prepareWorkspace copies the entire skillsDir tree (incl. .git) — EACCES crash and multi-MB annotate prompts on large skill stores

[bbudiono]

Setup: backpass 0.1.8, .backpassrc.json with skillsDir pointing at a large shared skills store (~720 skill dirs, several under git).

What happens: prepareWorkspace copies the whole skillsDir tree into the workspace, including .git object directories. Two failure modes observed:

1. Read-only git objects → EACCES crash during the copy.
2. The copied tree inflates the measured change list (27,039 "changes" in our run), producing a ~2.9MB annotate prompt that kills the model session.

Workaround: point skillsDir at a small repo-local dir instead of the shared store.

Suggested fix: exclude .git/VCS internals (or respect ignore files) during the workspace copy, and cap or summarize the annotate change list beyond a threshold so a large skills tree cannot blow up the prompt.

[kunchenguid]

Speaking as Kun's firstmate:

Confirmed on latest main (d8cbdb68ca20a9ad6626810e0c24a576e43223c7, backpass 0.1.14, 2026-08-29 17:27 UTC / 10:27 AM PT Aug 29). Filed against 0.1.8; the copy path was not repaired by #72 (it now walks every project skill root via skillDirs, same walkFiles). No open PR helps it. Labeling ready-for-pr and leaving the issue open. No implementation from this triage.

What is real on main. prepareWorkspace (src/workspace.js) rms the staging dir, then walkFiless every skillDirs tree with no skip for .git / VCS internals / unreadable entries, and copyFileSyncs each file. walkFiles recurses into every directory, including .git/objects. copyFileSync (and the follow-up readFileSync(..., "utf8") into originals) is uncaught: a read-only git object throws EACCES and cli.js prints the stack as a non-UserError. That matches failure mode 1.

The read path does not do this. loadSkills only opens SKILL.md (or a top-level .md) one level deep and fail-softs a read error. So analysis, the skill index, and the budget see the memory surface; the staging copy sees the whole store, including object packs.

measureWorkspace then diffs every original, and renderChangesForPrompt dumps every measured change with hunk lines and no size cap. A 720-dir store with nested .git trees produces a multi-MB annotate prompt (failure mode 2). The 27,039-change figure is the author's measured run; I did not replay a 720-dir store here. The code path that would emit it is present.

Workaround (point skillsDir at a small repo-local dir) is real. Related but not a duplicate: #32 (shared/fleet skill libraries as a feature); #76 (the unnamed AcpxError after the prompt blows up). #72 made this worse on main, not better: every resolved project skill root is copied the same way.

Contract-class (for a future PR, not a merge vote): restore if walkFiles skips .git / VCS internals (and treats copy EACCES as a named skip or UserError), so the staging copy stays the memory file plus skill content. The existing test ("skill directories are copied whole") keeps supporting files such as notes.txt; object stores were never the memory surface. new-default if the PR also stops copying supporting files and stages only SKILL.md — that replaces the specified "copied whole" contract. opt-in if ignore-file respect or a copy-mode flag is the only change.

Do not silently cap or summarize the annotate change list. VISION: "the run never silently truncates to fit" and "never papered over with a weaker path." If the measured set is still huge after VCS is excluded, fail loudly with a measured size (same named-failure hole as #76).

VISION per-rule

• Evidence is the only currency: aligns. No quote, corroboration, or two-session-bar change. .git objects are not evidence.
• The human owns the weights: aligns. No write-path, review-gate, or default-review change.
• Nothing the model says is taken on faith: aligns. Measurement still comes from the workspace diff. A 27k-change list of VCS internals is a number that does not describe the memory surface.
• The budget is the constraint: aligns. Does not make the memory file easier to grow. The inflated annotate prompt wastes the synthesis step on noise.
• It reads what you already have and owns nothing: aligns. Copying a shared store's .git into .backpass/synthesis/ is holding a copy of internals that are not the memory surface. Accuracy over coverage: walking everything bought junk and a crash.
• Failure is loud and named: aligns. Today's EACCES is an unnamed stack dump; the follow-on death is a bare AcpxError (propose failure surfaces as bare AcpxError exit 1 — no prompt-size hint when the synthesis prompt is too large #76). Root fix is skip VCS. A silent annotate cap would paper over the remaining blow-up.
• Scope: aligns. Staging the memory surface is the backward pass.

A change here should keep supporting skill files copyable and fail loud if a still-huge measured set would blow the prompt. Do not paper over a large tree by sending the model a quieter subset.

[kunchenguid]

Speaking as Kun's firstmate:

Stale restamp. Rechecked on latest main (2b8ee2125ac620e6e65540f6df86dfcecfc2a58e, backpass 0.1.21, 2026-09-12 13:56 UTC / 6:56 AM PT Sep 12). Still real. No open PR helps it. Leaving ready-for-pr. No implementation from this triage.

What changed since the Aug 30 stamp. EACCES on copy is no longer a fatal crash: prepareWorkspace now try/catches copyFileSync/chmodSync/readFileSync into unstageable + READ_ONLY_UNREADABLE warn and continues (src/workspace.js; test in test/workspace.test.js). That closes failure mode 1 as a hard crash. Failure mode 2 remains: walkFiles still has no .git/VCS skip, so a real nested object store is still walked and measured; renderChangesForPrompt (src/proposal.js) still dumps every measured change with no overall change-count or prompt-byte gate. #113 (symlink skill staging) only avoids walking a symlink target subtree; it does not skip a real nested .git.

Contract-class (for a future PR): restore if walkFiles skips .git/VCS internals so the staging copy stays the memory file plus skill content (supporting files like notes.txt stay copyable). new-default if the PR also stages only SKILL.md and stops copying supporting files. Do not silently cap or summarize the annotate list — VISION: never silently truncate to fit; if the measured set is still huge after VCS exclusion, fail loud with a measured size (same named-failure hole as #76).

VISION per-rule

• Evidence is the only currency: aligns. .git objects are not evidence.
• The human owns the weights: aligns. No write-path or review-gate change.
• Nothing the model says is taken on faith: aligns. A 27k-change list of VCS internals still does not describe the memory surface.
• The budget is the constraint: aligns. Inflated annotate prompt wastes the synthesis step on noise.
• It reads what you already have and owns nothing: aligns. Copying store internals into .backpass/synthesis/ is still holding junk that is not the memory surface.
• Failure is loud and named: aligns for the soft EACCES path now; remaining blow-up still feeds the unnamed AcpxError in propose failure surfaces as bare AcpxError exit 1 — no prompt-size hint when the synthesis prompt is too large #76. Silent annotate cap would paper over it.
• Scope: aligns. Staging the memory surface is the backward pass.