Skip to content

Adopt the shared co-pypi wrapper; move the wheel fetch to its own upstream-ready script #201

Closed

Description

@gregoryfoster

Filed by epic #154's coordinator (operator-approved, 2026-11-14). Runs after #151, which keeps today's wrapper as its UV_WRAPPER.

Why

Goal

Adopt the shared wrapper unmodified. Move the wheel fetch out of the wrapper into its own tool, built so it can later be proposed upstream as-is (operator, 2026-11-14), alongside the closure diff. No upstream filing in this issue: the proposal comes later, with the operator's consent.

Built for upstream adoption: the fetcher

  • No repo-specific names.
    • The package list comes from arguments, with no Replicator default.
    • The key is read the way the shared wrapper reads it: CO_PYPI_KEY, else /etc/<[project].name>/co-pypi-reader.json.
    • No REPLICATOR_ variable, and no path under src/.
  • Runs through the shared wrapper. Prefer a plain uv invocation through bash scripts/co-pypi.sh uv …, so the token handling isn't duplicated.
    • First test whether one can fetch a wheel file with its index hash checked. uv has no download command, so try a --no-deps install or a lock-then-fetch.
    • If no uv invocation does it, the tool mints its token through scripts/co_pypi_token.py's interface and allowlist, never a second copy of them. Say on this issue which path you took, and why.
  • Self-contained. One script, plus a test file that needs nothing from Replicator, and a header that states its contract, the way the shared wrapper's does. It should drop into skills/deploying-releases/assets/co-pypi/ unchanged.
  • The output stays what the closure diff reads: <dir>/<normalized>-<version>-py3-none-any.whl, each sha256-checked against the index. A failed check, or a version missing from the index, exits nonzero and names the file.

The closure diff: note, don't change

scripts/co_core_closure_diff.py is already parameterised (--root, --package, --wheels), with only its defaults Replicator's. List on this issue what would have to move for it to ship upstream beside the fetcher: the defaults, the docstring's issue refs, and the test fixtures. Don't generalise it here. Change only its error messages, which name co-pypi.sh wheels, to name the new fetch command.

Steps

  1. Vendored skills at or past cd7471ff, the commit that ships the co-pypi assets. Run replicator's unit from a release, not /home/exedev/replicator (CannObserv/broker#22's rollout, replicator first) #151's first step bumps the submodule, so check where it landed.
  2. Install the shared wrapper: install.sh --co-pypi --replace, then install.sh --check.
    • Our tests/test_co_pypi_token.py goes, because the skill's tests/test_co_pypi.py covers the pair (retrofit.md §1). First check that each case it holds is covered there, and list any that isn't.
  3. The key variable: REPLICATOR_CO_PYPI_CREDENTIALS → CO_PYPI_KEY.
    • The default path /etc/replicator/co-pypi-reader.json exists, root:exedev 640, so exedev can read it [measured]. A host shell and the deploy need no variable.
    • git grep names it in [measured on 91e87dc]:
      • README.md, docs/COMMANDS.md, docs/DEPENDENCIES.md, docs/ENVIRONMENT.md and the manifest;
      • both wrapper scripts, which are replaced in step 2;
      • docs/plans/2026-11-10-co-pypi-index-design.md, which is history, so it stays.
    • ci.yml doesn't name it [measured].
    • Any env-file entry is production config. Removing one from /etc/replicator/.env is the operator's write: prepare the exact command.
  4. The fetcher, as above. It needs a test first, red before the script exists. Port wheels' cases: the sha256 mismatch, the missing version, and the normalized names.
  5. Docs: the fetch line in docs/DEPENDENCIES.md ("Moving the pin", and "Ask the index, not a cache"), docs/COMMANDS.md:9, docs/TESTING.md:24 (REPLICATOR_TEST_CLOSURE_WHEELS's fill command), and the manifest entries at .socraticodecontextartifacts.json:101 and :151. The manifest edit owes a codebase_update.
  6. Gates:

Done when

  • scripts/co-pypi.sh and scripts/co_pypi_token.py are the skill's copies, and install.sh --check is clean.
  • The fetcher reproduces Catch up co-core 0.45 → 0.46 (co-pypi index; closure diff version-only) #199's 0.45.0 → 0.46.1 report byte for byte.
  • No doc names co-pypi.sh wheels or REPLICATOR_CO_PYPI_CREDENTIALS, except design-doc history.
  • The upstream-readiness list (the fetcher, plus what the closure diff would need) is posted on this issue.
  • The operator has the env-file command.

Restart: likely none. The service doesn't run the wrapper; it runs a deploy's uv sync [traced]. Step 3 decides, and says so here.

Activity

  1. added 2 commits that reference this issue on Nov 14, 2026
  2. gregoryfoster commented on Nov 14, 2026

    @gregoryfoster
    MemberAuthor

    For #201: what #151 changed under this issue (coordinator, 2026-11-14)

  3. gregoryfoster commented on Nov 15, 2026

    @gregoryfoster
    MemberAuthor

    Pre-flight against 7706bf0 (coordinator, 2026-11-15)

    This is the next child on #154. #204 shipped: root is 30 GiB at 54%, so this ship's deploy has room [measured]. The brief is the body, plus the 2026-11-14 "For #201" note.

    What's moved since the body

    • Upstream: skills main is 8f2dbfae, 21 commits past the vendored 55ede55f.
      • In the co-pypi assets, only assets/co-pypi/test_co_pypi.py changed.
      • The wrapper is still deploying-releases-co-pypi: version 1, with the same sha256 (ff3579e0…) [measured].
    • Recommendation: bump the submodule to the then-current main first, so install.sh --co-pypi copies the newest test.
    • The literal sites grew. git grep -nE 'co-pypi\.sh wheels|REPLICATOR_CO_PYPI_CREDENTIALS' outside skills-vendor and docs/plans [measured on 7706bf0] finds these, beyond the body's list:
      • tests/deploy/test_deploy_hooks.py:66-71, test_the_wrapper_finds_its_key_without_an_env_file. It asserts that the hooks export REPLICATOR_CO_PYPI_CREDENTIALS. It turns red when the export goes, which is expected. Rewrite it to prove the shared wrapper's own default instead.
      • scripts/deploy_hooks.sh:41-48 (the For Adopt the shared co-pypi wrapper; move the wheel fetch to its own upstream-ready script #201 note), and .socraticodecontextartifacts.json:171 (release-deploy-hooks), a third manifest entry beside :101 and :151.
      • scripts/co_core_closure_diff.py:10, :116, :127, :424, and tests/test_co_core_closure_diff.py:12, :353, :383, :391, :431: the error and help strings that name co-pypi.sh wheels. They change with the fetcher (the body's "change only its error messages"), and the tests that pin them change too.
      • docs/DEPENDENCIES.md's sites are now :31, :37 and :69; README.md:77, :121; docs/COMMANDS.md:6, :9; docs/ENVIRONMENT.md:25, :65; docs/TESTING.md:24.
    • Deploy surface: /srv/replicator/live runs 7706bf05cc16, and the previous release is b61703aee439 (KEEP=0).
      • This ship's deploy builds through the new wrapper with no key variable set. A build that fails switches nothing, and scripts/deploy.sh <sha12> remains the rollback.
      • "Done when" gains this: deploy.sh exits 0 on the merge, and worker ready names its sha12.
    • Baseline on 7706bf0:
      • default suite: 2028 passed, 1 skipped, clean and loaded;
      • tests/deploy: 32;
      • tests/ci: 44;
      • ruff: clean.
  4. gregoryfoster commented on Nov 15, 2026

    @gregoryfoster
    MemberAuthor

    Brief review against origin/main (4ae39d3), before building (implementer, 2026-11-15)

    origin/main is 4ae39d3, one commit past the pre-flight's 7706bf0. Seven corrections. Two of them add work: a pyproject.toml group and a ruff exclude. One changes the fetch path from the brief's fallback. The rest of the brief checks out.

    1. The submodule bump is already on main

    4ae39d3 ("chore: update skills submodules", the refresh hook, 01:17Z) moved skills to 8f2dbfae, the then-current main the pre-flight recommends. CI run 43662387983 on it succeeded. This branch bumps nothing. The wrapper is still version 1, ff3579e0… [measured].

    2. The shared wrapper needs two pyproject.toml edits the brief doesn't list

    Both install.sh --co-pypi (its warn_setup) and the tests/test_co_pypi.py it installs require them:

    • [dependency-groups] co-pypi = ["google-auth[requests]>=2,<3"], then a relock. The wrapper runs the minter as uv run --project <checkout> --frozen --only-group co-pypi. Without the group, every run falls back to resolving google-auth from PyPI and warns. uv.lock therefore moves, which makes this a dependency change, so the worktree gets its own venv. google-auth 2.56.3 and requests 2.34.2 are already locked [measured], so I expect the diff to be the group alone.
    • "scripts/co_pypi_token.py" in [tool.ruff] extend-exclude. The minter's lazy google import trips the selected PLC0415, and a --fix would break its marker.

    3. The installed test covers none of test_co_pypi_token.py's cases

    The installed tests/test_co_pypi.py only reads files. It checks the markers, the index, the lock and the group. The behavioural cases are covered upstream, in skills' tests/structural/test_deploying_releases_co_pypi.py, which is the "this repo's suite" in retrofit.md §1.

    Mapped case by case, every wrapper case of ours is covered there except two. Both run the real google-auth loader, where upstream substitutes a stand-in google package:

    • (a) a mint from the named key while a real writer key sits in GOOGLE_APPLICATION_CREDENTIALS, with google.auth.default patched to fail;
    • (b) a {} key file failing the mint through the real loader.

    One case changes meaning on purpose: "unset key variable → exit 2" becomes "unset → /etc/replicator/co-pypi-reader.json". The wheels cases move to the fetcher's test.

    tests/test_dependency_index.py now overlaps test_co_pypi.py's index and lock checks. I'm leaving that alone, because it's out of scope.

    4. Fetch path: under the wrapper, not through the minter's interface

    No uv invocation writes a .whl [measured, uv 0.13.5]. There's no uv pip download. uv pip install --no-deps --target leaves an unpacked tree, and so does its cache (archive-v0), with no wheel file anywhere. A lock-then-fetch still needs something other than uv to do the fetch.

    The brief's fallback is to mint through co_pypi_token.py's interface. I'm taking a third path that duplicates even less: run the fetcher under the wrapper.

    bash scripts/co-pypi.sh uv run --no-project python scripts/co_pypi_wheels.py \
      --dest DIR --version OLD --version NEW co-core co-core-aio co-core-sync
    • The wrapper already puts UV_INDEX_CO_PYPI_USERNAME/_PASSWORD into uv's allowlisted environment, and uv run hands that environment to the fetcher. So the fetcher reads the same pair uv would.
    • It calls nothing in the minter. It has no key lookup, no google-auth and no allowlist of its own. The key is read exactly as the wrapper reads it, because the wrapper is what reads it.
    • This stays inside the minter's own documented accepted limit: "whatever that uv spawns inherits the pair".
    • The index URL comes from the [[tool.uv.index]] named co-pypi, the name the two variables encode.
    • Run without the pair, the fetcher exits 2 and names the wrapper command.

    5. Four more sites the brief's regex misses

    git grep on 4ae39d3 finds these as well:

    • docs/DEPENDENCIES.md:36-50. It cites tests/test_co_pypi_token.py, and its "One key, no fallback … exits 2" no longer holds.
    • docs/ENVIRONMENT.md:65: "there is no fallback".
    • docs/ARCHITECTURE.md:67: the scripts list, which gains the fetcher.
    • .socraticodecontextartifacts.json:150: the path entry whose description is :151.

    /etc/replicator/.env holds the variable (one line). The repo .env doesn't [measured]. The operator's command will follow with the build.

    6. The real-wheel gate needs 0.36.0 and 0.37.0 too

    test_reproduces_136_for_0_36_to_0_37 reads those two versions. So the new fetcher fills REPLICATOR_TEST_CLOSURE_WHEELS with 0.36.0 and 0.37.0, and fetches 0.45.0 and 0.46.1 separately for the dry run. The reference for "byte for byte" is the pin comment's verbatim block in pyproject.toml, which #199 checked against the script.

    7. The deploy's first use runs the main checkout's venv

    deploy.sh calls bash $SRC/scripts/co-pypi.sh with $SRC set to the checkout and the working directory set to the release.

    • The shared wrapper's --project is its own parent directory. So the ship's deploy first runs uv run --project /home/exedev/replicator --frozen --only-group co-pypi against the main checkout's dev .venv. That run is inexact and google-auth[requests] is already installed, so I expect a no-op.
    • It then execs the release's uv sync --locked --no-dev.
    • co-pypi isn't a default group, so the release venv is unchanged.
    • No worker restart is owed by the env-file line either. The worker never reads the variable, and the deploy restarts the worker anyway.

    Proceeding with the build in worktree 201-shared-co-pypi, on 4ae39d3.

  5. gregoryfoster commented on Nov 15, 2026

    @gregoryfoster
    MemberAuthor

    Built: PR #205, gates on dd94444 (implementer, 2026-11-15)

    Six commits on 201-shared-co-pypi, based on 4ae39d3. The brief corrections above still stand. The fetch path is the one correction 4 describes: the fetcher runs under the wrapper and touches nothing in the minter.

    What changed

    • The shared wrapper. install.sh --co-pypi --replace installed it from skills 8f2dbfae, and install.sh --co-pypi --check reports current 1 1.
      • pyproject.toml gains co-pypi = ["google-auth[requests]>=2,<3"] and the ruff exclude for the minter.
      • The lock diff is that group alone, 9 lines.
      • The relock ran through the new wrapper with an empty environment (env -i PATH HOME), so it minted from the default key path.
      • tests/test_co_pypi_token.py is gone. Its coverage mapping, with the two real-loader gaps, is in correction 3.
    • The fetcher, scripts/co_pypi_wheels.py.
      • The test came first: tests/test_co_pypi_wheels.py failed 2 and errored 30 with no script, then passed 32 once it existed.
      • It ports wheels' cases: the sha256 mismatch, the missing version, the off-host token and redirect, the dropped connection, and the normalized names.
      • It adds cases for a missing digest, a missing package, a run outside the wrapper, a missing index, usage errors, an end-to-end run through main, and the script run as a command.
      • Two cases hold its portability: it names no repo, and it imports only the standard library.
    • The closure diff. Only its fetch-naming strings changed: the two error messages, the --wheels help, and the docstring's example.
      • It also had a docstring line the brief's regex missed (:39, "the wrapper's wheels verb"), which now names the fetcher.
      • The two errors now print the whole command to run, with the directory, version and packages filled in.
    • The deploy hooks export no key. test_the_wrapper_finds_its_key_without_an_env_file was rewritten test-first.
      • It checks from an empty environment that the hooks export no *CO_PYPI* variable.
      • It checks that the installed minter's key_path({}) is /etc/replicator/co-pypi-reader.json.
      • It was red against the old hooks.
    • The docs: README, COMMANDS, DEPENDENCIES, ENVIRONMENT, TESTING, ARCHITECTURE, and the manifest at :101, :151 and :171.
      • git grep -nE 'co-pypi\.sh wheels|REPLICATOR_CO_PYPI_CREDENTIALS' finds nothing outside docs/plans, skills-vendor and one negative assertion in a test.

    Gates

    The full output of each gate is saved on the host under ~/gate-logs/201/.

    Gate Baseline 4ae39d3 dd94444
    default suite, clean (env -i) 2028 passed, 1 skipped 2042 passed, 1 skipped, 99 % coverage
    default suite, loaded 2028 passed, 1 skipped 2042 passed, 1 skipped
    closure-diff real-wheel test, REPLICATOR_TEST_CLOSURE_WHEELS filled by the new fetcher (skips) 38 passed
    tests/deploy, clean and loaded 32 32
    tests/ci 44 44
    ruff check and format clean clean
    co-pypi.sh uv lock --check, empty environment n/a clean, with no fallback warning
    install.sh --co-pypi --check n/a current

    The counts reconcile: 2028 − 32 (test_co_pypi_token.py) + 14 (test_co_pypi.py) + 32 (test_co_pypi_wheels.py) = 2042. The one skip is the real-wheel test, which needs the directory variable.

    Wheels [measured]:

    • The new fetcher, run through the new wrapper with an empty environment, fetched 0.36.0, 0.37.0, 0.45.0 and 0.46.1 for all three packages.
    • All 12 files are byte-identical to the old wheels verb's, which I fetched before the swap.
    • The three 0.46.1 wheels match uv.lock's sha256.
    • Live failures behave as designed: version 0.99.0 exits 1 with co_core-0.99.0-py3-none-any.whl is not on …/co-core/ and writes nothing, and a run outside the wrapper exits 2 and names it.

    Dry pin move: co_core_closure_diff.py 0.45.0 0.46.1 --wheels <new fetcher's dir> is byte-identical (909 bytes) to the verbatim block in pyproject.toml's pin comment, which is #199's report. The new script sits in scripts/, one of the roots, and leaves the closure unchanged.

    CI run 43663540530 on dd94444 succeeded in all four jobs: test, lint, gcs and browser.

    Upstream-readiness list (nothing filed)

    The fetcher, scripts/co_pypi_wheels.py plus tests/test_co_pypi_wheels.py, is ready as-is.

    • It names no repo, and a test holds that.
    • It uses the standard library only, and a test holds that too.
    • Packages and versions come from arguments.
    • It gets the key only through the wrapper.
    • The index is the [[tool.uv.index]] named co-pypi.
    • Its header states its contract and its exit codes.
    • Its test looks for the script beside itself (the skill's assets/co-pypi/) or in ../scripts/ (a consumer), and needs no network, key or repo.

    What adoption would still need, all of it on the skill's side:

    1. Markers. Add the deploying-releases-co-pypi version and sha256 lines. install.sh --co-pypi would then carry three files, or two plus an optional third. Until this repo reinstalls with --replace, --check would call its copy foreign.
    2. A ruff exclude, like the minter's. The fetcher is formatted to this repo's ruff (line length 100). A consumer with other settings would reformat a marked file, so warn_setup and test_co_pypi.py's ruff check would have to cover it too.
    3. Where its test goes. It's behavioural, so upstream it belongs with tests/structural/test_deploying_releases_co_pypi.py, not in the consumer's file-read test.
    4. The two real-loader cases that correction 3 lists as uncovered upstream could go in at the same time. These are the real google-auth mint with a writer key in ADC, and a malformed key.

    The closure diff, scripts/co_core_closure_diff.py: what would have to move. This is noted only, and nothing here has changed.

    1. Defaults. DEFAULT_ROOTS = ["src", "scripts"] is this repo's layout, and DEFAULT_PACKAGES is co-core's three. Upstream, both should be required flags, or roots could default to the project's own package.
    2. The docstring's refs and framing: Keep the co-core import-closure diff as a script, not a rebuild per pin move #145, Catch up co-core 0.19.6 → 0.36.0 ahead of the generation release #140, Resolve co-core from the co-pypi Artifact Registry index, retiring the GCS wheelhouse (#169 part 1) #183, Adopt the shared co-pypi wrapper; move the wheel fetch to its own upstream-ready script #201, the sentence that says "Replicator", and the pyproject.toml pin-comment convention its output is shaped for.
    3. _normalize only turns - into _. The fetcher normalizes by the wheel spec (lowercase, runs of -_. → _). They agree for co-core's names. A general tool needs the fetcher's rule, or it will look for files the fetcher never writes.
    4. Test fixtures.
      • tests/test_co_core_closure_diff.py imports from scripts.co_core_closure_diff import …. Upstream it would load the script by path, as the fetcher's test does.
      • The synthetic wheels are built in tmp_path and portable, but they use co-core's names.
      • test_reproduces_136_for_0_36_to_0_37 stays here. It seeds from this repo's live src/ and scripts/, asserts this repo's counts (57 files, six not identical), and reads REPLICATOR_TEST_CLOSURE_WHEELS.
    5. It's already standard library only (argparse, ast, difflib, zipfile), and it already names the new fetch command.

    Operator: the env-file line

    /etc/replicator/.env line 7 is REPLICATOR_CO_PYPI_CREDENTIALS, and its value is the wrapper's default path [measured by comparison, not printed]. Line 6 above it is a stale comment, # SA key for the wheelhouse mirror (ExecStartPre sync_wheelhouse.py).. Nothing reads either line any more. The shared wrapper ignores the variable and leaves it off its allowlist, and the worker never read it. So no restart is owed, and the order relative to the deploy doesn't matter.

    sudo cp -p /etc/replicator/.env /etc/replicator/.env.bak-201
    sudo sed -i -e '/^# SA key for the wheelhouse mirror (ExecStartPre sync_wheelhouse\.py)\.$/d' \
                -e '/^REPLICATOR_CO_PYPI_CREDENTIALS=/d' /etc/replicator/.env
    sudo grep -nE 'CO_PYPI|wheelhouse' /etc/replicator/.env || echo "gone"   # expect: gone
    sudo stat -c '%U:%G %a' /etc/replicator/.env                           # expect the same as the .bak's
    sudo diff /etc/replicator/.env.bak-201 /etc/replicator/.env | grep -c '^<'   # expect: 2
    sudo rm /etc/replicator/.env.bak-201

    Skipped verification

    • ty: it doesn't gate, per AGENTS.md.
    • -m integration, -m gcs and -m browser locally: nothing under src/ changed. CI's gcs and browser jobs cover the latter two, and nothing covers integration in this round.
    • The manifest's codebase_update: SocratiCode indexes the main checkout, so it runs after the merge.
    • The deploy through the new wrapper: that's the ship's end-to-end proof. It hasn't run yet.

    Decisions to flag

    • AGENTS.md isn't edited. It says "BUILD_ID is the one deliberate exception" to the REPLICATOR_ prefix. CO_PYPI_KEY isn't a Replicator setting, and nothing here sets it, so ENVIRONMENT.md says that where the variable is documented.
    • tests/test_dependency_index.py overlaps test_co_pypi.py's index and lock checks. I left it alone because it's out of scope.
    • The ship's first deploy runs the main checkout's dev venv. The wrapper's --project is the checkout, and it runs uv run --frozen --only-group co-pypi there (correction 7). That sync is inexact, and google-auth[requests] is already installed there, so I expect no change.
  6. added 4 commits that reference this issue on Nov 15, 2026
  7. gregoryfoster commented on Nov 15, 2026

    @gregoryfoster
    MemberAuthor

    Shipped: a8900f5, deployed and live (2026-11-15)

    PR #205, fast-forwarded onto main (4ae39d3..a8900f5, 10 commits). Its Closes #201 became Refs before the push, so this closes only now, after the deploy.

    • The shared wrapper is in, unmodified: install.sh --co-pypi --check reports current 1 1 on main.
    • The fetcher is scripts/co_pypi_wheels.py. Its 12 real wheels are byte-identical to the old wheels verb's, and the dry pin move reproduces Catch up co-core 0.45 → 0.46 (co-pypi index; closure diff version-only) #199's report byte for byte (both in the Built comment above).
    • The deploy hooks export no key, and the docs name neither co-pypi.sh wheels nor the old key variable outside docs/plans. git grep on main finds only a test asserting the old name is absent.
    • The review (df459e0..a8900f5) changed four things:
      • CO_PYPI_KEY is documented as read by neither env file, instead of sitting under the service's.
      • The README's override table drops the unset row.
      • TESTING.md links to the fetch instead of quoting the command without the wrapper. That also brought the file back under its context budget.
      • The fetcher turns an unusable --dest or a non-UTF-8 index page into an exit-1 error that names it, where both used to end in a traceback. Its two tests failed first.

    Proof

    • Gates on a8900f5:
      • pre-ship: 2044 passed, 1 skipped (2042 + the review's 2);
      • the real-wheel closure test: 38 passed, its wheels re-fetched live and byte-identical;
      • tests/deploy and tests/ci: 76 passed;
      • ruff: clean.
    • CI on main: run 43674137985 succeeded in all four jobs (lint, test, gcs, browser).
    • The deploy, end to end:
      • scripts/deploy.sh a8900f5… exited 0 and logged "worker ready" on a8900f5d15a2 (pid 49915).
      • The deploying shell had no *CO_PYPI* variable at all. The wrapper mints a token before every uv it runs, so the release's uv sync --locked --no-dev minted from the default key path. That's the first production use of the new wrapper.
      • Live is releases/a8900f5d15a2, and its co-pypi.sh is the marked copy (ff3579e0…).
      • co-core and co-core-aio are 0.46.1 in the release.
      • The unit is active, NRestarts=0, and its journal shows no warning since the restart.
    • Restart: the deploy did the only one. The env-file line needs none.

    Follow-up:

    • The operator: removing lines 6–7 of /etc/replicator/.env (the old key variable and its stale wheelhouse comment) is still yours. The exact command is in the Built comment above. Nothing reads either line, so this can happen at any time and owes no restart.
    • Upstream: propose the fetcher, and later the closure diff, to deploying-releases with the operator's consent. The readiness list is in the Built comment, and nothing has been filed.

    Known limits:

    • -m integration wasn't run this round; nothing under src/ changed.
    • The manifest's re-embed: the capped driver update was refused because this session's server held the index lock. That server's watcher ran the incremental update instead, and codebase_context_search returns the new DEPENDENCIES.md and COMMANDS.md text.
  8. gregoryfoster commented on Nov 15, 2026

    @gregoryfoster
    MemberAuthor

    Follow-ups done (2026-11-15)

  9. gregoryfoster commented on Nov 15, 2026

    @gregoryfoster
    MemberAuthor

    Verified (coordinator, epic #154, 2026-11-15T04:3xZ)

    Everything below was read during this check-in [measured]:

    • Range: 7706bf0..a8900f5, 11 commits, all fast-forward. The skills-refresh hook's bump 4ae39d3 came first, then the 10 commits of PR #201 Adopt the shared co-pypi wrapper; fetch wheels with their own script #205.
    • CI: all four jobs (test, lint, gcs, browser) passed on a8900f5.
    • Live:
      • /srv/replicator/live points to releases/a8900f5d15a2, and the previous release is 7706bf05cc16. b61703aee439 was pruned (KEEP=0).
      • The worker is pid 49915. Its cwd is the release, it has 0 mappings under /home, and NRestarts=0.
      • BUILD_ID=a8900f5d15a2, and the worker was started at 03:52:41Z.
      • Since the start: 18 INFO lines, plus 6 non-JSON systemd lines and check_redis_floor (Redis 7.0.15).
      • The policy replay read 511 messages across 10 hosts. Capture and persist are enabled.
      • journalctl -t replicator-failure is empty, and no unit has failed.
      • The release carries co-core, co-core-aio and co-core-sync at 0.46.1.
    • The wrapper:
      • The live co-pypi.sh is byte-identical to main's.
      • Both files carry the deploying-releases-co-pypi version 1 markers (ff3579e0…, f6c3fedd…).
      • install.sh --co-pypi --check reports current 1 1 against skills 8f2dbfae.
    • The env file: /etc/replicator/.env has no line matching CO_PYPI or wheelhouse. It is root:exedev 640, no .bak-201 is left, and a loaded shell exports no *CO_PYPI* variable.
    • The fetcher, end to end, run by me:
      • I ran env -i PATH HOME bash scripts/co-pypi.sh uv run --no-project python scripts/co_pypi_wheels.py for 0.36.0, 0.37.0, 0.45.0 and 0.46.1. It exited 0 and wrote 12 wheels.
      • The three 0.46.1 wheels match uv.lock's sha256.
      • With those wheels, tests/test_co_core_closure_diff.py ran 38 passed.
      • The 0.45.0 → 0.46.1 dry run is 909 bytes, and every line of it appears in pyproject.toml's pin comment.
    • Suites on a8900f5:
      • default suite: 2044 passed, 1 skipped, both clean (env -u REPLICATOR_REDIS_URL) and loaded;
      • tests/deploy and tests/ci: 76 passed;
      • ruff check and ruff format --check: clean.
    • Context artifacts: codebase_context_search returns Adopt the shared co-pypi wrapper; move the wheel fetch to its own upstream-ready script #201's new text from DEPENDENCIES.md (the fetcher paragraph) and COMMANDS.md (the fetch command). That discharges the epic's manifest re-embed debt from Run replicator's unit from a release, not /home/exedev/replicator (CannObserv/broker#22's rollout, replicator first) #151.
    • Disk: root is at 51.82%, with 12.74G free.

    Carried to the epic:

    Nothing is owed. This issue is closed already.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions