Skip to content

fork-ci release: carry the files the heavy and desk kit installers read, and test every kit against the release - #920

Merged
scuttletoby merged 2 commits into
mainfrom
fix/ci-release-paths
Nov 22, 2026
Merged

scuttletoby merged 2 commits into
mainfrom
fix/ci-release-paths

Conversation

@scuttletoby

Copy link
Copy Markdown
Owner

What

  • .github/fork-ci-release.json paths now also carries the files two kit installers read from outside fork-kit/:
    • heavy: packages/shared/src/capacity/capacity-rule.json, capacity-format.json, capacity-vectors.json (only these three, not the folder's README or readings file: heavy-install copies exactly these).
    • desk: .github/scripts/fork-fleet-desk-agent.py, -desk-seed.py, -desk-wire.py, fork-fleet-hook.sh, fork-fleet-desk-agent.service, fork-fleet-refresh.service/.timer, fork-fleet-seed.service/.timer and .github/fork-fleet.json (the release's own fleet config, used when no --fleet-config is given).
  • New check in .github/scripts/fork-ci-release.test.py: it builds this commit's release file set with release_entries (exactly as the release does), writes those files to a folder, and runs each kit installer's own file check on it, with the installer loaded from that folder: read_source over source_files() (agent-guards, cleaner, golive, schedule-kick), SOURCE_FILES through source_file (heavy, model-guard), SOURCE_FILES (gh-apps), CODE (pc-agents), and missing_sources (desk). A kit folder with an installer (sbin/) that has no row fails too, so a new kit cannot be skipped. One more check proves the test fails when the capacity files are left out.
  • desk-install: the inline missing-file list is now missing_sources(source), which install uses and the test calls (same behaviour).
  • fork-kit/pc-agents/tests/setup.test.py: its ReleaseHolds test (pc-agents only, checked the config paths by prefix) is removed; the new shared check covers pc-agents from the real release tree.
  • Docs: fork-ci/README.md and the fork-ci-release.py header say the release carries the files the kit installers read, and that the test enforces it.

The review kit has no installer (used from a checkout), so it has no row.

Why

Since #786 heavy-install reads the capacity rule files from packages/shared/src/capacity/, which no release held, so every ci/vN since then (ci/v21 included) refuses at install: heavy-install: ... capacity/capacity-rule.json is missing. Checking every kit found the same gap in desk: desk-install refuses any release because ten of its files were never released. The kits' own tests run from a full checkout, so neither gap showed in CI.

Test

  • On origin/main's config (test commit only, before the fix), python3 .github/scripts/fork-ci-release.test.py exits 1:
    FAIL release holds what the desk kit's installer reads
         | .github/scripts/fork-fleet-desk-agent.py; .github/scripts/fork-fleet-desk-wire.py; .github/scripts/fork-fleet-desk-seed.py; .github/scripts/fork-fleet-hook.sh; .github/fork-fleet.json; .github/scripts/fork-fleet-desk-agent.service; .github/scripts/fork-fleet-refresh.service; .github/scripts/fork-fleet-refresh.timer; .github/scripts/fork-fleet-seed.service; .github/scripts/fork-fleet-seed.timer (add the path(s) to .github/fork-ci-release.json paths)
    FAIL release holds what the heavy kit's installer reads
         | capacity/capacity-rule.json; capacity/capacity-format.json; capacity/capacity-vectors.json (add the path(s) to .github/fork-ci-release.json paths)
    2 failed
    
  • With the fix: all passed: 86 ok, 0 skipped.
  • fork-kit/desk/tests/install.test.py: 58 tests OK. fork-kit/pc-agents/tests/setup.test.py: 60 tests OK.
  • node .github/scripts/fork-seam-check.cjs --config-ref origin/main origin/main HEAD: exit 0. vp fmt --check on the touched json/md: clean.

State question

  • Release runs again with no change: the release tree is the same as the branch head's, so it skips and writes nothing (existing behaviour, unchanged). The first release after this merge holds the new files once (ci/v22 or next).
  • Stops half-way: the release push is one atomic push of branch and tag, so nothing half-written; the new test only writes temp folders and never touches the checkout.
  • Existing / old state, consumer pinned to an old ci/vN: old tags never move, so heavy and desk still cannot install from ci/v21 or older; install from the first release after this. Consumers of the actions and reusable workflows see only extra files (about 420 KB) in the downloaded tree; every path they use stays the same, and the release still has no symlinks and no self-starting workflows.

General vs project

General: the paths and the test are about the fork's own kits and release; no project, person or machine fact. The test discovers every kit folder with an installer, so a future kit (or a new file an installer reads) is covered without project config.

🤖 Generated with Claude Code

scuttletoby and others added 2 commits November 22, 2026 05:06
… release (desk: missing_sources)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ty rule, fleet desk scripts)

heavy-install reads capacity-rule.json, capacity-format.json and capacity-vectors.json from
packages/shared/src/capacity, and desk-install reads the fleet desk scripts, units and
.github/fork-fleet.json; none were in the release paths, so both refused on every release.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@scuttletoby

Copy link
Copy Markdown
Owner Author

lens: merge-path · reviewer: lens agent (Opus 5.5) · read at: 659ed69
verdict: safe to merge
covered: the whole diff: .github/fork-ci-release.json, fork-ci-release.py (docstring only), the new kit check in fork-ci-release.test.py, desk-install missing_sources, the removed pc-agents ReleaseHolds test, fork-ci/README.md; against release_entries, check_workflows, verify_download, the release flow, the kits' bootstraps and each installer's own file check.
checks: ran fork-ci-release.test.py (86 ok), fork-kit/desk/tests/install.test.py (58 OK) and fork-kit/pc-agents/tests/setup.test.py (60 OK) on the head. Proved it fails on main: the head with main's fork-ci-release.json fails desk (the 10 fleet files) and heavy (the 3 capacity files). Mutations on the head, each caught: a new kit folder with sbin/ and no KITS row; a new pc-agents CODE source outside the paths; a new heavy SOURCE_FILES entry (capacity-readings.json) outside the paths. Read the PR's own CI log: the new rows ran and passed in "Workflow lint and upstream drift", which runs on every same-repo PR.
state: runs again: the release tree is built the same way, so a re-run finds the same tree and skips; the test writes only to its own temp folder. Stops half-way: the release push is still one atomic push (unchanged); a stopped test run leaves only its temp folder. Old pins: an existing ci/vN keeps its old tree, so desk and heavy installs from an old pin still refuse a missing file. The first main release after this merge (the tree changes, so it makes ci/vN+1) fixes it, and hosts move to the fix only when they re-pin to that sha. Empty or broken state fails: no fork-kit/ in the release fails every KITS row, an installer that crashes or prints unreadable output fails its row, and an unknown how raises an error, which also fails the row.
blocker: none
major: none
minor: .github/scripts/fork-ci-release.test.py:430-440: the heavy, model-guard, gh-apps and pc-agents rows run their own copy of each installer's check loop. The file lists and heavy's source_file are the real ones; only the loop and the path joins are copies. If an installer changes how it finds a file, the test can drift from it. Desk now calls the installer's own missing_sources; one such helper per installer would remove the copies (a follow-up).
minor: .github/scripts/fork-ci-release.test.py:476: a new kit counts only when it has an sbin/ folder. A kit whose installer lives somewhere else (for example only a bootstrap.sh) is not flagged. The KITS rows also assume the --source each bootstrap passes (the kit folder, or the release's top folder for desk and pc-agents); nothing checks that against the bootstraps.
minor: .github/scripts/fork-ci-release.test.py:500: every failed row ends with "(add the path(s) to .github/fork-ci-release.json paths)", also the "no row in KITS" and "failed to run" failures, where that advice is wrong.

@scuttletoby

Copy link
Copy Markdown
Owner Author

lens: state · reviewer: lens agent (Opus 5.5) · read at: 659ed69
verdict: safe to merge
covered: the whole diff; heavy-install and desk-install end to end (every file they open from --source, and every sibling file the copied scripts load at run time: capacity_rule.folder(), heavy_facts.load(), fork-fleet-desk-agent/seed module loads, fork-fleet-walls HERE files, the gate hook's --config); fleet-route's --config input (now that .github/fork-fleet.json is in the release)
checks: fork-ci-release.test.py 86 ok; fork-kit/desk tests (install 58, windows 31) and fork-kit/pc-agents tests (check 36, setup 60, update 35) all OK. Built this head's release folder with release_entries exactly as the release does and, with the installers' test PREFIX: (1) heavy: a plain old /usr/local/bin/heavy, then ci/v6's installer with --grace 45, then this release: config kept (grace 45, slice), replaced/ kept, versions/ kept, current and both links switched by rename, self-check OK (22 paths); stopped after each of the 10 steps and run again: each rerun 0 and self-check OK; run again: no change; the installed copy runs heavy --help, capacity --help, facts --help and t3-capacity --help from /, and capacity_rule.folder() is the version folder's own capacity/ (nothing read from outside it); rollback to the ci/v6 installer and forward again keep the config. (2) desk: missing_sources() is [] and stage() on the release folder passes its own self-test (walls os, the gate loads with the released fork-fleet.json, bash -n, desk agent --help; 23 files). Spot-check: this test file with main's fork-ci-release.json fails exactly "desk" and "heavy".
state: install runs again: no change (heavy copy_tool reuses a complete version folder; desk restages). Stops half-way: heavy reruns finish from every step; desk's swap and SOURCE-last order is unchanged by this PR. Old machine on ci/v6: upgrades in place, config and replaced/ kept. Empty machine: same path as the dry install. Release runs again: release_entries refuses when a listed path is gone from main, so a later rename of a fleet file blocks the release until the config follows (the new test names it). Consumer pinned to an old ci/vN (ci/v21): heavy-install refuses at check, "capacity/capacity-rule.json is missing", before writing anything; the version in use is unchanged. desk refuses the same way before any write.
blocker: none
major: none
minor: the new release test runs each installer's missing-file check only, not its self-test: a staged script that starts loading a new sibling file (desk-seed loads fork-fleet.py and the desk agent by path) would pass this test and fail at install. Running desk stage() and heavy copy_tool plus bin/heavy --help on the release folder (both work with PREFIX, as above) would close that; .github/scripts/fork-ci-release.test.py:403
minor: kit_release_problems() finds an installer only by a sbin/ folder; a future kit with its installer elsewhere gets no row and no check; .github/scripts/fork-ci-release.test.py:475

@scuttletoby
scuttletoby merged commit e38ed1f into main Nov 22, 2026
30 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant