ -1,6 +0,0 @@
-Fix SSRF parser-differential bypass (GHSA-g23j-2vwm-5c25). URLs containing
-backslash, whitespace, or ASCII control bytes are now rejected upfront by the
-SSRF validator and notification-URL validator; hostname extraction switched
-from `urllib.parse.urlparse` to `urllib3.util.parse_url` so the validator and
-the HTTP client agree on destination by construction. Credit: @Fushuling,
-@RacerZ-fighting.
diff --git a/changelog.d/2606.bugfix.md b/changelog.d/2606.bugfix.md
deleted file mode 100644
index 4fc5918133..0000000000
--- a/changelog.d/2606.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-Fix progress page log panel sizing so it uses CSS layout instead of JavaScript height calculations.
diff --git a/changelog.d/2609.bugfix.md b/changelog.d/2609.bugfix.md
deleted file mode 100644
index fd6ad578fb..0000000000
--- a/changelog.d/2609.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-Progress page current-task text now preserves line breaks and wraps long generated task messages instead of collapsing them into a hard-to-read wall of text.
diff --git a/changelog.d/2610.bugfix.md b/changelog.d/2610.bugfix.md
deleted file mode 100644
index 8c03420622..0000000000
--- a/changelog.d/2610.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-Live research log entries now keep the same newest-first visual order as loaded log history instead of appearing at the bottom of the log panel.
diff --git a/changelog.d/3544.bugfix.md b/changelog.d/3544.bugfix.md
deleted file mode 100644
index c9fca194db..0000000000
--- a/changelog.d/3544.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-**Journal-quality dashboard ("Your Research") no longer shows orphan papers from deleted research sessions.** Deleting a research session previously left behind `Paper` rows whose `PaperAppearance` had cascade-deleted, so the dashboard kept counting them. The aggregate top-200 query and the predatory-blocked count now exclude papers that have no remaining appearances. Per-research views were already correct.
diff --git a/changelog.d/3670.breaking.1.md b/changelog.d/3670.breaking.1.md
deleted file mode 100644
index bab61922d0..0000000000
--- a/changelog.d/3670.breaking.1.md
+++ /dev/null
@@ -1 +0,0 @@
-**`llamacpp` provider migrated from in-process to HTTP.** No longer loads `.gguf` files in-process via `llama-cpp-python` — connects to `llama-server`'s OpenAI-compatible HTTP endpoint instead. Removed settings: `llm.llamacpp_model_path`, `llm.llamacpp_n_batch`, `llm.llamacpp_n_gpu_layers`, `llm.llamacpp_f16_kv`. New settings: `llm.llamacpp.url` (default `http://localhost:8080/v1`), `llm.llamacpp.api_key` (optional, for auth proxies). Migration: run `llama-server -m <your-model.gguf>` and the default URL picks it up; alternatively use the `openai_endpoint` provider pointed at the same URL.
diff --git a/changelog.d/3670.breaking.md b/changelog.d/3670.breaking.md
deleted file mode 100644
index 9aaa147258..0000000000
--- a/changelog.d/3670.breaking.md
+++ /dev/null
@@ -1 +0,0 @@
-**`llm.model` no longer auto-fills `gemma3:12b`.** Pre-1.7 installs silently downloaded a multi-GB Ollama binary on first use. The field is now empty by default — pick a model in **Settings → LLM**, or research will fail loudly with a clear `ValueError`. Existing users with a saved `llm.model` setting are unaffected; fresh installs see an inline warning in Settings (`aria-live="polite"`) and a clear error on research start. Affects all built-in providers (`openai`, `anthropic`, `ollama`, `lmstudio`, `llamacpp`, `google`, `openrouter`, `ionos`, `xai`).
diff --git a/changelog.d/3670.feature.md b/changelog.d/3670.feature.md
deleted file mode 100644
index 802ed78ae2..0000000000
--- a/changelog.d/3670.feature.md
+++ /dev/null
@@ -1 +0,0 @@
-**LM Studio optional API key.** New `llm.lmstudio.api_key` setting for LM Studio instances with authentication enabled.
diff --git a/changelog.d/3773.feature.md b/changelog.d/3773.feature.md
deleted file mode 100644
index 864e01eb19..0000000000
--- a/changelog.d/3773.feature.md
+++ /dev/null
@@ -1 +0,0 @@
-**Release notes now use [towncrier](https://towncrier.readthedocs.io) news fragments.** Each PR drops one tiny `changelog.d/<id>.<category>.md` file (categories: `breaking`, `security`, `feature`, `bugfix`, `removal`, `misc`). At release prep, the maintainer runs `pdm run towncrier build --version <X.Y.Z> --yes`, which renders the fragments into `docs/release_notes/<X.Y.Z>.md` and removes them. Replaces the previous shared-file workflow, which scaled poorly at LDR's PR throughput. See `changelog.d/README.md` for the convention.
diff --git a/changelog.d/3800.bugfix.md b/changelog.d/3800.bugfix.md
deleted file mode 100644
index 0b8db2db5b..0000000000
--- a/changelog.d/3800.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-Fix the model dropdown in Settings to populate for every auto-discovered LLM provider, not just Ollama / OpenAI / Anthropic / OpenAI-Compatible-Endpoint. LM Studio, Llama.cpp, Google Gemini, xAI, OpenRouter, and IONOS were silently dropped by the frontend (`processModelData`) even when the backend returned their models correctly.
diff --git a/changelog.d/3800.security.md b/changelog.d/3800.security.md
deleted file mode 100644
index b1710e9813..0000000000
--- a/changelog.d/3800.security.md
+++ /dev/null
@@ -1 +0,0 @@
-Fix LM Studio model detection (#3800) — the settings dropdown was silently dropping models for every auto-discovered provider except Ollama / OpenAI / Anthropic / OpenAI-Compatible-Endpoint, including LM Studio, Llama.cpp, Google Gemini, xAI, OpenRouter, and IONOS. While investigating, also fixed a separate credential-leak class where the route's auto-discovery loop sent every other provider's configured API key in the `Authorization` header to the local LM Studio / llama-server endpoint on every settings-page load.
diff --git a/changelog.d/3816.bugfix.md b/changelog.d/3816.bugfix.md
deleted file mode 100644
index 11324dd770..0000000000
--- a/changelog.d/3816.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-Fix file descriptor exhaustion (#3816) caused by `ChatOllama._async_client` not being released alongside `_client`. `_close_base_llm()` now closes both the sync and async httpx clients owned by `ChatOllama`, addressing the eventpoll-FD growth seen in long-running deployments.
diff --git a/changelog.d/3825.bugfix.md b/changelog.d/3825.bugfix.md
deleted file mode 100644
index effac224c8..0000000000
--- a/changelog.d/3825.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-Fix the research details "View Journals" button so it opens the scoped journal-quality dashboard instead of a missing page.
diff --git a/changelog.d/3827.bugfix.md b/changelog.d/3827.bugfix.md
deleted file mode 100644
index c9fa275a7c..0000000000
--- a/changelog.d/3827.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-Fix Download Manager failure after 20–30 PDFs with `UNIQUE constraint failed: documents.document_hash` and the cascading PendingRollbackError that prevented further downloads.
diff --git a/changelog.d/3849.bugfix.md b/changelog.d/3849.bugfix.md
deleted file mode 100644
index 7dbabfbce6..0000000000
--- a/changelog.d/3849.bugfix.md
+++ /dev/null
@@ -1,5 +0,0 @@
-Fix login/registration over `http://localhost:5001` on Docker Desktop.
-The session cookie's `Secure` flag is now driven purely by the request
-protocol (HTTPS) rather than by source-IP heuristics, so HTTP requests
-from any IP — including Docker Desktop's NAT gateway — now keep their
-cookies and CSRF tokens.
diff --git a/changelog.d/3870.bugfix.md b/changelog.d/3870.bugfix.md
deleted file mode 100644
index d62f04b357..0000000000
--- a/changelog.d/3870.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-Fix Ollama embedding indexing aborting with `"input length exceeds the context length"` (HTTP 500) by exposing a configurable `embeddings.ollama.num_ctx` setting. The new field appears on the embedding settings page when the Ollama provider is selected and ships with a default of 8192 tokens, which covers chunk sizes up to ~5000 characters for `nomic-embed-text`, `mxbai-embed-large`, `bge-m3`, and `qwen3-embedding`.
diff --git a/changelog.d/3926.bugfix.md b/changelog.d/3926.bugfix.md
deleted file mode 100644
index d0bcb66896..0000000000
--- a/changelog.d/3926.bugfix.md
+++ /dev/null
@@ -1 +0,0 @@
-LangGraph agent strategy error messages now include the exception type, and the `'str' object has no attribute 'model_dump'` failure (issue #3897) is translated into a rich on-result-page recommendation listing the three possible causes — proxy/shim, serving stack, or model — with concrete remediation steps for each.
diff --git a/changelog.d/4012.feature.md b/changelog.d/4012.feature.md
deleted file mode 100644
index d7566195f3..0000000000
--- a/changelog.d/4012.feature.md
+++ /dev/null
@@ -1 +0,0 @@
-**New citation format: source-tagged with global numbering.** Reports can now render citations as `[arxiv-1]`, `[openai.com-2]`, `[arxiv-3]` — the source tag identifies what kind of source each citation is (short URLClassifier tag for known academic sources, cleaned domain for generic web URLs, or the collection name for local RAG/library hits, e.g. `[my-papers-4]`), while the number stays the original global counter so labels never collide and inline citations match the bibliography order. Opt in via `report.citation_format → "Source-tagged with global numbering"`; the existing default remains `Numbers with hyperlinks [1]`.
diff --git a/changelog.d/README.md b/changelog.d/README.md
index b96643a6fb..f0bb7bd71b 100644
--- a/changelog.d/README.md
+++ b/changelog.d/README.md
@@ -1,10 +1,12 @@
# Release-notes news fragments
Each PR with user-visible behavior change drops one tiny markdown file in
-this directory. At release prep time the maintainer runs
-`pdm run towncrier build --version <X.Y.Z> --yes`, which renders the
-fragments into `docs/release_notes/<X.Y.Z>.md` and removes them. The
-release workflow surfaces that file in the published GitHub release body.
+this directory. The release workflow renders the fragments into
+`docs/release_notes/<X.Y.Z>.md` automatically at release time, surfaces
+that file in the published GitHub release body, and opens a follow-up PR
+(`chore/post-release-cleanup-<X.Y.Z>`) that persists the rendered file
+and deletes the consumed fragments. Maintainers do not need to run
+towncrier by hand.
This per-PR fragment model replaces the old approach where every
contributor edited a shared `docs/release_notes/<version>.md` — that
@@ -69,29 +71,38 @@ release page?* If no, skip the fragment.
## Release flow (for maintainers)
-```bash
-# As part of release prep:
-pdm run towncrier build --version 1.6.9 --yes
-# bump __version__.py to 1.6.9
-git add src/local_deep_research/__version__.py docs/release_notes/1.6.9.md
-git commit -m "chore(release): bump to 1.6.9 + render news fragments"
-git push
-```
+The release workflow does the towncrier render itself — there's nothing
+to run before merging the version bump. The flow is:
+
+1. Merge the version-bump PR (or push a `v*.*.*` tag).
+2. The `create-release` job sparse-checks-out `changelog.d/`, runs
+ `towncrier build --yes --version <X.Y.Z>` in a throwaway runner
+ workspace, and uses the rendered file as input to the AI summary
+ plus the published GitHub release body.
+3. The `cleanup-changelog` job re-runs the render against the same
+ commit (`github.sha`) and opens a `chore/post-release-cleanup-<X.Y.Z>`
+ PR that persists `docs/release_notes/<X.Y.Z>.md` and deletes the
+ consumed fragments. Squash-merge it.
+
+Towncrier is configured under `[tool.towncrier]` in `pyproject.toml`:
+`single_file = false` plus `filename = "docs/release_notes/{version}.md"`
+makes each release land in its own per-version file.
-Towncrier (configured under `[tool.towncrier]` in `pyproject.toml` with
-`single_file = false` and `filename = "docs/release_notes/{version}.md"`):
+### Previewing locally
-1. Reads fragments from `changelog.d/`.
-2. Renders them into `docs/release_notes/1.6.9.md`.
-3. `git rm`s the consumed fragments (deletion staged for commit).
-4. Stages the new release-notes file via `git add`.
+To see what the next release's notes will look like without touching
+anything (no fragment deletion, no file write):
+
+```bash
+pdm run towncrier build --draft --version <X.Y.Z>
+```
-To preview without touching anything (e.g., while iterating on a
-fragment locally):
+If you do want a local dry-run that actually writes the file (e.g. to
+sanity-check rendering of a tricky fragment), pass `--keep` so the
+fragments are not deleted:
```bash
-pdm run towncrier build --draft --version 1.6.9
+pdm run towncrier build --keep --version <X.Y.Z>
```
-The release workflow reads `docs/release_notes/<version>.md` for the
-human-narrative input to the published release body.
+`git restore` afterwards undoes both the file write and any deletions.
diff --git a/docs/RELEASE_GUIDE.md b/docs/RELEASE_GUIDE.md
index b6732d3813..5a7b1ec79f 100644
--- a/docs/RELEASE_GUIDE.md
+++ b/docs/RELEASE_GUIDE.md
@@ -24,10 +24,19 @@ and short-circuits everything downstream).
- **Version**: Read from `src/local_deep_research/__version__.py` by
the `version-check` job; the tag is `v<version>`.
- **Release body**: Composed by `.github/workflows/release.yml` from three sources:
- 1. **AI TL;DR** generated by OpenRouter (`vars.AI_MODEL`).
- 2. **Hand-written notes** from `docs/release_notes/<version>.md` — produced
- from contributor-supplied `changelog.d/*.md` fragments by
- `pdm run towncrier build` at release prep (see [Release-notes flow](#-release-notes-flow-towncrier-news-fragments) below).
+ 1. **AI narrative** generated by OpenRouter (`vars.AI_MODEL`, default
+ `moonshotai/kimi-k2-thinking`). The model receives the rendered
+ hand-written notes, the auto-generated PR list, every PR's title +
+ body (batched via one GraphQL call), and the diff between the
+ previous release tag and this one (filtered to drop lockfiles,
+ generated docs, SBOM, static assets, and binary patches; capped at
+ 700k chars).
+ 2. **Hand-written notes** from `docs/release_notes/<version>.md` —
+ rendered from contributor-supplied `changelog.d/*.md` fragments by
+ the workflow itself at release time (see
+ [Release-notes flow](#-release-notes-flow-towncrier-news-fragments)
+ below). No manual `pdm run towncrier build` is required before
+ merging the bump.
3. **Auto-generated PR list** from GitHub's generate-notes API,
label-categorized.
- **No duplicates**: If a release for `v<version>` already exists, the
@@ -50,26 +59,29 @@ Code owners (defined in `.github/CODEOWNERS`):
### For Regular Releases:
-1. **Render news fragments** (run from repo root):
- ```bash
- pdm run towncrier build --version <X.Y.Z> --yes
- ```
- Towncrier writes `docs/release_notes/<X.Y.Z>.md` (per the
- `{version}`-templated `filename` in `pyproject.toml`'s
- `[tool.towncrier]`) and `git rm`s the consumed `changelog.d/*.md`
- fragments. Preview first with `--draft` (writes nothing). Skip if no
- fragments exist for this release — the workflow tolerates a missing
- per-version file (warns and proceeds with auto-notes only).
-2. **Bump version** in `src/local_deep_research/__version__.py`
-3. **Commit both** in one go:
- ```bash
- git add src/local_deep_research/__version__.py docs/release_notes/<X.Y.Z>.md
- git commit -m "chore(release): bump to <X.Y.Z> + render news fragments"
- git push
- ```
-4. **Open PR** if not on a release branch already, get approval from code owners
-5. **Merge to main** → Release automatically created
-6. **Approve publishing** in GitHub Actions (PyPI/Docker)
+1. **Bump version** in `src/local_deep_research/__version__.py` (or
+ merge the auto-bump PR opened by `.github/workflows/version_check.yml`).
+2. **Merge to main** → Release automatically created. The release
+ workflow renders fragments from `changelog.d/*.md` into
+ `docs/release_notes/<X.Y.Z>.md`, composes the body (AI narrative +
+ rendered changelog + auto PR list), and publishes the GitHub release.
+3. **Approve publishing** in GitHub Actions (PyPI/Docker).
+4. **Merge the cleanup PR** opened automatically by the
+ `cleanup-changelog` job (titled
+ `chore: clear changelog fragments for <X.Y.Z>`). It persists
+ `docs/release_notes/<X.Y.Z>.md` and removes the consumed fragments
+ from `changelog.d/`. Squash-merge — the diff has no review value
+ beyond a sanity check that the rendered notes look right.
+
+To preview the rendered notes locally before merging the bump:
+
+```bash
+pdm run towncrier build --draft --version <X.Y.Z>
+```
+
+(`--draft` writes nothing.) Skip if no fragments exist for this release
+— the workflow tolerates a missing per-version file (warns and proceeds
+with auto-notes plus AI summary only).
### For Hotfixes:
@@ -135,26 +147,31 @@ typo'd category doesn't silently vanish at render time. See
[`changelog.d/README.md`](../changelog.d/README.md) for the full
convention.
-### Maintainer side (release prep)
+### Maintainer side
-```bash
-pdm run towncrier build --version <X.Y.Z> --yes
-```
+The release workflow handles the render. There is nothing to run
+manually before merging the version bump.
+
+`.github/workflows/release.yml` (`create-release` job):
+1. Sparse-checks-out `changelog.d/` + `pyproject.toml`, installs
+ `towncrier~=24.8`.
+2. Runs `towncrier build --yes --version <X.Y.Z>` against the runner's
+ throwaway workspace. Towncrier writes the rendered output to
+ `docs/release_notes/<X.Y.Z>.md` (per the `{version}`-templated
+ `filename`; `single_file = false` makes this a per-release file
+ rather than appending to a master CHANGELOG) and removes the
+ consumed fragments locally.
+3. Reads the rendered file as input to the AI summary and as the
+ "hand-written notes" section of the published GitHub release body.
+
+Persistence to `main` happens in the `cleanup-changelog` job, which
+re-runs the same render against the release commit (`github.sha`) and
+opens a `chore/post-release-cleanup-<X.Y.Z>` PR with the deletions and
+the rendered file. Squash-merge it.
-What towncrier does (configured in `pyproject.toml` under
-`[tool.towncrier]`):
-1. Reads fragments from `changelog.d/`.
-2. Writes the rendered output to `docs/release_notes/<X.Y.Z>.md` (per
- the `{version}`-templated `filename`; `single_file = false` makes
- this a per-release file rather than appending to a master CHANGELOG).
-3. `git rm`s the consumed fragments (deletion staged for commit).
-4. Stages the new release-notes file via `git add`.
-
-The release workflow (`.github/workflows/release.yml`) then `cat`s
-`docs/release_notes/<X.Y.Z>.md` into the published release body. If the
-file is missing (maintainer forgot to run towncrier), the workflow logs
-a warning and proceeds with the AI TL;DR + auto PR list only — no hard
-failure, but the hand-written narrative is lost.
+If `changelog.d/` is empty (maintenance release with no fragments), the
+render step is skipped cleanly and the release proceeds with the AI
+narrative + auto PR list only — no hard failure.
### Preview without committing
@@ -167,12 +184,10 @@ Useful while iterating on a fragment locally.
## 📊 Release Checklist
-- [ ] News fragments rendered: `pdm run towncrier build --version <X.Y.Z> --yes`
-- [ ] Version updated in `__version__.py`
-- [ ] Render + bump committed together
-- [ ] Changes tested in PR
+- [ ] Version updated in `__version__.py` (or auto-bump PR merged)
- [ ] Code owner approval received
- [ ] CI tests passing
- [ ] Merge to main completed
- [ ] Release automatically created
- [ ] PyPI/Docker publishing approved
+- [ ] `chore: clear changelog fragments for <X.Y.Z>` PR merged
diff --git a/docs/release_notes/1.6.10.md b/docs/release_notes/1.6.10.md
new file mode 100644
index 0000000000..ebaaad11f3
--- /dev/null
+++ b/docs/release_notes/1.6.10.md
@@ -0,0 +1,85 @@
+
+### 💥 Breaking Changes
+
+- **`llm.model` no longer auto-fills `gemma3:12b`.** Pre-1.7 installs silently downloaded a multi-GB Ollama binary on first use. The field is now empty by default — pick a model in **Settings → LLM**, or research will fail loudly with a clear `ValueError`. Existing users with a saved `llm.model` setting are unaffected; fresh installs see an inline warning in Settings (`aria-live="polite"`) and a clear error on research start. Affects all built-in providers (`openai`, `anthropic`, `ollama`, `lmstudio`, `llamacpp`, `google`, `openrouter`, `ionos`, `xai`). ([#3670](https://github.com/LearningCircuit/local-deep-research/pull/3670))
+- **`llamacpp` provider migrated from in-process to HTTP.** No longer loads `.gguf` files in-process via `llama-cpp-python` — connects to `llama-server`'s OpenAI-compatible HTTP endpoint instead. Removed settings: `llm.llamacpp_model_path`, `llm.llamacpp_n_batch`, `llm.llamacpp_n_gpu_layers`, `llm.llamacpp_f16_kv`. New settings: `llm.llamacpp.url` (default `http://localhost:8080/v1`), `llm.llamacpp.api_key` (optional, for auth proxies). Migration: run `llama-server -m <your-model.gguf>` and the default URL picks it up; alternatively use the `openai_endpoint` provider pointed at the same URL. ([#3670](https://github.com/LearningCircuit/local-deep-research/pull/3670))
+
+### 🔒 Security
+
+- Fix LM Studio model detection (#3800) — the settings dropdown was silently dropping models for every auto-discovered provider except Ollama / OpenAI / Anthropic / OpenAI-Compatible-Endpoint, including LM Studio, Llama.cpp, Google Gemini, xAI, OpenRouter, and IONOS. While investigating, also fixed a separate credential-leak class where the route's auto-discovery loop sent every other provider's configured API key in the `Authorization` header to the local LM Studio / llama-server endpoint on every settings-page load. ([#3800](https://github.com/LearningCircuit/local-deep-research/pull/3800))
+- Fix SSRF parser-differential bypass (GHSA-g23j-2vwm-5c25). URLs containing
+ backslash, whitespace, or ASCII control bytes are now rejected upfront by the
+ SSRF validator and notification-URL validator; hostname extraction switched
+ from `urllib.parse.urlparse` to `urllib3.util.parse_url` so the validator and
+ the HTTP client agree on destination by construction. Credit: @Fushuling,
+ @RacerZ-fighting.
+- Hardened SSRF defenses against AWS ECS task metadata
+ (`169.254.170.2`, `169.254.170.23`), Tencent Cloud (`169.254.0.23`),
+ and AlibabaCloud (`100.100.100.200`) metadata endpoints — these are
+ now always blocked alongside the existing AWS IMDS / Azure / OCI /
+ DigitalOcean entry (`169.254.169.254`). Redacted credentials, path,
+ and query from URL-rejection logs (operators with grep/regex tooling
+ on the rejection log lines will see authority-only `scheme://host:port`
+ instead of full URLs going forward).
+- SSRF defense-in-depth: block IPv6 transition prefixes that can wrap
+ private IPv4 destinations on hosts with kernel sit0/NAT64 routes.
+
+ - `2002::/16` (6to4, RFC 3056 — deprecated by RFC 7526)
+ - `64:ff9b::/96` (NAT64 well-known, RFC 6052)
+ - `64:ff9b:1::/48` (NAT64 local-use, RFC 8215 — same SSRF threat class
+ as the WKP; missing it earned a HackerOne bounty against
+ ssrf_filter)
+ - `2001::/32` (Teredo, RFC 4380)
+ - `100::/64` (IPv6 discard, RFC 6666)
+ - `::/96` (IPv4-Compatible IPv6, RFC 4291 §2.5.5.1 — DEPRECATED 2006;
+ same SSRF threat class as the transition prefixes)
+
+ The metadata-IP block is hardened against IPv6-wrapped IMDS access:
+ when an IPv6 destination falls in a NAT64 prefix, the embedded IPv4 is
+ extracted and matched against `ALWAYS_BLOCKED_METADATA_IPS`, so
+ `[64:ff9b::a9fe:a9fe]` cannot reach 169.254.169.254 even on a NAT64
+ host.
+
+ Operators on IPv6-only deployments using DNS64+NAT64 (where outbound
+ IPv4 traffic is synthesized through `64:ff9b::/96`) can opt back in via
+ the env-only setting `security.allow_nat64`
+ (`LDR_SECURITY_ALLOW_NAT64=true`). The opt-in is scoped strictly to
+ the two NAT64 prefixes — 6to4, Teredo, and discard remain blocked
+ unconditionally, and the IMDS embedded-IPv4 carve-out still applies.
+
+### ✨ New Features
+
+- **LM Studio optional API key.** New `llm.lmstudio.api_key` setting for LM Studio instances with authentication enabled. ([#3670](https://github.com/LearningCircuit/local-deep-research/pull/3670))
+- **Release notes now use [towncrier](https://towncrier.readthedocs.io) news fragments.** Each PR drops one tiny `changelog.d/<id>.<category>.md` file (categories: `breaking`, `security`, `feature`, `bugfix`, `removal`, `misc`). At release prep, the maintainer runs `pdm run towncrier build --version <X.Y.Z> --yes`, which renders the fragments into `docs/release_notes/<X.Y.Z>.md` and removes them. Replaces the previous shared-file workflow, which scaled poorly at LDR's PR throughput. See `changelog.d/README.md` for the convention. ([#3773](https://github.com/LearningCircuit/local-deep-research/pull/3773))
+- **New citation format: source-tagged with global numbering.** Reports can now render citations as `[arxiv-1]`, `[openai.com-2]`, `[arxiv-3]` — the source tag identifies what kind of source each citation is (short URLClassifier tag for known academic sources, cleaned domain for generic web URLs, or the collection name for local RAG/library hits, e.g. `[my-papers-4]`), while the number stays the original global counter so labels never collide and inline citations match the bibliography order. Opt in via `report.citation_format → "Source-tagged with global numbering"`; the existing default remains `Numbers with hyperlinks [1]`. ([#4012](https://github.com/LearningCircuit/local-deep-research/pull/4012))
+- Context-overflow warnings (front-page banner, result-page banner, completion toast) now show a clickable action that takes you straight to the relevant diagnostics — historical warnings link to the context-overflow metrics page, while per-research warnings link to that research's overflow section on its details page.
+- Context-overflow warnings now include structured fields (`research_id`, `model`, `provider`) and detect both input-only and total-context (input+output) overflow. Four distinct tags (`[provider-confirmed]`, `[total-context]`, `[estimated]`, `[estimated-total-context]`) make different overflow causes distinguishable in logs and alerts. Hosted providers (OpenAI, Anthropic, OpenRouter) get estimation-based detection so their truncation events are no longer invisible. The high-context VRAM warning now links directly to the context-overflow metrics page.
+- OpenAI-compatible runtime errors (LM Studio / vLLM / llama.cpp server / OpenRouter / custom endpoint) now surface a message that names the provider, configured base URL, and model. The existing `Error type: <code>` token convention is extended with seven `openai_*` codes (`openai_connection_refused`, `openai_timeout`, `openai_auth`, `openai_permission_denied`, `openai_model_not_found`, `openai_bad_request`, `openai_unknown`), and `ErrorReporter` maps each to the right `ErrorCategory`. The original exception text is preserved in a `Details:` suffix; userinfo embedded in a base URL (`https://u:key@host/v1`) is stripped before display.
+- The `/metrics/context-overflow` page is reorganised around per-research truncation diagnosis: redundant aggregate sections (Token Usage Overview, Token Usage Over Time, Model Usage Breakdown, Context Limit Distribution doughnut, Token Usage by Phase) are dropped in favour of empty-state messages and a recoloured and re-shaped scatter chart that buckets points by context-utilisation ratio (green circle / amber triangle / red diamond — shape gives a redundant signal for colourblind users) rather than per-limit dashed reference lines. Tooltips now explicitly label requests with unknown context limits or unknown token counts. The main `/metrics` dashboard gains a small Context Overflow summary panel linking to the full diagnostics page; on aggregation failure the panel shows `—` rather than a misleading `0%`.
+- Upload rate limits are now configurable via `LDR_SECURITY_RATE_LIMIT_UPLOAD_USER` and `LDR_SECURITY_RATE_LIMIT_UPLOAD_IP`. The default cap is raised from `10 per minute;100 per hour` to `60 per minute;1000 per hour`, unblocking bulk RAG library uploads.
+
+### 🐛 Bug Fixes
+
+- Fix progress page log panel sizing so it uses CSS layout instead of JavaScript height calculations. ([#2606](https://github.com/LearningCircuit/local-deep-research/pull/2606))
+- Progress page current-task text now preserves line breaks and wraps long generated task messages instead of collapsing them into a hard-to-read wall of text. ([#2609](https://github.com/LearningCircuit/local-deep-research/pull/2609))
+- Live research log entries now keep the same newest-first visual order as loaded log history instead of appearing at the bottom of the log panel. ([#2610](https://github.com/LearningCircuit/local-deep-research/pull/2610))
+- **Journal-quality dashboard ("Your Research") no longer shows orphan papers from deleted research sessions.** Deleting a research session previously left behind `Paper` rows whose `PaperAppearance` had cascade-deleted, so the dashboard kept counting them. The aggregate top-200 query and the predatory-blocked count now exclude papers that have no remaining appearances. Per-research views were already correct. ([#3544](https://github.com/LearningCircuit/local-deep-research/pull/3544))
+- Fix the model dropdown in Settings to populate for every auto-discovered LLM provider, not just Ollama / OpenAI / Anthropic / OpenAI-Compatible-Endpoint. LM Studio, Llama.cpp, Google Gemini, xAI, OpenRouter, and IONOS were silently dropped by the frontend (`processModelData`) even when the backend returned their models correctly. ([#3800](https://github.com/LearningCircuit/local-deep-research/pull/3800))
+- Fix file descriptor exhaustion (#3816) caused by `ChatOllama._async_client` not being released alongside `_client`. `_close_base_llm()` now closes both the sync and async httpx clients owned by `ChatOllama`, addressing the eventpoll-FD growth seen in long-running deployments. ([#3816](https://github.com/LearningCircuit/local-deep-research/pull/3816))
+- Fix the research details "View Journals" button so it opens the scoped journal-quality dashboard instead of a missing page. ([#3825](https://github.com/LearningCircuit/local-deep-research/pull/3825))
+- Fix Download Manager failure after 20–30 PDFs with `UNIQUE constraint failed: documents.document_hash` and the cascading PendingRollbackError that prevented further downloads. ([#3827](https://github.com/LearningCircuit/local-deep-research/pull/3827))
+- Fix login/registration over `http://localhost:5001` on Docker Desktop.
+ The session cookie's `Secure` flag is now driven purely by the request
+ protocol (HTTPS) rather than by source-IP heuristics, so HTTP requests
+ from any IP — including Docker Desktop's NAT gateway — now keep their
+ cookies and CSRF tokens. ([#3849](https://github.com/LearningCircuit/local-deep-research/pull/3849))
+- Fix Ollama embedding indexing aborting with `"input length exceeds the context length"` (HTTP 500) by exposing a configurable `embeddings.ollama.num_ctx` setting. The new field appears on the embedding settings page when the Ollama provider is selected and ships with a default of 8192 tokens, which covers chunk sizes up to ~5000 characters for `nomic-embed-text`, `mxbai-embed-large`, `bge-m3`, and `qwen3-embedding`. ([#3870](https://github.com/LearningCircuit/local-deep-research/pull/3870))
+- LangGraph agent strategy error messages now include the exception type, and the `'str' object has no attribute 'model_dump'` failure (issue #3897) is translated into a rich on-result-page recommendation listing the three possible causes — proxy/shim, serving stack, or model — with concrete remediation steps for each. ([#3926](https://github.com/LearningCircuit/local-deep-research/pull/3926))
+- The Context Overflow summary panel on `/metrics` now honors the dashboard's research-mode filter (Quick / Detailed / All) the same way the rest of the panels do.
+
+### 📝 Other Changes
+
+- Added `LDR_DISABLE_RATE_LIMITING` as the canonical name for disabling HTTP rate limiting, matching the `LDR_` env-var convention. The legacy `DISABLE_RATE_LIMITING` still works but emits a deprecation warning. Note: this is distinct from `LDR_RATE_LIMITING_ENABLED`, which controls the adaptive search-engine rate limiter — a different subsystem (#3905).
+- Migrated CI workflows to the canonical `LDR_DISABLE_RATE_LIMITING` env var (added in #3936) and closed a test-isolation gap in `test_enabled_by_default` that did not clear the canonical var.
+- Removed unused chart-rendering JavaScript on the context-overflow analytics page (token-usage-over-time stacked bar, model-token-stats table, context-limits-distribution doughnut, and the configured-limits list). These functions stopped being called when the page was reorganised to focus on truncation diagnosis (#3792); deleting them now drops ~200 lines of dead code without any user-visible change.
+- The prerelease Docker workflow now also re-points the floating `prerelease` tag at each new RC manifest. Testers can pin `image: localdeepresearch/local-deep-research:prerelease` in their compose file and run `docker compose pull && docker compose up -d` to fetch the latest RC without editing the tag each cycle. Versioned tags (`prerelease-vX.Y.Z-<sha>`) are still published for reproducible testing.