`pull_request_review` runs on fork PRs, which includes all backport PRs,
get a read-only token and no secrets, so giteabot cannot write lgtm
labels and statuses there. A no-op `giteabot-review` workflow now
triggers giteabot through `workflow_run`, which gets both. This allows
retiring the legacy fly.io webhook bot.
Part of https://github.com/go-gitea/giteabot/issues/15
The API filter `head_sha` on `GET /repos/{owner}/{repo}/actions/runs`
selects runs by `commit_sha`, so `commit_sha` needs an index. For a
action_run table with 212k rows:
- Without the index: the query read 212k rows, and the API request took
35-52 s.
- With the index: the query read 94 rows in 0.14 s, and the API request
took 2-4 s.
---------
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
MSSQL's default READ COMMITTED makes reads wait on writers, so the
runner pickup deadlocks with concurrent claims, flaking
`TestCreateTaskForRunnerConcurrentClaim`.
- Enable `READ_COMMITTED_SNAPSHOT` on MSSQL so it reads like PostgreSQL
and MySQL
- Read the pickup cursor before claiming, a lost claim could skip
waiting jobs
- Add tests that fail without consistent READ COMMITTED
Performance: Writes on MSSQL now also store the previous row version in
tempdb, the same versioning cost PostgreSQL and MySQL always pay, and
Azure SQL enables it by default. Reads no longer block on writers, and a
32-runner pickup stress test ran 2.5x faster with it.
---------
Signed-off-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: Giteabot <teabot@gitea.io>
Replace citation-js with a Go port of ruby-cff, which GitHub uses for
"Cite this repository", rendering APA and BibTeX server-side and
dropping about 770KB of JS. Output matches GitHub on 1568 of 1571
real-world files, the rest are improvements over GitHub.
- `CITATION.cff` wins over `CITATION.bib`, matched case-insensitively
and through symlinks
- `CITATION.bib` is offered as-is, without APA
Signed-off-by: silverwind <me@silverwind.io>
Vue files only got `eslint-plugin-vue` rules, so stylistic and
TypeScript rules never ran on them, and `import-x` could not see cycles
between `.ts` and `.vue` files.
- Apply the main ESLint config to `.vue` files
- Let `import-x` parse `.vue` files, which surfaced a cycle between
`repo-findfile.ts` and `RepoFileSearch.vue`
- Fix the resulting lint issues
`vue-eslint-parser` is added as a direct dependency because `import-x`
resolves parsers by package name, see
https://github.com/un-ts/eslint-plugin-import-x/issues/381.
On Linux and ChromeOS, Chromium resolves the `math` font family to Latin
Modern Math, which neither installs, so MathML renders with a text font
and brackets and large operators don't stretch, see
https://issues.chromium.org/issues/40069293. The new `--fonts-math`
variable keeps `math` first, so browsers that always resolve it keep
their font. Only Chromium falls through to the math fonts Linux and
ChromeOS install by default:
- `STIX Two Math`: Fedora, and `fonts-stix` on Ubuntu 26.04
- `DejaVu Math TeX Gyre`: Debian 13 and openSUSE
- `Noto Sans Math`: Fedora and ChromeOS, last because Debian and Ubuntu
ship an older version without a `MATH` table
Math fonts also have smaller x-heights than UI fonts, so MathML rendered
smaller than KaTeX and the surrounding text in every browser, see
https://github.com/w3c/mathml-core/issues/41. `font-size-adjust:
ex-height 0.52` scales whichever math font is used to KaTeX's x-height.
KaTeX output is unchanged.
Fixes: https://github.com/go-gitea/gitea/issues/39489
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: bircni <75789103+bircni@users.noreply.github.com>
Co-authored-by: silverwind <me@silverwind.io>
`UpdatePushMirror` used `AllCols()`, so a sync could overwrite columns
changed concurrently (e.g. `interval`) with stale values. It now updates
only `last_update` and `last_error`, and is renamed to
`UpdatePushMirrorSyncStatus` to match.
Co-authored-by: Giteabot <teabot@gitea.io>
Co-authored-by: silverwind <me@silverwind.io>
Co-authored-by: Claude (Opus 5) <noreply@anthropic.com>
The `gitea.com/go-chi/session` package only exists for Gitea, so it
moves into `modules/session` to fix its bugs directly. Fixes the flake
in
https://github.com/go-gitea/gitea/actions/runs/36726154500/job/109923538400.
- Sessions are only written back when changed, so a read-only request
can't revert a concurrent change or restore a logged-out session, like
https://github.com/go-macaron/session/commit/ae808a4a4660c802965c834299ab08f167effd12
- The session cookie is only set once a session holds data
- Every backend refreshes the expiry on load and file sessions are
written atomically
- Also fix https://github.com/go-gitea/gitea/issues/36176
## ⚠️ BREAKING ⚠️
* the `mysql`, `postgres`, `couchbase` and `memcache` session providers
are removed, use `file`, `db` or `redis` instead
* login-related cookies are renamed to `gitea_session` and
`gitea_remember`, if you'd like to use the old names, set `COOKIE_NAME`
and `COOKIE_REMEMBER_NAME` in app.ini
---------
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
The `gitea.com/go-chi/cache` package only exists for Gitea, so it moves
into `modules/cache`.
- `ITEM_TTL = -1` disables caching as documented
- Sub-second TTLs round up instead of never expiring
- The Redis adapter no longer grows a `MacaronCache` hash
- Use two-queue cache for in-memory cache
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
The `gitea.com/go-chi/captcha` package only exists for Gitea,
so it moves into `modules/imagecaptcha`.
- Reloading an expired challenge shows a new image instead of a broken one
- Every answer is consumed on its first check
- OpenID registration stops after a failed captcha
---------
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
The `gitea.com/go-chi/binding` package only exists for Gitea, so it
moves into `modules/web/binding` to fix its bugs directly. Split out of
https://github.com/go-gitea/gitea/pull/39504.
- GET and HEAD always bind the query
- JSON `null` slice elements and nested `TrimSpace` fields bind
correctly
- Integer fields reject out-of-range values instead of wrapping
- An empty JSON body binds nothing and an unknown binding rule is an
error
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: bircni <bircni@icloud.com>
MariaDB 11.6.2+ defaults `innodb_snapshot_isolation` to `ON`, which
fails REPEATABLE READ transactions with error 1020 when a row they write
changed after their first read. Gitea's background work like push
processing writes the same rows, so merges, issue closes and workflow
runs fail sporadically.
- Use READ COMMITTED on MySQL and MariaDB, like PostgreSQL and MSSQL
- Update xorm to v1.4.3
Replaces: https://github.com/go-gitea/gitea/pull/39494
Fixes: https://github.com/go-gitea/gitea/issues/39492
---------
Signed-off-by: silverwind <me@silverwind.io>
Replace time.After with a stoppable time.Timer in BatchChecker.CheckPath.
Avoid accumulating up to 6 pending 5-second timers per file on the normal path.
Preserve the existing per-attribute timeout behavior.
---------
Signed-off-by: Calvin Tjoaquinn <calvintjoa23@gmail.com>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Speeds up two slow git paths. Results match `main` apart
from the `log.follow` fix.
- Activity top authors skip `--numstat` and an unused `rev-list
--count`, they only need names and emails
- Subdirectory listings pass only the directory as pathspec, which
already matches all its entries
- Directory listings pass `--no-follow` to `git log`, a configured
`log.follow` disabled parent rewriting and gave wrong last commits
| Benchmark | main | PR | Change |
|---|--:|--:|--:|
| Top authors, gitea, 1 month | 632 ms | 21 ms | -97% |
| Top authors, gitea, 1 year | 3216 ms | 68 ms | -98% |
| Top authors, tea | 62 ms | 12 ms | -80% |
| Listing, git `Documentation/technical` | 189 ms | 91 ms | -52% |
| Listing, gitea `options/license` | 329 ms | 208 ms | -37% |
| Listing, gitea `templates/repo` | 294 ms | 223 ms | -24% |
Tested with unit and sqlite integration tests in default and `gogit`
builds on git 2.25 and 2.56, and by comparing listing results and cache
writes with `main` on randomized histories and the gitea, tea and git
repos. Benchmarks are medians of 8 interleaved macOS runs.
Commit status list orders only by `created_unix`/`updated_unix`, which
have 1-second resolution while CI often posts many statuses per second.
With LIMIT/OFFSET paging, databases (e.g. PostgreSQL using a Sort plan)
may order tied rows differently per page, so `GET
/repos/{owner}/{repo}/commits/{ref}/statuses` returns some statuses
twice and never returns others.
This became visible after https://github.com/go-gitea/gitea/pull/36521
made requests without `page` paginated. Clients like Renovate that page
until `X-Total-Count` can miss a context's newest status and see a stale
`pending`, blocking automerge.
Fix: add `index` (unique per commit) as a tiebreaker to the
timestamp-based orders.
Co-authored-by: silverwind <me@silverwind.io>
Aligns the npm registry with what npm, pnpm and yarn expect:
1. Raise the publish body cap from
https://github.com/go-gitea/gitea/pull/37890 to 256 MiB like npmjs,
larger bodies get 413
2. Pick the tarball attachment by name, `npm publish --provenance`
failed at random
3. Store and serve `libc`, so mismatched glibc/musl optional binaries
are skipped
4. Treat root `*.gyp` files as an install script, like npm does
5. Always serve a `latest` dist-tag, yarn and pnpm fail without it
6. Take top-level metadata from `latest` and drop the per-version readme
7. Serve tarballs at the npmjs path `/<name>/-/<file>`, former URLs keep
working
8. Add ETag revalidation for metadata, `npm ping` and `npm whoami`
Tested with npm 12.1, pnpm 12.4, yarn 1.22 and yarn 4.18.
---------
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
The example shows `REPOSITORY_AVATAR_FALLBACK_IMAGE =
/img/repo_default.png`, but public files moved under `/assets` in
https://github.com/go-gitea/gitea/pull/15219 and nothing serves `/img/`
anymore. The value is also used as-is without the sub-path, so even
`/assets/img/repo_default.png` 404s when `ROOT_URL` has one. Leave the
key empty and document the real default
`{AppSubURL}/assets/img/repo_default.png` from
`modules/setting/picture.go`.
Signed-off-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: silverwind <me@silverwind.io>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Lets users approve an OAuth2 scope change on an existing grant instead
of failing with `a grant exists with different scope`.
- Approving a different scope updates the existing grant. Issued tokens
follow immediately, since their scope is read from the grant.
- Confidential and trusted apps show the consent page when the scope set
changes, instead of silently reusing the old grant.
- An omitted `scope` reuses the existing grant's scope, like GitHub.
- The consent page lists newly added scopes.
Fixes: https://github.com/go-gitea/gitea/issues/38940
Co-authored-by: bircni <bircni@icloud.com>
Co-authored-by: Giteabot <teabot@gitea.io>
Co-authored-by: silverwind <me@silverwind.io>
Align job and `runs-on` validation with github.com, as implemented by
the parser in https://github.com/actions/runner. A job without `runs-on`
could be claimed by any runner, so a job meant for a container could run
on the host.
- Jobs without `runs-on` fail with `Required property is missing:
runs-on`, called workflows included
- Unknown job keys and callers (`uses:`) mixed with steps-only keys like
`runs-on` are rejected
- Empty, null and nested `runs-on` values are rejected
- A `runs-on` evaluating to such a value fails only that job
- Called workflows are validated at run creation, an invalid one fails
the run as an invalid workflow file
- Zero labels (`runs-on: []` or `{}`) never match a runner, including
jobs queued before upgrading
<img width="960" alt="image"
src="https://github.com/user-attachments/assets/e746ce5a-b711-4e8b-aab8-81336ff53d86"
/>
**Behavior Change:** workflows that omit `runs-on`, use unknown job keys
or mix `uses` with `runs-on` stop running until fixed.
---------
Co-authored-by: bircni <bircni@icloud.com>
Co-authored-by: silverwind <me@silverwind.io>
Three commented defaults in `custom/conf/app.example.ini` differ from
what Gitea actually uses, so the file says they default to something
else:
- `MINIMUM_KEY_SIZE_CHECK`: example `false`, code `true`
(`modules/setting/ssh.go:55`, read at `:152`).
- `SSH_SERVER_HOST_KEYS`: example lists `ssh/gitea.rsa, ssh/gogs.rsa`,
code also loads `ssh/gitea.ed25519` and `ssh/gitea.ecdsa`
(`modules/setting/ssh.go:57`).
- `[queue] DATADIR`: example `queues/`, code `queues/common`
(`modules/setting/queue.go:33`), which the comment above it already
states.
Co-authored-by: silverwind <me@silverwind.io>
Several handlers skipped checks that their sibling routes or settings already enforce. This brings them in line.
- Push mirror API honors `DISABLE_NEW_PUSH` and checks the caller's permission
- Media API serves small files with the usual content headers
- Issue attachment API ignores comment attachments
- Push-to-create respects `FORCE_PRIVATE`
- Profile feeds and follow actions respect `ENABLE_FEED` and owner visibility
- Tag delete route refuses release tags
- Refresh token grant only accepts refresh tokens
- Gitea migrations bound the source's page size
Co-authored-by: bircni <bircni@icloud.com>
On a repository's Actions tab, runs are sorted newest first on initial
page load. After the first auto refresh (added in #38329, every 3
seconds while runs are active and every 12 seconds otherwise), the same
runs may appear in a different order and move again as their status
changes.
Example with four runs (Gitea 28.0.0, SQLite):
```
page load: #10 success, #9 failure, #8 success, #7 running
after refresh: #8 success, #10 success, #9 failure, #7 running
```
To reproduce, open the Actions tab of a repository with runs in
different statuses and wait for an auto refresh. On SQLite, the runs may
be regrouped by status, with each group ordered oldest first.
`preparePartialRefreshRuns` reloads the runs currently shown on the page
using `GetRunsByRepoAndID`. That query has no `ORDER BY`, while the
initial page load uses `FindRunOptions.ToOrders` and sorts by index
descending.
With SQLite, the query planner used the `(repo_id, status)` index, so
the returned row order differed from the original page order. Since the
query has no explicit ordering, this behavior is database-dependent. I
have not tested MySQL or PostgreSQL.
This change orders `GetRunsByRepoAndID` by index descending, the same
order `FindRunOptions.ToOrders` uses for the initial page load. The
refresh only reloads the runs already on the page, so they come back in
the original order, with or without filters and on any page.
The other caller of `GetRunsByRepoAndID`, run approval, does not depend
on result ordering.
Testing:
- Added `TestPreparePartialRefreshRunsKeepsRequestedOrder`. Without the
fix, runs 794, 793, 792, 791 are returned as 791, 792, 794, 793; with
the fix, the test passes.
- `go test` passes for `./routers/web/repo/actions/`,
`./models/actions/` and `./services/actions/`.
- `go vet` and `golangci-lint v2.13.2` pass for the changed packages.
- Manually tested by building Gitea 28.0.0 with this patch and running
it on our SQLite instance. The runs list keeps its newest-first order
across auto refreshes. The official 28.0.0 binary reproduces the
reordering.
AI-assisted: drafted with Claude Code (claude-opus-5-5), reviewed by me.
---------
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Set receive.fsckObjects=true in Gitea's internal global git config so
the receiving git process rejects bad, malicious or duplicate objects at
push time, before Gitea ever stores them.
Assisted-by: Codet:claude-opus-4-8
---------
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
Co-authored-by: silverwind <me@silverwind.io>
Introduces gitproxy module which spawns a small forward proxy as scanner
for git calls
Replaces hostmatcher with matchlist which supports port rules
Deprecates ALLOWED_DOMAINS/BLOCKED_DOMAINS and ALLOW_LOCALNETWORKS
settings in migration in favor of full names we have in security
configs.
Removes `external` preset in favor of lax/strict modes, strict mode
requiring explicit ports if they aren't standard http/s ones.
Breaking changes:
- `external` preset no longer works as deny rule. To enforce that, use
`strict` mode and allow ranges to connect to
- Wildcards are no longer accepted in IP addresses
- `*` is no longer allowed as entry in lists
- domain rules now use curl like syntax `*.example.com` matching
subdomains but not `example.com`, `example.com` matching itself and all
subdomains. `example.*` is not a valid rule
- In the default `lax` mode, `[security] ALLOWED_HOST_LIST` no longer
restricts public hosts, set `EGRESS_MODE = strict` to keep an exclusive
list. A startup warning flags this
- Invalid list entries are logged at startup, invalid
`BLOCKED_HOST_LIST`/`BLOCKED_DOMAINS` entries stop it
Docs: https://gitea.com/gitea/docs/pulls/557
Signed-off-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: silverwind <me@silverwind.io>
Co-authored-by: bircni <bircni@icloud.com>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Fixes https://github.com/go-gitea/gitea/issues/38705
Fixes https://gitea.com/gitea/runner/issues/1232
Jobs of a called workflow saw `gitea.event_name` as `workflow_call` and
`gitea.event.inputs` replaced by the caller's `with:`, so a condition
like `gitea.event_name == 'push'` never held in them, and a dispatched
run's own inputs were lost there.
They now keep the caller's trigger event and their `inputs` are the
run's `workflow_dispatch` inputs overlaid with the caller's `with:`, as
on GitHub. For example, with `workflow_dispatch` inputs `{target: prod,
debug: true}` and caller's `with: {target: dev}`, the called workflow's
`inputs` are `{target: dev, debug: true}`.
A runner cannot resolve these inputs itself, so they are sent in a new
`gitea_workflow_call` context entry, with the original event name and
inputs for the runner to restore. For more details, see the runner PR:
https://gitea.com/gitea/runner/pulls/1250
Co-authored-by: silverwind <me@silverwind.io>
Fixes several gaps in the approval of fork pull request runs:
1. Approving a run that was cancelled while awaiting approval revived
its cancelled jobs. Such a run is no longer treated as awaiting approval
by the merge box, run page, approve actions and API, and rerunning it
approves it.
2. Approval no longer revives jobs cancelled while the run was pending,
no longer lets two jobs sharing a concurrency group cancel each other,
and re-emits the run so jobs needing a cancelled job get resolved.
3. An unapproved run applies its workflow-level concurrency only once
approved.
4. For workflows from the pull request, both the event actor and the
pull request author must be trusted to skip approval. Workflows from the
default branch, like `issue_comment`, still only check the actor.
---------
Co-authored-by: silverwind <me@silverwind.io>
silverwind