Repository navigation
Bug hunt ledger: pip / requirements.txt #309
Mikola Lysenko (mikolalysenko)
started this conversation in
General
Replies: 55 comments
|
[agent] 2026-09-30: pip / requirements.txt bug-hunt run This is the first run, so the ledger started empty and there were no Tested: main Setup: a local mock patch API (the same shape as the Pipenv, Poetry and Hatch routines) served a patched Cells
Issues
False positives ruled out
Probe runs
I couldn't delete the probe branches: the git proxy hangs up on ref-delete pushes, the same as the siblings. A maintainer should delete Next
|
0 replies
0 replies
|
[agent] 2026-10-01: maintainer note: test global ( This is a maintainer request, not a run report. Add it to the top of the backlog and keep it there until the cells below are covered. Ask: make sure we correctly scan global installs when Where pip / requirements.txt puts global installs: What to check (prove each with a real global install, not by reading source):
Add OS × pip / requirements.txt version cells for |
0 replies
|
[agent] 2026-10-01: pip / requirements.txt bug-hunt run Tested: main Re-triageMain hasn't moved since the last run, so #376, #409, #410 and #412 can't have changed. No comments. Cells
Issues
Needs a maintainer decision (not filed)
False positives ruled out
Probe runs
I couldn't delete Next
|
0 replies
|
[agent] 2026-10-01: pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triage
Cells
Issues
False positives ruled out
Probe runs
I couldn't delete Next
|
0 replies
|
[agent] 2026-10-01: pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triage
Cells
Issues
False positives ruled out
Probe runs
I couldn't delete Next
|
0 replies
|
[agent] 2026-10-02: pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triageMain hasn't moved since the last run (#478 for #475, #503 for #328 and #481 are still open PRs), so #409 / #410 / #412 / #475 have nothing new to re-check. No comments. Cells
Issues
False positives ruled out
Probe runs
I couldn't delete Next
|
0 replies
|
[agent] 2026-10-02: pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triageMain hasn't moved. #412 / #523 are claimed by draft PR #530; #475 (PR #478) and #328 (PR #503) are still open. Nothing to re-check. Cells
Issues
False positives ruled out
Next
|
0 replies
|
[agent] 2026-10-02: pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triageMain hasn't moved and no human replies are on #309, so nothing needed re-checking. Open: #409, #410, #412, #475, #523, #542. Cells (Linux)
Sandbox hazard (again): agent IssuesNone filed, commented on or closed. False positives ruled out
Next
|
0 replies
|
[agent] Janitor: ledger drift. The coverage matrix still lists these issues as
This is a heads-up only. The janitor never edits ledgers. Generated by Claude Code |
0 replies
|
[agent] Janitor: ledger drift. This ledger still lists these issues as failing, but they are now closed:
Please re-check them and update the matrix on your next run. Generated by Claude Code |
0 replies
|
[agent] 2026-10-02: pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triage
Cells (Linux, pip 24.0 / py3.11)
Issues
False positives ruled out
Next
|
0 replies
|
[agent] 2026-10-03: pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triagemain hasn't moved, so #604, #542, #410 and #409 stand as last run. I didn't comment. #604 was re-observed (vendored lock-only Cells (Linux)
Issues
False positives and limits ruled out
Next
|
0 replies
|
[agent] 2026-10-03 (08:23 UTC): pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triagemain hasn't moved, so #638, #604, #542, #410 and #409 stand as last run. I didn't comment. The #410 sole-pin refusal was re-observed on the hosted → vendored takeover ( Cells (Linux)
Issues
False positives and limits ruled out
Next
|
0 replies
|
[agent] 2026-10-03 (14:18 UTC): pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triagemain hasn't moved and no fix PR touches the open pip issues, so #668, #638, #604, #542, #410 and #409 stand. I didn't comment. Cells (Linux)
Issues
False positives and limits ruled out
Next
|
0 replies
|
[agent] 2026-10-08 (UTC afternoon): pip / requirements.txt bug-hunt run Tested: main Setup:
Re-triage
New cells (Linux)
IssuesNone filed, commented on or closed this run. False positives ruled out
Next
|
0 replies
|
[agent] 2026-10-08 (UTC evening): pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.11 and 20.3.4 / py3.8 (uv venvs). Vendored runs used a throwaway, uncommitted test binary serving Re-triage (fixes that merged today)
New cells (Linux)
Issues
False positives ruled out
Next
|
0 replies
|
[agent] 2026-10-08: handover from the uv bug-hunt routine (#310) On main printf 'six # runtime\nattrs\n-c constraints.txt\n' > requirements.txt
printf 'six==1.16.0\nattrs==23.2.0\n' > constraints.txt
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt
git add -A && git commit -qm init
socket-patch scan --mode hosted --yes # exit 0; `six` → `six @ <socket url>#sha256=…`; pip installs patched (ok)
socket-patch rollback --yes --json # exit 0, success, hosted.reverted [six]
git diff
# -six # runtime
# +six==1.16.0 # runtimeExpected: the rollback restores the user's original line byte-identically ( Generated by Claude Code |
0 replies
|
[agent] 2026-10-09: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.11 and 20.3.4 / py3.8 (uv venvs). A standalone Python hosted mock served the batch, by-package, package grant, view and wheel routes, with a real patched six 1.16.0 wheel. The v4.0.0 release binary and a build of 03b9418 (the parent of #1152) were used for bisecting. The mocks were stopped after the run. No probe branches: every cell was OS-independent. Re-triage
Cells (Linux)
Issues
Handover from uv (20261008T213301Z-from-uv)
False positives ruled out
Next
Generated by Claude Code |
0 replies
|
[agent] 2026-10-09: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.13, 24.0 / py3.12 and 20.3.4 / py3.8 (uv venvs). A logging batch mock (records the purls requested, answers empty) on 127.0.0.1, and an empty Re-triage
Cells (Linux, lock-only)
Issues
False positives ruled out
Next
Generated by Claude Code |
0 replies
|
[agent] Janitor: drift. #1119 was closed as completed on 2026-10-09 00:55Z by #1152 ( Generated by Claude Code |
0 replies
|
[agent] 2026-10-09: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.13 and 20.3.4 / py3.8 (uv venvs). A logging batch mock on 127.0.0.1 (records the purls requested, answers empty), Re-triage
Cells (Linux, lock-only)
Issues
False positives ruled out
Next
Generated by Claude Code |
0 replies
|
[agent] 2026-10-09: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.13 and 20.3.4 / py3.8 (uv venvs), pip-tools on pip 25.2. A throwaway Python mock of Re-triage
Cells (Linux)
Issues
False positives ruled out
Next
Generated by Claude Code |
0 replies
|
[agent] Janitor: ledger drift. These issues are now closed but this ledger still lists them as Generated by Claude Code |
0 replies
|
[agent] 2026-10-12: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.13 and 20.3.4 / py3.8 (uv venvs), pip-tools 7.6.2 on pip 25.2. A fresh standalone Python mock (batch, by-package, package grant with sha256 + sha512, view with real Probe branches: blocked. Re-triage
Cells (Linux)
Issues
False positives ruled out / not filed
Next
Generated by Claude Code |
0 replies
|
[agent] 2026-10-13: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.13 (uv venvs), pip 24.0 / Debian py3.11 ( Probe branches: none pushed. Deletion was denied on run 38, and a probe branch must be deleted before the run ends. Everything ran on Linux. The nine old Re-triage
Cells (Linux)
Issues
False positives ruled out / not filed
Next
Generated by Claude Code |
0 replies
|
[agent] 2026-10-13: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.13, pip 24.0 / py3.11, pip 20.3.4 / py3.8 (uv venvs, seeded), virtualenv 21.4.3 over Debian Probe branches: still blocked. Re-triage
Cells (Linux)
Issues
False positives ruled out / not filed
Next
Generated by Claude Code |
0 replies
|
[agent] 2026-10-14: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.13, pip 25.0.1 / py3.12, pip 24.0 / py3.11, pip 20.3.4 / py3.8 (uv venvs). A fresh standalone Python mock (batch, by-package, view with git-sha256 + inline blobs, blob route, package grant with sha256 + sha512 SRI, hosted wheel route) served patched copies of the real wheels for Probe branches: still blocked. Re-triage
Cells (Linux)
Issues
False positives ruled out / not filed
Next
Generated by Claude Code |
0 replies
|
[agent] 2026-10-14: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 / py3.13, 25.1.1 / py3.12, 23.0.1, 21.3.1 / py3.11, 20.3.4 / py3.8, plus CPython 3.14.6 and free-threaded 3.14.6t (uv venvs, Probe branches: still blocked. Re-triage
Cells (Linux)
Issues
False positives ruled out / not filed
Next
Generated by Claude Code |
0 replies
|
[agent] 2026-10-15: pip / requirements.txt bug-hunt run Tested: main Setup: real pip 26.2.1 on CPython 3.11, with the repo's own Probe branches: none pushed (deletion is still blocked, see earlier entries). Linux only. Re-triage
Cells (Linux)
Issues
False positives ruled out
Next
Generated by Claude Code |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
[agent] Progress ledger for the scheduled pip / requirements.txt bug-hunt routine (label pm:pip).
Last run: 2026-10-15 (forty-third run), main
e76cb9b, latest release v5.0.1. New: #1567 (vendored requirements.txt rewrite re-joins a mixed CRLF/LF file with one terminator;vendor --revert/removeexit 0 but can't restore it). #1104 still fails. Main had no PyPI changes since306ec81. Probe branches are still blocked (none pushed). Open: #1567, #1104, #977, #819, #740, #668, #638.Coverage matrix
-rinclude, lock-only==1.16)-g(--user)six == X)six @ mirror)==1.16)--json)vendor)vendor --dry-runover hostedget --mode vendoredover hosted-cby a sibling--prune)pip lockpylock.toml (hosted / vendored)__pycache__--max-new-patches(hosted)get <name>PyPI spelling--max-new-patches(vendored)${VAR}-rinclude (lock-only)-ron one line (lock-only)-rinclude (vex / rollback)apply --check(agent, #1029)scan --pruneliveness (#1050)pip freeze-rinclude (--require-hashes)\continuation at EOF (lock-only)redirect_requirements_takeover_unreachable, verified 2026-10-05)pip lockneeds pip ≥ 25.1)6fe81ad)--global-prefix, unwritable)-rinclude, marker, spellings, transitive, URL includes; 2026-10-06)-r,-csupersede, pip-tools regen pass 2026-10-07)6fe81ad)pip install/pip sync)redirect_requirements_takeover_unreachable, verified 2026-10-05)b96a785: hosted refusescandidate_file_unreadable, lock-only decodes)--prune); moved-into-include / root-level sibling revert fixed #867 (verified 2026-10-08 ona845bf9); subdirectoryrequirements/*.txtfail #11679c43dfc, byte-identical; re-check the takeover and the pip 25.1 lock);pylock.<name>.toml, sdist-only, + requirements.txt,--dry-runpass (2026-10-05)__pycache__+-OO)-spelling, retained / floor /enabled: falsepass)6fe81ad)b96a785)a845bf9: vex attests, rollback byte-identical, both-rorders)VIRTUAL_ENV, downgrade skip,-g; 2026-10-08)file:///drift-kept (known non-bug); revert / remove / takeover with the file inrequirements/fail #1167--user)Scripts/+Lib/)--user,%APPDATA%\Python)pip 20.3.4 on py3.13 is blocked (no
distutils).setupwas removed in v5, so the old setup column (#377, #378) is retired; both issues are closed.Run forty-three (2026-10-15, pip 26.2.1 / py3.11, main
e76cb9b): vendored over a mixed CRLF/LF requirements.txt fails → #1567 (LF-majority: untouched CRLF lines become LF; CRLF-majority: the pin's own LF becomes CRLF;vendor --revertandremoveexit 0 but don't restore the file; pip installs patched). #1104 re-checked: still fails.Run forty-two (2026-10-14 evening, pip 26.2.1 / py3.13, 25.1.1 / py3.12, 23.0.1, 21.3.1, 20.3.4 / py3.8, CPython 3.14.6 and 3.14.6t, main
306ec81): #1524 verified fixed (vendor --revert,removeand the hosted takeover of a requirements.txt + pylock.toml entry restore both files; pip 26.2.1 and 25.1.1). Mixed entries also pass with the pin in a-rinclude, two pylocks (pylock.toml+pylock.dev.toml), uvpip compile --format pylock.toml, and a pylock deleted by the user. Vendoredrepairrebuilds a deleted, corrupted or whole-dir-deleted wheel (fresh--no-indexinstall patched), a mixed entry's wheel, and a manifest-less checkout; a ledger-less one refusesvendor_ledger_missing(exit 1). Format-control lines (--no-binary :all:/six,--only-binary :none:,--no-index,--prefer-binary) pass hosted and vendored. Split-line continuations (six==\/1.16.0,six \/==X, marker on the next line,\then a blank line,# c \swallowing the next line) pass hosted, vendored, install and revert. Agent / hosted / vex on CPython 3.14 and free-threaded 3.14t (lib/python3.14t) pass.Run forty-one (2026-10-14, pip 26.2.1 / py3.13, 25.0.1 / py3.12, 24.0 / py3.11, 20.3.4 / py3.8, main
313d532): dotted / underscored PyPI names with odd spellings pass hosted (plain, extras + markers, pip-compile hashed shapes under--require-hashes), agent and vendored (byte-identical revert), with vex and rollback; a unicode + space project path passes agent / hosted. Backlog 1h found #1524: a vendored requirements.txt plus a laterpip lock .pylock.toml leaves a mixedpython-lockentry that revert / remove / hosted takeover can't unwind.Run forty (2026-10-13 evening, pip 26.2.1 / py3.13, 24.0 / py3.11, 20.3.4 / py3.8, virtualenv 21.4.3 over Debian py3.11, main
d4af25a): #1486 regression pass (hosted hashed continuation andpip lockpylock.toml, install patched, vex attests); pyproject +uv.lock+uv exportrequirements.txt vendored pass (all three wired once,uv sync --lockedandpip install --require-hashespatched, byte-identical revert; backlog 3k done); virtualenv-made system-site venv hosted pass (stale warning names the Debian dir, vex omits);--grouphosted = #638 class; sha512-hashed continuation hosted pass, rollback restores the PyPI sha256 form (installs). #1104 / #977 still fail.Run thirty-nine (2026-10-13, pip 26.2.1 / py3.13, 24.0 / py3.11 (Debian), 20.3.4 / py3.8 (uv-managed base), pip-tools on 26.2.1, main
15c2c83): #409 verified fixed: hosted over a system-site venv now warnsredirect_pypi_stale_installandvexomits (not_applied) for a base copy in Debian/usr/lib/python3/dist-packages, inPYTHONUSERBASE, and in a uv-managed basesite-packages(stdlib venv anduv venv --system-site-packages); afterpip install --force-reinstallthe venv copy shadows andvexattests. Agent mode over a system-site venv now patches the base copy (root: apply + rollback byte-identical; non-root:apply_failedPermission denied, exit 1, vex omits). #1437 verified for pip: a vendor ledger beside a hosted pin ismode_ambiguous(exit 2, nothing written) for plain, CRLF, BOM, hashed continuation, marker, comment,Six, extras, a-rinclude,pylock.tomlandpylock.dev.toml, onscanand bareget(name, purl, uuid). Hosted supersede A → B over a pip-compile--generate-hashesfile passes (--require-hashesinstall, vex names B), and rollback restores the original two-hash block byte for byte.Run thirty-eight (2026-10-12, pip 26.2.1 / py3.13 and 20.3.4 / py3.8, pip-tools 7.6.2, main
7236190): #1365 and #604 verified fixed. The #1381 shared grammar passes hosted, vendored, vex and real pip for the space-form--hash, space-less markers, paren pins with markers, markers on continuation lines, upper-case names and long--config-settings. A crash at eachgroup_commitfailpoint during a vendored → hosted takeover recovers requirements.txt but leaves an orphan vendored unit (not filed, SIGKILL scope). Lock-only missessix==X -C a=b(not filed, exotic). #1167 holds with realpip freeze > requirements/lock.txt.Run thirty-seven (2026-10-09 evening, pip 26.2.1 / py3.13 and 20.3.4 / py3.8, pip-tools on pip 25.2, main
7eec31b): #542 verified fixed (mirror /file:/// sdist direct refs refused withredirect_requirements_direct_reference, file byte-identical, twice). #1034 target grammar passes over a hosted pin:rollback/removewithsix,Six,SIX, versionless and versioned purls (any case) and the uuid all restoresix==1.16.0;six_is correctly not found. Hosted over a CR-only / stray-CR / FF root refuses (redirect_requirements_entry_not_found/_version_ambiguous), nothing written. Vendored over a CR-only root appends a(transitive)line that pip 20.3.4 / 26.2.1 install patched;vendor --revertleaves the file pip-equivalent (final CR becomes LF). Realpip-compile --generate-hashesoutput: hosted rewrite, patched install on pip 20.3.4 / 26.2.1, byte-identical rollback. Revert residual probe: a rootrequirements.pip,requirements.in,Requirements-Prod.TXTor extensionless file naming the vendored wheel does not keep it (only*.txt/*.lock); not filed, see backlog question 0f.Run thirty-six (2026-10-09 afternoon, pip 26.2.1 / py3.13 and 20.3.4 / py3.8, main
40de3d5): #1249 verified fixed (EOF\with no newline, CRLF, and a-rinclude).-r dev.txt \as the last line, plus-e ./--editable ./pkgbeside a pin, pass lock-only. New fail #1281: bare-CR / FF / NEL / U+2028 line breaks (v4.0.0 finds only the first pin; main finds none). Noted in #1281 but not filed separately:six==1.16.0 \\(two backslashes) at EOF. pip'sstrip("\\")removes both and v4.0.0 discovers it, but main drops it.Run thirty-five (2026-10-09, pip 26.2.1 / py3.13, 24.0 / py3.12 and 20.3.4 / py3.8, main
e03a666): #1212 verified fixed (four codecs; root file and-rinclude). New fail #1249 (dangling\on the last line, every mode, root and include, LF and CRLF). A lock-only parser sweep passes: a continuation inside the name or version, a# …\comment line, a tab-spaced pin, a marker without spaces, uppercase names,_/.names normalised to-, extras, and hashed continuations at EOF.Run thirty-four (2026-10-09, pip 26.2.1 / py3.11 and 20.3.4 / py3.8, main
793edd4): #1119 verified fixed (a latin-1 coding line with a non-ASCII comment is discovered lock-only). Hosted over a non-UTF-8 latin-1 root refuses loudly (candidate_file_unreadable, exit 1, nothing written). A UTF-8 root with a latin-1-rinclude rewrites, attests (vex) and rolls back fine. An unmodelled-codec coding line (iso-8859-15 / cp1250 / latin-9 / mac-roman / gbk) on an ASCII file fails #1212 (first bad 793edd4 / #1152). #604 still fails. The uv handover (hosted rollback of a-c-constrained unpinned line restoressix==1.16.0) is backlog question 0c; not filed.Run thirty-three (2026-10-08 evening, pip 26.2.1 / py3.11 and 20.3.4 / py3.8, main
a845bf9): the #867 and #1086 fixes are verified. A vendor line in an unincluded subdirectory requirements file (moved there, or written bypip freeze > requirements/lock.txt) loses its wheel onvendor --revert,removeand the hosted takeover (#1167). The root-levellock.txtcontrol is kept.Run thirty-two (2026-10-08 afternoon, pip 26.2.1 / py3.11 and 20.3.4 / py3.8, main
3b4ac84): #1050 prune liveness passes (bumped pin reverted; quoted include, missing wheel kept). Agent mode in a stdlib venv withlib64 → libpasses. Hosted →pip freezeregen re-scans byte-identical and rolls back. Hosted hashed root + hashed include under--require-hashespasses. A lock-only PEP 735--groupproject finds nothing (documented). #1104, #977 still fail.Run thirty-one (2026-10-08 morning, pip 26.2.1 / py3.11, 24.0 / py3.12 and 20.3.4 / py3.8, main
b96a785): #721 and #1028 fixes verified.apply --check(#1029) passes on venv, multi-venv, externalVIRTUAL_ENV, downgrade,-gand egg-info. Hostedredirect.patches[]rows and #1035's uuid / purl remove and rollback pass. Latin-1 + coding line lock-only fails (#1119).Run thirty (2026-10-08, pip 26.2.1 / py3.11 and pip 20.3.4 / py3.8, main
d7f8679): vendored marker split in one file (#929) passes, and so does its vendored → hosted takeover. A marker split across a-rinclude fails (#1104); hosted handles it. The vendored → hosted takeover over a drifted vendor comment refuses cleanly (redirect_vendored_revert_failed). Vendored root pin + the same pin in-c constraints.txtpasses (install, vex, byte-identical revert).Run twenty-one (2026-10-05, pip 26.2.1 / py3.11, main
9c43dfc): #858 atomic writes (file mode, CRLF, revert byte-identity) and #865 digests (hashed hosted / vendored requirements.txt with--require-hashes, vendored pylock hashes) pass; vendored continuation grammar (six==\+ newline, a comment ending in\) passes.Commands covered on Linux: scan (all modes), get (hosted, agent
-g), rollback, remove, vendored takeover, vex (hosted with the mock origin, vendored,-g), repair, list, concurrent runs. Also covered (2026-10-02): free-threaded CPython 3.13t venvs, a.venvsymlink to an out-of-tree venv, a 16-case hosted grammar sweep with fresh installs and rollbacks, virtualenv layouts, pip-toolspip-compile --generate-hashes/pip-syncover a hosted file plus rollback,-cconstraints with an unpinned root,pip install --targettrees,--jsonenvelopes of rollback / remove failures, and interrupted (SIGTERM / SIGKILL) hosted scans. Run ten (2026-10-03, pip 26.2.1 / py3.13 and pip 20.3.4 / py3.8): an 18-shape hosted and vendored grammar sweep with freshpip install -rplus rollback and reinstall, vendored lock-onlyscanover-r/-rX/--requirement=/ CRLF / subdirectory includes,-cconstraints, and grant-token re-pin: all pass or refuse explicitly. Run eight (pip 24.0 / py3.11 and pip 20.3.4 / py3.8): PEP 503 name normalisation (_/-/./ case, venv and lock-only), duplicate and marker-split pins, BOM, no trailing newline,--no-index+--find-links, per-line--config-settings,--require-hasheswith multiple and sha384 hashes,-e/ VCS neighbours (lock-only +vex), symlinked root and include files, out-of-root includes,--prefixtrees: all pass or refuse explicitly.Backlog
1k. #1567 follow-ups: a mixed CRLF/LF
-rinclude, the superseding re-vendor (plan_rewire), and the hosted → vendored takeover over a mixed file (hosted keeps per-line endings, so it should be clean). Re-verify once fixed.1j. #1281 was closed not planned (2026-10-09). Hosted and vendored over CR-only roots were checked on run 37 and are fail-closed or install patched; nothing left to do unless the maintainer reopens it.
vexattests a hosted pin (six@1.16.0) from the lockfile basis when the venv holds a different version (six 1.15.0), with no warning (see the 20261002T142728Z entry). (b) The non--gno-venv fallback to global site-packages (20261001T083942Z). (c) Hosted rollback restores an unpinnedsix(or-c-constrained root) assix==1.16.0, because hosted mode is stateless; should the rewrite refuse unpinned roots, or keep them unpinned on restore? The uv routine handed over the same case on 2026-10-08 (entry 20261008T213301Z-from-uv), and it still waits on a maintainer call. (e) resolved: Fix remove/rollback missing PyPI name spellings (#1024) #1025 fixed non-canonical purls inremove/rollback(verified 2026-10-07 on05ecc6e). (d) Vendored mode wires a root(transitive)line but leaves a directsix==Xpin in a siblingrequirements-dev.txt, sopip install -r requirements-dev.txtalone stays unpatched (same class as Vendored mode in a Pipenv project wires only Pipfile.lock and silently leaves a sibling requirements.txt unpatched, and the hosted → vendored takeover reverts that file's hosted pin to plain PyPI #612 / PyPI hosted scan lacks a specific missing-requirements warning after redirect_unconfirmed reporting #638). (g) Plain-pip hostedredirect_pypi_stale_installfor a base-interpreter copy in a--system-site-packagesvenv says "Reinstall from the rewritten lock … or uninstall this package before installing": plainpip install -ris a no-op over the base copy andpip uninstallrefuses ("outside environment"); only--force-reinstall/--ignore-installedworks. This is the pip twin of Pipenv Pipenv stale-install remedy can't replace a base-interpreter copy in a--site-packagesvenv:pip uninstallrefuses it andpipenv synckeeps it, so following the warning leaves the package unpatched #1450 (fixed by Fix Pipenv remedy for base-interpreter copies (#1450) #1455 for Pipenv only). Not filed, because it matches CLI_CONTRACT.md's generic remedy and vex fails closed (2026-10-13). (f) The vendored revert residual probe (vendor/pypi.rsis_export_name) reads only*.txt/*.lock(case-sensitive) at the root and in subdirectories, so a rootrequirements.pip(legacy convention),requirements.in,Requirements.TXTor extensionless requirements file that installs from the vendored wheel doesn't keep it:vendor --revertdeletes the wheel (exit 0) andpip install -rfrom that file then fails. Not filed: under the v5 normal-toolchain scope no standard tool writes the vendored path into such a file (pip-compile output is.txtand is guarded).1a. Since #1152, a plain-ASCII requirements.txt whose coding line names an unmodelled codec (iso-8859-15, cp1250, gbk…) is read as absent: scan finds nothing (exit 0), and vex / rollback can't see an existing hosted pin #1212 lock-only is verified fixed (2026-10-09). Still open: hosted / vex / rollback and the vendored planner over such a file. Lock-only requirements.txt discovery drops a pin whose last line ends in a dangling
\continuation (six==1.16.0 \at EOF): scan exits 0 with "No patches", but pip installs it #1249 was verified fixed on 2026-10-09 (lock-only). Still open: check that hosted surfacesredirect_requirements_continuationon a lock-only checkout.-rtree (false "not pinned to ==1.16.0", exit 1), so a fresh install stays unpatched #1104, Vendored requirements.txt can't be reverted once the user touches the# socket-patch vendor:comment:vendor --revertdrift-keeps forever (exit 0), and the suggested "re-vendor" remedy is a no-op or adds a duplicate line #977, Agent-mode apply of a same-size PyPI patch in the same second aspip installleaves pip's__pycache__bytecode valid, so Python keeps running the unpatched code while apply reports it patched #819, Hosted and vendored rewrites of a requirements.txt that another file uses as-cconstraints break everypip installon pip 20.3–21.0 ("Links are not allowed as constraints"), with no warning #740,vendor --dry-runover a hosted requirements.txt pin previews apypi_requirement_not_pinnedfailure (exit 1), but the realvendortakes it over and succeeds #668 and PyPI hosted scan lacks a specific missing-requirements warning after redirect_unconfirmed reporting #638 as fixes land (In a--system-site-packagesvenv, pip keeps the base interpreter's unpatched copy after a hosted rewrite, no stale-install warning fires, andvexattests it as patched #409 verified fixed on Linux 2026-10-13; macOS / Windows re-verify needs a probe). Vendored requirements.txt still refuses a marker-split package when the other version's branch sits in a different file of the-rtree (false "not pinned to ==1.16.0", exit 1), so a fresh install stays unpatched #1104 /vendor --dry-runover a hosted requirements.txt pin previews apypi_requirement_not_pinnedfailure (exit 1), but the realvendortakes it over and succeeds #668 weren't re-checked on run 39 (no vendoring-service mock this run). Verify the merged fixes for Hostedrollback,removeand the vendored takeover refuse a requirements.txt whose only requirements are hosted pins (six==1.16.0alone can be patched but never unpatched) #410,get <name>doesn't PEP 503-normalise PyPI names, soget typing_extensionsorget ruamel.yamlreports "No packages matching" (exit 0) for an installed, patchable package #926, socket.ymlignorePackages/packagesandscan --packagedon't PEP 503-normalise PyPI names, soignorePackages: ["typing_extensions"]is silently ignored and the package is patched anyway #910 and Vendored requirements.txt fromuv pip compile --universalrefuses a marker-split package (six==1.16.0 ; python < 3.12+six==1.17.0 ; python >= 3.12) as "not pinned to ==1.16.0", while --dry-run previews would_vendor and hosted / vendored pylock handle the same split #928 with real pip. Poetry hosted ⇄ vendored mode switch is refused, and blames a "user-authored" source that socket-patch wrote itself #328 / Hosted requirements.txt rewrite skips PEP 440-equivalent pins likesix==1.16for an installed 1.16.0, soscanexits 0 and pip installs the unpatched release (regression from v4.0.0) #475 re-verify on pip 20.3.4 / 26.x is still open; Vendored requirements.txt never picks up a superseding patch: the re-vendor to a new uuid fails with pypi_requirements_already_vendored (exit 1), though--dry-runpreviewswould_revendorand the contract says it re-vendors automatically #765 on pip 20.3.4. Poetry hosted ⇄ vendored mode switch is refused, and blames a "user-authored" source that socket-patch wrote itself #328 / Hosted requirements.txt rewrite skips PEP 440-equivalent pins likesix==1.16for an installed 1.16.0, soscanexits 0 and pip installs the unpatched release (regression from v4.0.0) #475 re-verify on pip 20.3.4 / 26.x is still open.1b. Agent-mode apply of a same-size PyPI patch in the same second as
pip installleaves pip's__pycache__bytecode valid, so Python keeps running the unpatched code while apply reports it patched #819 on-g/--usersite-packages. (The rollback variant is in the Agent-mode apply of a same-size PyPI patch in the same second aspip installleaves pip's__pycache__bytecode valid, so Python keeps running the unpatched code while apply reports it patched #819 body; the pip-lock variants were done 2026-10-05; the new-file case became Agent-mode rollback / remove of a PyPI patch that added a file in a new directory leaves the empty directory in site-packages, so Python still imports it as a namespace package #838.)1g. done 2026-10-08: a hashed include beside a hashed hosted root under
--require-hashespasses on pip 26.2.1 and 20.3.4.1h. done: Vendored requirements.txt can't be reverted or removed once a later scan also wires a pylock.toml: the entry flips to flavor
python-lock, whose revert calls its ownrequirements_linewiring "unexpected" drift (exit 0, nothing restored) #1524 fixed by Fix vendored PyPI revert of entries wired across flavors (#1524) #1526 and verified 2026-10-14 evening (include pin, two pylocks, uv-written pylock, pylock deleted). A vendor line moved by hand into a-cconstraints file drift-keeps (fail-safe, user edit). Still open: a uv.lock (not pylock) added after vendoring requirements.txt.1i. Make the vendored-to-hosted takeover atomic #1039 crash cells done on 2026-10-12 (they recover, with an orphan unit left). Still open: a Windows backslash vendored path vs the residual probe's forward-slash needle.
1f. Fix hosted PyPI pinning platform-only wheels (#701, #932) #984 follow-up:
cp311-none-anygrants stay redirectable by design. Check pip 3.12 against such a hosted pin, and a build-tag wheel name inwheel_platform_from_filename.1e. Lock-only requirements.txt discovery skips
-rincludes that are quoted, backslash-escaped or use${VAR}(-r "dev reqs.txt",--requirement="dev.txt",-r ${DIR}/dev.txt), so the scan exits 0 with "No patches" while pip installs the include's unpatched pins #994 follow-ups: done 2026-10-08. A quoted include's pin is vendored in place (no root(transitive)line), andvexand prune see it.1d.
get <name>follow-ups afterget <name>doesn't PEP 503-normalise PyPI names, soget typing_extensionsorget ruamel.yamlreports "No packages matching" (exit 0) for an installed, patchable package #926 (fuzzy prefix picks such asget ruamelvsruamel.yaml.clib;policy_bypassedunder Fix PyPI package specs ignoring PEP 503 spellings (#910) #911). The vendored supersede under cap 0 and the deferred hashed-include pin passed on 2026-10-06; next is a supersede with the pin also in a-cconstraints file.1c. Hosted rollback and remove always refuse a
pip lockpylock.toml ("no sibling registry package shows the registry and artifact fields"), because pip writes[[packages.wheels]]tables, not uv's inlinewheels = [...]#804 follow-ups: re-check the hosted → vendored takeover over apip lockfile and a pip 25.1 lock on main.3k. done 2026-10-13: pyproject +
uv.lock+uv exportrequirements.txt vendored wires each file once and reverts byte for byte.pip freeze >(UTF-16) file end to end (blocked while probe branches can't be deleted: on 2026-10-12git push --deletefailed and was then denied, and nine oldbughunt/pip/*branches are still on the remote).3b. UTF-32 BOM: re-check with the Hosted and lock-only scans treat a UTF-16 requirements.txt (what Windows PowerShell's
pip freeze >writes) as absent: exit 0, no warning, and pip keeps installing the unpatched pin #721 fix (Fix UTF-16 requirements.txt silently skipped (#721) #724 decodes UTF-32 BE; an LE BOM matches the UTF-16 LE prefix first, as in pip's own BOM order). The PEP 263 coding line is Lock-only scans still drop a requirements.txt pip decodes through a PEP 263 coding line (latin-1): "No pypi packages found", exit 0, while the same project with a venv refusescandidate_file_unreadable#1119; after it's fixed, check the vendored wet run andvexover it.3d. pip ≥ 25.1
--group: lock-only finds nothing (documented, see Known non-bugs). Hosted with a venv is still untested (expectredirect_unconfirmed, the PyPI hosted scan lacks a specific missing-requirements warning after redirect_unconfirmed reporting #638 class).3c. PyPy venvs (
lib/pypy3.X/site-packagesis not matched by the crawler'spython3.*glob): blocked in the sandbox (uv can't download PyPy), and outside the CPython scope.-g) mode: Homebrew / PEP 668 interpreters, the py launcher with several interpreters, pipx venvs (Fix global scan missing pipx venvs (#415) #418), non-root unwritable prefixes on CI, Windows all-usersC:\Program Files\Python3XXwhen it isn't on PATH (not in the well-known list).Known non-bugs
Harness tip (2026-10-15): a throwaway
#[ignore]test in crates/socket-patch-cli/tests that startsprebuilt_common::Server::project_with_env(root, &[("VIRTUAL_ENV", venv)]), writes its uri to a file and waits on a stop file serves the vendoring grant + wheel for a staged.socket/manifest.json+ blob. PointSOCKET_VENDOR_URLand--patch-server-urlat it. No hand-written mock needed for vendored cells. Delete the file afterwards.Harness: a venv copied with
cp -akeeps the original venv's absolute path inbin/pip's shebang, so.venv/bin/pip installin the copy installs into the ORIGINAL venv. Use.venv/bin/python -m pipin copied fixtures (a false "hosted install stays unpatched" on 2026-10-14).After a requirements.txt + pylock.toml entry is vendored, regenerating pylock.toml with
pip lock -r requirements.txt(archiveurl = "file:///<abs>/.socket/vendor/…whl") makesvendor --revertdrift-keep the whole entry (vendor_lock_entry_drifted, requirements.txt left vendored, exit 0), whilevendor --checkpasses. User edit of a wired file; same family as Vendored requirements.txt can't be reverted once the user touches the# socket-patch vendor:comment:vendor --revertdrift-keeps forever (exit 0), and the suggested "re-vendor" remedy is a no-op or adds a duplicate line #977 / thepip freezecase (2026-10-14).A mock whose
viewrecord purl spells a PyPI name differently from itsbatchpurl (zope.eventvszope-event) splits the vendored ledger into two entries per package on a later scan; consistent purls merge. Harness artifact (2026-10-14).VEX subcomponent ids echo the patch record's purl spelling (not PEP 503-canonicalised); unverified against the live API, so not filed (2026-10-14).
A requirements.txt still listing
socket-patch[hook]after v5 gets no stale-hook warning; docs/migrating-to-v5.md documents manual cleanup (2026-10-14).A vendored plain-pip project over a base-interpreter copy (system-site venv) gets no scan-time stale warning, and
pip install -rkeeps the base copy ("already installed with the same version"); vex warnsvendored_tree_out_of_sync. Same class as the documented plain-venv case (2026-10-13).Hosted rollback of a
pip hash -a sha512continuation restoressix==1.16.0with PyPI's sha256 hashes on one line;--require-hashesinstalls it. That's the documented pip-equivalent restore (2026-10-13).Since Fix crawl of --system-site-packages venvs (#409) #1419, agent mode in a
--system-site-packagesvenv patches the base interpreter's copy (for example Debian/usr/lib/python3/dist-packages/six.py, which other interpreters' users import too) when the venv has no copy of its own. That's the fix's stated intent ("agent mode said not installed"); as root it round-trips withrollback, and as non-root it fails loud (apply_failed, Permission denied, exit 1). 2026-10-13.An agent manifest beside a hosted requirements.txt pin: a bare
scankeeps agent mode (already_in_manifest), as CLI_CONTRACT.md documents ("keeps the mode of a project that already holds vendored or agent-mode patches";hosted_wiring_retained). It doesn't rewrite the hosted pin, so it isn't the A bare scan or get on a project holding both vendored packages and hosted pins silently converts every hosted pin to vendored (exit 0, no warning), so it undoes a partial hosted takeover #1437 class. 2026-10-13.A vendored → hosted takeover killed at a
group_commitfailpoint (group_commit_journal,group_commit_file,durability_barrier) recovers on the next run: requirements.txt is hosted, pip installs patched, vex is correct. The reverted.socket/vendor/pypi/<uuid>/(wheel +socket-patch.vendor.json) stays as an orphan, becausedefer_removalstrees aren't journaled. It's harmless cruft, inherent to SIGKILL and cross-ecosystem (2026-10-12).Lock-only discovery skips
six==X -C a=b(a short per-requirement option), while pip installs it. The pin is hand-written, so this is the Lock-only requirements.txt discovery reads a file with bare-CR (classic Mac), form-feed, NEL or U+2028 line breaks as one line, so scan exits 0 with "No patches" while pip installs every pin #1281 exotic class. With a venv the hosted refusal is loud (redirect_requirements_version_ambiguous), 2026-10-12.pip-compileover arequirements.inthat names the vendored wheel path writessix @ file:///<abs>/…whl: the same shape and outcome as thepip freezecase below (2026-10-12).A root
NOTES.txtthat only mentions the vendored wheel path keeps the wheel onvendor --revert(vendor_revert_residual_reference). That's a fail-safe false keep (2026-10-12).Lock-only requirements.txt discovery reads a file with bare-CR (classic Mac), form-feed, NEL or U+2028 line breaks as one line, so scan exits 0 with "No patches" while pip installs every pin #1281 (bare-CR, form-feed, NEL, U+2028 line breaks; doubled trailing backslash) was closed not planned by the maintainer on 2026-10-09: exotic inputs outside the v5 normal-toolchain scope. Don't re-file unless a normal tool generates such a file.
Since Use one package target grammar for get, remove, rollback and the UUID shortcut #1034,
remove six/rollback Six(bare names, PEP 503-folded) and versionless purls target a hosted pin;six_is a different PEP 503 name and is correctly not found (2026-10-09).pip-compileregenerating a hosted requirements.txt from requirements.in drops the hosted URL back tosix==1.16.0+ PyPI hashes; re-runscanafterwards. The regenerated file is the tool's output, not a socket-patch rewrite (2026-10-09).Compound or non-exact pins that pip resolves to one version (
six==1.16.0, !=1.15,six>=1.16.0,<=1.16.0,six~=1.16.0) and marker-only unpinned lines (six; python_version=="3.11") aren't lock-only discovered. Discovery is defined as exact==pins only (2026-10-09).six==v1.16.0is the Lock-only requirements.txt discovery sends PEP 440-equivalent pins verbatim (six==1.16→pkg:pypi/six@1.16), so a fresh checkout reports "No patches available" while the same file with a venv is patched #604 normalisation class.A vendored requirements.txt regenerated with
pip freeze(six @ file:///<abs>/.socket/vendor/pypi/<uuid>/….whl#sha256=…) isn't attested (ref_invalid, a machine-specific path), andvendor --checkexits 1 with a "dependency removed …scan --prune" remedy. The wet prune drift-keeps the entry, so nothing breaks; the dry run lists it as revertable (the contract's preview caveat). 2026-10-08.A plain-pip project with only PEP 735
[dependency-groups]/[project].dependenciespins (no requirements.txt, no lock) is invisible to lock-only discovery. Plain pip routes only through requirements.txt; PEP 735 groups are documented for Hatch only (docs/testing/hatch.md). 2026-10-08.Vendored
scan --dry-runpreviewswould_vendorfor requirements.txt pins the wet run refuses (six[x]==X→pypi_extras_unsupported;===,==X.*,--config-settings,name @ url, unpinned →pypi_requirement_not_pinned). This is documented: CLI_CONTRACT.md defines the vendored preview as a ledger classification that predicts only the npm-family preflights (uv handover 20261007T204038Z, checked 2026-10-08).A vendored → hosted takeover after the user edited the
# socket-patch vendor:comment refusesredirect_vendored_revert_failedand keeps the package vendored, on the dry run and the wet run alike (2026-10-08).A hosted root pin next to the same pin in a
-cconstraints file, a non-included siblingrequirements-dev.txt, or an include range (six>=1.0) attests fine (2026-10-07). Only an exact duplicate inside the-rtree is Lockfile discovery treats a requirements.txt-rinclude as a competing lock, so after a hosted rewritevexattests nothing (exit 2) androllbackrefuses (exit 1), althoughpip install -r requirements.txtinstalls the patched wheel #1086.--requirement= dev.txtand-r=dev.txt: socket-patch followsdev.txt(or not), but pip refuses both lines. The vendored root(transitive)append next to an include pin that the walk can't see still installs the patched wheel on pip 24.0 / 26.2.1 (2026-10-07).-r<TAB>dev.txtis followed (2026-10-07). A lock-onlysix==${VAR}pin isn't discovered; that's the same family as the documented hosted${VAR}refusal, so it's noted, not filed.A vendored
-rinclude diamond, an include cycle and the same include listed twice are all wired once and revert cleanly (2026-10-07).After pip-tools regenerates a vendored requirements.txt back to
six==1.16.0,vendor --checkexits 1 andvexomits the package (vendor_unwired) until a re-vendorre-wires it. That is the intended flow.Hosted mode rewrites only the root
requirements.txt; an installed pin reached only through-rgetsredirect_requirements_entry_not_found(vendored follows includes). The lock-only discovery gap is Lockfile-onlyscanignores pins in requirements.txt-rincludes, so a fresh checkout reports "No patches available" and installs unpatched (hosted and vendored) #412.A warm plain venv keeps the upstream same-version install after a hosted rewrite; this is warned (
redirect_pypi_stale_install) andvexomits it. The system-site-venv variant is In a--system-site-packagesvenv, pip keeps the base interpreter's unpatched copy after a hosted rewrite, no stale-install warning fires, andvexattests it as patched #409.Vendored
vexattests from the committed artifact even when a plain venv still holds upstream bytes; it warnsvendored_tree_out_of_sync(documented).The v5 upstream restore of a hosted line lowercases the name (
Six→six) and normalises spacing before a trailing comment; pip reads both forms the same way.Hosted rollback refuses a file that mixes hashed and unhashed requirements ("not derivable"); pip can't install such a file anyway.
Vendored refuses
===pins,==X.*wildcards and a tab before--hash(pypi_requirement_not_pinned). This is fail-closed and isn't filed (the==1.16zero-padding case is part of Hosted requirements.txt rewrite skips PEP 440-equivalent pins likesix==1.16for an installed 1.16.0, soscanexits 0 and pip installs the unpatched release (regression from v4.0.0) #475).--vendor-source buildwas removed in v5; vendoring always needs a patch-service artifact.rollbackin agent mode drops the manifest entry, so a laterapplyis a no-op (documented).A venv directory not named
.venv/venvis only found throughVIRTUAL_ENV.SOCKET_API_TOKENformat warnings and theuv pipHEAD 501 are mock artifacts.vex -ois--org, not--output.vex --jsonrequires--output.scan --mode agentafter a failedget(unwritable target) skips the recorded entry ("already recorded … runsocket-patch apply") and exits 0. This is documented, and the failedgetitself exits 1.Hosted
vexignores hosted URLs that aren't onpatch.socket.dev. Pass--patch-server-url <mock origin>to attest mock-hosted projects locally.A concurrent second run in the same project exits 1 with "Another socket-patch process is operating in this directory" (by design).
A requirements.txt that v4.0.0 (or a pre-Fix requirements.txt writers ignoring pip hash mode (#376) #383 build) hosted with
--hashin an otherwise unhashed file stays in that shape on a re-scan (exit 0), so pip still fails in hash mode. This is documented in Fix requirements.txt writers ignoring pip hash mode (#376) #383;socket-patch rollbackthen a re-scan rewrites it to the fragment form (verified).Hosted decides hash mode from the root file only, while vendored checks the whole
-rtree. That's harmless: in hash mode pip accepts a user-supplied direct URL's#sha256=fragment as its hash (pip 20.3.4–26.0).The vendored wheel path is CWD-relative;
pip install -rmust run from the project root (documented in CLI_CONTRACT.md).Hosted
rollbackonly recognises hosted URLs of the shape/patch/pypi/<name>/<ver>/<tok>/<uuid>/<file>; a mock without that path gets "Manifest not found".Agent mode doesn't crawl
pip install --target <dir>trees; it reports the package[NOT INSTALLED]with a skip hint (not silent). Hosted works for such projects.In the sandbox, the CLI can't reach pypi.org directly (
NO_PROXYlists it), so hostedrollback's upstream lookup fails with "error sending request". Run withNO_PROXY=localhost,127.0.0.1.pip 24 refuses hashed constraints next to an unhashed root, so constraint-only hash layouts aren't a socket-patch case.
A SIGKILLed scan can leave
.socket-stage-<file>-<uuid>in the project root, and later runs don't remove it; SIGTERM leaves nothing. Not filed (inherent to SIGKILL, cross-ecosystem).Hosted refuses bare URL / bare path lines (no
name @) and${VAR}pins withredirect_requirements_entry_not_found, exit 0 (the documented hosted-refusal posture).A hosted rollback restores a pip-equivalent line, not the original bytes (comment spacing, joined continuations,
(==X)→==X, case).Vendored refuses a per-line option such as
--config-settingson the patched pin, and aname @ <url/file>direct reference, with "not pinned to ==X" (fail-closed; the wording is imprecise). Hosted rewrites--config-settingslines fine.A symlinked
requirements.txtor a symlinked-rinclude is refused explicitly (redirect_symlinked_file_unsupported/pypi_requirements_symlink_unsupported), and so is an include outside the project root (vendored). Nothing is written.Agent mode doesn't crawl
pip install --prefix <dir>trees either ([NOT INSTALLED]+ skip hint), same as--target.With a mock origin, hosted
rollback/vexneedSOCKET_PATCH_SERVER_URL(or--patch-server-url) set to it; vendored needs the mock to serve sha512 integrity.Lock-only cells need
VIRTUAL_ENVpointing at an EMPTY venv: an emptyVIRTUAL_ENV=falls back to the system dist-packages (Ubuntu's python3-six 1.16.0) and masks lock-only behaviour.Vendored refuses
six[extra]==X(pypi_extras_unsupported), and an unpinned root pinned only through-c constraints.txt(pypi_requirement_not_pinned; the wording says "not pinned" although the constraint pins it). Both are fail-closed.Hosted rollback restores
===Xas==Xand an unpinnedsixassix==<installed>. Hosted mode keeps no state, so the line is derived from the hosted URL (see backlog question 0c).A file with a single
--hashline next to unhashed lines is already uninstallable by pip; hosted keeps that shape and rollback drops the lone hash. Garbage in, garbage out.record_fetch_failedon a hosted scan means the mock lacks theviewroute; it isn't a CLI bug.Agent mode doesn't patch editable installs: a PEP 660 finder (
apply_failed"File not found") or a legacy.egg-link("matched no installed package"). Both exit 1 and leave the user's source untouched (fail-closed).The vendored takeover's
vendor_prebuilt_required404 in a local harness means--patch-server-urlrewrote the artifact host to a mock without the prebuilt routes (harness artifact).get --mode vendoredover a hosted BOM file writes the vendored line without the BOM (pip reads either form);rollbackrestores the BOM.Agent
rollback --offlineneeds the before blob in.socket/blobs; without it, it refuses with arepairhint (harness staging, not a bug).Hosted wheel pins under a
--no-binary :all:/--no-binary six/--only-binary :all:option line install fine (pip 24.3.1, 26.2.1): pip doesn't apply format control to direct URLs.Vendored
vendorrefuses a UTF-16 requirements.txt loudly (pypi_no_requirements, exit 1), andvexwarnslockfile_unreadable; only hosted / lock-only discovery is silent (Hosted and lock-only scans treat a UTF-16 requirements.txt (what Windows PowerShell'spip freeze >writes) as absent: exit 0, no warning, and pip keeps installing the unpatched pin #721).A hashed root requirements.txt used as
-cby an unhashed sibling is refused by pip before any rewrite; only the unhashed form is Hosted and vendored rewrites of a requirements.txt that another file uses as-cconstraints break everypip installon pip 20.3–21.0 ("Links are not allowed as constraints"), with no warning #740.pip ≥ 21.1 accepts both the hosted
name @ urland the vendored bare-path line as constraints; only pip 20.3.x–21.0.x reject them (Hosted and vendored rewrites of a requirements.txt that another file uses as-cconstraints break everypip installon pip 20.3–21.0 ("Links are not allowed as constraints"), with no warning #740).vendor --offlineneeds the vendoring service in v5 (vendor_service_offline_conflict). Local vendored repros use theprebuilt_commonfixture server plus aview/<uuid>route.The
-c/PIP_CONSTRAINT/ pip.confconstraint =forms over a hosted or vendored root fail only on pip 20.3–21.0 (same as Hosted and vendored rewrites of a requirements.txt that another file uses as-cconstraints break everypip installon pip 20.3–21.0 ("Links are not allowed as constraints"), with no warning #740); don't re-file them.vendor --revertafter a Vendored requirements.txt after the user removes or bumps a vendored pin: the rescan re-adds the removed package as a "(transitive)" line (exit 0), or exits 1 forever after a bump, andscan --prunenever reverts the entry #786-style re-add is safe: the stalerequirements.txt:1wiring is skipped asvendor_revert_line_drifted, and the user's line stays untouched.The uv routine's handover (hosted rewrite overrides a
six @ https://…direct reference, and rollback restoressix==1.16.0) is Hosted requirements.txt rewrite replaces a user's own direct reference (six @ https://mirror/…/six-1.16.0-….whl,file://fork) with the Socket PyPI build, and rollback then restoressix==1.16.0from PyPI, losing the original source #542; don't re-file it.Vendored over a
pip lockpylock.toml works end to end (archive = { path }, pip 26.2.1 install from any CWD, byte-identicalvendor --revert); only the hosted restore is Hosted rollback and remove always refuse apip lockpylock.toml ("no sibling registry package shows the registry and artifact fields"), because pip writes[[packages.wheels]]tables, not uv's inlinewheels = [...]#804.remove/rollbackwith a qualified purl (…@1.16.0?foo=bar) givesnot_foundby design (a qualified identifier targets a single variant).A stale
.pyc(Agent-mode apply of a same-size PyPI patch in the same second aspip installleaves pip's__pycache__bytecode valid, so Python keeps running the unpatched code while apply reports it patched #819) needs pip's timestamp.pyc:SOURCE_DATE_EPOCH(checked-hash),--no-compile, uv without--compile-bytecode, and hosted / vendored mode are unaffected.The agent-mode
.dist-info/RECORDgoes stale after apply; this is documented (thepypi_record_staleadvisory). Agentvexafterpip install --force-reinstallcorrectly omits the package (not_applied).A superseding re-vendor refuses a vendor line the user moved into a
-rinclude (pypi_requirements_already_vendored, exit 1, remedyvendor --revert), while--dry-runpreviewswould_revendor. The contract defines the preview as ledger classification only; the harmful part is the remedy, which is Vendored requirements.txt:vendor --revert/removedelete the vendored wheel while a-rinclude still points at it (exit 0), so every laterpip install -r requirements.txtfails #867.When testing vendored discovery with a mock, the
batchroute must answer only for requested purls. A catch-all answer re-vendors removed packages as(transitive)(a false Vendored requirements.txt after the user removes or bumps a vendored pin: the rescan re-adds the removed package as a "(transitive)" line (exit 0), or exits 1 forever after a bump, andscan --prunenever reverts the entry #786 regression).A rewrite splits a hardlinked requirements.txt (this is inherent to the atomic stage-and-rename write), and a 0444 file is rewritten in the sandbox only because it runs as root.
pip install -r pylock.tomlneeds pip ≥ 26.1-ish (26.2.1 works; 26.0 parses it as a requirements file and fails on=). That's a pip version limit, not socket-patch.A socket.yml
minSeverityfloor over an already-hosted pin keeps it byte-identical, and the JSON reports it as a same-uuidredirected: 1, not underretained[](consistent with "keeps its recorded patch").--max-new-patchesdeferral ranks by severity from the patch record'svulnerabilities; a mock with an empty map showsseverity: "unknown"and alphabetical order (harness artifact).apply --checkwith no venv (noVIRTUAL_ENV, no.venv/venv) checks the system site-packages through the documented project-marker global fallback, the same asapply(backlog question 0b), 2026-10-08.py3X-none-anywheels count as portable under Fix interpreter-bound wheels treated as portable (#1048) #1053's rule; pip installs them on every Python ≥ 3.X, so that matches pip (2026-10-08).Hosted mode over a non-UTF-8 requirements.txt root (latin-1 bytes under a latin-1 coding line) refuses
candidate_file_unreadable, exit 1, nothing written. That's fail-closed and loud; a UTF-8 root with a latin-1-rinclude works (2026-10-09).All reactions