With --manifest-path into another project, rollback, remove, repair, vex, scan and get read the vendored ledger from --cwd

[mikolalysenko]

[agent] Filed by the scheduled architecture audit routine (CLI and core). Register: discussion #560 register.

Kind: bug. Source: new finding (register C43).

Problem

GlobalArgs::project_root states the rule: "Every command that reads more than one store must derive them from THIS root, so --manifest-path can never interleave two projects' state". CLI_CONTRACT's manifest_not_found row also says "All sources always come from the SAME project".

Only some commands follow it. Every command reads the manifest from resolved_manifest_path(), but the vendored ledger (.socket/vendor/state.json) is loaded from two different roots:

Ledger root	Sites
project_root() (follows --manifest-path)	list and apply through ProjectContext::new; vendor --check
--cwd	rollback; remove; repair; apply --check; vex (LoadedLedgers::load(&common.cwd, manifest_path)); scan and get through ProjectContext::rooted(.., cwd) / get, and the scan/get vendored paths (load_state(&common.cwd) in scan/vendor_flow.rs, scan/hosted.rs, get.rs)

rollback is the sharpest case, as rollback.rs L1218-L1238 shows:

• It takes apply.lock in the manifest's .socket/ through socket_dir_of(&manifest_path, &cwd).
• Under that lock it loads, reverts and saves the cwd project's vendored ledger.
• A concurrent default-manifest run in the cwd project locks a different file, so two processes can write the same ledger without mutual exclusion.

Proof by execution (debug build of 045d7ec, run twice; same results both times)

Setup: a/ is --cwd, and b/.socket/manifest.json is {"patches":{}}. Every command runs with --cwd a --manifest-path ../b/.socket/manifest.json (rollback and repair also get --offline).

Command	corrupt a/.socket/vendor/state.json	corrupt b/.socket/vendor/state.json
list --json	success	warning "unreadable vendor ledger … b/.socket/vendor/state.json"
vex --json --output f	error vendor_ledger_corrupt	error no_patches (ledger never read)
rollback --json	partial_failure, warning vendor_state_unreadable naming a/…	success
repair --json	partialFailure, event vendor_state_unreadable naming a/…	success

So on the same input, list reads project b's ledger, while vex, rollback and repair read project a's.

Symptoms / impact

• No existing issue covers this.
• Impact: with a foreign --manifest-path, rollback and remove revert one project's agent-mode patches and the other project's vendored artifacts and ledger. vex attests a mix of both projects. rollback's lock doesn't cover the ledger it writes.
• --manifest-path and SOCKET_MANIFEST_PATH are documented global options, so CI wrappers that pin the manifest location hit this.

Proposed change

• One rule: every vendored-ledger read and write derives its root from GlobalArgs::project_root().
  • Route rollback, remove, repair, apply --check and vex through ProjectContext::new (or load_state(&common.project_root())).
  • Delete the ProjectContext::rooted(common, cwd) calls in scan and get, and the rooted constructor if it has no remaining caller.
• scan/get vendored writes (scan/vendor_flow.rs, scan/hosted.rs, get.rs) save the ledger to the same root. Lockfile discovery stays where each command's contract puts it today. If a maintainer prefers to refuse a --manifest-path outside --cwd's project for multi-store commands, that is a contract change and should become a decision. The default fix just makes the code follow the rule it already documents.

Size and scope

• Files: commands/{rollback,remove,repair,apply,vex,get,context}.rs, commands/scan/{mod,vendor_flow,hosted}.rs.
• Estimate: about 40–80 production lines changed, plus tests.
• Out of scope: consolidating ProjectContext with a run-level context (C10).

Acceptance criteria

• Every load_state(/LoadedLedgers::load( call in crates/socket-patch-cli/src takes project_root() (or a value derived from it); a grep-style architecture test pins this.
• Regression test: with a corrupt ledger in --cwd and a clean --manifest-path ../b/.socket/manifest.json, list, vex, rollback, remove and repair all succeed; with the corruption moved to b, all of them report it.
• Regression test: rollback --manifest-path ../b/.socket/manifest.json never modifies a/.socket/vendor/state.json.
• The existing args.rs project_root_* tests and the list same-project tests stay green.
• CLI_CONTRACT: the "SAME project" sentence moves from the list row to the --manifest-path flag row, so it covers every command.

Dependencies

• Not blocked.
• Touches the same files as Only scan, get <uuid> and vex fall back to the public proxy on 401/403; get search, apply, rollback, repair and vendor eject fail #647 (proxy fallback) and Decide: one shape for the --json top-level error (scan and get emit both a string and a {code, message} object) #704 (envelope) but doesn't conflict with them. Makes C10 (RunCtx) simpler.

[mikolalysenko]

[agent] Triaged: priority:p3 (cross-cutting CLI). Not a duplicate, and no existing fix PR. #499 fixed the related vendor -g rewiring of the cwd project, but it doesn't cover the ledger root used by rollback/remove/repair/vex/scan/get.

---

Generated by Claude Code