# PROP-056 — Scraped project export and in-place VibeVM removal {#root} @status:impl/done [p01] @fact:STATUS-LINE This PROP defines and implements the deterministic operation that turns a VibeVM-managed consumer into a self-contained native project with no selected VibeVM repository residue. It freezes the contract, plan, typed rewrite, verification, export, in-place transaction, recovery and evidence semantics for epoch-1 Windows execution; portable planning remains available and other mutating hosts fail closed. @status:impl/done ## 1. Owner mandate and outcome {#mandate} [p02] @fact:OWNER-MANDATE The owner requires a project that may be authored and maintained with VibeVM and then delivered without VibeVM repository artifacts: the `vibevm/` tree, project `.vibe/` state, `vibe.toml`, `vibe.lock`, discipline policy/index artifacts, managed harness blocks and source-level traceability metadata as selected by policy. The operation must preserve the native build. Ambiguity is expressed once in a project contract, after which inventory, rewriting, deletion, health checks, rollback and proof are algorithmic. @status:spec/done [p03] @fact:OWNER-CONTRACT-RULING The default contract lives at `vibevm/scrape/contract.toml`, may be replaced by an explicit `--contract` path, and—when contained in the project and selected for removal—is the last project payload file removed. Standard Cargo, npm, Maven and Python/pip health checks are declarative; a snapshotted user-defined verifier is the escape hatch for domain checks. @status:spec/done ## 2. Relationship to existing laws {#related} - [p04] @fact:RELATED-ALGORITHMIC-REFACTORING PROP-031 supplies the write-side law: typed operation, deterministic plan, no free-form LLM rewriting, and post-operation gating. Scrape is one composed refactoring over source, manifests, generated state and project topology. @status:spec/done - @fact:RELATED-LAYOUT PROP-052 defines the `vibevm/` root. Scrape is the explicit terminal exception that removes that root from an exported consumer; it does not reintroduce the retired pre-PROP-052 Vibe layout. @status:spec/done - @fact:RELATED-CLEAN PROP-053 `vibe clean` removes regenerable dependency and boot products while retaining authored project state and `vibe.lock`. Scrape is not clean: it removes the Vibe relationship itself and therefore requires a contract, native health proof and recoverable commit. @status:spec/done - @fact:RELATED-LIFECYCLE PROP-054 supplies typed lifecycle planning, injected environment authority, artifact identity and durable transaction precedents. Scrape reuses those principles but has a distinct external journal because project `.vibe/` is removable. @status:spec/done - @fact:RELATED-DISCIPLINES The Rust, TypeScript and Go AI-Native stacks remain authoring disciplines. Scrape removes their registered metadata and tool wiring; it never erases useful product code merely because a discipline helped create it. @status:spec/done ## 3. Problem, boundary and vocabulary {#problem} [p05] @fact:SCRAPE-TERM-BOUNDARY Scrape is the terminal cleaning operation defined here: it removes the selected VibeVM repository layer while preserving the native product. The command vocabulary deliberately reserves attach and detach for a future live-tool relationship such as connecting and disconnecting a debugger. Scrape is never an alias for that future runtime operation, and no detach spelling is accepted by this command. @status:spec/done [p06] @fact:SCRAPE-PROBLEM Blindly deleting Vibe paths is not safe. A Rust consumer may path-depend on the `specmark` proc macro inside `vibevm/vibedeps`; build scripts or CI may call `vibe bin`; generated product source may live beside disposable generated state; authored specifications may share a tree with generated boot output. A path name alone cannot decide disposition. @status:spec/done - [p07] @fact:SCRAPED-TREE A scraped tree is a project tree for which every contract-selected Vibe path, dependency, managed marker, source annotation, tool script and environment reference is absent or explicitly retained under a neutral identity, and whose required native health panel is accepted after scraping under the declared strict or no-regression baseline policy. A retained non-pass verdict is never described as an unconditional pass. @status:spec/done - @fact:PRODUCT-PRESERVATION Product preservation means: all untouched files are byte-identical; every authored source rewrite is metadata erasure-equivalent; every declared relocation preserves exact bytes and mode; native manifests and lockfiles describe the scraped dependency graph; the health oracle accepts the result. @status:spec/done - @fact:ZERO-RESIDUAL-SCOPE Zero residue means zero registered Vibe identities, paths and markers selected by this contract. It does not mean deleting an ordinary word such as “vibe” from prose, nor undoing architectural choices learned while using a discipline. @status:spec/done - @fact:ALGORITHMIC-BOUNDARY The owner supplies policy: delete, keep, relocate, strip, retain, required checks and acceptable baseline mode. The engine supplies all mechanics and verdicts. Any state not decidable from the sealed contract and observed tree is a blocker, never a guess or best-effort success. @status:spec/done [p08] @fact:NON-GOALS Schema 1 does not reverse the historical influence of VibeVM on product architecture, infer replacements for arbitrary runtime libraries, mutate Git history, publish artifacts, or silently remove user-global deployments. Schema 1 does not discover, report or remove receipt-owned user-global deployments; the operator runs the existing explicit undeploy operation separately when external removal is desired. @status:spec/done [p09] @fact:SCRAPE-EXECUTION-PLATFORM-E1 Epoch-1 contract parsing, planning and contract checking are portable. Mutating `--output`, `--in-place` and `--recover` execution is enabled only on Windows, where the implementation proves handle-relative exclusive directory creation, no-replace publication, write/delete-denying namespace seals and restartable cleanup. Linux, macOS and other hosts return the typed blocker `scrape-platform-unsupported` before journal creation or project mutation. Their wire/state-machine shape remains portable, but the product does not claim a weaker check-then-rename implementation as support. @status:spec/done [p10] @fact:WINDOWS-NAMESPACE-DURABILITY-E1 Windows epoch 1 does not claim POSIX directory fsync or a persistent-media namespace-flush guarantee that the public handle-relative rename/delete APIs do not provide. Every namespace change instead runs under an already-durable external WAL intent, uses the native held-source/expected-identity atomic primitive, is re-observed as exactly its sealed before or after state, and is replayed on recovery. This is reported as journal-recoverable namespace evidence, never as synced or write-through evidence. A verified Windows scrape therefore carries reduced assurance; a future strict media-durable profile remains unsupported until the platform supplies a compatible capability-relative primitive. Path-based `MoveFileEx` is not substituted because it would discard the no-follow/root-capability authority. @status:spec/done ## 4. Command surface {#surface} - [p11] @fact:SCRAPE-PLAN-COMMAND `vibe scrape --plan [--contract ] [--output | --in-place]` is read-only. With neither execution mode, it plans the in-place form because that is the larger risk surface; it writes no project, journal, quarantine or report file. @status:impl/done - @fact:SCRAPE-EXPORT-COMMAND `vibe scrape --output [--contract ]` creates a scraped export and never changes the source tree. The destination must be absent; adoption of an existing directory is forbidden. @status:impl/done - @fact:SCRAPE-INPLACE-COMMAND `vibe scrape --in-place [--contract ]` executes the journaled in-place transaction. Interactive use shows the exact modified-file and deletion census; unattended use requires `--assume-yes` after a green plan. @status:impl/done - @fact:SCRAPE-RECOVER-COMMAND `vibe scrape --recover --path ` settles the one pending user-local transaction. Recovery reads the journaled contract and plan, never the possibly removed project contract. A new scrape refuses while recovery is pending. @status:impl/done - @fact:SCRAPE-CONTRACT-COMMANDS `vibe scrape contract init` writes a conservative default contract only when absent; `vibe scrape contract check` parses and plans it read-only. Existing contracts are never silently refreshed by install or a language discipline. @status:impl/done - @fact:SCRAPE-MODE-EXCLUSIVITY `--output` and `--in-place` are mutually exclusive. `--recover` accepts neither mode nor contract because the durable journal already fixes both. @status:spec/done ## 5. Contract discovery, schema and snapshot {#contract} [p12] @fact:DEFAULT-CONTRACT The default is the project-root literal `vibevm/scrape/contract.toml`. Every operand inside it is project-root-relative regardless of where an alternate contract lives. Contract location never changes mutation authority. @status:spec/done - [p13] @fact:ALTERNATE-CONTRACT An explicit `--contract` may name a portable project-relative literal or an explicit absolute regular file. A contained contract may be removed under the contract-last law. An external contract is read-only and can never be deleted or make any external path writable. Absolute selection walks and pins every parent no-follow; contained/external identity is decided by capability ancestry, never lexical prefix. @status:spec/done - @fact:CONTRACT-STRICTNESS The TOML schema rejects unknown tables, fields, tagged kinds and enum values. `schema` must equal 1; every id is a unique portable token across classification, rewrite, relocation, assertion and health rows. @status:spec/done - @fact:CONTRACT-READ-SAFETY The contract is opened no-follow as a regular single-link file, size-bounded to 1 MiB and read stably through one held handle. A symlink, reparse point, hard link, special file, identity change or second-pass mismatch refuses. @status:spec/done - @fact:CONTRACT-SNAPSHOT Before the first mutation the engine durably stores exact contract bytes, digest, canonical parsed value, expanded inventory, prepared rewrite outputs, verifier snapshots and plan hash in the user-local transaction. Recovery never rereads or reparses the source contract. @status:spec/done - @fact:CONTRACT-LAST For a contained contract, `commit.contract = "delete-last"` reserves it from every earlier wildcard batch. It is the last project payload file moved to quarantine; only deterministic pruning of now-empty ancestor directories follows. An external contract requires `commit.contract = "preserve"`. @status:spec/done - @fact:CONTRACT-LAST-VALIDATION A contained delete-last contract must resolve to exactly one regular effective-delete file. It cannot be effectively kept/generated, a rewrite target, relocation source/destination, baseline target or ordinary deletion entry. Broad patterns may select it only because the engine reserves its exact identity into the last slot before classification ordering. Any overlap that would produce another disposition refuses. @status:spec/done [p14] @fact:CONTRACT-EXAMPLE-LEAD Schema-1 example for a mixed Rust/TypeScript/Go consumer: @status:spec/done [p15] ```toml schema = 1 id = "org.example.scraped-release" [policy] unclassified = "refuse" links = "refuse" concurrent_change = "refuse" [scope] closed_roots = ["vibevm", ".vibe"] outside = "implicit-keep" [commit] contract = "delete-last" [[classify]] id = "remove-vibevm-tree" kind = "delete" patterns = ["vibevm", "vibevm/**"] owner = "vibe" proof = "contract-assertion-v1" modified = "delete" require_match = true [[classify]] id = "remove-project-vibe-state" kind = "delete" patterns = [".vibe", ".vibe/**", "vibe.toml", "vibe.lock"] owner = "vibe" proof = "contract-assertion-v1" modified = "delete" require_match = false [[classify]] id = "remove-discipline-state" kind = "delete" patterns = [ "conform.toml", "specmap.toml", "specmap.json", "*-ai-native-conform-baseline.json", "discipline/**", ] owner = "vibe" proof = "contract-assertion-v1" modified = "delete" require_match = false [[classify]] id = "keep-harness-docs" kind = "keep" patterns = ["AGENTS.md", "CLAUDE.md", "GEMINI.md"] owner = "project" require_match = false [[rewrite]] id = "remove-harness-blocks" kind = "managed-block-remove-v1" paths = ["AGENTS.md", "CLAUDE.md", "GEMINI.md"] marker = "vibevm" matches = "zero-or-one-per-file" [[rewrite]] id = "strip-rust-specmark" kind = "rust-specmark-strip-v1" patterns = ["**/*.rs"] exclude = [".git/**", "vibevm/**", ".vibe/**", "target/**"] forms = ["scope", "spec", "verifies", "cell"] matches = "zero-or-more" [[rewrite]] id = "remove-rust-specmark-dependency" kind = "cargo-package-remove-v1" manifests = ["Cargo.toml", "crates/*/Cargo.toml"] package = "core-ai-native-specmark" aliases = ["specmark"] matches = "zero-or-more" [[rewrite]] id = "remove-vibe-workspace-excludes" kind = "toml-array-values-remove-v1" path = "Cargo.toml" table = ["workspace"] key = "exclude" values = ["vibevm", "vibevm/vibedeps"] matches = "zero-or-more" [[rewrite]] id = "strip-typescript-spec-tags" kind = "typescript-spec-comments-strip-v1" patterns = ["src/**/*.ts", "src/**/*.tsx"] matches = "zero-or-more" [[rewrite]] id = "strip-go-spec-directives" kind = "go-spec-directives-strip-v1" patterns = ["**/*.go"] exclude = [".git/**", "vibevm/**", ".vibe/**"] matches = "zero-or-more" [[assert]] id = "vibe-paths-absent" kind = "paths-absent-v1" patterns = ["vibevm", "vibevm/**", ".vibe", ".vibe/**", "vibe.toml", "vibe.lock"] [[assert]] id = "vibe-cargo-paths-absent" kind = "cargo-path-prefix-absent-v1" manifests = ["Cargo.toml", "crates/*/Cargo.toml"] prefixes = ["vibevm/", ".vibe/"] [health] baseline = "strict" before_failure = "refuse" after_failure = "rollback" parallel = false network = "tool-offline" max_stdout_bytes = 1048576 max_stderr_bytes = 1048576 max_result_bytes = 1048576 termination_grace_seconds = 5 [[healthcheck]] id = "rust" kind = "cargo" root = "." build = "build" workspace = true locked = true all_targets = true tests = "required" profile = "dev" features = [] timeout_seconds = 900 [[healthcheck]] id = "web" kind = "npm" root = "web" manager = "npm" lockfile = "package-lock.json" install = "none" build_script = "build" tests = "required" test_script = "test" timeout_seconds = 900 when = { path_exists = "package.json" } [[healthcheck]] id = "domain-health" kind = "custom" root = "." source = "vibevm/scrape/health/domain_health.py" snapshot = [ "vibevm/scrape/health/domain_health.py", "vibevm/scrape/health/lib/**", ] interpreter = "python" argv = ["--phase", "{phase}", "--result", "{result}"] protocol = "vibe-health-json-v1" network = "deny" reads = ["**"] writes = [] spawn = true timeout_seconds = 120 ``` ### 5.1 Closed schema-1 shape {#schema-shape} [p16] @fact:SCHEMA-TOP-LEVEL Required singleton tables are `policy`, `scope`, `commit` and `health`. Required scalars are `schema = 1` and one nonempty portable `id`. The repeated tables are `classify`, `baseline`, `rewrite`, `relocate`, `assert` and `healthcheck`. At least one classify, one assertion and one required after healthcheck must exist. Every omitted optional collection is empty; there are no open extension tables. @status:spec/done - [p17] @fact:SCHEMA-POLICY `policy` requires `unclassified = "refuse"`, `links = "refuse"` and `concurrent_change = "refuse"`; schema 1 admits no other values. `scope` requires a nonempty set of portable literal `closed_roots` and `outside = "implicit-keep"`. Missing closed roots are observations, and are errors only when a covering classify row requires a match. @status:spec/done - @fact:SCHEMA-COMMIT `commit.contract` is `delete-last` for a contained contract and `preserve` for an external contract. The contained/external observation selects which value is legal; the contract cannot weaken it. @status:spec/done - @fact:SCHEMA-WHEN Schema-1 `when` is an optional deny-unknown table containing exactly one required portable literal `path_exists`, resolved below that check's root. No environment, command or OS predicate is inferred here; the future common applicability feature may replace this shape only by an explicit schema epoch. @status:spec/done [p18] @fact:SCHEMA-CLASSIFY Every `[[classify]]` requires `id`, a nonempty canonical set of `patterns`, `kind`, `owner` and `require_match`. Keep requires `owner = "project"` and forbids `proof`/ `modified`. Delete/generated require `owner = "vibe"`, proof and modified policy. Proof is exactly `contract-assertion-v1 | sha256-v1 | vibe-generated-v1`; modified is exactly `refuse | keep | delete`. Variant-inapplicable fields are rejected. @status:spec/done [p19] @fact:SCHEMA-BASELINE Each `[[baseline]]` has one portable literal `path` and one lowercase `sha256:<64-hex>`. Paths are globally unique. Exactly one baseline is required for each regular file selected by a `sha256-v1` rule; a baseline selected by no such rule is an error. @status:spec/done [p20] @fact:SCHEMA-MATCH-CARDINALITY Every rewrite requires one `matches` enum. Set-level adapters admit exactly `zero-or-more | one-or-more | exactly-one`. Per-file adapters admit exactly `zero-or-one-per-file | exactly-one-per-file`. A kind's row below fixes which family it uses; integer/free-text cardinalities are rejected. @status:spec/done [p21] | Rewrite kind | Required variant fields | | --- | --- | | @fact:SCHEMA-RW-MANAGED `managed-block-remove-v1` @status:spec/done | @fact:SCHEMA-RW-MANAGED-FIELDS Literal `paths`, registered `marker`, and `matches = zero-or-one-per-file or exactly-one-per-file`. @status:spec/done | | @fact:SCHEMA-RW-RUST `rust-specmark-strip-v1` @status:spec/done | @fact:SCHEMA-RW-RUST-FIELDS Nonempty `patterns`, optional `exclude`, nonempty unique `forms` drawn from scope/spec/verifies/cell, and `matches = zero-or-more or one-or-more`. @status:spec/done | | @fact:SCHEMA-RW-CARGO `cargo-package-remove-v1` @status:spec/done | @fact:SCHEMA-RW-CARGO-FIELDS Manifest patterns, exact Cargo `package`, optional alias allowlist and set-level `matches`. Observed aliases outside a nonempty allowlist refuse rather than disappear. @status:spec/done | | @fact:SCHEMA-RW-NODE `node-package-remove-v1` @status:spec/done | @fact:SCHEMA-RW-NODE-FIELDS Required literal `package_json` and `lockfile`, selected `manager = npm or pnpm or yarn`, nonempty `packages` array, optional arrays of component-array `script_paths` and `config_paths`, plus set-level `matches`. @status:spec/done | | @fact:SCHEMA-RW-GOMOD `go-module-remove-v1` @status:spec/done | @fact:SCHEMA-RW-GOMOD-FIELDS Required literal `go_mod`, optional literal `go_sum`, nonempty exact `modules` array and set-level `matches`. Replace/exclude/tool directives are part of the same parsed graph. @status:spec/done | | @fact:SCHEMA-RW-TOML `toml-array-values-remove-v1` @status:spec/done | @fact:SCHEMA-RW-TOML-FIELDS One literal `path`, component-array `table`, scalar `key`, nonempty exact `values` and match cardinality. @status:spec/done | | @fact:SCHEMA-RW-LANGUAGE `typescript-spec-comments-strip-v1` / `go-spec-directives-strip-v1` @status:spec/done | @fact:SCHEMA-RW-LANGUAGE-FIELDS Nonempty `patterns`, optional `exclude` and match cardinality; the registered grammar is adapter-owned, not contract-provided. @status:spec/done | | @fact:SCHEMA-RW-JSON `json-member-remove-v1` @status:spec/done | @fact:SCHEMA-RW-JSON-FIELDS One literal `path`, component-array `object`, exact member names and match cardinality. @status:spec/done | | @fact:SCHEMA-RW-TEXT `text-exact-replace-v1` @status:spec/done | @fact:SCHEMA-RW-TEXT-FIELDS One literal `path`, required complete-file `sha256`, UTF-8 `before`/`after` literals and positive exact `occurrences`. @status:spec/done | [p22] @fact:SCHEMA-LOCK-CONSEQUENCE Cargo, Node and Go native lock changes are derived consequences of explicit package/module removal rows, never independent inferred authority. Each adapter records the selected manager and exact graph delta. A graph change with no authorizing rewrite row refuses. @status:spec/done [p23] @fact:SCHEMA-RELOCATE-ASSERT Every relocation requires `id`, literal `from`/`to`, `conflict = "refuse"` and boolean `required`. Assertion kinds are closed: `paths-absent-v1` requires patterns; `text-literal-absent-v1` requires path patterns plus literal needles; `cargo-path-prefix-absent-v1` requires manifest patterns plus prefixes; `language-metadata-absent-v1` requires language plus source patterns; `dependency-identities-absent-v1` requires manager, manifest and identities. @status:spec/done [p24] @fact:SCHEMA-HEALTH-COMMON `health` requires baseline, before/after failure policy, parallel flag, network policy, positive stdout/stderr/result caps and a positive `termination_grace_seconds`. These limits are authored explicitly in schema 1—there are no hidden process or protocol defaults. Schema 1 admits only `before_failure = "refuse"`, `after_failure = "rollback"` and `parallel = false`; these fields are explicit evidence, not open strategy extension points. Every `[[healthcheck]]` requires unique id, kind, literal root, positive timeout, optional strict `when` and optional per-check network override; health root alone may use the special literal `.`. Cargo requires `build = check | build`, booleans workspace/locked/all_targets, `profile = dev | release`, a possibly empty feature array and tests mode. npm requires manager, literal lockfile, `install = none | ci`, tests mode, and exactly one nonempty build_script or typecheck_script; test_script is required unless tests is skip. Maven requires `runner = wrapper-first | explicit`, goal, offline boolean and tests mode. Python requires interpreter, nonempty source_roots, dependency_check/build booleans and tests mode; test_runner is required unless tests is skip. Fields not named for a variant are forbidden. @status:spec/done [p25] @fact:SCHEMA-HEALTH-CUSTOM Custom health requires source, nonempty snapshot patterns containing that source, explicit interpreter, argv, protocol, read/write patterns, spawn boolean and network policy. JSON protocol requires the complete `{result}` placeholder exactly once. Result and scratch paths are implicit transaction-owned writes and MUST NOT be repeated in `writes`. Exit-code protocol forbids `{result}`; optional diagnostics must live below `{scratch}`. @status:spec/done [p26] @fact:SCHEMA-JTD-WIRES The normative machine shape is encoded as strict JTD wire schemas for `scrape_plan/e1`, `scrape_report/e1` and `scrape_health_result/e1` in the first implementation atom. Those schemas must reproduce this section's required/optional/forbidden sets and closed enums exactly; generated types are the sole JSON reader/writer and handwritten shadow DTOs are forbidden. @status:spec/done ## 6. Path, glob, inventory and classification law {#paths} [p27] @fact:PATH-ROOTING Every mutation operand is resolved from the pinned project root. No contract field can name an absolute mutation path, parent traversal, drive, UNC path, URL or device. The explicit alternate contract source is the sole read-only external-path exception. @status:spec/done - [p28] @fact:GLOB-GRAMMAR Schema 1 owns one glob grammar: `*` means zero or more non-separator characters within one component; `**` means zero or more complete components and is legal only as a whole component. `?`, character classes, braces, negation, escapes, `***`, embedded `**`, backslashes, colons, empty components and dot components are rejected. @status:spec/done - @fact:NO-AMBIENT-GLOB The filesystem and shell never expand a pattern. The engine enumerates no-follow logical paths, renders forward slashes, matches case-sensitively, byte-sorts and deduplicates. Authored pattern order has no semantic effect. @status:spec/done - @fact:INVENTORY-CLOSED Every entry beneath a selected deletion/classification root is inventoried. Any walk error, non-UTF-8 name, unsafe component, link/reparse point, special file or changed identity is a failure rather than absence. Schema 1 requires `policy.unclassified = "refuse"`. @status:spec/done - @fact:CLASSIFICATION-UNIVERSE The engine enumerates the whole pinned project for pattern matching and residual proof, excluding only `.git/` internals and externally owned transaction state. Every existing entry at or below a literal `scope.closed_roots` member must be explicitly classified; an unmatched entry there refuses. Outside closed roots, unmatched paths are implicit project-owned keep. Explicit classify/rewrite/assert patterns may still select them. Rewrite targets are effectively kept unless an explicit classification conflict says otherwise. @status:spec/done - @fact:NO-RECURSIVE-BLIND-DELETE Project-selected trees are never passed to an ambient recursive delete. Files are classified and quarantined individually; directories are removed only when independently observed empty, deepest first. A kept descendant prevents ancestor deletion. @status:spec/done - @fact:GIT-METADATA-PROTECTED `.git` and every descendant are non-addressable mutation paths in schema 1. A classify, rewrite, relocation, baseline, health-write or assertion operand that can effectively match them after include-minus-exclude evaluation refuses. Every broad source selector must therefore exclude `.git/**`; exclusion from inventory alone is never permission or a silent no-op. @status:spec/done - @fact:CLASSIFICATION-LATTICE Classification uses a fixed lattice, not first-match order: keep plus generated is invalid; keep explicitly shields delete; generated proof dominates an overlapping delete; same-kind overlap is allowed only when owner, proof and modified policy are identical. Every path receives one effective disposition. @status:spec/done ### 6.1 Ownership and modification {#ownership} [p29] @fact:FILE-CLASSES The plan classifies observations as `generated-owned`, `managed-region`, `authored-metadata`, `authored-product` or `unknown`. Path spelling alone never proves a class. @status:spec/done - [p30] @fact:OWNERSHIP-PROOFS Vibe ownership is proven by a matching receipt/artifact record, an exact registered generated marker plus recorded digest, a contract SHA-256 baseline, or explicit `contract-assertion-v1`. A failed ownership proof always refuses; `modified = "delete"` cannot waive absent ownership. @status:spec/done - @fact:MODIFIED-POLICY Modification policies are `refuse`, `keep` and `delete`. Refuse is the default. Keep reclassifies the path and updates ancestor disposition. Delete quarantines the modified bytes but must appear separately in human confirmation and JSON evidence. @status:spec/done - @fact:MODIFICATION-STATE Observed modification is exactly `unmodified | modified | unknown | not-applicable`. SHA-256 proof compares the contract baseline; generated proof obtains its expected digest from the exact receipt/artifact record or registered generated transaction that names the path. Contract assertion has no baseline and therefore yields unknown. Refuse rejects modified and unknown; keep preserves either; delete accepts either only as the owner's explicit lossy disposition. Not-applicable belongs only to keep. @status:spec/done - @fact:AUTHORED-SPECS Authored files under `vibevm/vibespecs` are not generated merely because of their directory. A full zero-residue contract must explicitly delete them or relocate them to a neutral home. No built-in policy silently decides their value. @status:spec/done - @fact:PRODUCT-CODE-SURVIVES Generated cells, seams, tests, models and other useful application source are authored product once accepted into the project. They remain unless an exact contract relocation or deletion row names them. Origin in a discipline is not deletion authority. @status:spec/done ### 6.2 Relocation {#relocation} [p31] @fact:RELOCATION-CONTRACT A `[[relocate]]` row names one exact source and one exact destination; schema 1 admits no relocation glob. Source descendants must be effectively kept, destination must be outside every deletion root, all relocations are nonoverlapping and acyclic, and `conflict = "refuse"` is the only collision policy. @status:spec/done [p32] @fact:RELOCATION-PHYSICS In-place relocation is a capability-relative same-filesystem rename with journaled parent creation. Cross-filesystem copy-and-delete is not an implicit fallback. Export mode stages the corresponding byte/mode-identical copy inside its new destination tree. @status:spec/done ## 7. Typed rewrite algebra {#rewrites} [p33] @fact:TYPED-ONLY Globs select candidates; they never rewrite bytes. Authored source and structural manifests are changed only by a closed typed adapter. Generic regex replacement, shell rewrite hooks, implicit format conversion and whole-document reserialization are forbidden. @status:spec/done [p34] @fact:REWRITE-PRECONDITION Every planned rewrite records adapter epoch, parsed node identity, original byte span, complete before digest, prepared after bytes/digest, expected match cardinality and reason. Apply accepts only the journaled before state; recovery accepts only exact before or exact after. A third state is concurrent mutation and refuses. @status:spec/done [p35] | Schema-1 kind | Exact responsibility | | --- | --- | | @fact:REWRITE-MANAGED-KIND `managed-block-remove-v1` @status:spec/done | @fact:REWRITE-MANAGED-RESP Remove one registered, whole-line, correctly paired managed block while preserving every surrounding byte except deterministic adjacent-newline repair. @status:spec/done | | @fact:REWRITE-RUST-KIND `rust-specmark-strip-v1` @status:spec/done | @fact:REWRITE-RUST-RESP Remove proven Specmark scope macros, attributes and now-unused import leaves through the Rust syntax tree; preserve attributed product items. @status:spec/done | | @fact:REWRITE-CARGO-KIND `cargo-package-remove-v1` @status:spec/done | @fact:REWRITE-CARGO-RESP Remove resolved package identities/aliases from every dependency table, workspace inheritance, target table, feature edge, patch and replace entry after proving no remaining use. @status:spec/done | | @fact:REWRITE-TOML-KIND `toml-array-values-remove-v1` @status:spec/done | @fact:REWRITE-TOML-RESP Use a formatting-preserving TOML editor to remove exact scalar array members at a typed table/key path and exact cardinality. @status:spec/done | | @fact:REWRITE-TS-KIND `typescript-spec-comments-strip-v1` @status:spec/done | @fact:REWRITE-TS-RESP Remove only registered, grammar-valid discipline JSDoc tag nodes; retain descriptions and unrelated tags. @status:spec/done | | @fact:REWRITE-GO-KIND `go-spec-directives-strip-v1` @status:spec/done | @fact:REWRITE-GO-RESP Remove grammar-valid `//spec:` directive lines from parsed Go comment groups without touching strings, block prose, `//go:` or build/lint directives. @status:spec/done | | @fact:REWRITE-JSON-KIND `json-member-remove-v1` @status:spec/done | @fact:REWRITE-JSON-RESP Remove exact registered dependency/script/config members through a strict JSON object path and formatting-preserving emitter. @status:spec/done | | @fact:REWRITE-NODE-PACKAGE-KIND `node-package-remove-v1` @status:spec/done | @fact:REWRITE-NODE-PACKAGE-RESP Remove explicit Node package/script/config identities and reconcile exactly one selected native lock manager as a derived graph operation. @status:spec/done | | @fact:REWRITE-GO-MODULE-KIND `go-module-remove-v1` @status:spec/done | @fact:REWRITE-GO-MODULE-RESP Remove explicit Go module/tool identities and reconcile go.mod/go.sum without changing an unrelated module edge. @status:spec/done | | @fact:REWRITE-EXACT-TEXT-KIND `text-exact-replace-v1` @status:spec/done | @fact:REWRITE-EXACT-TEXT-RESP Escape hatch for an unsupported textual format: one literal file, required complete-file SHA-256, literal before/after text and exact occurrence count. No regex, capture, encoding guess or glob target. @status:spec/done | ### 7.1 Rust and Specmark {#rust} - [p36] @fact:RUST-ALIAS-RESOLUTION The adapter derives every local dependency name whose Cargo package identity is `core-ai-native-specmark`; it never assumes the alias is `specmark`. Qualified paths and explicit imports are resolved. Glob imports, macro re-exports, shadowed aliases or ambiguous identity refuse before mutation. @status:spec/done - @fact:RUST-ERASURE A valid `scope!` expands to no product tokens; `spec`, `verifies` and `cell` preserve the item and add metadata/ rustdoc only. The after AST with registered metadata erased must equal the before AST under the same erasure. Invalid pre-scrape macro grammar refuses rather than becoming a newly compiling program. @status:spec/done - @fact:RUST-CARGO-COVERAGE Cargo removal covers normal, dev, build, target-specific and workspace dependency tables; member `workspace = true`; `dep:` and alias-feature references; patch and replace tables; examples, benches and build scripts. A dependency is removed only after the residual resolver proves no use. @status:spec/done - @fact:RUST-LOCK `vibe.lock` is removable Vibe state; `Cargo.lock` is native project state. If Cargo.lock existed, a typed lock-reconciliation step runs after manifest rewriting and a final `--locked` check must pass. A previously lockless library stays lockless unless the contract explicitly requests creation. @status:spec/done - @fact:RUST-RETAIN-MODE Schema 1 has no verified retain-annotations mode: a remaining registered Specmark macro, package identity or source reference is residue and refuses scrape. A future neutralization adapter would have to relocate a complete self-contained macro toolchain, change its registered identity/content and repoint every Cargo edge. Merely moving the shipped proc-macro—or copying it without its grammar/workspace dependencies—is not neutral scraping. @status:spec/done ### 7.2 TypeScript and Go {#typescript-go} [p37] @fact:TYPESCRIPT-SCRAPE TypeScript metadata is parsed as JSDoc attached to AST nodes. In a mixed comment only registered tags whose payload parses under the discipline grammar leave; description, unrelated tags, indentation, order and line-ending style remain. Matching text in strings, templates, regexes, Markdown or ordinary comments is not a candidate. npm/pnpm/Yarn manifest and lock changes occur only under an explicit `node-package-remove-v1` row and are its derived graph consequence; ambiguous simultaneous managers or unrequested graph changes refuse. @status:spec/done [p38] @fact:GO-SCRAPE Go metadata is parsed with `go/parser` and `go/token`. Only complete registered `//spec:` directive lines leave their comment group. Product prose, strings/raw strings, block comments, generated headers, `//go:`, build constraints and lint directives remain. Native `go.mod`/`go.sum` reconciliation removes only dependencies made unreachable by an explicit `go-module-remove-v1` row. A module graph delta without that authority refuses. @status:spec/done ### 7.3 Managed blocks {#managed} [p39] @fact:MANAGED-BLOCK-SAFETY A managed block is removed only when exact registered begin/end markers form the expected number of nonoverlapping, correctly ordered pairs owned by the scraping provider. Missing ends, orphan ends, duplicate ids, illegal nesting, string-embedded pseudo-markers or digest drift refuse. Structural formats use their parser, never comment-like marker removal. @status:spec/done ## 8. Verification and user-defined health {#health} [p40] @fact:HEALTH-IS-ORACLE Verification is part of the contract, not an informal script run after deletion. The engine deterministically resolves, snapshots, invokes and judges the declared oracle. It proves exactly that the declared oracle accepted; it cannot infer that an incomplete user oracle covers every product requirement. @status:spec/done - [p41] @fact:HEALTH-BEFORE-AFTER The same canonical ordered health plan runs before and after. Strict baseline requires every check to pass; no-regression requires an admissible structured before verdict and its subset/severity relation after. A non-admissible before state causes no project mutation; an unacceptable after state rolls in-place work back. Executable, arguments, verifier snapshot, environment identity and policies are phase-identical except journaled phase/root/scratch values. @status:spec/done - @fact:HEALTH-TESTS-MODE Each preset declares `tests = "skip" | "if-present" | "required"`. Required with no discoverable test target is a preflight failure. If-present records a typed skip, never pass. Skip is allowed but the final report labels verification reduced. @status:spec/done - @fact:HEALTH-NO-RETRY No check retries by default. Timeout, cancellation, process-tree loss, unavailable executable, malformed protocol, undeclared effect or nonzero exit is an execution failure and never baselineable. A valid structured result written by an exit-0 custom verifier may carry a health-fail verdict; that is evidence eligible for explicit no-regression comparison, not an execution failure. @status:spec/done ### 8.1 Built-in presets {#presets} [p42] | Kind | Build/dependency checks | Tests | | --- | --- | --- | | @fact:HEALTH-CARGO-KIND `cargo` @status:spec/done | @fact:HEALTH-CARGO-BUILD Resolve one Cargo executable and run `check` or `build` with declared workspace, all-targets, features, profile, locked and offline flags. `CARGO_TARGET_DIR` is phase scratch. @status:spec/done | @fact:HEALTH-CARGO-TEST Run `cargo test` with the same graph selectors when not skipped. @status:spec/done | | @fact:HEALTH-NPM-KIND `npm` @status:spec/done | @fact:HEALTH-NPM-BUILD Validate `package.json`/selected lock, optionally run explicit `ci`, then require the declared build or typecheck script. Missing build is not pass. Windows invokes the resolved Node plus npm-cli asset rather than reparsing through npm.cmd. @status:spec/done | @fact:HEALTH-NPM-TEST Run the exact declared test script; never use a silent `--if-present` shortcut for required tests. @status:spec/done | | @fact:HEALTH-MAVEN-KIND `maven` @status:spec/done | @fact:HEALTH-MAVEN-BUILD Resolve Maven wrapper metadata or an explicit Maven installation, record Java/launcher identity, and run batch/no-transfer-progress offline `verify`. Module target writes are phase-owned. @status:spec/done | @fact:HEALTH-MAVEN-TEST Required/if-present runs normal verify tests; skip adds the adapter's exact `-DskipTests` policy and reports reduction. @status:spec/done | | @fact:HEALTH-PYTHON-KIND `python-pip` @status:spec/done | @fact:HEALTH-PYTHON-BUILD Use an explicit interpreter/environment; run `python -m pip check` and compile selected files without importing project modules. A package-shaped project carrying `pyproject.toml`, `setup.cfg` or legacy `setup.py` must also run offline `python -m build --no-isolation` for full assurance; explicitly omitting it forces `assurance = "reduced"`. Bytecode/build output is phase scratch. @status:spec/done | @fact:HEALTH-PYTHON-TEST Run the declared test runner, initially pytest. Required with no runner is failure. @status:spec/done | [p43] @fact:PRESET-NO-INSTALL-BY-DEFAULT Presets use already available dependencies by default. Network access and dependency installation require explicit contract choices. Missing offline inputs are failures with a recipe, never permission to fetch silently. @status:spec/done ### 8.2 Custom verifier {#custom} [p44] @fact:CUSTOM-VERIFIER-SOURCE A custom verifier names a root-contained regular source/native executable, explicit interpreter and structured outer argv. The engine never adds an implicit shell or accepts an inline shell command string, implicit shebang choice or interpolated argv. Explicit Python, Node, Bash, PowerShell or direct-native interpreters are allowed because the snapshotted script is the owner's code-execution authority; shell-language behavior inside an explicitly selected Bash or PowerShell script is visible custom code, not engine reparsing. @status:spec/done [p45] @fact:CUSTOM-VERIFIER-SNAPSHOT Before the before phase, the engine snapshots the exact declared source bundle—bytes, modes, paths, sizes and digests—into the external transaction. Both phases execute that snapshot. A verifier stored under `vibevm/` therefore remains available after contract-last; undeclared dynamic imports from its former source directory fail. @status:spec/done - [p46] @fact:CUSTOM-ARGV Arguments are arrays. Placeholders `{root}`, `{phase}`, `{scratch}` and `{result}` are allowed only as an entire argument and expand to exactly one argument. JSON protocol requires `{result}` exactly once. The engine precreates private result/scratch parents outside the product view; writes there are implicit transaction effects. @status:spec/done - @fact:CUSTOM-EXIT-PROTOCOL Under `protocol = "exit-code"`, exit 0 is pass and every other terminal state is fail. Output is diagnostic only. @status:spec/done - @fact:CUSTOM-JSON-PROTOCOL Under `protocol = "vibe-health-json-v1"`, exit 0 plus one atomically written, size-bounded UTF-8 result object with protocol 1 is required. Its `pass | warn | fail` status is a valid health verdict: strict baseline admits only pass, while no-regression may compare warn/fail by stable findings. Missing, malformed, duplicate, oversized, wrong-version or nonzero-exit output is protocol/execution failure. Stdout can never spoof the result. @status:spec/done [p47] ```json { "protocol": 1, "status": "pass", "summary": "domain health passed", "findings": [ { "id": "stable/check-id", "severity": "info", "message": "bounded diagnostic", "evidence": "root-relative/path:line" } ], "metrics": { "tests": 42 } } ``` [p48] @fact:CUSTOM-FINDING-LAW Finding ids are nonempty, unique within one report and stable across phases. Severity is exactly `info < warning < error`. Pass carries no warning/error finding; warn carries at least one warning and no error; fail carries at least one error. A contradictory status/finding set is protocol-error. @status:spec/done ### 8.3 Executable identity, environment and effects {#execution} - [p49] @fact:EXECUTABLE-IDENTITY Every executable, interpreter and adapter asset is resolved before mutation to canonical absolute path, stable file identity, digest, size and mode. Version evidence is tagged by the closed vocabulary `probe | content`; epoch-1 Windows uses the exact content digest as the version identity rather than executing an unjournaled probe during read-only planning. After uses the same held identity. PATH shadowing, replacement, disappearance or retargeting fails closed. @status:spec/done - @fact:SCRAPED-ENVIRONMENT The child environment is allowlisted and removes Vibe-specific variables and Vibe-managed bin directories. Reports contain variable names and value digests, never secret values. A check inherits the global network policy unless it overrides it. Policy is exactly `deny`, `tool-offline` or `inherit`; deny must be OS-enforced or refuse. Tool-offline is legal only for built-in package-manager adapters, adds their native offline flags and reports network unverified beyond those flags. Custom checks default to deny, cannot select tool-offline, and must explicitly select inherit for unconfined network. @status:spec/done - @fact:HEALTH-EFFECTS Checks declare read, write, process-spawn and network effects. In both phases every allowed write and known build output is redirected to phase scratch or an enforced copy-on-write view and discarded afterward; no healthcheck can mutate the source or delivered tree. The writes list authorizes paths inside that isolated view, never direct product writes. A custom verifier defaults to `writes = []`. Any undeclared project create, modify, delete, hardlink write or symlink escape is failure. Network is failure under deny; tool-offline/inherit instead report their explicitly narrower assurance. @status:spec/done - @fact:WINDOWS-E1-HEALTH-ENFORCEMENT The Windows epoch-1 mutating backend runs both before and after health in freshly materialized disjoint exact copies under the identity-bound external transaction workspace. The actual source/delivered root is protected and completely re-proven immediately before and after each child, but is never the child cwd or `{root}`. The phase copy is also completely re-proven after execution, so a write to cwd/`{root}` fails health while leaving the product root unchanged. This different-path proof yields `assurance = reduced`, never masquerades as COW, and is allowed for custom code only for the narrow declared profile `protocol=exit-code, network=inherit, reads=["**"], writes=[], spawn=true` executed from the sealed bundle. Any stricter custom effects, denied network or structured result require a stronger backend and remain a plan blocker. @status:spec/done - @fact:HEALTH-OUTPUT-BOUND Stdout and stderr are drained concurrently. The report retains total byte count, full-stream digest, truncation/UTF-8 state and an explicit `redacted` flag. Epoch 1 always emits `redacted = true` with empty head/tail fields, so `truncated = false` never implies that excerpts were retained; storage never grows beyond declared caps and no secret is printed. A structured custom result is independently limited by `max_result_bytes`; stdout cannot consume or enlarge that budget. @status:spec/done - @fact:HEALTH-TIMEOUT-CANCEL Timeout or cancellation terminates the exact process tree. The prepared plan records `termination_mode = graceful-then-forced | forced-tree`. Unix sends a graceful group signal and forces after the explicitly declared `termination_grace_seconds`; Windows epoch 1 uses the non-escapable Job Object's `forced-tree` termination because it has no honest portable graceful broadcast. Either terminal is a health execution failure—never an accepted reduced result—and every pipe/process is reaped. Before-phase cancellation leaves the project untouched; after-phase cancellation enters rollback. @status:spec/done ### 8.4 Baseline comparison {#baseline} [p50] @fact:STRICT-BASELINE `baseline = "strict"` is the default: all required before checks and all required after checks pass. An execution, environment or protocol failure is never baselineable. @status:spec/done [p51] @fact:NO-REGRESSION-BASELINE `baseline = "no-regression"` is admitted only for structured verifiers with stable unique finding ids: every before pass remains pass, after findings are a subset of before findings, and no severity increases under `info < warning < error`. A well-formed exit-0 warn/fail report is a comparable health verdict; an opaque nonzero exit can never become an accepted baseline. @status:spec/done [p52] @fact:NO-REGRESSION-ASSURANCE Any accepted after verdict other than pass forces `assurance = reduced`. Its complete retained finding ids and severities appear in JSON and the human headline; a no-regression comparison may authorize delivery but cannot relabel existing product illness as full health. @status:spec/done ## 9. Residual and preservation proof {#residual} [p53] @fact:RESIDUAL-CLASSES The final verifier reports zero unresolved members of: source metadata, dependency identity, manifest path, managed marker, generated artifact, tool config, tool script, selected spec URI, environment reference, lock entry and link escape. @status:spec/done - [p54] @fact:RESIDUAL-LAYERS Proof combines language-aware scans, parsed manifest/dependency graphs, exact registered path/identity scans and a supplemental bounded lexical scan. A lexical prose hit is a disposition blocker, never auto-deletion. @status:spec/done - @fact:ERASURE-EQUIVALENCE For each authored Rust, TypeScript and Go source, `erase_registered_metadata(parse(before)) == parse(after)`. Untouched files have identical bytes; relocations have identical bytes/mode under the new path. @status:spec/done - @fact:LOCK-GRAPH-PROOF Native manifest and lock adapters record before and after dependency graphs. Only dependencies made unreachable by planned removals disappear; shared transitive dependencies remain. @status:spec/done - @fact:LOCK-GRAPH-E1 Epoch-1 implements Cargo.lock v3/v4 reconciliation when the removed package has one unambiguous root identity and records exact before/after/removed graph evidence. npm, pnpm, yarn and Go native lock reconciliation remains a typed plan blocker until each manager has an equally exact graph proof. @status:impl/done - @fact:PROOF-OUTSIDE-TREE Canonical proof is emitted to stdout and the user-local transaction/report home, never into the scraped project whose zero-residue assertion it would contradict. @status:spec/done ## 10. Scraped export mode {#export} - [p55] @fact:EXPORT-SOURCE-READONLY Export holds the source project in an enforced read-only/COW view. The requested output must be disjoint from the source—neither equal, descendant nor ancestor. The engine pins its parent no-follow and creates an exclusive sibling candidate on the output filesystem; inability to enforce source immutability refuses. @status:spec/done - @fact:EXPORT-COPY-SET All project files outside selected deletion roots are copied byte/mode-identically except typed rewrites. The exact engine-reserved exclusion is `.git` and its descendants; authoritative transaction state is external and therefore not in the inventory. Ambient Git ignore or cache heuristics are never consulted: every other observed regular project file is copied unless the contract explicitly deletes it, so untracked/ignored product inputs cannot disappear by accident. @status:spec/done - @fact:EXPORT-PUBLISH After candidate structural proof, publish uses an atomic no-replace rename onto the still-absent requested output; a raced occupant or platform without no-replace guarantee refuses. The output is still transaction-owned and rollback-capable. Full residual and after health run in the external exact-copy view while the requested final path is protected by complete pre/post reproof. Failure removes only the exact owned output back into stage; before that rename the engine re-enumerates the complete no-follow descendant set and requires exact equality with the journaled candidate paths, identities, digests and modes. An extra, missing or changed descendant is a third state and refuses, preserving the concurrent data in place. The source remains unchanged. @status:spec/done - @fact:EXPORT-RECOVERY Export uses the same external journal with states `prepared → candidate → published-pending-verify → verified → cleanup-pending → complete`. Before verified, recovery discards or quarantines only the transaction-owned candidate/output and never publishes it as success; verified states roll forward through cleanup/report. Recovery validates the same complete descendant state, never only the root-directory identity; any third state refuses. @status:spec/done ## 11. In-place transaction and recovery {#inplace} [p56] @fact:EXTERNAL-JOURNAL The authoritative journal and lock are user-local and outside the project, for example `~/.vibe/scrape/transactions///`. Project `.vibe/` is removable and can never contain the only journal, recovery intent, verifier snapshot or rollback bytes. The journal location is implementation-owned, not contract-configurable, and must itself be proven outside the pinned project root. @status:spec/done [p57] @fact:SAME-VOLUME-QUARANTINE Removed and replaced bytes live in an exclusively created, identity-pinned sibling quarantine under the project parent on the same filesystem. A pre-existing, linked, replaced or unverifiable quarantine refuses and is never adopted or deleted. The chosen name is durable before creation. @status:spec/done 1. [p58] @fact:TX-STEP-RECOVER Acquire the user-local project transaction lock and settle any prior journal before loading a new contract. @status:spec/done 2. @fact:TX-STEP-PLAN Pin root and parent; read/validate the contract; inventory; prepare rewrites, relocations, native lock updates, assertions and health identities; calculate one canonical plan hash. @status:spec/done 3. @fact:TX-STEP-PREPARE Persist contract/verifier snapshots, canonical plan, full before identities, complete rollback map and prepared after bytes in the external journal; fsync it before executing any child. No project byte or quarantine exists yet. @status:spec/done 4. @fact:TX-STEP-BEFORE Run before health from the sealed verifier bundle in an enforced read-only/COW view whose writes are all redirected to phase scratch. Strict requires pass; no-regression admits only its specified structured baseline. If isolation cannot be enforced, refuse. A before refusal mutates no source-project byte. @status:spec/done 5. @fact:TX-STEP-QUARANTINE After an admissible baseline, durably select the name and creation intent, exclusively create and identity-check the same-volume quarantine, then persist its root identity and complete manifest before entering mutating state. A crash after create but before that first identity seal is fail-closed manual recovery: the existing name is never adopted from spelling alone. @status:spec/done 6. @fact:TX-STEP-REWRITE Copy and fsync every rewrite before-image into quarantine while its visible name remains present, then atomically replace that name with the prepared after-image. Reconcile native locks and apply exact relocations. Platforms lacking an atomic replace primitive for the target refuse; no move-then-create missing-name window is claimed atomic. @status:spec/done 7. @fact:TX-STEP-REMOVE Move every non-contract removal into quarantine in canonical file order, then prune only observed-empty non-contract directories. @status:spec/done 8. @fact:TX-STEP-PRECOMMIT Run residual/preservation proof with a contained delete-last contract as the sole permitted residue; an external preserved contract receives zero project-tree exemption. Any failure rolls back. @status:spec/done 9. @fact:TX-STEP-CONTRACT Durably enter the rollback-capable `contract-boundary` state. Its tagged action is: contained/delete-last moves the contract as the last payload and prunes empty ancestors; external/preserve performs no project mutation. Both then run zero-exemption residual proof and after health in the isolated view. After it exits, re-observe the complete scraped tree digest and residual proof before verified commit. Crossing the contract boundary is not commit. @status:spec/done 10. @fact:TX-STEP-COMMIT After all gates pass, durably mark verified, release project mutation ownership, copy the canonical report to the stable user-local report home, then clean the exact verified quarantine and transaction journal. Cleanup failure leaves a verified transaction in cleanup-pending state for roll-forward cleanup; it never reopens product mutation. @status:spec/done [p59] @fact:ROLLBACK-DIRECTION The durable states are `prepared → before-passed → mutating → contract-boundary → verified → cleanup-pending → complete`, with `rolling-back → rolled-back | rollback-failed` as the inverse branch. The contract-boundary record carries exactly `delete-last-moved | external-preserved`. Failure or recovery from any state before verified restores in reverse journal order, including restoration of a moved contained contract, and proves the original bytes/modes. Verified and cleanup-pending only roll forward through report/quarantine/journal cleanup. Any path matching neither journaled before nor after is a third state and refuses rather than overwriting concurrent/user work. @status:spec/done [p60] @fact:GIT-INDEPENDENCE Git is not the rollback engine. Scrape works in non-Git projects and preserves untracked/ignored files according to the contract. A dirty repository is allowed only when every affected path's observed digest is sealed in the plan; unrelated concurrent changes remain detectable and refuse at the affected boundary. @status:spec/done ## 12. Plan, report and observable outcomes {#plan-evidence} [p61] @fact:PLAN-CONTENT Human and JSON plans are projections of one typed value. They include project/contract identities, mode, every matched path and effective class, ownership proof, modification state, rewrite before/after digests, relocation graph, native lock changes, residual assertions, exact health argv/executable identities, test modes, contract-last order, quarantine requirement, blockers and summary counts. @status:spec/done [p62] @fact:PLAN-JSON-SHAPE The plan wire is `scrape_plan/e1`. Its required top-level members and enums are fixed below; item/rewrite/health variant members are the generated tagged unions from §5.1. Digests are lowercase `sha256:<64-hex>`, sizes/counts are nonnegative integers, and every array is already in canonical execution order. @status:spec/done [p63] ```json { "schema": 1, "command": "scrape", "mode": "in-place", "plan_id": "sha256:...", "project": { "display_root": "C:/work/project", "tree_digest": "sha256:..." }, "contract": { "display_path": "vibevm/scrape/contract.toml", "contained": true, "sha256": "sha256:...", "action": "delete-last" }, "items": [], "rewrites": [], "relocations": [], "native_lock_changes": [], "assertions": [], "healthchecks": [], "contract_boundary": { "kind": "delete-last", "path": "vibevm/scrape/contract.toml", "empty_ancestors": ["vibevm/scrape", "vibevm"] }, "blockers": [], "summary": { "keep": 0, "rewrite": 0, "relocate": 0, "delete_unmodified": 0, "delete_modified": 0, "delete_unknown": 0, "delete_last": 1 } } ``` [p64] @fact:PLAN-CONTRACT-BOUNDARY The required `contract_boundary` tagged union is either `{"kind":"delete-last","path":, "empty_ancestors":[...]}` or exactly `{"kind":"preserve"}`. Preserve has no path/ancestor members; delete-last requires both. Variant leakage is rejected by JTD. @status:spec/done [p65] @fact:PLAN-DETERMINISM Plan display may carry absolute root, contract and executable paths; the separately generated hash projection excludes timestamps, transaction nonce, project-root display path, external contract source path, user-local journal location and quarantine name. It includes contract/verifier bytes, project-relative operands, executable content/version identity, platform, mode and—for export—the canonical requested output identity because that is a mutation target. The strict JTD identity object contains no maps/floats and is encoded in schema field order with canonical array order before SHA-256. @status:spec/done [p66] @fact:REPORT-CONTENT The canonical external JSON report records before/after root digests, unchanged-file hashes, deleted artifacts with provenance, typed rewrites, relocations, dependency graphs, every health command step/argv/terminal/output digest/finding, skipped tests, residual counts, actual apply/rollback/recovery work, bounded diagnostic events and final outcome. Human output is rendered from the same data. Outcome is exactly `verified | refused | rolled-back | rollback-failed`; orthogonal `assurance = full | reduced` and `cleanup = complete | pending` fields prevent an outcome-enum explosion. The headline renders all three. @status:spec/done [p67] @fact:NO-FALSE-SUCCESS Unavailable verification, truncated/missing structured evidence and rollback failure cannot yield a verified outcome. Explicitly skipped/if-present checks may yield `outcome = verified, assurance = reduced`; committed quarantine/report cleanup residue yields `outcome = verified, cleanup = pending` and a recovery command. Neither is rendered as unqualified verified/full/complete. @status:spec/done [p68] @fact:STABLE-REPORT-HOME Before transaction-journal cleanup, the canonical report is atomically copied to `~/.vibe/scrape/reports/.json` (or emitted only to stdout when the operator explicitly disables retention). The report home is outside the project and distinct from the ephemeral transaction directory. @status:spec/done ## 13. Security and authority invariants {#security} - [p69] @fact:SEC-CONTAINMENT All project reads/writes are capability-relative under pinned root/parent identities. No path normalization, glob, relocation, managed marker or verifier write may escape. @status:spec/done - @fact:SEC-NO-FOLLOW Links, reparse points, special files, unexpected hard links and name/handle identity changes fail closed at inventory and again before mutation. @status:spec/done - @fact:SEC-NATIVE-NOFOLLOW-BOUNDARY Epoch-1 Windows crosses a hand-written native ABI in exactly five narrowly reviewed functions: lease-entry, remove-expected, create-directory, rename-no-replace and NTSTATUS conversion. No safe standard-library or cap-std primitive preserves the required retained-handle, no-follow, no-replace and race semantics for those operations. Each function owns its inline SAFETY proof and exact tests. This is not a broad crate or file exemption. A second consumer or an expanded native ABI is the trigger to revisit a dedicated audit crate. @status:spec/done - @fact:SEC-CUSTOM-CODE A custom health verifier is user-authorized code with the same trust consequence as a project build script. The plan names its exact snapshot/interpreter/effects before confirmation; packages cannot inject an executable scrape hook. Epoch 1 exposes only the disjoint copy through cwd and `{root}`; arbitrary malicious host-absolute writes are outside its sandbox guarantee and therefore require the same operator trust as any other authorized build script. @status:spec/done - @fact:SEC-SECRETS Contract, plan, argv, output and reports may contain variable names and paths but never credential values. Custom children receive only allowlisted environment. Output redaction and bounds precede persistence. @status:spec/done - @fact:SEC-NO-SELF-DELETION The running Vibe executable, user-local journal and machine cache are outside project mutation authority. Scrape removes repository integration, not the operator's installed tool. @status:spec/done ## 14. Required verification matrix {#acceptance} - [p70] @fact:ACCEPT-SCHEMA Strict parse/write tests cover every tagged variant, unknown field/kind/enum, duplicate id, missing required member, wrong schema and canonical plan identity. @status:spec/done - @fact:ACCEPT-GLOBS Glob tests cover literal, component star and whole-component double-star; reject question/class/brace/negation/ escape/triple-star/embedded-double-star, absolute/dot/backslash/colon spellings; prove no shell/host expansion and deterministic order. @status:spec/done - @fact:ACCEPT-CLASSIFICATION Fixtures cover clean generated, modified generated, explicit authored deletion, keep shielding, invalid keep/generated overlap, unclassified members, authored specifications and kept descendants preventing ancestor deletion. @status:spec/done - @fact:ACCEPT-RUST Rust goldens cover qualified/renamed aliases, explicit/grouped imports, scope/spec/verifies/cell, every Cargo dependency table, workspace inheritance, features and lock pruning across src, tests, examples, benches, build.rs and arbitrarily nested workspace members. Negatives preserve same-named local macros, strings/comments, re-exports, glob imports and invalid macro grammar. End-to-end REDs prove that a still-present path dependency blocks tree deletion before mutation, and a still-present macro/attribute blocks dependency removal rather than becoming rustc E0433 after publication. @status:spec/done - @fact:ACCEPT-TS-GO TypeScript/Go goldens cover mixed comments, empty metadata comments, CRLF/indentation preservation and non-target strings/prose/compiler/build/lint directives; product source/tests remain native-buildable. @status:spec/done - @fact:ACCEPT-MANAGED-RELOCATE Managed-block fixtures cover valid, absent, orphaned, reversed, duplicate, nested and embedded markers. Relocation covers destination collision, overlap, cycle, cross-filesystem refusal, byte/mode equality and ancestor retention. @status:spec/done - @fact:ACCEPT-PRESETS Epoch-1 has a real hermetic Cargo before/after smoke plus exact argv and discovery unit coverage. Equivalent real npm, Maven and Python preset smokes remain planned and unsupported profiles remain typed blockers. @status:impl/plan - @fact:ACCEPT-CUSTOM-HEALTH Epoch-1 ships the sealed Windows exit-code/inherit/reads-all/writes-none/spawn-allowed custom profile with bundle, argv, output, timeout, cancel and drift tests. Stronger structured, denied-network and restricted-effect profiles remain planned behind explicit capability blockers. @status:impl/plan - @fact:ACCEPT-BASELINE Strict before-red refuses with zero writes; before-green/after-red restores exact tree; admissible structured no-regression before-fail/after-subset yields reduced assurance; a new or worsened after finding rolls back; opaque nonzero never baselines. @status:spec/done - @fact:ACCEPT-TRANSACTION Fault injection at every journal, stage, rewrite, relocation, quarantine, contract-last, assertion, health, rollback and cleanup boundary proves restart idempotence, correct direction, third-state refusal and exact restoration. Before and after verifier cwd/`{root}` writes are trapped in the external exact copy/scratch, both the copy and delivered tree digest are re-proven after health, and ambient TMP/TEMP inside the project cannot redirect workspace creation into the product tree. @status:spec/done - @fact:ACCEPT-EXPORT-RACES Export REDs cover output equal to/inside/above source, destination race, unavailable no-replace rename, crash before/after publication, after health in an exact disjoint copy while the final location remains sealed, and extra/missing/changed descendants before rollback. Concurrent output data is never captured into transaction stage. @status:spec/done - @fact:ACCEPT-E2E Epoch-1 end-to-end fixtures cover real Rust export/in-place, Specmark erasure, Cargo.lock reconciliation, green before/after, zero selected residue, external report and safe second-scrape refusal. The mixed TypeScript/Go expansion remains planned. @status:impl/plan - @fact:ACCEPT-HOST-DOGFOOD A future host fixture covers this repository's direct `vibevm/vibepacks` Cargo path wiring and proves the scraper reports non-Specmark product tool dependencies as blockers rather than deleting them into a broken build. @status:impl/plan ## 15. Implementation decomposition {#implementation} 1. [p71] @fact:IMPL-A A — strict contract model, default generator, path/glob parser, inventory/classification, canonical plan and human/JSON rendering. @status:impl/done 2. @fact:IMPL-B B — typed rewrite/relocation algebra, managed blocks, Rust/Cargo, TypeScript/npm and Go/module adapters, Cargo v3/v4 native-lock reconciliation and residual proof. npm/pnpm/yarn/Go manager-native lock rows remain explicit blockers. @status:impl/done 3. @fact:IMPL-C C — prepared Cargo/npm/Maven/Python presets, the exact-copy Windows Cargo backend, the narrow sealed custom profile, effect/output/process control and baseline comparison. Chains without complete runtime capabilities fail closed. @status:impl/done 4. @fact:IMPL-D D — read-only source export transaction, candidate publication and export evidence. @status:impl/done 5. @fact:IMPL-E E — external journal/lock, same-volume quarantine, in-place contract-last commit, rollback and recovery. @status:impl/done 6. @fact:JOURNAL-E2-TRANSITION The pre-public internal transaction journal moved from handwritten epoch 1 JSON to the strict generated `scrape-transaction-journal/e2` root. Old epoch-1 state refuses with an actionable restart recipe and is not migrated automatically; the public scrape plan, report and health-result roots remain byte-identical at epoch 1. Root and named-object member order is schema-declared; inline standard JTD discriminator arms use the pinned generator's deterministic canonical field order. @status:impl/done 7. @fact:HEALTH-PERSISTENCE-E2-TRANSITION The pre-public prepared-health snapshot and canonical transaction health evidence use the strict generated `scrape-prepared-health-snapshot/e2` and `scrape-verification-health-evidence/e2` roots. An old handwritten prepared-health shape refuses with the restart recipe rather than being guessed forward. Live file identities and separately journaled verifier bundle bytes remain outside the plan snapshot and are reattached only after their recorded digest, byte count, mode and path identity agree. The public scrape plan, report and custom health-result roots remain byte-identical at epoch 1; their health finding/report-row vocabulary has one generated shared authority. @status:impl/done 8. @fact:IMPL-F F — the bounded epoch-1 evidence includes transaction/store fault matrices, composed production recovery, real Rust Cargo/custom smokes, Windows races and explicit non-Windows fail-closed probes. Wider ecosystem and host-dogfood rows remain planned above. @status:impl/done ## 16. Foundational decisions {#decisions} ### 16.1 Contract over heuristics {#decision-contract} [p72] @fact:DECISION-CONTRACT Decision. One strict project contract carries every policy choice; observation only evaluates it. @status:spec/done [p73] @fact:WHY-CONTRACT Why. Generated, authored and accepted-product states can share names. Making policy data is the only way to keep execution deterministic without guessing. @status:spec/done [p74] @fact:REJECT-CONTRACT Rejected. Delete-known-names, Git-ignore-as-ownership, prompt-per-file and LLM classification at execution time: each is order/environment dependent or non-reproducible. @status:spec/done [p75] @fact:REVISIT-CONTRACT Revisit when. A stable package-contributed recipe protocol can compile into the same closed plan without executing package code during scrape. @status:spec/done ### 16.2 Typed rewrites over regex and shell {#decision-typed} [p76] @fact:DECISION-TYPED Decision. Native parsers own source/manifests; literal exact-replace is the only textual escape hatch. @status:spec/done [p77] @fact:WHY-TYPED Why. Identity, cardinality and erasure equivalence are mechanically provable only over syntax nodes and exact preimages. @status:spec/done [p78] @fact:REJECT-TYPED Rejected. Regex source rewrites, arbitrary cleanup hooks and whole-file TOML/JSON regeneration: they damage product text or hide the changed surface. @status:spec/done [p79] @fact:REVISIT-TYPED Revisit when. Add a typed adapter when a real unsupported format has a parser and preservation oracle. @status:spec/done ### 16.3 Built-in health plus explicit custom oracle {#decision-health} [p80] @fact:DECISION-HEALTH Decision. Common ecosystems use declarative presets; unusual domain checks use a sealed custom verifier with structured argv and protocol. @status:spec/done [p81] @fact:WHY-HEALTH Why. Standard operations should not require one-off scripts, while no finite preset catalogue can express every project's definition of healthy. @status:spec/done [p82] @fact:REJECT-HEALTH Rejected. One shell string, after-only testing, silent missing test scripts, automatic retry and treating exit 0 as proof beyond the declared oracle. @status:spec/done [p83] @fact:REVISIT-HEALTH Revisit when. A new ecosystem demonstrates a stable build/test vocabulary deserving a built-in adapter. @status:spec/done ### 16.4 Export default, recoverable in-place option {#decision-transaction} [p84] @fact:DECISION-TRANSACTION Decision. Export is the safest delivery form; explicit in-place uses an external journal and same-volume quarantine with contract-last commit. @status:spec/done [p85] @fact:WHY-TRANSACTION Why. Project `.vibe` cannot recover its own deletion, Git does not cover every file, and multi-file mutation has no filesystem-wide atomic rename. Journaled before/after states make interruption decidable. @status:spec/done [p86] @fact:REJECT-TRANSACTION Rejected. Journal under project `.vibe`, Git reset as rollback, delete-first uninstall, blind recursive removal, external configurable journal, and cleaning quarantine before verified commit. @status:spec/done [p87] @fact:REVISIT-TRANSACTION Revisit when. A platform offers a stronger atomic directory-tree primitive without weakening cross-platform recovery semantics. @status:spec/done ## 17. History {#history} [p88] @fact:HISTORY-COMMISSIONED 2026-09-04: commissioned directly by the owner after auditing how Rust Specmark, TypeScript JSDoc and Go directive metadata behave when the Vibe tree disappears. The owner then extended the request to an explicit default/alternate contract and algorithmic standard/custom health checks, and ordered this feature ahead of every remaining lifecycle R task. @status:spec/done [p89] @fact:HISTORY-EPOCH1-WINDOWS 2026-09-05: implemented and ratified the Windows epoch-1 product surface: generated plan/report wire, external journal/workspace, capability mutation, export, in-place contract-last, rollback/recovery, exact-copy health, Cargo and the narrow custom profile. Recorded the remaining manager-native lock, wider preset/custom, mixed-language and host-dogfood work as explicit planned scope. @status:impl/done