Skip to content

Handoff: the state of the work

  • Heads: master is 49d26b879, the merge of #858 (2026-11-18 01:19 JST), and the tag v2.7.12, below; before it, 14bdc5fa6, the merge of #853 (2026-11-17 21:10 JST), after dd6f452fc, the merge of #852 (19:40 JST), both below; before them, ea531ad88, the merge of #851 (2026-11-17 17:10 JST), below; before it, 7f0f1c564 (16:11 JST), after e84d90bd0 (16:10 JST), the two merges from the private forks of the advisories fixed in 2.7.11 (below), and the tag v2.7.11; before them, 4e6494d1c, the merge of #843 (04:02 JST), below, and 2c7e833a0, the merge of #831 (2026-11-16 18:47 JST), below; before it, 8bfe5f987 (17:22 JST), a merge of two commits for 2.7.10 (below), and the tag v2.7.10; before it, fc44a3cfe, the merge of #829 (15:57 JST), and 22ef61915, the merge of #827 (04:06 JST), after 69fd3b758, the merge of #824 (02:09 JST), both below; before them, a7800d865, the merge of #822 (2026-11-15 23:33 JST) and the tag v2.7.9, below; before it, 8f204e061, the merge of #819 (20:18 JST), and 11f80acfa, the merge of #816 (16:51 JST), after a745fe8be, the merge of #817 (16:34 JST), all below; before them, 976015c48, the merge of #813 (06:57 JST) and the tag v2.7.8, after 31d5c75f3, the merge of #814 (05:55 JST), and 7f4063399, the merge of #812 (05:20 JST), all below. next is e8643df33, the merge of #856 (2026-11-17 22:18 JST), which brought master at 14bdc5fa6 into next; before it, 45cf3ae52, the merge of #854 (2026-11-17 19:42 JST), after 4c78c56a3, the merge of #850 (18:19 JST), which brought master at ea531ad88 into next; before them, e9c214de2, the merge of #849 (2026-11-17 17:52 JST), after b2c312f50, the merge of #848 (16:16 JST), both below; before them, the merges of #847 (06:48 JST), #844 (04:38 JST), #835 (04:25 JST), #836 (02:27 JST), #834 (01:23 JST), #833 (2026-11-16 23:55 JST) and #832 (21:28 JST), and 7cbe1664a, the merge of #830 (19:30 JST), which brought master at 2c7e833a0 into next. This entry is written in #859, the open pull request to next that merges master at 49d26b879 into next. v2.x mirrors master. The latest published release is 2.7.12: the tag v2.7.12 is on 49d26b879, and the release was published on 2026-11-18 at 02:56 JST, with the release notes docs/releases/v2.7.12.md and .ja.md, in the form of “The GitHub release” in docs/releases/README.md. Right after the release, site.yml ran on next (run 44142470100, success), and the home page of ngx.mruby.org names v2.7.12 (Latest release: v2.7.12). The push of the tag ran the whole build matrix (run 44142399585): 17 cells and ci-ok passed. Before it, 2.7.11: the tag v2.7.11 is on 7f0f1c564, and the release was published on 2026-11-17 at 16:26 JST, with the release notes docs/releases/v2.7.11.md and .ja.md, in the form of “The GitHub release” in docs/releases/README.md. Right after the release, site.yml ran on next (run 44038818769, success), and the home page of ngx.mruby.org names v2.7.11 (Latest release: v2.7.11). The push of the tag ran the whole build matrix (run 44038769049): 17 cells and ci-ok passed. Before it, 2.7.10: the tag v2.7.10 is on 8bfe5f987, and the release was published on 2026-11-16 at 17:46 JST, with the release notes docs/releases/v2.7.10.md and .ja.md, in the form of “The GitHub release” in docs/releases/README.md. Right after the release, site.yml ran on next (run 43834045818, success), and the home page of ngx.mruby.org names v2.7.10 (Latest release: v2.7.10). The push of the tag ran the whole build matrix (run 43833989570): 17 cells and ci-ok passed. Before it, 2.7.9: the tag v2.7.9 is on a7800d865, and the release was published on 2026-11-15 at 23:53 JST, with the release notes docs/releases/v2.7.9.md and .ja.md (#816, #819, #822). Its text is the form of “The GitHub release” in docs/releases/README.md: the section “Behavior changes: read before upgrading” of the English file, the same section of the Japanese file, and links to the two files, with the relative links of the notes given as links to the files at the tag v2.7.9 (the releases v2.7.7 and v2.7.8 copied the whole files instead). Right after the release, site.yml ran on next (run 43726844028, success: the build, the deploy and the check of the published site), and the home page of ngx.mruby.org names v2.7.9 (Latest release: v2.7.9). The push of the tag ran the whole build matrix (run 43726820701): 17 cells and ci-ok passed.
  • The check before the 2.7.11 tag, recorded here because it needed no pull request (“Before the tag” in docs/releases/README.md). The pull requests and merges into master since v2.7.10: #831, no entry (it adds the advisory id to docs/releases/v2.7.10.md and .ja.md only); #843, no entry (the pin of the CodeQL action in two workflows and a Dependabot group, which a contributor or a packager does not need to know); and the two merges from the private forks, e84d90bd0 (the delete of a header) and 7f0f1c564 (the assignment through Nginx::Var), which are items 2 and 1 of “Behavior changes: read before upgrading” of v2.7.11.md, with their tests under “Build and test changes”. The English and the Japanese files were compared: the same sections, and the same two entries with the same lines in the same order. MODULE_VERSION of master is 2.7.11. The callgrind comparison of v2.7.10 (8bfe5f987) with a local commit that has the tree of 7f0f1c564 (tree 012fb1376), in the Docker image ngx-mruby-dev:22.04 with sh test/perf/compare.sh (N=20000 WARMUP=2000), has all 14 scenarios ok; the largest change is sub_request, from 47886 to 47847 Ir per request without GC (-39, -0.08%). The whole build matrix ran on master at 7f0f1c564 (run 44036969023, workflow_dispatch): 17 cells and ci-ok (manual) passed. Before the fixes were merged, the nightly run on master at 4e6494d1c (run 43998959200) passed every cell and ci-ok (valgrind).
  • e84d90bd0 (master, for 2.7.11, from the private fork of the advisory GHSA-2gcp-h74q-c9g6): deleting a request or response header (Nginx::Headers_in#delete, Nginx::Headers_out#delete and the update by []=) removes the element in place, without moving the other elements, keeps headers->last and headers->nalloc consistent with the parts of the list, and matches the exact name. Its test is test/t/cases/header_del.rb (port 18105). It also adds docs/releases/v2.7.11.md and .ja.md and sets MODULE_VERSION of master to 2.7.11.
  • 7f0f1c564 (master, for 2.7.11, from the private fork of the advisory GHSA-r7v5-479r-hmqr): Nginx::Var (the accessor and #set) fills the value of a variable that has a set_handler, such as $args or $limit_rate, before it calls the set_handler. Its test is test/t/cases/var_setter_init.rb (port 18104).
  • #850 brings both to next. src/http/ngx_http_mruby_request.c and src/http/ngx_http_mruby_var.c merge without a conflict, with the same added and removed lines as on master. The only conflict is src/http/ngx_http_mruby_module.h, where next keeps its include of src/common/ngx_mruby_code.h and MODULE_VERSION 3.0.0. The changes of #843 were already on next through #844.
  • The advisories GHSA-2gcp-h74q-c9g6 (Medium, 2.7.10 and earlier, patched in 2.7.11; the delete of a header) and GHSA-r7v5-479r-hmqr (Low, 2.7.10 and earlier, patched in 2.7.11; the assignment through Nginx::Var) were published on 2026-11-17 at 16:33 and 16:34 JST. #851 (master, documentation) links them in the two entries of docs/releases/v2.7.11.md and .ja.md, and brings the Japanese file to the conventions of “The Japanese file” in docs/releases/README.md; it was merged on 2026-11-17 at 17:10 JST as ea531ad88. The English and the Japanese parts of the text of the GitHub release v2.7.11 link the two advisories as well. #850 brings the two files to next.
  • #858 (master, for 2.7.12): Nginx::Headers_in#[]= and Nginx::Headers_out#[]= set the lowcase_key and the next of the header that they add, so that nginx’s proxy and grpc modules read no unset value when they look the request headers up; the lowcase_key is as many zero bytes as the name, and the hash stays 1. Its test is test/t/cases/headers_in_proxy.rb with test/conf/conf.d/74-headers-in-proxy.conf (ports 18099 and 18100). It adds the second entry under “Behavior changes: read before upgrading” and an entry under “Fixes” of docs/releases/v2.7.12.md and .ja.md. Merged on 2026-11-18 at 01:19 JST as 49d26b879.
  • #859 (next, open; its author continues it, and the session that coordinates the work merges it): merges master at 49d26b879 (2.7.12, #858) into next. src/http/ngx_http_mruby_request.c merges without a conflict, with the three changes of #858 in ngx_mrb_set_request_header; the only conflict is the list of ports of AGENTS.md, which gets 18099-18105.
  • The check before the 2.7.12 tag, recorded here because it needed no pull request (“Before the tag” in docs/releases/README.md): the pull requests merged into master after v2.7.11 are #851 (no entry; the ids of the advisories in the 2.7.11 notes and the conventions of the Japanese file), #852 (no entry; corrections of the 2.7.11 notes and the list of ports of AGENTS.md), #853 (entry 1 under “Behavior changes: read before upgrading”, “Fixes”, “Build and test changes”, “Upgrading”) and #858 (entry 2 under “Behavior changes”, “Fixes”, “Build and test changes”, “Upgrading”). The English and the Japanese files have the same sections and entries, and MODULE_VERSION of master is 2.7.12. The callgrind comparison of v2.7.11 (7f0f1c564) with 49d26b879 (sh test/perf/compare.sh in the image ngx-mruby-dev:22.04, N=20000 WARMUP=2000) has all 14 scenarios ok; the largest change is headers, from 34933 to 35022 Ir per request without GC (+89, +0.26%). Every cell ran on 49d26b879 in the run (run 44139934199, workflow_dispatch): 17 cells of the build matrix and ci-ok (manual) passed.
  • The published advisory GHSA-2gcp-h74q-c9g6 was updated on 2026-11-17 at 19:45 JST with the description of one more path that 2.7.11 fixes (the delete of the first header of a full last part); its severity and its versions did not change.
  • #852 (master, documentation): corrections of docs/releases/v2.7.11.md and .ja.md, and ports 18104 and 18105 in the list of ports of AGENTS.md; merged on 2026-11-17 at 19:40 JST as dd6f452fc.
  • #853 (master, for 2.7.12): a HEAD, 204 or 304 response through an mruby output header filter or body filter keeps its connection, as #848 and #849 did on next, and the branch of the body filter for a response without a length passes nginx’s later calls on, so that a header that needs more than one write goes out in full. It also stops passing NULL to the copy of an empty body in ngx_http_mruby_read_body, which next already did. It creates docs/releases/v2.7.12.md and .ja.md and sets MODULE_VERSION of master to 2.7.12. Its tests are test/t/cases/head_header_filter.rb and header_only_keepalive.rb (port 18112). Merged on 2026-11-17 at 21:10 JST as 14bdc5fa6.
  • #856 (next, merged on 2026-11-17 at 22:18 JST as e8643df33): merges master at 14bdc5fa6 (#852, #853) into next, and makes the branch of ngx_http_mruby_body_filter for a response without a length pass nginx’s later calls on for a response that ends at the header (ngx_http_next_body_filter(r, NULL)), the change of #853 that #849 did not have. Its first commit adds the case of #853 with limit_rate 50 to test/t/cases/head_header_filter.rb.
  • #854 (next, section 2.3 and decision 3 of docs/proposals/core-alpha1.md): one function, ngx_mrb_var_assign, assigns the variables of Nginx::Var#<name>=, Nginx::Var#set and the end of an mruby_set handler that waited, and an assignment to a variable that nginx does not know raises NameError, one to a variable that nginx does not let scripts change or keeps no value of raises RuntimeError. Its test is test/t/cases/var_assign.rb (port 18107). It is ready: the local checks are quoted in its body, and an agent that did not write it reviewed it with no “must fix” item. The CI run on its head, the run of the perf job and the merge are recorded on the pull request; the session that opened it continues it.
  • The check before the 2.7.10 tag, recorded here because it needed no pull request (“Before the tag” in docs/releases/README.md): the release notes v2.7.10.md and .ja.md list under “Fixes” #827, #829 and the fix of a shared file handler recompiled by a location without mruby_cache (below), with their tests under “Build and test changes”; #824 changes documentation only. The callgrind comparison of v2.7.9 with the tree of dd147ab83 has all 14 scenarios ok; the largest change is ruby_call_10, from 36203 to 36214 Ir per request without GC (+11, +0.03%). The two commits of that fix were merged into master as 8bfe5f987 without a pull request that runs CI, so the whole build matrix ran on master at 8bfe5f987 after the merge and before the tag (run 43831182947, gh workflow run test.yml --ref master): 17 cells and ci-ok (manual) passed, and the step “Check the valgrind reports” of build (1.31.8) passed with ERROR SUMMARY: 0 errors from 0 contexts. The release was published after that run.
  • #829 (master, for 2.7.10): a location with mruby_add_handler on and mruby_cache on ended the worker on a request to a script under it, because the code that mruby_add_handler makes from the request path on every request ran as cached and reached the fiber without a compiled proc. The content handler now runs that code as not cached. Its test is test/t/cases/add_handler_cache.rb (port 18120).
  • 8bfe5f987 (master, for 2.7.10; ecf876042 and dd147ab83): a file-based mruby handler shared by locations with different mruby_cache settings (a server-level handler without the cache keyword, inherited by a location with mruby_cache on and one without) keeps its recompiled proc valid for the location that caches it; before, that location could run a proc that was no longer valid. master keeps the recompiled proc with mrb_gc_register. Its test is test/t/cases/shared_handler_cache.rb (port 18121).
  • #830 brings both to next. The change of #829 to the content handler merges as on master. For the shared handler, next keeps its own code: since #825, ngx_mrb_code_reload in src/common/ngx_mruby_code.c puts a recompiled proc in the GC root slot of its code, so the change of master to ngx_http_mruby_state_reinit_from_file and ngx_mrb_code_t, which are not on next, is not brought in; the test of master comes to next and runs on next’s code. MODULE_VERSION of next stays 3.0.0.
  • The advisory GHSA-2hc3-r8g9-9f9x (Low, 2.7.9 and earlier, patched in 2.7.10) was published on 2026-11-16 at 18:14 JST. It is the shared file handler above, fixed on master by 8bfe5f987; on next, the GC root slot of #825 keeps the recompiled proc, as the previous item says. #831 (master, documentation) links it in the two entries of docs/releases/v2.7.10.md and .ja.md on that fix, and on 2026-11-16 at 18:48 JST the English and the Japanese parts of the text of the GitHub release v2.7.10 got the same reference. #830 brings the two files to next.
  • #825 (next, merged on 2026-11-16 at 18:03 JST) moves the code of a directive (ngx_mrb_code_t), its compile and the reload of a handler file without cache, which the http and the stream modules had as two copies, into src/common/ngx_mruby_code.{h,c}, the row of section 2.6 of the proposal for the rest of the core before alpha.1. The compiled proc is kept in a slot of the GC root holder until the mrb_state is closed, and a reload replaces it there. No behavior changes. Its test is test/t/cases/code_reload.rb (port 18125).
  • #826 (next, merged on 2026-11-16 at 13:52 JST): the execution context and the phases of the Ruby API, the row of section 2.1 of the proposal for the rest of the core before alpha.1.
  • #832 (next): Nginx::Request#store, the Hash of a request that every handler and filter of the request gets and that stays after an internal redirect, the row “Per-request store” of section 2.1 of the proposal for the rest of the core before alpha.1. Its test is test/t/cases/request_store.rb (port 18094).
  • #834 (next, merged on 2026-11-17 at 01:23 JST as 5b65b1277): the directive table and one duplicate check, the row of section 2.5 of the proposal for the rest of the core before alpha.1. The directives with Ruby code of each module are rows of an X-macro table, and ngx_mrb_code_set sets the code of every such directive of both modules. Behavior change: nginx -t refuses a stream hook given twice in one context (decision 4). Its test is test/t/cases/stream_duplicate_hooks.rb (nginx -t only, port 12398).
  • #835 (next, merged on 2026-11-17 at 04:25 JST as a739bda78): the filter handlers, the row of section 2.4 of the proposal for the rest of the core before alpha.1. The header filter handler has the type of nginx’s header filters, the merge of the location configuration copies the filter handlers with their code, and postconfiguration adds only the phase handlers that a directive uses. Behavior change: an output filter given at http or server level, or in a location, runs in the locations below it. Its test is test/t/cases/filter_inherit.rb (port 18108).
  • #836 (next, merged on 2026-11-17 at 02:27 JST as 558a2c3a8): the check of the published site after a deploy (site/scripts/screenshots-live.mjs in the verify-live job of site.yml) checks a browser context again, once, when its only errors are 5xx answers of ngx.mruby.org, after the URLs answer 200 (at most 240 seconds in a run). From 2026-10-29 to 2026-11-16, 8 of the 14 failed or re-run runs of site.yml on next failed on such a 503 alone. Its test is site/scripts/test-screenshots-live.mjs, in npm run check:ci.
  • #843 into master as 4e6494d1c on 2026-11-17 at 04:02 JST, and #844 into next as 89c61e3d2 on 2026-11-17 at 04:38 JST: github/codeql-action 4.39.3 for init, analyze and upload-sarif together, and a Dependabot group that updates these steps in one PR per branch.
  • #847 (next, merged on 2026-11-17 at 06:48 JST as f9ea754a5): in a location with an mruby header filter and no body filter, ngx_mruby no longer sets r->filter_need_in_memory, so nginx no longer reads the files of the responses there into memory before it sends them. Since #835, a header filter given at http or server level runs in every location below it, so one such filter made nginx read every static file of the server into memory (the note of the review of #835). With sendfile on, nginx sent such a file with sendfile already, from the memory buffers that kept its position. Behavior change: such a location no longer reads the files of its responses into the buffers of output_buffers. Its test is test/t/cases/filter_sendfile.rb, with the mode http_sendfile of the second nginx of test/t/cases/_second_instance.rb (port 18116). It replaces #846, closed, whose regression test passed before the fix.
  • #848 (next, merged on 2026-11-17 at 16:16 JST as b2c312f50): a HEAD request to a location with an mruby header filter and no body filter, or to a location with an mruby body filter whose response has no Content-Length, no longer closes its connection after the header. Both paths returned NGX_ERROR when nginx had set r->header_only for the HEAD response, and nginx closed the connection that the header kept alive; since #835, a header filter at http or server level put every HEAD request of the server on the first path. They now return the result of sending the header, as the known-length body filter path does. Behavior change, next only; the fix of the known-length path is on master as well (2.7.5). Its test is test/t/cases/head_header_filter.rb (port 18112).
  • #849 (next, merged on 2026-11-17 at 17:52 JST as e9c214de2): a 204 or 304 response that passes an mruby output header or body filter, such as the 304 of a conditional GET of a static file or return 204;, no longer closes its connection after the header. The three places that pass the header on (ngx_http_mruby_header_filter_send, the branch of ngx_http_mruby_body_filter for a response without a length, and ngx_mrb_finalize_body_filter) returned NGX_ERROR when nginx had set r->header_only for a request other than HEAD; they now return the result of sending the header for every method and pass no body on; the branch without a length marks the dropped buffers as sent, so that the copy filter can read the rest of a file larger than output_buffers instead of waiting for send_timeout (also for HEAD since #848). Since #835, a header filter at http or server level put every conditional GET of a static file that nginx answered with 304 on the first place. Behavior change, next only. Its test is test/t/cases/header_only_keepalive.rb (port 18112).
  • The advisory GHSA-8j8w-mcp5-xp3h (Low, patched in 2.7.8) was published on 2026-11-15 at 15:39 JST. #817 (documentation of master) links it in the entry of Nginx.redirect of docs/releases/v2.7.8.md and .ja.md, and makes line 23 of docs/releases/v2.7.7.ja.md use the word アドバイザリ. The texts of the GitHub releases carry the same changes: the release v2.7.8 links the advisory in English and in Japanese, and the Japanese part of the release v2.7.7 uses アドバイザリ. #818 brought the three files to next.
  • #816 (master, for 2.7.9): Nginx::Upstream#keepalive_cache, #keepalive_cache=, #server and #server=, and Nginx::Stream::Connection#upstream_server and #upstream_server= raise ArgumentError on an object that holds no upstream (one created with allocate, or one whose initialize raised); up to 2.7.8 they returned nil or ended the worker process. initialize keeps the context only when it found the upstream and its server. The case is test/t/cases/upstream_uninitialized.rb with test/conf/conf.d/97-upstream-uninitialized.conf (port 18093) and test/conf/conf.d/stream/53-upstream-uninitialized.conf (port 12397), the same ports on both branches. On next, the case builds the expected text of the exception with ArgumentError#inspect of the runner, as argument_types.rb does, because mruby 4.1 shows an exception as #<ArgumentError: message>. The entries are in docs/releases/v2.7.9.md and .ja.md, which #816 created. 2.7.9 is released (above).
  • #819 (master, for 2.7.9; tests and documentation, no file under src/): upstream_uninitialized.rb also calls the upstream methods on copies made with dup and clone of an object that .new returned, on the servers of ports 18093 and 12397: they raise ArgumentError, and the original keeps its upstream. The entries of #816 in docs/releases/v2.7.9.md and .ja.md and the sections of Nginx::Upstream.new and Nginx::Stream::Connection.new in docs/class_and_method/README.md name the copies. On master, the server of argument_types.rb moved from port 18200 to 18092 (its port on next), and worker_connections of test/conf/nginx.conf is 1024, as on next. “Writing tests” of AGENTS.md and docs/test/README.md say that a new case takes a port that is free on both branches (git grep on origin/master and origin/next), from 18080 to 18199 for HTTP and from 12345 to 12399 for stream, ports below 18110 included; #820 brought the rule to next, where it replaced the ranges 18110-18199 and 12353-12399 of those texts and of docs/ja/test/README.md.
  • #822 (master, for 2.7.9): MODULE_VERSION in src/http/ngx_http_mruby_module.h, which Nginx.module_version returns and the mechanism enabled line of error.log shows, is 2.7.9 instead of 2.7.1; 2.7.2 to 2.7.8 gave 2.7.1. A new assertion of test/t/cases/module_and_request.rb fails when the value differs from the newest docs/releases/vX.Y.Z.md, and docs/releases/README.md says that the pull request that creates that file sets the value, and that “Before the tag” checks it. The entries are in docs/releases/v2.7.9.md and .ja.md. On next, where the newest file is v3.0.0.md, #823 set the value to 3.0.0 and added the entry of that value to docs/releases/v3.0.0.md and .ja.md, and docs/ja/releases/README.md got the two passages of #822. 2.7.9, the release of #822, is published (above).
  • #824 (master, documentation): docs/install/README.md and README.md of master say that the stream module of 2.x needs nginx 1.25.5 or later since v2.7.0, and how to build 2.x with nginx 1.24 without --with-stream. On next, #828 keeps the text of next: the install guide of next is for 3.x, which supports nginx 1.30 and 1.31 and stops with nginx 1.24 also without --with-stream, and its “Prerequisites” already say what 2.x does with nginx 1.24; the section of README.md that #824 edits is not on next.
  • #827 (master, for 2.7.10): an SSI page that includes, after another include, a location with an mruby body filter whose body was written with Nginx.rputs or Nginx.echo gets its whole response; before, the response stopped until send_timeout. ngx_mrb_finalize_rputs and the body filter keep a context for a subrequest whose output nginx holds in r->postponed. The cases are three more pages of test/conf/conf.d/91-ssi-body-filter.conf in test/t/cases/ssi_body_filter.rb; the notes are in docs/releases/v2.7.10.md and .ja.md (under “Fixes”; no behavior change), and docs/directives/README.md has a paragraph on SSI includes. MODULE_VERSION of master is 2.7.10; on next it stays 3.0.0. #828 brought the fix to next, with the same condition in both places of src/http/.
  • The check before the 2.7.9 tag, recorded here because it needed no pull request (“Before the tag” in docs/releases/README.md): the pull requests merged into master from v2.7.8 (976015c48) to a7800d865 are #817, #816, #819 and #822. #816 has its entries under “Behavior changes”, “Fixes” and “Build and test changes” in v2.7.9.md and .ja.md; #819 adds the copies made with dup and clone to the “Behavior changes” and “Fixes” entries of #816, and has its own entries under “Build and test changes”; #822 has its entries under “Behavior changes”, “Build and test changes” and “Upgrading”; #817 changes the notes of 2.7.7 and 2.7.8 only and needs no entry. The first check, on 8f204e061, found two points, which #822 corrected: a line of v2.7.9.ja.md with parentheses around words that are not identifiers, and MODULE_VERSION in src/http/ngx_http_mruby_module.h, which was still 2.7.1. The check was repeated on a7800d865: every “Behavior changes” entry has every line of the template, the two files have the same sections and entries in the same order and say the same, and MODULE_VERSION is 2.7.9, the version of the tag. Nothing else needed a correction, so no check pull request was opened. The callgrind comparison of v2.7.8 with a7800d865 has all 14 scenarios ok; the largest change is sleep, from 15138 to 15139 Ir per request without GC (+1, +0.01%), and no other scenario changes by more than 1 Ir. The whole build matrix ran on the public master at a7800d865 before the tag (run 43725196090, gh workflow run test.yml --ref master): 17 cells and ci-ok (manual) passed, and the step “Check the valgrind reports” of the cell under valgrind (build (1.31.8)) passed with ERROR SUMMARY: 0 errors from 0 contexts.
  • The check before the 2.7.8 tag, recorded here because it needed no pull request (“Before the tag” in docs/releases/README.md): the pull requests merged into master from v2.7.7 (f736bae0c) to 976015c48 are #811, #812, #814 and #813. #813 has its entries under “Behavior changes”, “Fixes” and “Build and test changes” in v2.7.8.md and .ja.md; #811, #812 and #814 change documentation only and need no entry. The two files were compared, and nothing needed a correction, so no check pull request was opened. The callgrind comparison of v2.7.7 with 976015c48 has all 14 scenarios ok; the largest change is proxy_stream_plain_50, a scenario that runs no Ruby code, from 54311 to 54379 Ir per request without GC (+68, +0.13%). The whole build matrix ran on the public master at 976015c48 before the tag (run 43678321490, gh workflow run test.yml --ref master): 17 cells passed.
  • #812 and #814 (documentation of master) add to the release procedure what the 2.7.7 release showed (“Before the tag”, “The GitHub release” and “Security fixes” in docs/releases/README.md, AGENTS.md, docs/test/README.md and the PR template): the check before the tag opens a pull request only when something needs a correction, and is otherwise recorded in this file; every cell of the build matrix runs on the tip of the base before the tag; the session starts site.yml on next right after the release and checks the home page; the commit messages, test names and comments of a fix in an advisory’s private fork are neutral and name no advisory id; the private pull request checks the release notes it changes as 3.x pages, in a local commit of a checkout of next (#814), and runs the suite on the oldest nginx of the matrix with the static module and on the nginx of the dynamic cell; and the session checks the patched versions and the text of an advisory draft against the release before the advisory is published. On master, a pull request with build standard or full also runs the suite with the static module on the oldest nginx of the matrix (1.26.3), which its four cells do not run. next does not take that rule: its oldest nginx, 1.30.6, is the nginx of the dynamic standard cell. The merge of master into next writes the rest into next’s documents and their Japanese pages.
  • #813 (master, for 2.7.8): Nginx.redirect with a URL and a second argument that is not an Integer raises TypeError and makes no redirect, a third argument raises ArgumentError, and Nginx::Server#add_listener and Nginx::Stream.add_listener raise TypeError for a missing :address or one that is not a String, which stops nginx while it reads the configuration. Its case is test/t/cases/argument_types.rb with test/conf/conf.d/96-argument-types.conf. Its port is 18200 on master and 18092 on next, where the ports of the suite are 12345-12399 and 18080-18199, the ranges that TEST_SH_PORT_OFFSET moves. On next, the case builds the expected text of the exception with TypeError#inspect of the runner, because mruby 4.1 shows an exception as #<TypeError: message>. The entries are in docs/releases/v2.7.8.md and .ja.md.
  • The release 2.7.7: the tag v2.7.7 is on f736bae0c, and the release was published on 2026-11-15 at 02:06 JST. Two advisories were published with it at 02:43 JST: GHSA-cjxv-4q9m-62pf (Moderate; the setters of Nginx::Request, the path of Nginx::Async::HTTP.sub_request and Nginx::Headers_in/out#[]= read an argument that is not a String without checking its type; they now raise TypeError) and GHSA-pmjr-f94r-j47g (Low; the store of Nginx::Upstream#server= and Nginx::Stream::Connection#upstream_server= across a reload with master_process off). Both fixes were merged from their advisory pages into master (7a2724265, f736bae0c) and into next (4fbe72294, fef20e671, in next’s form, with the stores in src/common/ngx_mruby_upstream.c and the test on port 18172 instead of 18192, which exception_log.rb uses on next). The tree of f736bae0c is the tree of the head of the second fix’s pull request, which contains the head of the first. The check before the tag: the pull requests merged into master from v2.7.6 to cc829db6e (master before the fixes) are #800, #801, #803, #804, #807 and #809; the release notes needed no correction for them. The callgrind comparison of v2.7.6 with 7a72306bb has all 14 scenarios ok; the largest change is -2 instructions per request without GC. The whole build matrix at 7a72306bb (run 43594266917) failed in the four cells of nginx 1.26.3: nginx before 1.27.1 does not accept a stream server without a content handler, which the stream servers that #803 added to test/conf/conf.d/stream/52-exception-log-script-methods.conf were. #809 gave them a return directive (a change of the test configuration only), and its run on its head passed in all 17 cells. The push of the tag ran the whole matrix at f736bae0c (run 43619287702): 17 cells passed. #811 (91499f767) added the links of the two advisories to docs/releases/v2.7.7.md and .ja.md, and reworded two lines of the Japanese file whose last line held two characters on the site (check-japanese-lines.mjs); the release body links both advisories in English and Japanese. The site: site.yml was started on next after the release at 03:27 JST (run 43626723963, success), and the home page of ngx.mruby.org names v2.7.7 since 03:31 JST.
  • #807 (merged into master on 2026-11-14 at 19:02 JST as 7a72306bb) brings the decision of 2026-11-12 under “Decisions” (local checks first) to the documents of master: AGENTS.md, .github/PULL_REQUEST_TEMPLATE.md, docs/test/README.md and docs/DEVELOPMENT.md. Its two reviews found that the local list missed the files job and gem-heads, that CodeQL ran in CI with no reason written down, and that the condition of the sanitizer and valgrind runs (“C code and build changes”) was not defined; #807 fixed these on master. #808 (merged on 2026-11-14 as dc6a6a6cd) merged it into next and made the same fixes in next’s text: the check of the added files and gem-heads in the local list, CodeQL as the one check kept in CI with its reason, the sanitizer and valgrind runs when test/ci/changes.sh gives build standard or full, the OpenSSL of the docker forms (3.0.2 of Ubuntu 22.04) against the OpenSSL 3.5.9 of the cells, and how to read vm.mmap_rnd_bits from a container. On master, the auto-ssl case is a second check kept in CI for changes outside its paths; next has no auto-ssl gem or case.
  • The site: the deploy of ngx.mruby.org works again. The run of site.yml on the push of 9cf8f8a4e (run 43188350847) had failed in node scripts/check-japanese.mjs on docs/releases/v2.7.6.ja.md and docs/ja/releases/README.md. #804 (base master) and #805 (base next) fixed both pages, and #805 added to “Site CI” of docs/test/README.md (and its Japanese page) that a merge of master into next runs npm run check:ci on its own head, and to condition 2 of “Review and merge” in AGENTS.md that the session that merges reads the run of site.yml that the merge started. The run of site.yml on the push of 246877177 (run 43572439640) succeeded.
  • #803 (merged into master on 2026-11-14 at 17:00 JST as 66b284a94) brings to 2.x the protection of the inspect and backtrace calls that ngx_mruby makes when it writes an exception to error.log (the 2.x counterpart of #796 and #799), with its cases (test/t/cases/exception_log_script_methods.rb, ports 18091, 12383 and 12389) and the entries of docs/releases/v2.7.7.md and .ja.md. Its re-review found no must fix and no should fix. #806, which merged it into next, keeps next’s src/ (which has the same protection in src/common/ngx_mruby_log.c), runs the new cases on next’s code, and adds the three ports to the list of AGENTS.md. Its body has the state and the checks.
  • The release 2.7.6: the tag v2.7.6 is on a8ef5ecb8 (2026-11-12 15:53 JST), and the release is published. Two advisories were published with it at 15:58 JST: GHSA-r6wq-r8rh-6rc3 (Moderate; an HTTP log handler’s Nginx.rputs/Nginx.echo could write an extra response on a persistent connection) and GHSA-79vp-37cx-547v (Moderate; an mruby content handler or body filter in an SSI subrequest could split the parent HTTP response). The release body links both advisories in English and Japanese. #801 (298e0dd11) added the two advisory ids to v2.7.6.md, .ja.md and the release body; its independent review found no must-fix, the 2.x pages check passed, and ci-ok passed. #800 (aa3d9798a) added to the “Security fixes” of docs/releases/README.md how a security fix is merged from the advisory page and the temporary protected-lines bypass that merge needs. Both #801 and #800 are documentation only, and reached next with #802. The check before the tag: the pull requests merged into master from v2.7.5 to the release are #768, #772, #773, #774, #775, #778, #781, #784, #786, #790 and #792; each has its entry under “Behavior changes”, “Fixes” or “Build and test changes” in v2.7.6.md/.ja.md, or changes nothing a configuration, script or build observes (docs, CI and test.sh), so no separate public check pull request was opened. The whole build matrix ran at 65ce972e1 (master before the security fixes) in run 43018470114 (success); the tag push ran the whole matrix at a8ef5ecb8 in run 43171298562 (success). The callgrind comparison of v2.7.5 (801d78b48) with the release tree (master + both security fixes, 13195fc71) has every scenario ok, the largest change ruby_call_10 at +0.84% without GC (below the 3% WARN threshold); the fixes add a subrequest branch to three finalization sites and one read-body check and touch no measured scenario’s hot path. The two fixes were merged from their advisory pages (the log-handler fix, then the SSI subrequest fix) with the protected-lines admin bypass set to always for the merges and restored to “pull request only” afterwards.
  • The release 2.7.5: the release was published on 2026-11-04 at 14:49:05 JST. Its tag points to 801d78b4831965326adb1f9b216a99740b5ab92e. GHSA-wx6w-gpg3-cmc3 was published at 15:08:25 JST (Moderate, CVSS 3.1: 5.3): additional datagrams could restart a sleeping UDP stream handler and postpone its pending wait. The advisory confirms 2.7.4 as affected and 2.7.5 as fixed; earlier versions have not been established. The release and advisory contain both English and Japanese text. Public access and reference links were checked. The site’s release-version update passed. The GitHub release body links the advisory in both languages. The release-note references for master were merged in #763 at 15:58:32 JST as 1a2e97ee1aa1426ade91d833be6472842b81aa04. Its independent review and ready CI passed; the automatic site update started. The 2.7.5 changes reached next in #765.
  • The release 2.7.3: the tag v2.7.3 is on 591147b01 (2026-11-01 00:08 JST), and the release is published. Two advisories were published with it: GHSA-86wx-fcwx-m3vh (High; the stream Nginx::Stream::Async.sleep use-after-free, fixed by #645) and GHSA-8r78-rxqq-5gp5 (Low; a string with a NUL byte in Nginx::Upstream#server= and Nginx::Stream::Connection#upstream_server=, fixed by 591147b01, entry 9 of the 2.7.3 release notes). #699 added the two advisory ids to the 2.7.3 release notes.
  • The release 2.7.4: the tag v2.7.4 is on e2aeb5ddd (2026-11-02 01:04 JST), and the release is published. One advisory was published with it: GHSA-xjh6-6vgm-56wj (Moderate; the stream module: the reads of a session whose code waits in Nginx::Stream::Async.sleep on a listen ... proxy_protocol, and proxy_protocol_addr for a connection without a PROXY source address). master has it as entries 1 and 2 of the 2.7.4 release notes (Japanese), which reach next with the merge of master at 0f865ed4a; next got the same fix in its own form, with the two entries in the 3.0.0 release notes. The check before the tag: the pull requests merged into master from v2.7.3 to the fix are #699 (the advisory ids of 2.7.3, docs only), #704 (CI, its entry under “Build and test changes”), #703 and #716 (docs only); none changes an input of the measured binary, and the fix changes src/stream/ only, which the callgrind lane does not build (--with-stream is not given), so 2.7.4 and 2.7.3 measure as the same binary. “Security fixes” in docs/releases/README.md asks for a public check pull request before the fix; none was opened, although v2.7.4.md had existed since #704 and the check could have been made before the fix. The check was made with the private pull request of the fix instead, and is recorded in #731, which adds the advisory id to master.
  • #729 (the terms of the AI gateway on the site) was merged at 01:33 JST, and #733, the should-fix items of its review, at 02:14 JST.
  • #731 (base master, 01:51 JST) and #732 (base next, 01:56 JST) are merged: the advisory id of 2.7.4, and the owner’s decision of 2026-11-02 that a session publishes the release after the owner’s go-ahead, written into AGENTS.md, CLAUDE.md, docs/releases/README.md and docs/DEVELOPMENT.md.
  • #734, the merge of master at 0f865ed4a into next, is merged (3dfe8d05f).
  • #737 (base next) is merged: site.yml runs on the push to next only, without the browser lanes, and npm run check:ci in site/ runs its checks and every browser check locally (the owner’s decision of 2026-11-02 under “Decisions”).
  • The core stabilization is done with #735, and its proposal is archived (the row of step 4 under “Queue”).
  • #739 (base master) is merged (5440e7a0e): site-v2.yml runs only on workflow_dispatch, and a pull request to master checks the 2.x pages locally. It is included in next through #746.
  • #740 is merged on master: a rewrite handler that returns Nginx::DECLINED after Nginx::Async.sleep keeps the status of the request.
  • #751, the repeated HTTP sleep regression, merged as 63c6b7152 at 03:41 JST. Its local validation, independent review, CI and publication checks passed.
  • #756 merged the HEAD connection fix into next as f1f102966 at 04:09 JST. Its combined suite passed 759 cases, local soak/performance/nginx-tests/site/example checks passed, and the independent review and all 15 CI checks passed. The automatic publication and exact public revisions were verified (checkpoint).
  • #754 merged into master as 382639258 at 04:18 JST. It updates the pinned workflow-analysis Action from v0.7.1 to v0.8.0 while retaining zizmor 1.31.1 and the existing policy. The PR records the local audit, independent review and CI.
  • #757 merged the workflow-analysis Action pin into next at 04:38 JST. Publication and cleanup are complete. The equivalent next proposal in #755 was closed.
  • #758 merged stream sleep measurement as 8beaaa381 at 05:10 JST. The 13-scenario soak, 16-scenario performance comparison, independent review, CI and publication checks passed; cleanup is complete (checkpoint).
  • #759 merged the public 2.7.5 release-note check as f058927cb at 05:30 JST. The seven preceding PRs and English/Japanese notes were compared. Local checks, the 14-scenario cumulative performance comparison, all 17 build configurations, independent review and publication checks passed; cleanup is complete (checkpoint). The check did not publish a release or create a tag.
  • #760 merged as 60ba72f86 at 14:16 JST. It preserves every copper CTA edge and balances the Japanese hero text and logo. The automatic site publication passed.
  • #765 merged master at 1a2e97ee1 into next as ec8efde7c on 2026-11-04 at 17:08 JST. It keeps the fiber state machine of next for the UDP waiting behavior and brings the expanded regression tests, the current references and both 2.7.5 release-note files.
  • #767 merged into next as ea3c5996e on 2026-11-09 at 18:33 JST. It adds the proposal for the rest of the core before alpha.1.
  • #766 merged into next as 46bf09566 on 2026-11-09 at 19:21 JST: the build_config.rb rewrite of the queue. ENV and Process come from mruby’s core gems, and mruby-vedis and mruby-localmemcache leave the default build. The owner accepted its perf row stream_sleep (FAIL, +6.05% without GC) on the PR. The replacement of SecureRandom and Digest is the queue row that follows it.
  • #768 merged into master as 40578dd63 on 2026-11-09 at 19:25 JST: the 2.x form of the protected conversion with to_s of an mruby_set value and of the values of the setter of Nginx::Var and Nginx::Var#set. Its entries are in docs/releases/v2.7.6.md and .ja.md. On 2.x, a value that cannot be converted leaves the variable empty before a wait, as an exception of the code does; next answers 500 there (#735).
  • #769 merged master at 40578dd63 into next. It keeps the implementation of next in src/, brings test/t/cases/var_value_to_s.rb on port 18189 with the expectations of next for the values that the handler returns, and rewrites the “Before” line of the #735 entry in the 3.0.0 release notes from what #768 ran on 2.x. Its merge commit is e8d9ff52f (2026-11-09, 20:27 JST).
  • #773 merged into master as 94d8cd84e on 2026-11-09 at 20:58 JST: docs/directives/README.md of 2.x says that an mruby_set handler that waits in Nginx::Async.sleep ends the request without a response, unless an earlier handler declined, and points to 3.0.0 or to a rewrite or access handler with Nginx::Var#set. The 2.x code is not changed.
  • #772 merged into master as a52a19dcd on 2026-11-09 at 21:14 JST: on 2.x, the timer of Nginx::Async.sleep runs the posted requests of the connection after every outcome, so a subrequest that the request posts after the sleep starts at once. Its test is test/t/cases/async_sub_request_after_sleep.rb (port 18190) and its entry is in docs/releases/v2.7.6.md and .ja.md.
  • #774 merged into master as 1001b8b20 on 2026-11-09 at 21:34 JST: the category comments of test/ci/changes.sh on master match those of next, and “The 2.x pages of the site” in docs/test/README.md names the Node.js of site/.node-version.
  • #770 merged into next as d3fb11439 on 2026-11-09 at 22:28 JST: mruby log handlers run before access_log and keep the status that was sent, and a local Nginx.redirect raises in a log handler.
  • #771 merged into next as 6196a49f1 on 2026-11-09 at 23:00 JST, the rest of the build_config.rb rewrite: Digest and SecureRandom from mrbgems/digest and mrbgems/secure-random on OpenSSL 3, and OpenSSL 3.5.9 in the build matrix of CI.
  • #775 merged into master as 915076528 and #776 into next as 6b642bb0e, both on 2026-11-09 at 23:40 JST: Nginx.rputs, Nginx.echo, Nginx.redirect and Nginx::Filter#body= raise TypeError when to_s of the argument does not return a String. Their test is test/t/cases/output_value_to_s.rb (port 18191), and their entries are in docs/releases/v2.7.6.md and v3.0.0.md with the Japanese files.
  • #777 merges master through 915076528 into next, with #772, #773, #774 and #775. It keeps the implementation of next in src/ and its test/t/cases/output_value_to_s.rb, brings the test of #772 on port 18190, and leaves the paragraph of #773 out of the docs/ of next, which describe 3.0.0.
  • #778 merged into master as af51f9dbe on 2026-11-10 at 01:28 JST: the two Japanese lines of docs/releases/v2.7.6.ja.md that #777 changed on next, and “The 2.x pages of the site” in docs/test/README.md without the number of the Node.js version.
  • #781 into master as 5bb40f5d9 and #782 into next as 9a09deaeb, both on 2026-11-10 at 15:00 JST: zizmor-action 0.8.1.
  • #784 merged into master as 13153ae22 on 2026-11-10 at 20:02 JST: in a location with an mruby body filter, a response that a handler wrote with Nginx.rputs no longer stops after its first write when nginx needs more than one write to send it (limit_rate, sendfile_max_chunk, a slow client). Its test is test/t/cases/body_filter_split_write.rb (port 18141) and its entry is in docs/releases/v2.7.6.md and .ja.md.
  • #786 merged into master as 10c4cd927 on 2026-11-10 at 23:13 JST: in a location with an mruby body filter, with or without an mruby header filter, a response no longer stops after its first write when an earlier handler wrote with Nginx.rputs and returned Nginx::DECLINED, so that its output stayed in the module context. The body filter run drops that context; ngx_http_mruby_body_filter_handler_inner now gives the request an empty context that passes the writer’s later calls on. Its test is test/t/cases/body_filter_leftover_chain.rb (port 18145) and its entry is in docs/releases/v2.7.6.md and .ja.md.
  • #790 merged into master as b9e42c3ad on 2026-11-11 at 04:31 JST: the README says that 2.x builds with nginx 1.28 and 1.26, and with nginx 1.24 only without the stream module (without --with-stream). #792 merged into master as 65ce972e1 on 2026-11-11 at 14:34 JST: the lines of test.sh that make the certificates run in a subshell, and test.sh stops and names the certificate (localhost or dummy) when openssl req fails. Neither adds an entry to the release notes.
  • #787 (merged into next as bbd88868f on 2026-11-11 at 00:01 JST) merged master through 10c4cd927 into next, with #786. The PR body says why the fix works the same on next, where ngx_mrb_finalize_rputs has more branches than on master.
  • #793 (open, to next) merges master through 65ce972e1 into next, with #790 and #792. README.md conflicted in the paragraph on the promotion; next’s text is kept, because next already says in the paragraph after it that 2.x with nginx 1.24 builds only without the stream module. The lines of test.sh that make the certificates came from master.
  • #785 (merged into next as 76ebd2fac on 2026-11-10 at 21:33 JST) merged master through 13153ae22 into next, with #778, #781 and #784. A test commit before the merge, test/t/cases/body_filter_split_write_status.rb (port 18142), shows the same stop on next for output sent with a status set before or after it (#779, #783); the merge applies the fix of #784 to the branch of ngx_mrb_finalize_rputs on next that sends the output. The PR body has the state, the checks and the other places that clear the module context.
  • #799 (merged 2026-11-12 17:11 JST at 1791b8bf1, base next) protects the calls of inspect and backtrace in the logging of an exception (src/common/ngx_mruby_log.c) with mrb_protect_error, as ngx_mrb_fiber_resume protects a resumed fiber, so that a script that redefines them so that they raise does not send that exception past the nginx functions when the handler runs inside another mruby call. An exception whose inspect raises or does not return a String is logged with its class (an entry under “Behavior changes” of docs/releases/v3.0.0.md and .ja.md). Its test is test/t/cases/exception_log_raising_methods.rb (port 18198 and the stream port 12382). The PR body has the state and the checks.
  • #796 (merged into next as e6d2efdae on 2026-11-12 at 02:20 JST) moves the logging of an exception of Ruby code, seven functions of the HTTP and the stream modules, and the two copies of Nginx.log into src/common/ngx_mruby_log.{h,c}, the row of section 2.8 of the proposal for the rest of the core before alpha.1. The lines of error.log keep their text and their level; stream also logs the backtrace in a build with --with-debug (its entry under “Behavior changes” of docs/releases/v3.0.0.md and .ja.md), and an element of a redefined backtrace that is not a String is skipped (a “Fixes” entry). Its tests are test/t/cases/exception_log.rb (ports 18192, 18193 and 12380) and a case of test/t/cases/config_hooks.rb. The PR body has the state and the checks.
  • #788 (merged into next as ae8df760e on 2026-11-11 at 01:51 JST) moves the checks, the parse and the per-peer store that Nginx::Upstream#server= and Nginx::Stream::Connection#upstream_server= shared as two copies into src/common/ngx_mruby_upstream.{h,c}, the row of section 2.7 of the proposal for the rest of the core before alpha.1. No behavior changes. The PR body has the state, the checks and the table of what is shared and what each setter keeps.
  • #780 (merged into next as f14cceaeb on 2026-11-10 at 02:48 JST) adds the directive mruby_variable of section 2.11 of the proposal for the rest of the core before alpha.1, the agent proxy row on variable declarations. Its test is test/t/cases/mruby_variable.rb (port 18171), and its entries are under “New features” of docs/releases/v3.0.0.md and .ja.md.
  • #761 merged the unchanged empty-subrequest response clarification as 8cbdf7748. #762 merged the capability-based agent model and reasoning-effort guidance as 4eb847421.
  • #750 added current 3.x documentation before the existing 2.x footer link in both languages. It merged as c440a15dc; local checks, independent review, CI and publication verification passed.
  • #752 merged into master as 9acbd8530 at 02:22:44 JST. A successful known-length HEAD response finalized by the output body filter keeps an otherwise reusable connection for the following GET. Author and independent suites matched: the test-only commit failed the three new filtered-HEAD cases, and the fix passed all 362 tests. All CI checks and both performance comparisons passed; the independent review has no outstanding findings. The automatic site publication and the public 2.x revision were verified (checkpoint).
  • #753 groups the Japanese homepage introduction into two paragraphs without changing its CSS or English text. Its full site checks, independent review, CI and published-page checks passed. The integration retains that component unchanged.
  • #747, the homepage reading-order and local-trial correction, merged at 2026-11-02 23:59:24 JST as fad853820. Its local checks, independent review and CI passed. The site run passed and published fad853820. The public documentation index matched that clean revision; twelve live-page checks and the English/Japanese desktop/mobile views passed (publication checkpoint).
  • #746 imported the earlier master changes and aligned the shared recording and coordination rules. Its Linux suite passed with nginx 1.31.8 (Total 753, OK 753, KO 0, Crash 0); its full local site check passed on 1f5cc482c in 446 seconds. Independent review, all four CI build cells, ci-ok, CodeQL and zizmor passed.
  • #748 aligned the master recording rules and summarized existing decisions in CI guidance and the English/Japanese 2.7.3 release notes. Its full local site check passed in 464 seconds without skips, and independent review and CI passed. #749 brought these changes into next, retaining its CI guidance and both release-note corrections.

This file records where the work on ngx_mruby stands, so that any session, clone or contributor can continue it: the open pull requests, the decisions that are not written down elsewhere yet, the work in progress and the queue.

  • It records state, not rules. The rules are in AGENTS.md. The plan for v3 and the owner’s decisions about it are in docs/proposals/v3-plan.md. Where this file disagrees with them, they win, with the one exception that AGENTS.md “Session handoff” sets: a dated row under “Decisions” holds over their older text until a pull request writes it there, but only when the row records an adopted decision, its date, scope, rationale and constraints, and the older rule it replaces. Proposals are labeled separately and change no rule.
  • A decision of the owner that is not written there yet is recorded under “Decisions”, in short, with its date, rationale, technical constraints and the place where it will be written. Where the sessions apply a decision with a procedure of their own, the row names that procedure separately.
  • A pull request on next updates this file in the same pull request when it changes what the file records, its own row and the queue item that it finishes included. A pull request on master cannot, because the file is only on next: the session that merged it into master opens a small pull request on next that updates the file, or updates the file in the merge of master into next if that comes first. A release is recorded in the same way by the first session that finds its tag missing here. Link pull requests and issues instead of copying their text.
  • Remove an item when its pull request is merged or closed, when its issue is closed, or when it is written in the place that its row names. The pull requests and issues keep the history (see the gh commands below).
  • The file lives on next. From a checkout of another branch, read it with git fetch origin next and git show origin/next:docs/HANDOFF.md.
  • The statements about branches, pull requests and issues were checked with gh and git at the time in the next heading. Check them again before you act on them: other sessions merge pull requests all the time.

These records state the agreed requirements, rationale and technical constraints, with dates. Proposals are identified separately from adopted decisions. Conversation quotations and attribution are not a substitute for checking whether a decision fits the evidence and requirements.

  • 2026-11-15, the version of the pre-releases of 3.0.0. On next, MODULE_VERSION in src/http/ngx_http_mruby_module.h is 3.0.0 for the whole 3.0.0 cycle, and the builds of the alpha, beta and rc pre-releases add no suffix to it. Scope: the value that Nginx.module_version returns and the line ngx_mruby/<version> (mruby/<version>) mechanism enabled that nginx logs while it reads the configuration. Rationale: the value keeps the form X.Y.Z, so scripts that compare it in that form and the assertion of test/t/cases/module_and_request.rb, which accepts only X.Y.Z, stay as they are; Ruby’s RUBY_VERSION also gives the numbers only in its preview and rc builds. Constraints: the value does not tell a pre-release build from the build of 3.0.0, or one pre-release from another; the tag or the commit of the build does. The decision can change before the 3.0.0 tag. It replaces no earlier rule (there was none), and it is what #822 wrote in “Who writes the entries, and when” and “Before the tag” of docs/releases/README.md; #823 wrote the same in docs/ja/releases/README.md and in the entry of v3.0.0.md and .ja.md.
  • 2026-11-12, local checks first. CI is the final check in a clean environment. Every kind of check that CI runs for a change is run locally by the author before the PR is marked ready, and its failures are found there: the suite with the static and the dynamic module, the suite under AddressSanitizer and UndefinedBehaviorSanitizer and under valgrind (for C code and build changes), the callgrind comparison, the soak, the steps of examples/start, nginx-tests, the site’s checks and the static analysis of the workflows. On macOS the Linux-only checks run in a Linux container. CI keeps only the spread to what one machine does not hold: the other nginx versions and link modes of the build matrix, the x86-64 runner and a clean environment. Rationale: a check whose first run is in CI is found failing there, and each fix costs a push and a CI wait; when the runner itself fails, a local result of the same check is what separates a code defect from a runner problem. #796 ran all of these on an arm64 Mac in Docker while three CI attempts on the same head stopped in the package installation. This replaces the text of 2026-10-31 and 2026-11-02 that sent the sanitizers, valgrind and the callgrind comparison to CI as environments a machine does not have; the limits on CI time and the base-move condition stay. Written in AGENTS.md (“Design principles”, “Branches and pull requests”, “Review and merge”, “Build and test”), the PR template and docs/test/README.md on next, and on master in #807. The following is the sessions’ procedure, not part of the decision of 2026-11-12: #807 (on master) and #808 (on next) settled it after the reviews of #807, and AGENTS.md “Branches and pull requests” and “Build and test” hold its current text. The author runs the check of the added files for every PR, and the suite with the static and the dynamic module, under the sanitizers and under valgrind, when test/ci/changes.sh gives build standard or full. Besides what one machine does not hold, CI keeps the checks whose reason AGENTS.md writes down: CodeQL on next, and CodeQL and, for changes outside its paths, the auto-ssl case on master.
  • 2026-11-03, agent assignment. Choose Codex models and reasoning effort by role and task complexity, using currently available capabilities rather than fixed model names. Routine work uses a sufficient faster, lower-cost model; difficult analysis uses stronger reasoning. The task leader keeps scope, ownership and validation gates, with independent review. The shared coordination rules apply across operating systems and tools; Codex model controls do not become requirements for other tools. This adds selection guidance to the existing coordination procedure without replacing its tests or review and merge gates. The rule is in Assigning agent work.
  • 2026-11-02, public records. Public commits, PRs, issues and documents contain only the relevant decision and technical requirements. Working documents and handoffs, whether public or private, record intent, reasons, constraints and decisions without quoting conversations or using attribution to justify a conclusion. Do not copy private conversation text, message identifiers or personal environment details. Link an already-public decision when available. Existing public git history is not rewritten without an explicit decision; any sensitive historical content is reported privately.
  • 2026-11-02, homepage follow-up. Review the entire page, including the hero, section introductions, use cases, local trial steps, documentation links and footer. Shared styles give each content role a consistent readable width, alignment and spacing. Continuous introductory prose stays in one column; individual use-case articles may share a row. The four documentation links form a compact ordinary list. Keep meaningful paragraphs, glossary terms and technical explanations without promotional language. Inspect the complete desktop and mobile pages independently; passing mechanical checks alone does not establish visual quality. The local trial follows prerequisites, startup, request, expected result and further instructions. The homepage PR can progress independently of the master-to-next merge.
  • 2026-11-02, development coordination. Codex continues PR preparation, independent review, eligible merges, handoff maintenance, release preparation and contributor work under AGENTS.md. Independent PRs may progress in parallel with explicit ownership, remaining work, blockers and next actions. Serialize dependencies, conflicting edits and competing expensive checks; use agents where they shorten completion time without duplicating work. Review and merge gates remain unchanged. Release publication needs permission for that release; advisory publication remains with the owner.
  • 2026-11-02, release publication. After the release check and permission for that release, the session publishes the release and its tag; the owner publishes advisories. #731 and #732 record the procedure in AGENTS.md, docs/releases/README.md and docs/DEVELOPMENT.md. The earlier permission of 2026-10-29 applied to 2.7.2 only and was carried out by #624.
  • 2026-11-02, CI. Local checks must be thorough; CI covers other build environments and deployment without repeating local browser checks. PRs run neither site.yml nor site-v2.yml. The push to next builds, checks and deploys the site; a push to master dispatches that run. npm run check:ci supplies the local site and browser checks. A changed base requires another PR run only when both sides changed code or tests. The build matrix, sanitizers, valgrind and performance lane remain. #737 and #739 implement this decision. The category comment remains in Queue.
  • 2026-11-02, terminology. Models, agents, their requests and the AI gateway use the terms in the glossary and section 7 of the site design. Japanese uses katakana for the glossary terms and preserves identifiers in English. The homepage heading follows the same terms.
  • 2026-11-01, site implementation. Use Astro and repository-owned components in place of Starlight, as section 5.4 of the site design describes (#700, #708, #713). Work that can be completed together is not split into unnecessary stages. The mobile layout pattern is 402px; root sizing follows section 4.4 of that design.
  • 2026-11-01, build coverage and next work. Code/test PRs run four standard matrix cells: nginx mainline static, stable dynamic, the sanitizer cell and the valgrind cell. Nightly runs, tag runs and the pre-tag check with full-matrix retain full coverage (#702, #704). The harness extension for multiple workers and reload precedes the core stability work. Discussions are enabled and the wiki is disabled.
  • 2026-11-01, handler semantics. Q1 to Q6 are recorded in section 9 of the core stabilization proposal (#710): support the specified waits, apply redirects at handler completion, preserve HTTP/1.x continuation after client close and discard a waiting fiber without running ensure when its request ends, execute the UDP handler once while it waits, raise RuntimeError for sub_request inside mruby_set while preserving sleep, and isolate handler results so a server rewrite without a result advances to the next phase.
  • 2026-10-31, local verification. Test locally before making a PR ready; CI is not the place to discover avoidable implementation errors (#692, #693). CodeQL excludes vendored upstream trees, whose existing alerts were dismissed as upstream code rather than reported without a confirmed defect and patch.
  • 2026-10-31, mruby performance. Keep mruby’s default GC settings after the measurements failed to remove the extra GC work without increasing other work. The accepted mruby 4.1 differences are recorded in #678 and “Known differences” in the test guide. #696 implements the specific 4.0-to-4.1 auth allowance in test/perf/accepted.txt; it does not authorize a 3.3-to-4.1 allowance before the first v3 tag. Other WARN/FAIL results still require the owner’s decision.
  • 2026-10-31, upstream reports. Report a confirmed mruby defect only when it warrants an upstream fix beyond ngx_mruby’s convenience and a patch can be provided. This belongs in AGENTS.md’s other-repository rules when that subsection is next updated.
  • 2026-10-31, completed 2.7.3 publication. The separately published stream advisory and the regression fix of #698 shipped with 2.7.3. The published releases and advisories are linked in State.
  • 2026-10-29, maintained gems. Update owner-maintained gems when that resolves compatibility; removal is acceptable when an underlying library is no longer maintained. AGENTS.md records this rule (#637, #639).
  • 2026-10-29, v3 scope. The decisions on mruby 4.1.0-rc2, removal of auto-ssl from next, agent-proxy work, subrequest empty responses and empty output filters are in items 1, 6 and 15 of section 7 of the v3 plan. GitHub releases include the behavior changes in English and Japanese. HANDOFF remains English with a Japanese notice.
  • 2026-10-04, async behavior, amended 2026-10-29. Support sleep in a content handler. On next, a rewrite handler with neither output nor an explicit result proceeds to the next phase (#720); master keeps the compatible 2.x behavior. The broader resume rewrite belongs on next, while master receives the minimum stability fixes. #715, #722 and #728 implement the main rewrite; waits inside another mruby call remain queued. The plan is updated as the remaining decisions are implemented.
  • Earlier recorded decisions. The mruby steps and bilingual “Why it changed” entries are recorded in the v3 plan and release-note guide (#561, #575). The Nginx::Var method-lookup and introspection choice is recorded in its archived proposal and v3 notes (#610). The master performance lane brings measurement and CI without Nginx::Debug (#601). The selected site specimens are recorded in the site design (#607).
  • #705 (base next, merged as fe9a3f2bf) translated six of the seven pages of NOT_TRANSLATED_YET in site/scripts/lib/site.mjs: the contributing page, development, the test guide, the releases page, security and the trust boundary. The handoff page stays in English with a Japanese notice, as the owner decided on 2026-10-29 (Q2 in section 7 of docs/proposals/site.md), so NOT_TRANSLATED_YET keeps only contributing/handoff. #703 (base master, merged as 6eae5a2df) rewrote the Japanese 2.x release notes. What is left: after master is merged into next, a pull request to next empties NOTES_ONLY in site/scripts/check-japanese-lines.mjs if the four release-note pages pass. The branch claude/site-step5-last holds an earlier translation that #705 and #703 drew on.

The closed issues (21 on 2026-10-29, among them #225 closed by #566, #320 by #565 and #519 by #562, and #206, #321, #502 and #200, closed on 2026-10-19; #200 after #574): gh issue list --state closed --search 'closed:>=2026-10-04' --limit 200. The reason is in the last comment of each issue, or in the pull request that closed it.

Taken up and still open: none.

Follow-up Target branch Waits for
The category comments of test/ci/changes.sh describe local site checks and next-push publication after the CI changes of 2026-11-02 (#737 and #739) master and next Updated on next with stream measurement, and on master in #774
Merge master into next after each merge into master. Model: #573. next Merged through 915076528 (#775) in #777. Merged through 13153ae22 (#778, #781, #784) in #785. Merged through 10c4cd927 (#786) in #787. Merged through 65ce972e1 (#790, #792) in #793, through aa3d9798a (v2.7.6, #801, #800) in #802, through 39bfae179 (#804) in #805, through 66b284a94 (#803) in #806, through 7a72306bb (#807) in #808, through 91499f767 (#809, v2.7.7 and #811) in #810, through 976015c48 (#812, #814, #813 and v2.7.8) in #815, through 11f80acfa (#817 and #816) in #818, which also raised worker_connections of test/conf/nginx.conf from 200 to 1024, through 8f204e061 (#819) in #820, through a7800d865 (#822) in #823, which also set MODULE_VERSION of next to 3.0.0, through 22ef61915 (#824, #827) in #828, and through 2c7e833a0 (2.7.10 and #831) in #830. Merged through ea531ad88 (#843, 2.7.11 and #851) in #850. Merged through 14bdc5fa6 (#852, #853) in #856. #859 merges through 49d26b879 (2.7.12, #858) and waits for its review and merge; the next merge of master after it waits for a merge into master
Public work records summarize requirements, rationale and constraints, distinguish decisions from proposals, and omit conversation quotations. Shared rules and next documents were aligned in #746; master rules and the corresponding English/Japanese 2.7.3 entries were aligned in #748 master and next #746 and #748 merged; #749 brought the remaining release-note corrections to next
The homepage shares coherent width, alignment and spacing across the hero, section introductions, use cases and footer; continuous introductions read in one column (see “Decisions”) next Merged in #747 as fad853820; site deployment and publication checks passed (checkpoint)
The fixed heading 「アップグレードの前に読む挙動の変更」 of the Japanese release notes ends with one character on its own line at 402 px, in v2.7.5.ja.md and in v2.7.4.ja.md. The heading and template are corrected in #745, with the previous anchors preserved master (reaches next by the merge) Merged in #745 and included in next through #746
Subrequests: the entry of #574 in docs/releases/v3.0.0.md and .ja.md explains the unchanged 500 from a subrequest content handler that returns Nginx::HTTP_OK without output, the 200 from nginx’s own return 200;, and why auth_request treats the retained 500 as an error instead of allowing the request next Merged in #761; local site checks and independent semantic and visual review passed
A known-length HEAD response finalized by an mruby output body filter keeps its connection for the following GET (the review of #574, should fix 1; #628 reproduced it) master, then next Fixed on master in #752 with failing-test and passing-fix commits, and integrated into next in #756. The combined suite passed all 759 tests; local validation, independent review, CI and publication checks passed. Both release lines and the current Filter reference describe the known-length HEAD scope
An end-to-end benchmark scenario: throughput and p50 and p99 latency with oha or h2load, planned in AGENTS.md “Design principles” and in section 8 of the plan, which also records the reference proxy end to end next Ready to start; where the results are recorded is decided when it is added (AGENTS.md)
Site: steps 1 to 4, 6 and 7 are merged (#591, #607, #614, #617–#623, #626, #629–#632, #653–#657), and the site is deployed from next. Step 5, the Japanese pages, is merged (#705 for six pages, #703 for the 2.x release notes on master); the handoff page stays English with a notice, and NOTES_ONLY of check-japanese-lines.mjs still lists four release notes (the subsection “Japanese pages of the site” of “Work in progress”) next for site/; master for 2.x sources Still open: the emptying of NOTES_ONLY once master is merged into next; the site on Astro without Starlight (5.4 of the site proposal, #700, the owner’s decision of 2026-11-01), whose first layer, the own shell beside Starlight behind NGX_MRUBY_SITE_SHELL=own, is merged in #708, and whose rest is #713 (from a Codex session, taken over by Claude Code for the browser acceptance), which also changes the iPhone width to 402px with the root size of 4.4; the 3.x development banner of 3.4; and the observability and certificate-lookup sample prerequisites of #619. Q9 is done (Discussions on, wiki off, 2026-11-01)
nginx-tests with the modules that its skipped files need: a build with the options of test.sh and --with-http_v2_module, --with-http_v3_module, --with-stream_ssl_module, --with-mail and the other modules that “Results of nginx-tests” in docs/test/README.md lists (in the CI run of #602, 189 of the 209 skipped files need a module that test.sh does not build), with and without ngx_mruby, in the nginx-tests workflow next Ready to start: #602 is merged
nginx-tests with ngx_mruby code configured: runs that add handlers or filters with code for every request through TEST_NGINX_GLOBALS_HTTP (the Pillar E row of the plan), compared with the run without code. The pull request decides which code; since #720, a rewrite handler that ends without Nginx.return goes on to the next phase on next next Ready to start: #602 is merged
Test harness rewrite (step 2 of section 6 of the plan; Pillar E): #635 (merged, c4e6ac6cc) made test.sh stop only the nginx that it started, by its pid, with no killall or pkill, and send SIGTERM again while the single nginx process still runs, because that process can miss a SIGTERM that arrives right before it waits for events (the run at 7926f72, where nginx stayed in epoll_wait() for 20 minutes after it logged the SIGTERM: “Stopping nginx” in docs/test/README.md has the cause and the measurements). #641 (merged, be559a017) added TEST_SH_PORT_OFFSET, which moves every port of the suite, so that two checkouts can run test.sh at the same time (“Ports” in docs/test/README.md). #706 added nginx with master_process on and several workers, and a reload in the middle of a case: test/t/cases/_workers.rb starts such an nginx from test/conf/workers/, and test/t/cases/workers.rb checks that each of four workers answers from an mrb_state of its own, and that across a SIGHUP the old workers finish an Nginx::Async.sleep and an Nginx::Stream::Async.sleep in flight while the new configuration answers (“Several workers and reloads” in docs/test/README.md). Still to do, in pull requests of their own: requests from CRuby instead of the client built into the test mruby for the other cases, and HTTP/2 and HTTP/3 clients. A should fix of the review of #635 waits for the pull request that next changes test/build_release.sh, an input of the perf job: its comment still says that test.sh kills every running nginx. The readiness wait of #593 is the first part, on next since the merge of master at 6b27f08da next Several workers and reloads: done by #706 (decision (3) of 2026-11-01). Requests from CRuby and the HTTP/2 and HTTP/3 clients: ready to start
CI runs only the jobs that a change needs: done on both branches by #676, #685, #686 and #692 on next, and #679, #690 and #693 on master; the owner’s decisions and rules are in AGENTS.md and docs/test/README.md. Left: a pull request to master that gives it the lines of test/ci/changes.sh and test/ci/changes_test.sh that next has (on master, a change of test/soak/mock_llm_test.rb alone runs that check only with the label perf); the first run of the whole build matrix through fromJSON (should fix 4 of the review of #686), which the nightly run under valgrind makes; and from the review of #693: local commands for the build of the 2.x pages in “The 2.x pages of the site” of docs/test/README.md on master (should fix 1 and 2), and on both branches, a sentence in “Required status check” that a draft run queued after the ready run cancels it in site-v2, codeql and zizmor, or a concurrency group that includes the draft state (should fix 4). #778 made two changes on master: “The 2.x pages of the site” in docs/test/README.md names site/.node-version instead of the number of the Node.js version (should fix 1 of the review of #774), and docs/releases/v2.7.6.ja.md takes the two edits of c41981dd5 in #777, so that the file is the same on both branches before the 2.7.6 tag. With #785, docs/test/README.md of next takes the same sentence, and docs/releases/v2.7.6.md and .ja.md are the same on both branches. Four cells: code-changing pull requests run four cells of the build matrix: done on both branches by #702 on next (bf1b742b7) and #704 on master (166341b93), with build standard of test/ci/changes.sh; every cell runs only nightly, on a tag, in the check pull request before a tag (the label full-matrix, which exists in the repository) and on “Run workflow” without base master, next Ready to start
nginx floor for v3 (item 2 of section 7 of the plan): drop the cells for nginx 1.26.3 and 1.28.3 from the build matrix of .github/workflows/test.yml and document the floor of 1.30 stable and 1.31 mainline next Done in #789, which also removes the nginx_version branches of src/ for nginx older than 1.25.5 and adds the entry of the floor to the 3.0.0 release notes. Each April a pull request moves the matrix and the docs to the new stable and mainline lines
2.x with nginx 1.24 and the stream module: the build stops while it compiles src/stream/ngx_stream_mruby_core.c, because nginx 1.24 has no ngx_stream_listen_opt_t; the build without the stream module completes (#789 measured both on master at 10c4cd927). Linux distributions such as Ubuntu 24.04 ship nginx 1.24. #824 wrote in docs/install/README.md and README.md of master that the stream module of 2.x needs nginx 1.25.5 or later since v2.7.0, and how to build 2.x with nginx 1.24 without --with-stream. No code for nginx 1.24 is added. This is how the sessions proceed, not a decision of the owner: nginx 1.24 is neither the stable nor the mainline line, and no cell of CI builds it, so code added for it could not be kept checked master Documented by #824; no code for nginx 1.24 and the stream module is planned
OpenSSL 3.5 and 4.0 (item 3 of section 7 of the plan): the cells of the build matrix of .github/workflows/test.yml build OpenSSL 3.5.9 in place of 1.1.1w since #771, whose Digest and SecureRandom need OpenSSL 3.0 or later, and configure reads the version of an OpenSSL 3 source tree. Left: a cell for OpenSSL 4.0, the floor of 3.5 in the docs and the build (the soak and perf jobs take the OpenSSL 3.0.2 of the runner), and a check for ENGINE and the removed APIs in ngx_mruby and the gems next Ready to start: #601’s change to .github/workflows/test.yml is on next since the merge of master at 6b27f08da
build_config.rb rewrite with explicit gems (step 3 of section 6; Pillar A): done. #771 replaces monochromegane/mruby-secure-random and iij/mruby-digest with mrbgems/secure-random (SecureRandom on RAND_bytes of OpenSSL) and mrbgems/digest (Digest on the EVP and EVP_MAC APIs of OpenSSL 3), as decided on 2026-11-09, keeping the classes and methods of the two gems; the differences are in its entry of the 3.0.0 release notes. The reasons of the decision: the list of mruby gems has no maintained replacement (Asmod4n/mruby-sysrandom replaces mruby-secure-random with another API, and nothing replaces mruby-digest), and gems in this repository are checked by the suite, CI and the perf job of ngx_mruby as they are. #766 did the rest of the row: iij/mruby-env and iij/mruby-process replaced by the core gems mruby-env and mruby-process (with mruby-signal), mruby-onig-regexp kept as item 1 of section 7 decides, and mruby-vedis and mruby-localmemcache out of the default build next Done: #771 merged as 6196a49f1 on 2026-11-09
mruby 4.1.0 (step 3 of section 6; item 1 of section 7). The owner decided on 2026-10-29 that next bundles 4.1.0-rc2 now and moves to 4.1.0 in a small subtree pull request when it is tagged. next bundles 4.1.0-rc2 since #678. The step to 4.1.0: ./update-mruby-subtree 4.1.0, vendor-mruby-prism if the gitlink of Prism moves, the lock, and the release notes next The tag 4.1.0 of mruby
Step 4 of section 6 of the plan, the core stability work of Pillar B (execution context, response finalization, fiber lifecycle, common layer, headers and variables, and the two agent proxy rows; #770 does the row on log handlers before access_log), then v3.0.0-alpha.1 next The core stabilization is done: the core stabilization proposal was implemented in #712, #714, #715, #720, #722, #727, #728 and #735 (the last on 2026-11-02), is archived, and the model is in How a handler waits of docs/DEVELOPMENT.md. Left in this step: the work designed in the proposal for the rest of the core before alpha.1 (the execution-context type, headers, Nginx::Var, the filter handler pointers, the table-driven directives, the common layer beyond the fiber, and the rest of the async resume rewrite; the row on log handlers before access_log, with the raise of a local Nginx.redirect in log handlers, is done by #770, the row on the execution context and the phases of the Ruby API by #826, merged on 2026-11-16, the row on the per-request store by #832, the agent proxy row on mruby_variable by #780, the row on common upstream switching by #788, merged on 2026-11-11, the row on common error logging by #796, merged on 2026-11-12, with its follow-up #799, merged on 2026-11-12, for the calls of inspect and backtrace in the logging, the row on common code loading by #825, merged on 2026-11-16, the row on the directive table and one duplicate check by #834, the row on the filter handlers by #835, and the row on a body with any status by #779, merged on 2026-11-10, with its follow-up #783, merged on 2026-11-10, for a status set before the output (decision 6), Nginx::HTTP_CLOSE with output (decision 7) and the should fix items 2 and 4 of the review of #779), one pull request per row of its section 4, and the check before the tag of its section 3. Its design decisions were made on 2026-11-09 and 2026-11-10 (section 5 of the proposal), and no point of the design is open. Started by the owner’s decision of 2026-11-01 to extend the harness so that test.sh tests master_process on, several workers and reloads, and then to rework the core fiber lifetime and response completion. #706 made that extension; the requests from CRuby and the HTTP/2 and HTTP/3 clients of the row above for the test harness rewrite stay there
nginx-tests required: the nginx-tests workflow records its results and fails only when a build, the fetch of the suite or prove fails. In the runs of 2026-10-28 and 2026-10-29 the build with ngx_mruby and the build without it have the same counts (511 files, 302 passed, 0 failed, 209 skipped). To do: the job fails when the build with ngx_mruby fails a file that the build without it passes, and the job joins the needs of ci-ok next Ready to start: the counts are the same in every run since #602
The perf job in the needs of ci-ok, and the “In CI” text of docs/test/README.md aligned with the rule of AGENTS.md “Design principles”, under which a session does not merge while the job reports FAIL or ERROR. On next, #794 does it: ci-ok needs perf, and a FAIL that the owner accepted on a pull request passes through a line with the new condition pull-request in test/perf/accepted.txt. The comparison judges the Ir per request without the GC and without the constant lookups that miss the constant cache of mruby, which made the stream_sleep FAIL of #766 that another run on the same commits did not repeat, except where the head clears that cache more often, or clears it and misses more lookups than the base (#795); then with the lookups counted. Left: the same on master, in a pull request to master (its test.yml, test/perf/perf.rb, test/perf/accepted.txt, AGENTS.md and docs/test/README.md) next, then master next: done by #794. master: ready to start
Soak: record Pss and Private_Dirty from /proc/<pid>/smaps_rollup beside VmRSS, then bound their slope over more windows, set after measuring (the Pillar E row on memory accounting). Today test/soak/soak.rb bounds VmRSS by a step and a total and checks the counters of Nginx::Debug.stats next Ready to start
Soak and perf for stream: the normal TCP stream_sleep scenario checks exact response bytes and EOF, observes stream_timers and stream_fibers, and collects the stream VM before memory samples. The performance build includes stream and counts completed sessions (section 7 of the core stabilization proposal) next Merged in #758 as 8beaaa381; local checks, independent review, CI, publication verification and cleanup passed
The case async_sleep_many.rb of section 6 of the core stabilization proposal: a handler that sleeps 1,000 times for 1 ms answers, and Nginx::Debug.stats shows the same timers and gc_root_fibers before and after next Merged in #751 as 63c6b7152; local checks, independent review, CI and publication checks passed. The normal test build enables the counters, and each HTTP exchange has a 60-second limit
The comments in src/ and test/ that name docs/proposals/core-stabilization.md (20 places, git grep -n core-stabilization -- src test): the proposal is now docs/proposals/archive/core-stabilization.md. To do: the next pull request that changes those files points them at the new path, or at the section “How a handler waits” of docs/DEVELOPMENT.md next The next code pull request that touches those files (a comment-only change of src/ would run the whole build matrix)
The ASan cell with -DNGX_DEBUG_PALLOC -DNGX_DEBUG_MALLOC (3.7 of the plan; the cell of test.yml sets only -fsanitize and -fno-omit-frame-pointer), so that ASan sees an access that runs from one pool allocation into the next. To do: a run of the suite with these flags, whose findings follow SECURITY.md, then the cell next, master Ready to start
A test build with ngx_otel_module (nginx/nginx-otel, built with CMake against the nginx source of test.sh) and an OTLP receiver for the case of the observability sample of 3.2 of the site proposal next The harness rewrite (the row above), of which this waits for what is left: the requests from CRuby instead of the client built into the test mruby, and the HTTP/2 and HTTP/3 clients. #706 added master_process on, several workers and reloads
Step 7 of section 6 of the plan, the agent proxy, parts (a) to (e); none exists on next next Parts (a) and (c) start “after the safety net of step 2, beside step 4” (step 7 of section 6 of the plan, decided on 2026-10-29 in item 15 of section 7), so they wait for what is left of that safety net: in the row above for the test harness rewrite, the requests from CRuby instead of the client built into the test mruby, and the HTTP/2 and HTTP/3 clients. #706 added master_process on, several workers and reloads, and its test/t/cases/_workers.rb runs the several workers that the test of N concurrent incr calls of the shared dictionary needs (section 8 of the plan). (b) and (d) also wait for step 4. The order inside the step is a proposal of the 2026-10-04 amendment until the owner decides it. The serves: notes on the rows of the plan name the uses of the ledger that each part serves
The AI uses of v3 beyond the agent proxy: the next work comes from the ledger of the uses, first its rows at designing (MCP tool authorization, whose first recipe uses two tenants and read and write mock tools, MCP tool profiles, MCP progressive retrieval, and Model API fallback, whose test is in section 8 of the plan), then its rows at direction. This row copies the ledger; when they disagree, the ledger holds next What is left of step 4 of the plan (the row of step 4 above); the fallback also waits for step 7 of the plan (the row above). No API for these uses is designed yet
Remove example/ (README.md, nginx.conf, proxy.rb, set.rb, set2.rb, unified_hello.rb): no CI job starts it, no page links it, and its /proxy location picks among test1.example.com to test3.example.com, which are unreachable; examples/start/ is checked by examples.yml next Done in #791: the directory and the sentence about example/nginx.conf in docs/proposals/site.md are removed; no other reference existed, and no release note is needed because no configuration, script or build reads the directory
What is left of the async resume rewrite (decision (c) of 2026-10-04, amended on 2026-10-29: the rewrite of how resumed handlers end): a wait that starts in a handler that nginx runs inside another mruby call (the location of an internal redirect that an mruby_set handler, a log handler or a filter makes inside the Ruby call), or after an internal redirect replaced the module context of the handler, is still resumed by its timer or by the end of its subrequest and ended by ngx_mrb_post_fiber in src/http/ngx_http_mruby_async.c; a content handler there does not keep the request, so the request ends at the sleep. The rest is done in #715, #722 and #728 (section 12 of the core stabilization proposal) next Done in #833, as section 2.9 of the proposal for the rest of the core before alpha.1 designs it: Nginx.redirect with a local URI raises in mruby_set handlers and output filters (decision 1 of 2026-11-09; #770 did log handlers), and ngx_mrb_post_fiber is removed, so every wait ends through the wake and nginx’s phase checker
The generated reference of the site (Pillar F of the plan): the directives from the ngx_command_t tables, the Ruby API with YARD and yard-mruby, the C API with Doxygen. site/scripts/check-reference.mjs checks, among other things, that each directive name has a heading in docs/directives/README.md; the pages are written by hand next Ready to start; no step of section 6 of the site proposal produces it yet
The “Before” line of the entry on mruby_set values that are not Strings in the 3.0.0 release notes next Done in #769. #768 ran the cases on master before its fix: a to_s that returns something other than a String ended the worker with a segmentation fault before and after Nginx::Async.sleep, and a value without to_s or whose to_s raises was read as a String although it was the exception, which stayed in the mrb_state of the worker. The line in docs/releases/v3.0.0.md and .ja.md now says this
The documentation group of the site footer links the current 3.x documentation before the other version in both languages, without duplicating the existing 3.x link on 2.x pages (4.10 of the site proposal) next Merged in #750 as c440a15dc; local checks, independent review, CI and publication verification passed
The 2.x form of the protected conversion of an mruby_set value (#735 on next) master Done in #768 (40578dd63, 2026-11-09 19:25 JST), with its entries in docs/releases/v2.7.6.md and .ja.md. On 2.x, a value that cannot be converted counts as an exception of the code: before a wait the variable is empty and the request goes on, after Nginx::Async.sleep the request gets 500. It reached next in #769, which keeps the form of #735
Nginx.rputs and Nginx.echo (ngx_mrb_rputs_inner in src/http/ngx_http_mruby_core.c, the function that both methods call), Nginx.redirect (ngx_mrb_redirect there) and Nginx::Filter#body= (ngx_mrb_set_filter_body in src/http/ngx_http_mruby_filter.c) call to_s on a value that is not a String and read the result with RSTRING_PTR without checking that it is a String. They run in Ruby code, so a to_s that raises raises in the script, but a to_s that returns another object is read as a String. master and next have the same code. Found while testing #768 (“Not changed, found while testing” in its body) master, next Done in #775 on master and #776 on next, which convert the argument with ngx_mrb_var_to_string() and raise TypeError when to_s does not return a String, as the setter of Nginx::Var does; each has its regression test in test/t/cases/output_value_to_s.rb (port 18191) before the fix and its entry under “Behavior changes” (v2.7.6 and v3.0.0). The setters of Nginx::Request (uri=, args= and the others of NGX_MRUBY_DEFINE_METHOD_NGX_SET_REQUEST_MEMBER_STR), Nginx::Headers_in#[]=, Nginx::Headers_out#[]= and the path of Nginx::Async::HTTP.sub_request, which these two PRs left, raise TypeError for an argument that is not a String since 2.7.7, on master and on next (GHSA-cjxv-4q9m-62pf). The two places left after 2.7.7 are done by #813 on master (for 2.7.8) and on next by the merge of master: the second argument of Nginx.redirect (TypeError for a redirect to a URL with a status that is not an Integer, and ArgumentError for a third argument) and :address of Nginx::Server#add_listener and Nginx::Stream.add_listener (TypeError when it is missing or not a String)
On 2.x, an mruby_set handler that ends normally after Nginx::Async.sleep in the location /set/ of test/conf/conf.d/83-var-value-to-s.conf ends the request without a response: the timer handler finalizes the request with the NGX_OK of ngx_mrb_finalize_rputs, and the phases do not go on (seen while testing #768; /sleep_with_proxy_set_code on port 18080 gets its response). On next, test/t/cases/async_set_var.rb checks that such a handler sets its variable and the request goes on (#715); the location of var_value_to_s has not been run with a valid value after the sleep on next master Done: the 2.x code is not changed, and #773 writes the limit in docs/directives/README.md of 2.x; 3.0.0 fixed it in #715. The decision of 2026-10-29 under “Decisions” puts the rewrite of the async resume on next and gives master the minimum stability fixes
test.sh makes its certificates with cd ${NGINX_INSTALL_DIR}/html && ... openssl req ... && cd -. When openssl req fails, set -e does not stop the script inside the && list, the shell stays in html, and the next command does not find test/t/cases/_wait_for_listen.rb, which hides the cause (seen in the CI runs of #771). To do: test.sh stops when a command of the certificates fails and says which one next, master Done: #792 on master and, by the merge of master in #793, on next
On 2.x, a rewrite handler that sleeps in Nginx::Async.sleep and returns Nginx::DECLINED, followed by an access handler of the same location that calls Nginx::Async::HTTP.sub_request and declines, leaves the request without a response (seen by the reviewer of #740 before and after its fix; without the sleep the answer is immediate). Assess on next first, where the fiber record of #714, #715 and #722 may already cover it; if 2.x still hangs, a fix with its regression test next, then master if needed Assessed on 2026-11-03: the normal composition completes with the next runtime of c440a15dc. A 100 ms sleep and its no-sleep control each returned 200, the expected static body and EOF in three requests bounded at 8 seconds (nginx 1.31.8, static suite build with GC stress and debug counters; runtime C statements unchanged). No next fix is indicated by these six requests. The 2.x follow-up is done in #772, whose test async_sub_request_after_sleep.rb also runs on next since #777
A body filter whose request lost its module context while the response was still being written: the review of #785 (should fix, “The other place that clears the module context”) found that the body filter drops the context when an earlier handler’s Nginx.rputs output is still in it, and that the response then stopped after its first write until send_timeout master, then next Done: fixed on master in #786 (10c4cd927), for locations with and without an mruby header filter, and brought to next by #787
Make an mruby body filter run once on a proxied response that goes out in one write after an earlier handler’s Nginx.rputs (should fix 4 of the review of #786). Today the empty last buffer that ngx_http_upstream_finalize_request sends runs the code of the body filter a second time on an empty body, and $sent_http_content_length logs 0 instead of the length. Since #786, a response sent in more than one write runs the code once and logs the length, so the logged value depends on the number of writes next (a behavior change, not for 2.x) Ready to start: a test of the one-write case with a proxied location of test/conf/conf.d/90-body-filter-leftover-chain.conf, a fix and an entry under “Behavior changes” of docs/releases/v3.0.0.md and .ja.md
Local checks first (the decision of 2026-11-12 under “Decisions”) on master: the same change to AGENTS.md (“Design principles”, “Branches and pull requests”, “Review and merge”, “Build and test”), .github/PULL_REQUEST_TEMPLATE.md and docs/test/README.md of master master Done: #807, merged on 2026-11-14 as 7a72306bb; its fixes reach next in #808
Candidates to add to the release procedure (docs/releases/README.md, “Before the tag” and “Security fixes”), seen in the 2.7.7 release: the pull requests of the fixes in the advisories’ private forks did not check the release notes they changed as 3.x pages (the check of “The 2.x pages of the site” in docs/test/README.md on master), so two lines of v2.7.7.ja.md failed check-japanese-lines.mjs first in #810; the step that starts site.yml on next after the release was missed until 03:27 JST, so the home page named 2.7.6 until then; the check before the tag was not a public pull request (as for 2.7.6, it is recorded in this file); the “patched versions” of an advisory draft are set before it is published; the commit messages of the fixes in a private fork name neither the advisory id nor the details of the defect; and the whole build matrix runs on the public master before the tag master (the procedure of 2.x releases), then next by the merge Done: #812 and #814 on master, merged on 2026-11-15; next takes them in the merge of master through 976015c48, without the local run on the oldest nginx, which next does not need (the state above)
An mruby body filter on a subrequest of SSI (<!--# include virtual=... -->): the case that the row named, an include with only a body filter and no Nginx.rputs (a note of the review of #786), was fixed by the fix of the advisory GHSA-79vp-37cx-547v, published with 2.7.6. Another case remained on master: an include after the first one, with a body filter and a body written with Nginx.rputs or Nginx.echo, waited until send_timeout. #827 fixed it for 2.7.10, and #828 brings the fix to next master, then next by the merge Done

Use Continuing shared work for startup, ownership, checkpoints, temporary files and resumption. The rules are in AGENTS.md; this file records the queue and decisions, while each PR records its current head and evidence. The procedure applies to Claude, Codex and human contributors alike.

Vulnerability reports and their fixes are handled privately as SECURITY.md describes; this file does not record them.