@ai-sdlc/orchestrator 0.9.0 → 0.10.0

package/dist/runtime/attestations.js

@@ -32,15 +32,27 @@
  * supports it natively.
  */
 import { createHash, generateKeyPairSync, sign, verify } from 'node:crypto';
+import { execFileSync } from 'node:child_process';
+import { cleanGitEnv } from './git-env.js';
 /**
  * The currently-accepted predicate schema versions. CI rejects any envelope
  * whose `payload.schemaVersion` is not in this allowlist — this is the
- * forward-compatibility hatch (we add a new version here when we change the
- * predicate shape, and CI keeps accepting v1 until we explicitly remove it).
+ * forward-compatibility hatch.
+ *
+ * AISDLC-103 (Verifier Phase 3) narrowed this to `['v3']` only:
+ *  - `v1` envelopes (pre-AISDLC-94, diffHash-only) are rejected.
+ *  - `v2` was never landed as a distinct schemaVersion — the AISDLC-94
+ *    `contentHash` and AISDLC-101 `contentHashV3` shipped under the v1
+ *    schemaVersion as additive optional fields during the dual- and
+ *    triple-hash soak windows.
+ *  - `v3` envelopes carry `contentHashV3` as a required field and DO NOT
+ *    carry `diffHash` or `contentHash` (the legacy hashes are forbidden;
+ *    a v3 envelope smuggling either field is rejected by
+ *    `validatePredicateShape`).
  *
  * Exported so the `verify-attestation` workflow can `import`/inline it.
  */
-export const ACCEPTED_SCHEMA_VERSIONS = ['v1'];
+export const ACCEPTED_SCHEMA_VERSIONS = ['v3'];
 /**
  * The DSSE PAE payload type for our predicate. DSSE spec mandates a payload
  * type URI — we use a project-controlled vendor URI rather than the
@@ -59,8 +71,11 @@ export const DSSE_PAYLOAD_TYPE = 'application/vnd.ai-sdlc.attestation+json';
 // value — never give the attacker a way to smuggle their payload past
 // us by burying it in our reason text.
 //
-// Mirror of `.ai-sdlc/schemas/attestation.v1.schema.json` — kept in
-// sync by the `validatePredicateShape` test ('schema mirror in sync').
+// Mirror of `.ai-sdlc/schemas/attestation.v3.schema.json` — the v3 schema
+// requires `contentHashV3` and forbids the legacy `diffHash` / `contentHash`
+// fields. AISDLC-103 (Verifier Phase 3) narrowed the schemaVersion allowlist
+// to `['v3']` only; envelopes carrying the legacy hashes (= v1/v2 envelopes
+// smuggling themselves into the v3 window) are rejected with a fixed reason.
 /** sha1 git commit (40 lowercase hex chars). */
 const SHA1_HEX = /^[0-9a-f]{40}$/;
 /** sha256 hex (64 lowercase hex chars). */
@@ -69,6 +84,13 @@ const SHA256_HEX = /^[0-9a-f]{64}$/;
 const ISO_8601 = /^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(?:\.\d+)?(?:Z|[+-]\d{2}:\d{2})$/;
 /** Free-form short identifier — letters, digits, dot, dash, underscore. */
 const SHORT_ID = /^[A-Za-z0-9._-]+$/;
+/**
+ * Semver-shape pattern for `pipelineVersion` (AISDLC-100.6). Accepts
+ * `MAJOR.MINOR.PATCH` and the optional `-prerelease` suffix used by npm
+ * tags (e.g. `0.1.0-rc.2`). Mirrors the schema's regex so JSON-Schema
+ * validators and the in-process shape validator agree.
+ */
+const SEMVER = /^[0-9]+\.[0-9]+\.[0-9]+(-[a-z0-9.]+)?$/;
 /**
  * `harnessNote` is the only field the operator can put long-form text
  * in. We allow letters/digits/punctuation/whitespace but reject CR/LF
@@ -76,13 +98,18 @@ const SHORT_ID = /^[A-Za-z0-9._-]+$/;
  */
 const SAFE_TEXT = /^[^\r\n]*$/;
 /**
- * Validate a parsed predicate against the v1 schema regex patterns.
+ * Validate a parsed predicate against the v3 schema regex patterns.
  *
  * Returns `null` when the predicate is shape-valid; otherwise returns
  * a static failure reason that does NOT embed any user-controlled
  * value (just the field path). This is the load-bearing property:
  * the malicious value never reaches the `reason` string, so it can't
  * propagate to GITHUB_OUTPUT or commit-status descriptions.
+ *
+ * AISDLC-103 (Verifier Phase 3): `contentHashV3` is now required, and the
+ * legacy `diffHash` / `contentHash` fields are FORBIDDEN — a predicate
+ * carrying either is treated as a v1/v2 envelope smuggling itself into the
+ * v3 window and rejected with a static reason.
  */
 export function validatePredicateShape(parsed) {
     if (parsed === null || typeof parsed !== 'object') {
@@ -112,11 +139,40 @@ export function validatePredicateShape(parsed) {
     if (typeof sha1 !== 'string' || !SHA1_HEX.test(sha1)) {
         return 'schema validation failed: subject.digest.sha1 does not match pattern';
     }
-    // diffHash + policyHash — 64 hex chars each.
-    for (const field of ['diffHash', 'policyHash']) {
-        const v = p[field];
+    // policyHash — 64 hex chars. Required.
+    {
+        const v = p['policyHash'];
         if (typeof v !== 'string' || !SHA256_HEX.test(v)) {
-            return `schema validation failed: ${field} does not match pattern`;
+            return 'schema validation failed: policyHash does not match pattern';
+        }
+    }
+    // AISDLC-103 (Phase 3): legacy `diffHash` (v1) and `contentHash` (v2)
+    // are FORBIDDEN in v3 envelopes. A predicate that claims `schemaVersion:
+    // 'v3'` but carries either field is a v1/v2 envelope smuggling itself
+    // into the v3 window — reject with a fixed reason that doesn't embed
+    // the bad value.
+    if (p['diffHash'] !== undefined) {
+        return 'schema validation failed: diffHash is forbidden in v3 envelopes (legacy v1 field)';
+    }
+    if (p['contentHash'] !== undefined) {
+        return 'schema validation failed: contentHash is forbidden in v3 envelopes (legacy v2 field)';
+    }
+    // contentHashV3 (AISDLC-101) — REQUIRED in v3 envelopes. Must be a
+    // 64-char hex sha256.
+    {
+        const ch3 = p['contentHashV3'];
+        if (typeof ch3 !== 'string' || !SHA256_HEX.test(ch3)) {
+            return 'schema validation failed: contentHashV3 does not match pattern';
+        }
+    }
+    // contentHashV4 (AISDLC-193.1) — OPTIONAL during the v3+v4 dual-write
+    // transition window. When present, MUST be a 64-char hex sha256;
+    // when absent (legacy v3-only envelopes signed before this field
+    // landed), the verifier falls back to the v3 ancestor walk.
+    if (p['contentHashV4'] !== undefined) {
+        const ch4 = p['contentHashV4'];
+        if (typeof ch4 !== 'string' || !SHA256_HEX.test(ch4)) {
+            return 'schema validation failed: contentHashV4 does not match pattern';
         }
     }
     // pluginVersion — short ID (no CR/LF, no `=`).
@@ -126,6 +182,17 @@ export function validatePredicateShape(parsed) {
         !SHORT_ID.test(pluginVersion)) {
         return 'schema validation failed: pluginVersion does not match pattern';
     }
+    // pipelineVersion (AISDLC-100.6) — optional. When present, must be a
+    // semver-shaped string (`MAJOR.MINOR.PATCH` with optional `-prerelease`).
+    // Absence is OK (legacy v1 envelopes signed before pipeline-cli existed
+    // / before Phase 6 landed). The verifier logs but does NOT enforce a
+    // specific version — see `scripts/verify-attestation.mjs`.
+    if (p['pipelineVersion'] !== undefined) {
+        const pv = p['pipelineVersion'];
+        if (typeof pv !== 'string' || pv.length === 0 || !SEMVER.test(pv)) {
+            return 'schema validation failed: pipelineVersion does not match pattern';
+        }
+    }
     // iterationCount — positive integer.
     const iterationCount = p['iterationCount'];
     if (typeof iterationCount !== 'number' ||
@@ -139,6 +206,27 @@ export function validatePredicateShape(parsed) {
     if (typeof harnessNote !== 'string' || !SAFE_TEXT.test(harnessNote)) {
         return 'schema validation failed: harnessNote contains forbidden characters';
     }
+    // harness (AISDLC-202.3) — optional envelope-level harness field.
+    // Absent on pre-202.3 envelopes — accepted for backward compatibility.
+    // When present, must be an object with a SHORT_ID `name` and an optional
+    // SEMVER `version`. Validated before interpolation to prevent injection.
+    const harness = p['harness'];
+    if (harness !== undefined) {
+        if (harness === null || typeof harness !== 'object') {
+            return 'schema validation failed: harness must be an object when present';
+        }
+        const h = harness;
+        const hName = h['name'];
+        if (typeof hName !== 'string' || hName.length === 0 || !SHORT_ID.test(hName)) {
+            return 'schema validation failed: harness.name does not match SHORT_ID pattern';
+        }
+        const hVersion = h['version'];
+        if (hVersion !== undefined) {
+            if (typeof hVersion !== 'string' || !SEMVER.test(hVersion)) {
+                return 'schema validation failed: harness.version does not match SEMVER pattern';
+            }
+        }
+    }
     // signedAt — ISO 8601.
     const signedAt = p['signedAt'];
     if (typeof signedAt !== 'string' || !ISO_8601.test(signedAt)) {
@@ -195,6 +283,144 @@ export const REQUIRED_REVIEWER_AGENT_IDS = Object.freeze([
     'test-reviewer',
     'security-reviewer',
 ]);
+/**
+ * Name-equivalence map for the reviewer-set completeness check (AISDLC-252).
+ *
+ * A "role" is satisfied when any of the listed agentIds is present in the
+ * envelope's reviewer set. This lets codex-harness variants (`code-reviewer-codex`,
+ * `test-reviewer-codex`) satisfy the same role as their Claude counterparts,
+ * enabling the bidirectional cross-harness review goal without requiring a
+ * redundant Claude review on Codex-reviewed PRs.
+ *
+ * Security stays Claude-only: `security-reviewer` has no codex variant per
+ * `feedback_subagent_model_selection.md` (Claude Opus for security reasoning
+ * depth is not yet validated for Codex o4-mini).
+ *
+ * The map is keyed by role name (= the canonical agentId), each value is the
+ * set of ALL agentIds that satisfy the role (including the canonical one).
+ *
+ * Frozen to discourage callers from mutating it.
+ */
+export const REVIEWER_ROLE_EQUIVALENCES = Object.freeze({
+    'code-reviewer': Object.freeze(['code-reviewer', 'code-reviewer-codex']),
+    'test-reviewer': Object.freeze(['test-reviewer', 'test-reviewer-codex']),
+    'security-reviewer': Object.freeze(['security-reviewer']),
+});
+/**
+ * When the implementer ran in Codex (`predicate.harness.name === 'codex'`),
+ * these reviewer roles MUST be satisfied by a reviewer whose `harness` field
+ * differs from `codex`. Per RFC-0010 §13.10 `requiresIndependentHarnessFrom`:
+ * code and test reviewers must come from a different harness than the
+ * implementer to preserve cross-harness independence.
+ *
+ * Security is excluded — it is always Claude-only regardless.
+ *
+ * Frozen to discourage callers from mutating it.
+ */
+export const INDEPENDENCE_REQUIRED_ROLES = Object.freeze([
+    'code-reviewer',
+    'test-reviewer',
+]);
+/**
+ * Regex matching the envelope self-exclusion path pattern
+ * `.ai-sdlc/attestations/<sha>.dsse.json`. Used to filter out the
+ * envelope file itself from the file collector for AISDLC-193.1
+ * `contentHashV4` and AISDLC-101 `contentHashV3` purposes.
+ *
+ * The chore-commit pattern signs the predicate at the dev-commit (HEAD
+ * BEFORE the envelope file exists), then the chore commit on top adds
+ * the envelope file at `.ai-sdlc/attestations/<sha>.dsse.json`. If the
+ * collector includes the envelope file in the hashed file set, the
+ * verifier (which runs against PR HEAD = dev-commit + chore commit)
+ * will see an EXTRA entry for the envelope that the signer never saw
+ * → mismatch even on direct PR HEAD without any rebase.
+ *
+ * The exclusion applies to the file COLLECTOR for HASHING purposes
+ * only. The verifier's chore-commit allowlist (`scripts/verify-attestation.mjs`
+ * `CHORE_COMMIT_PATH_ALLOWLIST`) STILL allows the envelope file in the
+ * chore commit's diff — that's a separate concern from "what is in the
+ * file set we hash."
+ *
+ * Anchored with `^...$` against the forward-slash-normalized path so
+ * an attacker cannot bypass with `./.ai-sdlc/attestations/x.dsse.json`
+ * or `foo/.ai-sdlc/attestations/x.dsse.json`. Note that git's
+ * `--name-only` always emits paths relative to the repo root with
+ * forward slashes, so the match is straightforward in practice.
+ */
+export const ATTESTATION_ENVELOPE_PATH_PATTERN = /^\.ai-sdlc\/attestations\/[^/]+\.dsse\.json$/;
+/**
+ * Predicate to determine whether a file path identifies an attestation
+ * envelope and should therefore be excluded from `contentHashV3` /
+ * `contentHashV4` file enumeration. Defensive about backslash
+ * normalization (Windows callers).
+ */
+export function isAttestationEnvelopePath(path) {
+    if (typeof path !== 'string')
+        return false;
+    const normalized = path.replace(/\\/g, '/');
+    return ATTESTATION_ENVELOPE_PATH_PATTERN.test(normalized);
+}
+/**
+ * The "shared churn" exclude list for `contentHashV4` (AISDLC-258).
+ *
+ * Files in this list are EXCLUDED from the v4 file collector in BOTH the
+ * signer (`collectChangedFileDeltaEntries`) and the verifier
+ * (`computeHeadContentHashV4` in `scripts/verify-attestation.mjs`). When
+ * a file appears in this list, changes to it after signing (e.g. from a
+ * merge-queue rebase that regenerated `pnpm-lock.yaml`) do NOT cause
+ * `contentHashV4` to mismatch, so the operator is never asked to re-sign
+ * just because a shared tooling file was regenerated automatically.
+ *
+ * **Security trade-off (operator-approved, 2026-05-10):** An attacker
+ * COULD slip malicious changes through these files undetected (the
+ * attestation would still pass even if the ignore-listed file was
+ * tampered). The operator accepted this risk because:
+ *  - None of these files contain reviewable hand-written code.
+ *  - `pnpm-lock.yaml` is generated from `package.json` (which IS hashed).
+ *  - `CHANGELOG.md` variants are auto-generated by release-please from
+ *    commit history (which IS hashed via the commit-level binding).
+ *  - `generated-schemas.ts` is generated from spec schemas (reviewed
+ *    separately in the spec/ PR that changed them).
+ *
+ * **DO NOT add to this list:** `package.json` (real dep changes are
+ * reviewable), source files, test files, configs, RFCs, or anything a
+ * human writes by hand. The list is intentionally narrow.
+ *
+ * Paths are exact matches against the forward-slash-normalized repo-relative
+ * path emitted by `git diff --name-only`. Patterns (globs/regex) are NOT
+ * supported to keep the list auditable — every entry must be exact.
+ *
+ * Exported so `scripts/verify-attestation.mjs` can import it from the
+ * orchestrator barrel and apply the same exclusions on the verifier side.
+ */
+export const CONTENTHASHV4_IGNORE_FILES = Object.freeze([
+    'pnpm-lock.yaml',
+    'CHANGELOG.md',
+    'pipeline-cli/CHANGELOG.md',
+    'orchestrator/CHANGELOG.md',
+    // NOTE: `reference/src/core/generated-schemas.ts` was REMOVED from this
+    // list per AISDLC-258 code-review CRITICAL finding. Even though it's
+    // auto-generated from spec/schemas/, it remains a `.ts` source file
+    // shipped in reference/dist. An attacker who hand-edited it post-signing
+    // would bypass attestation. Keep it in the hash; PRs that change schemas
+    // pay the re-sign cost (which is correct — the schema change IS reviewable).
+]);
+/**
+ * Predicate to determine whether a file path should be excluded from the
+ * `contentHashV4` computation because it is a "shared churn" file (see
+ * `CONTENTHASHV4_IGNORE_FILES`). Defensive about backslash normalization.
+ *
+ * Note: this predicate is intentionally separate from
+ * `isAttestationEnvelopePath` because the two exclusions serve different
+ * purposes and may diverge independently. Merge them only if the list
+ * becomes large enough to warrant a single unified predicate.
+ */
+export function isIgnoredForContentHash(path) {
+    if (typeof path !== 'string')
+        return false;
+    const normalized = path.replace(/\\/g, '/');
+    return CONTENTHASHV4_IGNORE_FILES.includes(normalized);
+}
 /**
  * Compute a sha256 hex digest. Single source of truth for the hashing
  * algorithm — every predicate field that ends in `Hash` flows through here.
@@ -206,19 +432,448 @@ export function sha256Hex(input) {
 export function sha1Hex(input) {
     return createHash('sha1').update(input).digest('hex');
 }
+/**
+ * Compute the rebase-tolerant `contentHash` (AISDLC-94) over a changed-file
+ * set. The canonical encoding is one line per entry, sorted ascending by
+ * path, with `<path>\t<blobSha>\n` per line. The whole string is sha256-ed.
+ *
+ * Why this beats `diffHash`:
+ *   - Rebasing PR-X onto a new `main` that already touched the same files
+ *     does NOT change the post-apply blob SHAs (assuming no conflict),
+ *     so `contentHash` stays stable across the rebase.
+ *   - A conflict resolution that picks different content WILL change the
+ *     blob SHA → `contentHash` changes → attestation correctly invalidated.
+ *   - Force-pushing a no-op edit (e.g. `git commit --amend --no-edit`) keeps
+ *     blob SHAs identical → `contentHash` stays stable.
+ *
+ * The deduplication step makes the function idempotent if a caller
+ * accidentally passes the same path twice (last-write-wins per path).
+ *
+ * Pure function. The caller (sign-attestation script) is responsible for
+ * gathering the file set (via `git diff --name-only` + `git ls-tree`).
+ */
+export function computeContentHash(entries) {
+    // Dedup by path (last entry wins) so callers passing the same file
+    // twice — e.g. an add+modify in two diff invocations — don't produce
+    // a different hash than a clean run.
+    const byPath = new Map();
+    for (const e of entries) {
+        if (typeof e?.path !== 'string' || e.path.length === 0) {
+            throw new Error(`computeContentHash: entry path must be a non-empty string`);
+        }
+        if (typeof e.blobSha !== 'string') {
+            throw new Error(`computeContentHash: entry blobSha must be a string for path ${e.path}`);
+        }
+        // Reject path entries containing the canonical-encoding delimiters
+        // (\t between path and sha, \n between lines). Without this, a
+        // single entry `{ path: 'a\tB1\nb', blobSha: 'B2' }` and the
+        // two-entry set `[{ a, B1 }, { b, B2 }]` produce the same canonical
+        // string and therefore the same hash — defeating the binding. Git's
+        // default config already disallows \n in tracked filenames on most
+        // platforms; we defend in depth here so the hash itself is injective
+        // regardless of what the caller hands us.
+        if (e.path.includes('\t') || e.path.includes('\n')) {
+            throw new Error(`computeContentHash: entry path must not contain tab or newline characters (got ${JSON.stringify(e.path)})`);
+        }
+        // Normalize: forward-slashes (git already emits forward-slashes
+        // regardless of platform but be defensive), lowercase blob SHA.
+        const normalizedPath = e.path.replace(/\\/g, '/');
+        byPath.set(normalizedPath, e.blobSha.toLowerCase());
+    }
+    const sorted = [...byPath.entries()].sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0));
+    const canonical = sorted.map(([path, sha]) => `${path}\t${sha}\n`).join('');
+    return sha256Hex(canonical);
+}
+/**
+ * Collect the changed-file set used to compute `contentHash` (AISDLC-94).
+ *
+ * Returns one `{ path, blobSha }` entry per file in
+ * `git diff --name-only <baseRef>...<headRef>` with the blob SHA from
+ * `git ls-tree -r <headRef> -- <path>`. Deleted files get an empty
+ * `blobSha` (the path still appears so the canonical encoding distinguishes
+ * "deleted" from "kept").
+ *
+ * `--no-renames` so a rename shows up as add+delete (= two entries) — that
+ * way a rebase that resolved a conflict by renaming differently produces a
+ * different hash. `-c core.quotepath=false` mirrors the verifier's git
+ * helper so unicode paths come back as raw UTF-8.
+ *
+ * Path entries containing `\t` or `\n` are rejected to keep the canonical
+ * encoding injective (mirrors the rejection in `computeContentHash`). Such
+ * paths are exceedingly rare in practice — git's default config disallows
+ * `\n` in tracked filenames on most platforms — but we defend in depth so
+ * malicious or pathological inputs can't smuggle entries past the binding.
+ *
+ * Extracted from the previously-duplicated helpers in
+ * `ai-sdlc-plugin/scripts/sign-attestation.mjs` so a single source of truth
+ * applies the same parsing + validation at every signing site.
+ */
+export function collectChangedFileEntries(baseRef, headRef, repoRoot, options = {}) {
+    const runGit = options.runGit ??
+        ((args, cwd) => execFileSync('git', args, {
+            cwd,
+            env: cleanGitEnv(),
+            encoding: 'utf-8',
+            maxBuffer: 64 * 1024 * 1024,
+        }));
+    let nameOnly;
+    try {
+        nameOnly = runGit([
+            '-c',
+            'core.quotepath=false',
+            'diff',
+            '--name-only',
+            '--no-renames',
+            `${baseRef}...${headRef}`,
+        ], repoRoot);
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        throw new Error(`collectChangedFileEntries: git diff --name-only failed: ${msg}`);
+    }
+    const paths = nameOnly.split('\n').filter((p) => p.length > 0);
+    const entries = [];
+    for (const path of paths) {
+        // Reject delimiters here too so the error surfaces at the enumeration
+        // site (cleaner than failing later inside computeContentHash).
+        if (path.includes('\t') || path.includes('\n')) {
+            throw new Error(`collectChangedFileEntries: path must not contain tab or newline characters (got ${JSON.stringify(path)})`);
+        }
+        // `git ls-tree -r <ref> -- <path>` returns blank when the path doesn't
+        // exist at <ref> (= deleted file). Empty blobSha is then used as the
+        // marker — see computeContentHash for canonical encoding.
+        let blobSha = '';
+        try {
+            const lsOut = runGit(['-c', 'core.quotepath=false', 'ls-tree', '-r', headRef, '--', path], repoRoot);
+            // ls-tree output: `<mode> <type> <sha>\t<path>` (one line per file).
+            const line = lsOut.split('\n').find((l) => l.length > 0);
+            if (line) {
+                const m = line.match(/^[0-9]+\s+blob\s+([0-9a-f]{40})\t/);
+                if (m)
+                    blobSha = m[1];
+            }
+        }
+        catch {
+            // ls-tree failed (path missing) → treat as deleted, leave blobSha=''.
+        }
+        entries.push({ path, blobSha });
+    }
+    return entries;
+}
+/**
+ * Compute the per-file-delta `contentHashV3` (AISDLC-101) over a set of
+ * `{path, baseBlobSha, headBlobSha}` triples. The canonical encoding is
+ * one line per entry, sorted ascending by path, with
+ * `<path>\t<fileDeltaHash>\n` per line, where
+ * `fileDeltaHash = sha256(baseBlobSha + ' -> ' + headBlobSha)`. The
+ * outer `contentHashV3` is the sha256 of the concatenated lines.
+ *
+ * Why per-file delta hashing — and what it adds vs. AISDLC-94's `contentHash`:
+ *   - `contentHash` (AISDLC-94) hashes the post-apply blob SHA per file.
+ *     If a sibling PR landed between OUR sign + OUR merge AND modified
+ *     the SAME file, the rebased file's HEAD blob SHA contains both the
+ *     sibling contribution AND ours → contentHash diverges (false reject).
+ *   - `contentHashV3` (AISDLC-101) hashes the (base, head) blob-pair
+ *     transition per file. Provides a stricter "we moved file F from blob
+ *     A to blob B" binding than just "we ended up at blob B". Any genuine
+ *     content change still flips the head blob SHA → fileDeltaHash flips
+ *     → contentHashV3 flips → reject (threat model preserved).
+ *
+ * This is the SECOND line of defense in the 3-layer rebase-tolerance
+ * plan (AISDLC-94 = Phase 1 verifier-side dual-hash, AISDLC-102 = Phase 1.5
+ * producer-side pre-sign rebase, AISDLC-101 = Phase 2 per-file delta).
+ * The verifier OR's all three legs during the triple-hash window.
+ *
+ * Path-delimiter rejection (\t / \n) mirrors `computeContentHash` so the
+ * canonical encoding stays injective regardless of caller input.
+ *
+ * Pure function. Idempotent against double-enumeration via dedup-by-path
+ * (last-write-wins per path), same as `computeContentHash`.
+ */
+export function computeContentHashV3(entries) {
+    // Dedup by path (last entry wins) so callers passing the same file
+    // twice — e.g. add+modify in two diff invocations — don't produce a
+    // different hash than a clean run.
+    const byPath = new Map();
+    for (const e of entries) {
+        if (typeof e?.path !== 'string' || e.path.length === 0) {
+            throw new Error(`computeContentHashV3: entry path must be a non-empty string`);
+        }
+        if (typeof e.baseBlobSha !== 'string') {
+            throw new Error(`computeContentHashV3: entry baseBlobSha must be a string for path ${e.path}`);
+        }
+        if (typeof e.headBlobSha !== 'string') {
+            throw new Error(`computeContentHashV3: entry headBlobSha must be a string for path ${e.path}`);
+        }
+        // Reject path entries containing the canonical-encoding delimiters
+        // (\t between path and delta hash, \n between lines). See the same
+        // rejection in `computeContentHash` for the injectivity rationale.
+        if (e.path.includes('\t') || e.path.includes('\n')) {
+            throw new Error(`computeContentHashV3: entry path must not contain tab or newline characters (got ${JSON.stringify(e.path)})`);
+        }
+        const normalizedPath = e.path.replace(/\\/g, '/');
+        byPath.set(normalizedPath, {
+            baseBlobSha: e.baseBlobSha.toLowerCase(),
+            headBlobSha: e.headBlobSha.toLowerCase(),
+        });
+    }
+    const sorted = [...byPath.entries()].sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0));
+    const canonical = sorted
+        .map(([path, { baseBlobSha, headBlobSha }]) => {
+        const fileDeltaHash = sha256Hex(`${baseBlobSha} -> ${headBlobSha}`);
+        return `${path}\t${fileDeltaHash}\n`;
+    })
+        .join('');
+    return sha256Hex(canonical);
+}
+/**
+ * Compute the BASE-INDEPENDENT per-file head-blob `contentHashV4`
+ * (AISDLC-193.1) over a set of `{path, headBlobSha}` pairs. The
+ * canonical encoding is `JSON.stringify(sorted-by-path-array-of-{path,
+ * headBlobSha}-objects)`, hashed with sha256.
+ *
+ * Why JSON-of-sorted-array (and not the v3 `<path>\t<fileDeltaHash>\n`
+ * canonical) for v4:
+ *   - JSON's quoting rules already cover delimiter injection
+ *     (a malicious path containing tab/newline can't smuggle through
+ *     because they round-trip as escape sequences). We still reject
+ *     such paths defensively so the canonical stays injective and
+ *     the on-the-wire representation is what readers expect.
+ *   - JSON is unambiguous about field ordering (stringify of a
+ *     `{path, headBlobSha}` literal always emits `path` first,
+ *     `headBlobSha` second — V8's object-key ordering is insertion
+ *     order, and we insert in this order in the .map() below).
+ *   - Easier to extend: future hash versions can add fields
+ *     (`mode`, `executable bit`, etc) to the entry objects without
+ *     breaking the canonical encoding scheme.
+ *
+ * Why this is BASE-INDEPENDENT (= the whole point):
+ *   - v3's per-file delta hashes the (base_blob, head_blob) pair.
+ *     When the merge queue rebases the PR onto current main (which
+ *     advanced past the merge-base the producer signed against), the
+ *     base blob SHA for any shared file changes → v3 invalidates.
+ *   - v4 hashes only `{path, headBlobSha}`. Whatever the rebase does
+ *     to the base ref or the merge-base, as long as the head blob SHA
+ *     (= the actual reviewed file content) is unchanged, v4 matches.
+ *   - The reviewer never approved "base_blob X → head_blob Y"; they
+ *     approved "the file contents at head_blob Y." v4 binds to that
+ *     directly.
+ *
+ * Threat model preserved:
+ *   - Genuine post-sign content tampering (someone amends the PR to
+ *     add unreviewed code) flips the head blob SHA → v4 hash flips →
+ *     verifier rejects. Same threat-model surface as v3.
+ *   - The signing key still has to be a trusted reviewer's; v4
+ *     doesn't change the signature/key flow, just what the predicate
+ *     binds to.
+ *
+ * Pure function. Idempotent against double-enumeration via dedup-by-path
+ * (last-write-wins per path), same as `computeContentHash` and
+ * `computeContentHashV3`.
+ */
+export function computeContentHashV4(entries) {
+    // Dedup by path (last entry wins) — see computeContentHash and
+    // computeContentHashV3 for the idempotency rationale.
+    const byPath = new Map();
+    for (const e of entries) {
+        if (typeof e?.path !== 'string' || e.path.length === 0) {
+            throw new Error(`computeContentHashV4: entry path must be a non-empty string`);
+        }
+        if (typeof e.headBlobSha !== 'string') {
+            throw new Error(`computeContentHashV4: entry headBlobSha must be a string for path ${e.path}`);
+        }
+        // Reject path entries containing JSON-control characters that we
+        // can't unambiguously round-trip. JSON.stringify would happily
+        // escape these, but our canonical form is supposed to be readable
+        // + reversible — defense in depth keeps the on-the-wire form
+        // injective regardless of caller input. Same rejection list as
+        // computeContentHash / computeContentHashV3 for consistency.
+        if (e.path.includes('\t') || e.path.includes('\n')) {
+            throw new Error(`computeContentHashV4: entry path must not contain tab or newline characters (got ${JSON.stringify(e.path)})`);
+        }
+        const normalizedPath = e.path.replace(/\\/g, '/');
+        byPath.set(normalizedPath, e.headBlobSha.toLowerCase());
+    }
+    const sorted = [...byPath.entries()]
+        .sort(([a], [b]) => (a < b ? -1 : a > b ? 1 : 0))
+        .map(([path, headBlobSha]) => ({ path, headBlobSha }));
+    return sha256Hex(JSON.stringify(sorted));
+}
+/**
+ * Collect the per-file-delta set used to compute `contentHashV3` (AISDLC-101).
+ *
+ * Returns one `{ path, baseBlobSha, headBlobSha }` entry per file in
+ * `git diff --name-only <baseRef>...<headRef>`. The base blob SHA is read
+ * from the *merge-base* of `<baseRef>` and `<headRef>` (which the `...`
+ * 3-dot diff range already targets — `A...B` diffs against
+ * `merge-base(A,B)`); the head blob SHA from `<headRef>`. Files newly
+ * added in the PR have empty `baseBlobSha`; deleted files have empty
+ * `headBlobSha`.
+ *
+ * Mirrors `collectChangedFileEntries`'s flag set (`--no-renames`,
+ * `core.quotepath=false`) for consistency with the other binding's file
+ * enumeration.
+ *
+ * Extracted so a single source of truth handles the two ls-tree lookups
+ * (one per endpoint) at every signing site (`sign-attestation.mjs`).
+ */
+export function collectChangedFileDeltaEntries(baseRef, headRef, repoRoot, options = {}) {
+    const runGit = options.runGit ??
+        ((args, cwd) => execFileSync('git', args, {
+            cwd,
+            env: cleanGitEnv(),
+            encoding: 'utf-8',
+            maxBuffer: 64 * 1024 * 1024,
+        }));
+    // Resolve the merge-base ONCE so each ls-tree below uses a stable
+    // commit (`<baseRef>` may be a moving ref like `origin/main`). The
+    // `A...B` diff range already targets merge-base(A,B), so reading
+    // base blob SHAs at the merge-base keeps the per-file delta
+    // semantically aligned with the file enumeration.
+    let mergeBase;
+    try {
+        mergeBase = runGit(['merge-base', baseRef, headRef], repoRoot).trim();
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        throw new Error(`collectChangedFileDeltaEntries: git merge-base failed: ${msg}`);
+    }
+    if (!/^[0-9a-f]{40}$/.test(mergeBase)) {
+        throw new Error(`collectChangedFileDeltaEntries: git merge-base returned non-SHA output: ${JSON.stringify(mergeBase)}`);
+    }
+    let nameOnly;
+    try {
+        nameOnly = runGit([
+            '-c',
+            'core.quotepath=false',
+            'diff',
+            '--name-only',
+            '--no-renames',
+            `${baseRef}...${headRef}`,
+        ], repoRoot);
+    }
+    catch (err) {
+        const msg = err instanceof Error ? err.message : String(err);
+        throw new Error(`collectChangedFileDeltaEntries: git diff --name-only failed: ${msg}`);
+    }
+    const paths = nameOnly.split('\n').filter((p) => p.length > 0);
+    const entries = [];
+    /**
+     * Resolve a file's blob SHA at a given ref via `git ls-tree -r`. Returns
+     * the empty string when the path doesn't exist at the ref (= the file
+     * was added in the PR for `mergeBase`, or deleted in the PR for `headRef`).
+     */
+    const resolveBlobSha = (ref, path) => {
+        try {
+            const lsOut = runGit(['-c', 'core.quotepath=false', 'ls-tree', '-r', ref, '--', path], repoRoot);
+            const line = lsOut.split('\n').find((l) => l.length > 0);
+            if (line) {
+                const m = line.match(/^[0-9]+\s+blob\s+([0-9a-f]{40})\t/);
+                if (m)
+                    return m[1];
+            }
+        }
+        catch {
+            // ls-tree failed (path missing at ref) → empty blob marker.
+        }
+        return '';
+    };
+    for (const path of paths) {
+        if (path.includes('\t') || path.includes('\n')) {
+            throw new Error(`collectChangedFileDeltaEntries: path must not contain tab or newline characters (got ${JSON.stringify(path)})`);
+        }
+        // AISDLC-193.1 envelope self-exclusion: the chore-commit pattern
+        // signs the predicate at the dev commit (HEAD before the envelope
+        // file exists), then a chore commit on top adds the envelope file
+        // at `.ai-sdlc/attestations/<sha>.dsse.json`. If the collector
+        // includes the envelope file in the hashed file set, the verifier
+        // (which runs against PR HEAD = dev-commit + chore commit) will
+        // see an EXTRA entry the signer never saw → contentHashV3 mismatch
+        // even on direct PR HEAD without any rebase.
+        //
+        // Applied to BOTH v3 (this collector) AND v4 (the v4 collector
+        // delegates to this one + projects to head-only) so existing v3
+        // envelopes that touched .ai-sdlc/attestations/ as part of their
+        // diff still work after the dual-write switchover. The verifier's
+        // chore-commit allowlist STILL allows the envelope file to appear
+        // in the chore-commit diff — the exclusion is for HASHING only.
+        if (isAttestationEnvelopePath(path))
+            continue;
+        // AISDLC-258: shared-churn exclude list. Files like `pnpm-lock.yaml`
+        // and `CHANGELOG.md` change in nearly every PR (auto-generated by
+        // tooling or release-please). When a merge-queue rebase regenerates
+        // them, their blob SHAs shift → v4 mismatches → operator must re-sign
+        // despite no hand-written code change. Excluding them here (and on the
+        // verifier side in `computeHeadContentHashV4`) prevents that loop.
+        // Security trade-off accepted by operator 2026-05-10 (see
+        // `CONTENTHASHV4_IGNORE_FILES` for full rationale).
+        if (isIgnoredForContentHash(path))
+            continue;
+        const baseBlobSha = resolveBlobSha(mergeBase, path);
+        const headBlobSha = resolveBlobSha(headRef, path);
+        entries.push({ path, baseBlobSha, headBlobSha });
+    }
+    return entries;
+}
+/**
+ * Project a v3 `ChangedFileDeltaEntry` set down to the v4
+ * `ChangedFileHeadEntry` shape (`{path, headBlobSha}`). Convenience
+ * for callers that already collected v3 deltas and want to dual-emit
+ * both hashes from the same file enumeration. Pure function.
+ *
+ * The envelope self-exclusion is enforced upstream by
+ * `collectChangedFileDeltaEntries`, so this projection is a simple
+ * field-pick — no path filtering needed here.
+ */
+export function projectDeltaEntriesToHeadEntries(deltas) {
+    return deltas.map((d) => ({ path: d.path, headBlobSha: d.headBlobSha }));
+}
 /**
  * Build the predicate payload from raw inputs. Pure function — no I/O,
  * no signing. The caller (`/ai-sdlc execute` Step 10) reads files and git
  * output, then hands them here.
+ *
+ * AISDLC-103 (Verifier Phase 3): always emits a v3 envelope. The caller
+ * MUST provide `changedFileDeltas` (use `[]` for no-op PRs); the legacy
+ * `diff` + `changedFiles` inputs were dropped along with the legacy
+ * `diffHash` + `contentHash` fields.
  */
 export function buildPredicate(inputs) {
     if (!/^[0-9a-f]{40}$/i.test(inputs.commitSha)) {
         throw new Error(`buildPredicate: commitSha must be a 40-char hex SHA-1, got ${inputs.commitSha}`);
     }
-    return {
-        schemaVersion: 'v1',
+    if (!Array.isArray(inputs.changedFileDeltas)) {
+        throw new Error(`buildPredicate: changedFileDeltas must be an array (pass [] for no-op PRs)`);
+    }
+    // Per-element shape guard catches producer-side bugs early — without this,
+    // a malformed delta would surface as an opaque contentHashV3 mismatch on the
+    // verifier side (different machine), making debugging much harder.
+    for (let i = 0; i < inputs.changedFileDeltas.length; i++) {
+        const delta = inputs.changedFileDeltas[i];
+        if (!delta || typeof delta !== 'object') {
+            throw new Error(`buildPredicate: changedFileDeltas[${i}] must be an object`);
+        }
+        if (typeof delta.path !== 'string' || delta.path.length === 0) {
+            throw new Error(`buildPredicate: changedFileDeltas[${i}].path must be a non-empty string`);
+        }
+        if (typeof delta.baseBlobSha !== 'string') {
+            throw new Error(`buildPredicate: changedFileDeltas[${i}].baseBlobSha must be a string`);
+        }
+        if (typeof delta.headBlobSha !== 'string') {
+            throw new Error(`buildPredicate: changedFileDeltas[${i}].headBlobSha must be a string`);
+        }
+    }
+    // AISDLC-193.1: derive v4 head-entry set from the v3 delta set so
+    // the file enumeration (and the envelope self-exclusion built into
+    // the v3 collector) is shared between both hashes by construction.
+    // Producers therefore can't accidentally compute v3 over one file
+    // set and v4 over another.
+    const headEntries = projectDeltaEntriesToHeadEntries(inputs.changedFileDeltas);
+    const predicate = {
+        schemaVersion: 'v3',
         subject: { digest: { sha1: inputs.commitSha.toLowerCase() } },
-        diffHash: sha256Hex(inputs.diff),
+        contentHashV3: computeContentHashV3(inputs.changedFileDeltas),
+        contentHashV4: computeContentHashV4(headEntries),
         policyHash: sha256Hex(inputs.policy),
         reviewers: inputs.reviewers.map((r) => ({
             agentId: r.agentId,
@@ -232,6 +887,23 @@ export function buildPredicate(inputs) {
         harnessNote: inputs.harnessNote,
         signedAt: inputs.signedAt ?? new Date().toISOString(),
     };
+    // AISDLC-100.6: include `pipelineVersion` only when the caller provided
+    // it. Omitted otherwise so envelopes signed in environments without
+    // pipeline-cli installed still round-trip identically through
+    // validatePredicateShape.
+    if (typeof inputs.pipelineVersion === 'string' && inputs.pipelineVersion.length > 0) {
+        predicate.pipelineVersion = inputs.pipelineVersion;
+    }
+    // AISDLC-202.3: include `harness` only when the caller provided it.
+    // Omitted on legacy / Claude Code paths so pre-202.3 envelopes round-trip
+    // cleanly through validatePredicateShape (which treats absence as back-compat).
+    if (inputs.harness && typeof inputs.harness.name === 'string' && inputs.harness.name.length > 0) {
+        predicate.harness = { name: inputs.harness.name };
+        if (typeof inputs.harness.version === 'string' && inputs.harness.version.length > 0) {
+            predicate.harness.version = inputs.harness.version;
+        }
+    }
+    return predicate;
 }
 /**
  * DSSE Pre-Authentication Encoding. Per the spec
@@ -379,8 +1051,47 @@ export function verifyAttestation(opts) {
             reason: `subject digest mismatch (envelope was signed for a different commit)`,
         };
     }
-    if (predicate.diffHash !== opts.expected.diffHash) {
-        return { valid: false, reason: 'diffHash mismatch (PR diff changed since attestation)' };
+    // AISDLC-193.1: v4-prefer, v3-fallback. The verifier prefers
+    // `contentHashV4` (base-independent — survives queue rebases) when
+    // BOTH the envelope and the expected state carry it. v3 is only
+    // consulted when the envelope is legacy v3-only OR when v4 is
+    // present but doesn't match (unusual — implies the head blobs
+    // genuinely changed, which we still reject below).
+    //
+    // Why prefer v4 absolute when both are present:
+    //   - v3 binds to (base_blob, head_blob) per file. The merge queue's
+    //     rebase moves the merge-base forward → base_blob changes for
+    //     shared files → v3 mismatches even though reviewed content is
+    //     unchanged. Net: required check fails on every queued PR.
+    //   - v4 binds only to head_blob per file. Reviewers approved the
+    //     content at head_blob; whatever the rebase did to base, v4
+    //     stays valid.
+    //
+    // Threat model preserved: a genuine post-sign content tampering
+    // (someone amends the PR to add unreviewed code) flips head_blob →
+    // v4 flips → verifier rejects.
+    const envelopeHasV4 = typeof predicate.contentHashV4 === 'string';
+    const expectedHasV4 = typeof opts.expected.contentHashV4 === 'string';
+    if (envelopeHasV4 && expectedHasV4) {
+        if (predicate.contentHashV4 !== opts.expected.contentHashV4) {
+            return {
+                valid: false,
+                reason: 'contentHashV4 mismatch (PR content changed since attestation)',
+            };
+        }
+        // v4 matched → skip the v3 check entirely. The producer's v3 was
+        // computed against a base ref that may have moved on by now (queue
+        // rebase, sibling overlap); the v4 match is the source of truth.
+    }
+    else {
+        // Legacy v3-only envelope OR caller did not supply expected.contentHashV4
+        // → fall back to v3. Same as pre-AISDLC-193.1 behavior.
+        if (predicate.contentHashV3 !== opts.expected.contentHashV3) {
+            return {
+                valid: false,
+                reason: 'contentHashV3 mismatch (PR content changed since attestation)',
+            };
+        }
     }
     if (predicate.policyHash !== opts.expected.policyHash) {
         return {
@@ -397,20 +1108,44 @@ export function verifyAttestation(opts) {
             };
         }
     }
-    // ── Reviewer-set completeness ────────────────────────────────
-    // Every attestation MUST cover all three required reviewers (code,
-    // test, security). Without this, a contributor could ship an
-    // attestation containing only `code-reviewer` and bypass the test
-    // and security review entirely.
-    const present = new Set(predicate.reviewers.map((r) => r.agentId));
-    for (const required of REQUIRED_REVIEWER_AGENT_IDS) {
-        if (!present.has(required)) {
+    // ── Reviewer-set completeness (AISDLC-252) ──────────────────────
+    // Every attestation MUST cover all three required reviewer ROLES (code,
+    // test, security). Each role is satisfied by ANY agentId in its
+    // equivalence group — so `code-reviewer-codex` satisfies the `code-reviewer`
+    // role, enabling cross-harness reviews without a redundant Claude review.
+    // Security stays Claude-only (no codex variant, per policy).
+    const presentIds = new Set(predicate.reviewers.map((r) => r.agentId));
+    for (const [role, variants] of Object.entries(REVIEWER_ROLE_EQUIVALENCES)) {
+        const satisfied = variants.some((v) => presentIds.has(v));
+        if (!satisfied) {
             return {
                 valid: false,
-                reason: `reviewer set incomplete: missing required reviewer '${required}'`,
+                reason: `reviewer set incomplete: missing required reviewer '${role}' (or any variant: ${variants.join(', ')})`,
             };
         }
     }
+    // ── Independence enforcement (AISDLC-252, RFC-0010 §13.10) ──────
+    // When the implementer ran in codex (`predicate.harness.name === 'codex'`),
+    // the code-reviewer and test-reviewer MUST NOT also be codex — that would
+    // defeat the cross-harness independence goal. Security is exempt because it
+    // is always Claude-only.
+    const implementerHarness = predicate.harness?.name?.toLowerCase();
+    if (implementerHarness === 'codex') {
+        for (const role of INDEPENDENCE_REQUIRED_ROLES) {
+            // Find the reviewer entry that satisfied this role.
+            const satisfyingVariants = REVIEWER_ROLE_EQUIVALENCES[role] ?? [];
+            const reviewerEntry = predicate.reviewers.find((r) => satisfyingVariants.includes(r.agentId));
+            if (reviewerEntry) {
+                const reviewerHarness = reviewerEntry.harness?.toLowerCase();
+                if (reviewerHarness === 'codex') {
+                    return {
+                        valid: false,
+                        reason: `independence violation: implementer harness is 'codex' but reviewer '${reviewerEntry.agentId}' also uses codex (requiresIndependentHarnessFrom per RFC-0010 §13.10)`,
+                    };
+                }
+            }
+        }
+    }
     return { valid: true, predicate, trustedReviewer: matchedReviewer };
 }
 /**