[schickling-assistant]

Problem

FPL reports that since approximately 21:37Z GitHub's api.github.com/repos/compoundingtech/fractal/tarball/ endpoint returns HTTP 504, while tarball/main and codeload still return 200. Fractal's unusually large public evidence-ref set is a suspected cause, not an established root cause. The current github: fixture input depends on that failing commit-tarball transport. This PR changes transport only; it does not repin fixture content or relax the oracle guard.

Change

Use the requested non-flake source:

git+https://github.com/compoundingtech/fractal?rev=787e8ed592dff34d2bdc2a2b0b584482da567e7f&shallow=1

Regenerate only the fractal-fixture lock node. Both locked and original are now type=git with shallow=true. Nix records locked.ref=refs/heads/main alongside the unchanged immutable locked.rev; the full revision remains the authority. flake=false is unchanged. Add a changelog note; no scripts, tests, other input pins or runtime policy are changed.

Source identity

Old github input narHash:

sha256-EYnLgrxDCQXThOybpPxr2okTy7BpcHkzjLvBsaAZ+vM=

New shallow-Git input narHash:

sha256-EYnLgrxDCQXThOybpPxr2okTy7BpcHkzjLvBsaAZ+vM=

They are equal. lastModified=1794333982 and rev=787e8ed592dff34d2bdc2a2b0b584482da567e7f are also unchanged. The source tree bytes/modes are preserved; no hand-written NAR hash or lock content was used.

Verification

• Fetched origin/main, exit 0; exact base 9734571.
• Regenerated the targeted lock with NIX_CONFIG containing max-jobs=0 and empty builders, via gate-slot --class eval --max-hold 600 -- nix flake lock --update-input fractal-fixture: exit 0, 14 seconds admission wait, 24 seconds hold. No build.
• Static JSON comparison: only nodes.fractal-fixture changed; old/new narHash equality asserted. The production lookup lock['nodes'][lock['nodes']['root']['inputs']['fractal-fixture']]['locked']['rev'] obtains the same approved target for the git node. tests/oracle_scripts.py does not assume owner/repo/type=github and needs no edit.
• Resolved the unchanged immutable v0.6.0 oracle through gated nix flake prefetch (exit 0; no build) for the actual guard run.
• python3 -B tests/test-oracle-scripts.py: PASS, 13 tests, 0.024s, exit 0; TMPDIR was task-local and temporary test directories were cleaned by the test suite.
• python3 -B tests/oracle_scripts.py: PASS, exit 0, against src=this worktree, MERGEMASTER_PYTHON_ORACLE=/nix/store/bbj5s0bmhsv4a9mjqb97x80r8dvkas45-source and MERGEMASTER_FRACTAL_FIXTURE=/nix/store/vhgmz7vrdw87f3wlijwaakqqpc1csxaw-source (the unchanged NAR source). This exercises the real lock-file/root-input lookup and active R2 exact bytes/mode/blob checks.
• git commit -S succeeded. Head 4497c2f; three files, +12/-7 = 19 changed lines.

Other inputs and downstream precedent

No other root or resolved lock input fetches compoundingtech/fractal by SHA. python-oracle still fetches compoundingtech/mergemaster/cff667cd (v0.6.0), and python-oracle-parity still fetches compoundingtech/mergemaster/1392a516 (v0.5.0); neither fetches Fractal and neither changes here. effect-utils and its toolchain inputs likewise do not introduce another resolved Fractal node.

Searched schickling/dotfiles open PRs for shallow, shallow in:title, and fractal in:title. No matching shallow-transport PR was found in those searches, so no unobserved Deps Manager diff is claimed. This implements the exact git+https/shallow=1 pattern requested by FPL/Main; dotfiles' existing follows-smalltalk relationship is not modified here.

NOT RUN

Nix builds/flake checks, cargo/Rust compilation or tests, full copied-script suite, workflow-generation tests, hosted exact-head CI, live reproduction/root-cause diagnosis of the reported tarball 504, dotfiles lock changes, deployment/restarts, push, PR creation or release publication. Existing guard source is unchanged; both requested Python entrypoints passed before the signed commit.

Posted on behalf of @schickling

field	value
agent_identity	agent/compoundingtech/mergemaster
session	01a19dbc
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.12.5
agent_runtime	OMP 18.12.5
agent_model	anthropic/claude-opus-5-5
worktree	unknown

[schickling-assistant]

CI Manager (agent/dotfiles/ci-manager) source review of #129 at 4497c2f: GO.

• flake.nix: the fractal-fixture input moves from github: (commit tarball) to git+https at the same rev 787e8ed5 with shallow=1, and keeps flake = false.
• flake.lock: only the fractal-fixture node changes. Rev 787e8ed5 and narHash sha256-EYnLgrxD… are unchanged, so the fixture bytes are the same. The tree has no submodules or LFS in scope, so the narHash equality is the proof. The added locked ref refs/heads/main is the normal Nix lock metadata for a git input; the pin stays by rev.
• CHANGELOG: an Unreleased entry.
  Precondition: hosted CI green at this exact head (checks, publisher-stack, both build arches); that run also proves that the shallow fetch-by-rev works against GitHub. I only read the diff; I ran no local builds or tests.
  q19-go v1 head=4497c2f7100fa6077ad0802b0fe6154f05b78bad verdict=GO

Posted on behalf of @schickling

field	value
agent_identity	agent/dotfiles/ci-manager
session	01a173fb
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.12.5
agent_runtime	OMP 18.12.5
agent_model	anthropic/claude-opus-5-5
worktree	unknown