The 19 fixed credential shapes in pre-edit-secrets.mjs were regex literals;
they now come from signatures/secret-egress.json in the vendored commons via
a new scanners/lib/secret-egress.mjs. Policy-injected custom patterns (entries
20+) are unchanged and still appended by the hook.
Measured before the swap, not assumed: all 19 positions compared for order,
name, regex source and flags, plus recompilation identity, against the literal
table sliced out of the module text. Zero divergences. Commons had reported
the same result; that was their measurement, so this one was run anyway.
STATE's expectation that the golden gate would go red on both table records
and file sha256 was wrong: pre-edit-secrets.mjs is in neither PINNED_FILES nor
WALKED_MODULES, so the table had no golden coverage at all and the swap moved
nothing. Rather than leave the vendored data with only behavioural coverage,
secret-egress.mjs joins WALKED_MODULES — walked, not pinned, since it inlines
no regex of its own. Golden diff was 19 ADDED, 0 CHANGED, 0 REMOVED, each
source byte-identical to the pre-swap literal; re-blessed. suite-counts.json
untouched.
Tests: coverage is derived from the loaded table, so an entry commons adds
cannot arrive without an end-to-end probe. All 19 now block through the real
hook and are asserted by label, which also pins the ordering contract (a
Bearer-wrapped JWT must report as the header). Mutating the vendored JSON
fires in both directions plus reorder: under-match (AKIA quantifier) reddens
3 hook tests + golden; over-match (Anthropic key truncated to its prefix)
reddens the false-positive probe + golden; moving the JWT entry ahead of the
Bearer entry reddens the ordering test.
Suite 2231 tests / 2223 pass / 6 skipped. The two parallel-run failures
(pre-compact size-cap, benchmark) pass alone — the known timing flakes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MGMv5ZTUhVzZtCCwRrNZG5
Bare/unquoted legacy OpenAI keys (no label assignment, no Bearer prefix)
slipped past the pre-write secret-detection hook. Added a pattern anchored
on the T3BlbkFJ base64 "OpenAI" watermark (vendor-documented shape),
avoiding the collision-prone bare sk-+48alnum form. Failing tests first,
full suite green (2184/0/6).
The originally planned source for this fix — porting two entries from
commons' secret-egress.json — turned out to be a false premise: that file
is a byte-identical copy of this hook's own table, not a superset. The two
missing names existed only as prose in commons' conformance/manifest.json,
describing a different repo's (the guard's) unpublished Python table.
gcp-service-account-json was measured NOT to be a gap (already covered by
the existing PEM-block pattern); openai-api-key-legacy was the one real
gap, closed here with a locally-authored pattern rather than an invented
"port". Commons notified via coord-send that their secret-egress.json
(count: 18) is now stale.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PYJX35KLH3rpS6pi7LHj8u
BREAKING CHANGE: the four LLM_SECURITY_* configuration env-vars deprecated in
v7.3.0 are removed. .llm-security/policy.json is now the only source:
LLM_SECURITY_INJECTION_MODE -> injection.mode
LLM_SECURITY_TRIFECTA_MODE -> trifecta.mode
LLM_SECURITY_ESCALATION_WINDOW -> trifecta.escalation_window
LLM_SECURITY_AUDIT_LOG -> audit.log_path
LLM_SECURITY_DEPRECATION_QUIET -> dies with the mechanism it silenced
Setting a removed var is now inert - it does not warn, and it does not
configure. Env-vars with no policy equivalent (PRECOMPACT_MODE,
PRECOMPACT_MAX_BYTES, UPDATE_CHECK, MCP_CACHE_FILE, IDE_ROOTS) are unaffected.
getPolicyValueWithEnvWarn and its one-shot stderr warning are deleted from
policy-loader.mjs, along with the module-scoped warned-var Set. The four call
sites collapse to getPolicyValue. getPolicyValue's JSDoc claimed "environment
variables ALWAYS take precedence" - it never read env itself, so that line
described the shim, and it is corrected rather than deleted.
User-facing hook strings that advertised a removed var as the escape hatch now
name the policy key instead: the inject-scan block reason, its warn-mode note,
both escalation-window advisories, and the trifecta block message. A blocked
user following the old text would have set a var that does nothing.
Tests. tests/lib/v8-env-removal.test.mjs is the regression gate and was written
failing first (8 of 12 red before the change). It pins the NEGATIVE - setting a
removed var does not alter the outcome - because that is the half that rots
silently: a re-introduced process.env read would leave every migrated positive
test green, since those configure through policy.json and never set the var at
all. One assertion walks hooks/scripts and scanners for `process.env.<removed>`
so the re-introduction is caught structurally, not only behaviourally.
The 44 env-driven test occurrences (18 inject-scan, 13+4 session-guard, 9
audit-trail) migrate to a throwaway .llm-security/policy.json via a new
runHookWithPolicy helper in hook-helper.mjs; audit-trail runs in-process, so it
supplies the same policy through CLAUDE_PROJECT_ROOT. The D3 mechanism tests in
policy-loader.test.mjs are deleted with the mechanism.
Suite 2039 tests, 2037 pass (+12 gate, -7 D3 mechanism). The 2 failures are the
known parallel-load timing flakes (pre-compact-scan size-cap,
pre-install-supply-chain F-3); both green when run isolated.
Remaining in Phase 3: posture-scanner TRIFECTA_MODE heuristic, riskScoreV1
removal, ghost-var cleanup, docs + migration note.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BB4vXvwvtW4dxbPRd6vsez
The PostToolUse hook read input.tool_output, but live Claude Code delivers the tool result as tool_response — so the indirect-injection scan on MCP tool output silently never fired outside the test harness (which sent tool_output). The hook now reads tool_response with a tool_output fallback for older harnesses/fixtures. +3 tests covering string and object tool_response and the precedence. Found via a live-session check during the v7.8.3 sweep; not one of the 52 MEDIUM-tier findings. Suite 2016/0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
Bumps plugin to 7.8.3 (package.json, plugin.json, README version badge) and syncs the documentation surfaces that had drifted: orchestrated scanner count 9/10/13 to 14 (#8; the WFL/workflow scanner was omitted from the reference list), posture categories 13 to 16 (#44), red-team scenarios 64 to 72 (#45), tests badge/prose 1822 to 2013 (#46). Corrects the pathguard hook-table matcher and header comment to Edit|Write (follows the #9 fix) and removes the two dangling ROADMAP.md references (#47; the roadmap is not a public file). Adds the v7.8.3 CHANGELOG entry, README Recent-versions row, and CLAUDE.md highlights. The '23 scanners' total is a separate module count and is left unchanged.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
#9 pathguard registered matcher Write only, so an Edit/MultiEdit to an existing protected file (settings.json, .env, .ssh, the hooks themselves) bypassed it entirely; matcher now Edit|Write (the script already reads only tool_input.file_path). #10 the primary trifecta detector's 20-entry window counted marker lines, so accumulated markers scrolled a real leg out (false negative); the window now counts tool-call entries. #13 pre-edit-secrets caught bare provider keys only inside a quoted label assignment; added anchored patterns for Anthropic sk-ant, OpenAI sk-proj, fine-grained github_pat_, Google AIza, and JWT eyJ (minimum-length guarded against prose false positives).
#12 the remote-pipe-to-shell block required a shell immediately after the first pipe, so xargs/sudo/tee/env interposition evaded it and the comment falsely claimed xargs was caught; broadened to reach a shell through intermediate segments while leaving shell-OR fallbacks unblocked. #11 (doc only): knowledge/owasp-skills-top10.md claimed pre-bash-destructive blocks persistence commands — it does not; corrected to mark persistence detection as unimplemented/future (the detector itself is deferred to v8). Suite 2004/0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
#18 dep-auditor called 'pip audit' (no such subcommand) so Python CVE detection was a permanent silent no-op; now spawns 'pip-audit' with an argv array. #19 it audited the scanner HOST's env with results mislabelled to the target's requirements.txt; now audits the target via -r and skips cleanly when absent.
#14 the offline npm blocklist was skipped for bare/range/tag installs because it used the (null) parsed-spec version; now re-checks the resolved version before the OSV network path. #15 non-hoisted nested lockfile keys derived the wrong package name (leading-only node_modules/ strip); now strips to the last segment. #48 lockfileVersion-1 nested dependency trees are now walked recursively.
#16/#17 the yarn.lock matcher paired two unassociated whole-file substrings with an unanchored pkg@ (false BLOCK of a legit package) and only matched Yarn Classic quoted versions (Berry known-malware allowed); rewritten as a per-entry parser that associates version to its own entry, anchors the name, and matches both Classic and Berry. Suite 1931/0.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
The BLOCK rule named "Filesystem root destruction" did not block the bare
root target it is named for. The target alternation ended in a shared `\b`,
and a word boundary cannot hold after `/` or `~` at end-of-command.
Measured against the shipped pattern:
bare `/` target NOT blocked `$HOME` target blocked
bare `~` target NOT blocked `/usr` target blocked
`/*` glob target NOT blocked
swapped flags (-fr) NOT blocked
sudo-prefixed NOT blocked
Only targets whose first character is a word character ever satisfied the
assertion, so the rule caught `/etc` but not bare root. These fell through
to WARN (exit 0) — advisory only, command executed.
- Pattern: `\b` moved onto the `$HOME` alternative alone, where it is
meaningful (it ends in a word char, so `$HOMEDIR` is still not
swallowed). Dropped from `/` and `~`. Nothing else changes: `/etc`,
`/home`, `./build` behave exactly as before.
- The old test file encoded this defect as expected behaviour, with a NOTE
claiming the pattern "requires separate flag groups (e.g. -f -r, not -rf
combined)". That diagnosis was wrong — the `/etc` case blocks fine with
merged flags. Comment replaced with the real root cause.
npm test: 1872/1872 green (1865 + 7 new).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TcQyMTQfyrsAapaCMPxTtQ
Session B of the security-fix track. Both sinks trusted untrusted strings at a
filesystem/subprocess boundary; same fix family as F-1.
F-2 (HIGH, arbitrary file write) — scanners/auto-cleaner.mjs: applyFixes() did
resolve(targetPath, f.file) with no containment, then wrote the cleaned content
back. f.file is untrusted (scanned-repo filenames, or a fully attacker-chosen
--findings file), so file: "../../.claude/settings.json" let the cleaner modify
files OUTSIDE the scanned tree. Add a prefix-containment check before grouping:
absPath must equal targetPath or start with targetPath + sep, else the finding
is refused and reported as skipped. (Documented residual gap: prefix containment
does not stop a symlink inside the tree pointing out — noted inline.)
F-3 (HIGH, command injection, pre-confirmation) — hooks/scripts/
pre-install-supply-chain.mjs: inspectNpmPackage ran execSafe(`npm view ${spec}
--json`), a shell string. spec derives from package tokens parsed out of the
scanned Bash command, so a metachar-bearing token reached the shell on
PreToolUse(Bash) — BEFORE the install, so it ran even if the user then denied
the command. Switch to spawnSync('npm', ['view', spec, '--json']) (no shell);
spec is passed as one argv element. The static `npm audit --json` execSafe call
is left as-is (no interpolation).
TDD (repro -> red -> green):
- tests/scanners/auto-cleaner-traversal.test.mjs: a "../secret.txt" finding must
not rewrite the outside file; contained files still get cleaned.
- tests/hooks/supply-chain-injection.test.mjs: `npm install $(>/abs/PWNED)`
(redirect-only $(...) survives normalizeBashExpansion + the whitespace split)
must not create the sentinel.
Closing gates: full node --test suite 1863/0 (was 1860; +3); gitleaks clean;
F-1 regression re-run GREEN (sink still closed).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01V3s6WnubSSrFjAQTLQdVbG
Wave C step C2: surface the cumulative-drift signal from
checkDescriptionDrift() (added in C1) as a separate MEDIUM advisory
with finding category mcp-cumulative-drift. Independent of the existing
per-update drift advisory — a slow-burn rug-pull that keeps each update
below the 10% per-update threshold but cumulatively drifts >=25% from
the sticky baseline now triggers the new advisory without ever crossing
the per-update bar.
The advisory references /security mcp-baseline-reset (added in C3) so
the user knows how to acknowledge a legitimate MCP server upgrade.
CLAUDE.md updates:
- post-mcp-verify hooks-table row mentions per-update + cumulative drift
- mcp-description-cache lib bullet documents baseline schema, history,
cumulative threshold policy key, and LLM_SECURITY_MCP_CACHE_FILE
override.
Tests: 2 new hook tests using LLM_SECURITY_MCP_CACHE_FILE for cache
isolation. Existing 68 still pass; total 70.
Plugin README and root marketplace README updates land in C3 alongside
the new /security mcp-baseline-reset slash command (combined Wave-C
doc update per plan §"Wave C — Touch" list).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds a scope-hopping detector to the npm install gate. When a user
installs `@<scope>/<unscoped>`, the hook now emits a MEDIUM warning
on stderr (exit 0, never blocks) if:
- `<unscoped>` matches a popular npm package (POPULAR_NPM, ~80
names from knowledge/top-packages.json), AND
- `<scope>` is not on NPM_OFFICIAL_SCOPES (built-in 22 entries) or
on policy.json `supply_chain.allowed_scopes`.
Why: an attacker publishing `@evilcorp/lodash` cannot squat the bare
`lodash` name, but they can register an unrelated scope and rely on
typo or copy-paste to trick installs. NPM_OFFICIAL_SCOPES anchors the
known-good scopes (@types, @reduxjs, @nestjs, …) so legitimate
installs stay silent.
Implementation:
- `scanners/lib/supply-chain-data.mjs`: exports POPULAR_NPM,
NPM_OFFICIAL_SCOPES, and `checkScopeHop(name, extraAllowedScopes)` —
pure function, no policy/network dependency, fully unit-testable.
- `knowledge/typosquat-allowlist.json`: mirrors NPM_OFFICIAL_SCOPES as
`npm_official_scopes`. A doc-consistency assertion ensures the two
lists never drift.
- `hooks/scripts/pre-install-supply-chain.mjs`: imports checkScopeHop,
reads `supply_chain.allowed_scopes` from policy, and pushes a
warning before existing compromised/audit checks.
Tests:
- 9 new cases in tests/hooks/pre-install-supply-chain.test.mjs:
TP @evilcorp/lodash, TP @attacker/express, allowlist @types,
allowlist @reduxjs, allowlist @modelcontextprotocol, FP unscoped
name not in top-100, bare unscoped name, policy override, defensive
non-string input, NPM_OFFICIAL_SCOPES <-> typosquat-allowlist.json
consistency.
Adds BLOCK_RULE for the malware-loader pattern:
echo|cat|printf <base64-blob> | base64 -d | <shell>
This is a common RCE delivery shape that bypasses static name-matching
gates by encoding the destructive command as a base64 blob. The new
rule fires only when the final pipe target is a shell interpreter
(bash, sh, zsh, dash, ksh) — base64 decoded into jq or any non-shell
consumer remains allowed.
5 new tests in pre-bash-destructive.test.mjs:
- 3 BLOCK cases (echo|base64|bash, printf|base64|sh, cat|base64|zsh)
- 2 FP probes (base64 -d -> jq passes; base64 -d alone passes)
Closes E9 in critical-review-2026-04-20.md.
The existing CRITICAL pattern in injection-patterns.mjs only fires when
a comment body contains AGENT/AI/HIDDEN markers. Adversaries can drop
the marker and still hide instructions inside <!-- ... --> for any
agent that reads page source. This generalizes the comment scan: every
comment body is HTML-entity-decoded and run through the full
injection rule set. The existing keyword-restricted pattern still
fires (defense-in-depth).
Emits at the strongest tier with category html-comment-injection.
+3 tests (65 → 68).
Refs: Batch B Wave 4 / Step 11 / v7.2.0
SVG containers carry text that is invisible in the rendered image but
fully parsed by an agent reading the source. <desc>, <title>,
<metadata>, and <foreignObject> are all valid surfaces for adversarial
injection.
Adds a per-element extractor inside the existing HTML-tag gate, gated
on /<svg[\s>]/i so it only fires for actual SVG content. Inner text is
HTML-entity-decoded then run through scanForInjection. Emits at the
strongest tier with category svg-element-injection.
+3 tests (62 → 65).
Refs: Batch B Wave 4 / Step 10 / v7.2.0
Adversarial payloads in markdown link title attributes (rendered as
tooltips, parsed by agents) bypassed the existing HTML-content checks
which gated on `<tag>` presence. Pattern: [text](url "title").
Adds linkTitleRegex extraction to the HTML-content block, runs each
captured title through scanForInjection, emits at the strongest tier
encountered with category markdown-link-title-injection.
+3 tests (62 → 62 in post-mcp-verify.test.mjs file, was 59).
Refs: Batch B Wave 4 / Step 9 / v7.2.0
Critical-review §4 E17 finding: pre-v7.2.0 the delegation-after-input
advisory fired only within a 5-call window. Attackers who deliberately
waited 6+ calls before delegating bypassed detection. Window was also
hardcoded — operators couldn't tune it for their environment.
Two coordinated changes:
1. LLM_SECURITY_ESCALATION_WINDOW env var (primary window override)
- parseInt(env) || getPolicyValue('trifecta', 'escalation_window', 5)
- Mirrors the established pattern from
LLM_SECURITY_TRIFECTA_MODE et al.
- Setting env=3 narrows; env=8 expands.
2. Secondary 20-call MEDIUM advisory (slow-burn variant)
- DELEGATION_ESCALATION_WINDOW_MEDIUM = 20 (hardcoded — same value
for all operators; tunable in a future patch if needed)
- checkEscalationAfterInput now returns `tier: 'primary'|'secondary'|null`
- formatEscalationWarning emits a different message for secondary —
mentions "slow-burn", references env-var, distinct from the
primary "DeepMind Category 4" framing
Hook reads max(WINDOW_SIZE, secondary+5) entries to cover the wider
window. Existing duplicate-suppression (`escalation_warning` state
entry) covers both tiers. Audit-trail event captures `tier` field.
Tests: +5 cases in tests/hooks/post-session-guard.test.mjs:
- secondary window catches 9-call distance (slow-burn)
- secondary boundary at exactly 20 calls
- primary regression guard (1-call distance)
- env=3 narrows primary (4-call distance becomes secondary)
- env=8 expands primary (7-call distance stays primary)
Updated existing test "does NOT trigger when input_source is >5 calls
ago" — now requires >20 calls (secondary window catches 6-20).
Suite: 1644 → 1672 (+28 from new tests + extended scope). All green.
CLAUDE.md hooks table updated to document both windows and the env var.
Closes A2 of v7.1.0 critical-review patch (docs/critical-review-2026-04-20.md):
- B4 (severity JSDoc): 4 critical = 93, not 90. Fixed in scanners/lib/severity.mjs:23
and CHANGELOG.md v7.0.0 tier description. The actual computation has always been
93 (70 + log2(5)*10 = 93.22 → round); only the docs were wrong.
- §5.4 co-monotonicity: new sweep test in tests/lib/severity.test.mjs over 15
representative count vectors. Asserts that (verdict, riskBand) agree under the
v7.0.0 contract for every case — catches future drift between riskScore tiers,
verdict cutoffs, and riskBand cutoffs. Includes a B4 anchor test (riskScore
{critical: 4} === 93) so doc/code drift fails loudly.
- B8 (CaMeL claims toned down): post-session-guard.mjs:646 comment block and
CLAUDE.md:184 Defense Philosophy bullet now describe the implementation
honestly — opportunistic byte-matching of truncated output fingerprints
(first 200 bytes, SHA-256/16-hex), not semantic data-flow tracking.
Trivially bypassed by mutation, summarisation, or re-encoding. Inspired by
CaMeL (DeepMind 2025), but not a CaMeL capability-tracking implementation.
Tests: 1495 → 1511 (+16: 15 sweep cases + 1 B4 anchor). All green.
Previously, `LLM_SECURITY_TRIFECTA_MODE=block` only exited 2 when the
detected trifecta was MCP-concentrated (all three legs via the same MCP
server) or involved sensitive-path + exfil. Distributed trifectas —
three legs originating from different tools, with a non-sensitive data
path and a non-sensitive exfiltration sink — were detected and warned
but not blocked. This mismatched the documented semantics of block mode
and gave operators a false sense of enforcement.
Change: remove the `(mcpInfo.concentrated || sensitiveExfil)` AND-gate
in the `TRIFECTA_MODE === 'block'` branch so any detected trifecta
blocks in block mode. Audit event `severity` still differentiates
critical (concentrated / sensitive-exfil) from high (distributed); the
blocked stderr message now explicitly names "Distributed trifecta:
three legs from different sources" when the confidence sub-signals
are absent.
Addresses critical review 2026-04-20 §2 B2 (HIGH) and §9 row 1
("enforces the Rule of Two").
Tests: 1 added (distributed trifecta in block mode now exits 2).
All 1495 tests pass.
The previous ENV regex `/[\\/]\.env\.[a-z]+$/` only matched a single
lowercase segment after `.env`. Multi-segment and mixed-case variants
such as `.env.production.local.backup`, `.env.stage-1.local`, and
`.env.CI.secret` slipped past the hook. Replaced with
`/[\\/]\.env(\.[A-Za-z0-9._-]+)*$/` which matches `.env` plus any
number of dot-separated alphanumeric/dot/hyphen/underscore segments.
`.envrc` (direnv config, no dot separator) is still allowed.
Addresses critical review 2026-04-20 §2 B1 (HIGH).
Tests: 7 added (6 new multi-segment BLOCK cases + 1 .envrc ALLOW).
All 1494 tests pass.
New policy-loader.mjs reads .llm-security/policy.json with deep-merge against
defaults that exactly match existing hardcoded values. Integrated into all 7 hooks:
- pre-prompt-inject-scan: injection.mode (env var still takes precedence)
- post-session-guard: trifecta.mode, window_size, long_horizon_window
- pre-edit-secrets: secrets.additional_patterns
- pre-bash-destructive: destructive.additional_blocked
- pre-write-pathguard: pathguard.additional_protected
- pre-install-supply-chain: supply_chain.additional_blocked_packages
- post-mcp-verify: mcp.volume_threshold_bytes, mcp.trusted_servers
Backward compatible: no policy file = identical behavior to v5.1.0.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
New audit-trail.mjs writes structured events to LLM_SECURITY_AUDIT_LOG path.
Integrated into post-session-guard at 6 warning emission points: trifecta,
escalation-after-input, data flow, volume threshold, slow-burn, behavioral drift.
No-op when env var not set — zero overhead for existing users.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add /ultraresearch-local for structured research combining local codebase
analysis with external knowledge via parallel agent swarms. Produces research
briefs with triangulation, confidence ratings, and source quality assessment.
New command: /ultraresearch-local with modes --quick, --local, --external, --fg.
New agents: research-orchestrator (opus), docs-researcher, community-researcher,
security-researcher, contrarian-researcher, gemini-bridge (all sonnet).
New template: research-brief-template.md.
Integration: --research flag in /ultraplan-local accepts pre-built research
briefs (up to 3), enriches the interview and exploration phases. Planning
orchestrator cross-references brief findings during synthesis.
Design principle: Context Engineering — right information to right agent at
right time. Research briefs are structured artifacts in the pipeline:
ultraresearch → brief → ultraplan --research → plan → ultraexecute.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
