@massu/core 1.15.4 → 1.15.6

package/CHANGELOG.md

@@ -10,6 +10,89 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 > redistributed under the **Apache License 2.0**. Full license + NOTICE:
 > `packages/core/assets/embedder/MODEL-LICENSE`.
 
+## [1.15.6] - 2026-07-14
+
+### Fixed — the golden path's Deep Security Audit phase was missing entirely
+
+`references/phase-3.5-security-audit.md` — a full adversarial security audit loop over every file
+changed in a golden-path run (parallel red-team agents, iterating until the audit converges to ZERO
+findings). Its own text: *"This phase is NEVER skipped. Security is non-negotiable regardless of
+change size, type, or scope."*
+
+    0.6.1    SHIPS phase-3.5-security-audit.md
+    1.0.0    does NOT ship it
+    1.15.5   does NOT ship it
+
+**It exists in no commit in massu's history** — no deletion commit, no changelog entry, no trace.
+From v1.0.0 through v1.15.5, every user's golden path ran with **no security audit phase at all**.
+The only surviving copy was in a consumer repo that had committed its installed commands months ago.
+
+Restored, wired as **Phase 3.5** (after simplification, before pre-commit, exactly where its own
+spec says it belongs), and added to the reference table.
+
+**Why every guard missed it, and what changed:** the reference-link check asks *"does every
+reference file have a link?"* — and a file DELETED OUTRIGHT has no link to be missing. A check that
+only sees BROKEN things is structurally blind to ABSENT ones. The anti-vacuity suite now asserts,
+per subsystem, that the file **EXISTS** *and* is **LINKED** *and* is **INVOKED** — a conjunction no
+deletion can satisfy. Proven by attack: deleting the spec file goes RED; deleting the phase while
+keeping the file goes RED.
+
+This is the fifth quality subsystem found missing from the golden path in one day, and the only one
+deleted so completely it left no evidence anywhere in the repository.
+
+## [1.15.5] - 2026-07-14
+
+### Fixed — CRITICAL: `install-commands` could destroy your customizations
+
+- **The installer recorded a hash for a file it did not write.** On meeting a command file with no
+  manifest entry, it kept the file (correct) and then recorded THAT FILE'S hash into the manifest —
+  a manifest whose entire meaning is "this is the hash of what massu installed". On the next run the
+  safe-upgrade path read that back, concluded the file was untouched, and **overwrote it**.
+  **Run 1 armed it; run 2 destroyed the work.** Measured against a real repo: 36 files of local
+  customization deleted on the second install, reported as success.
+
+  It now records **nothing** when it cannot prove authorship. The ambiguity is re-detected on every
+  run and the file is kept on every run. A customized file with no provenance now FREEZES: safe, but
+  stale. Upstream changes reach it via an explicit `rm <file> && massu install-commands`, or a future
+  three-way merge. Between "stale" and "your work is silently deleted", we choose stale.
+
+- **A test was protecting the bug.** `MANIFEST-06` asserted the fail-open as correct behaviour
+  ("Manifest seeded with EXISTING hash for kept"). The test agreed with the code because both were
+  written from the same wrong assumption. Assertion inverted.
+
+- **`scripts/quarantine-poisoned-manifest.mjs`** — repairs a manifest that already contains such a
+  hash. It does not re-derive what massu "would have written" (the installer resolves variants and
+  renders templates, so the installed bytes are a function of the repo's config); it runs the REAL
+  installer against a scratch copy and drops any entry that authorises an overwrite. Dry run by
+  default; backs the manifest up before writing.
+
+### Fixed — the golden path was missing the machinery that checks the work
+
+`8fff05d`, whose message announced only additions, silently **deleted** three quality subsystems.
+The spec files kept shipping; nothing invoked them. Restored and re-wired:
+
+- **Sprint contracts (2A.5)** — agree definition-of-done per item BEFORE implementing it.
+- **QA evaluator (2C)** — adversarial acceptance testing AGAINST those contracts.
+- **VR-VISUAL** — weighted 4-dimension scoring (threshold >= 3.0).
+- **Gap analyzer's 7th category** — sprint-contract compliance. Without it the loop could report
+  zero gaps while the work missed the definition of done it was built against.
+
+And a fourth that was **never wired at all**: **Phase 5.5 — Production Verification**, shipped in
+v0.6.0 and invoked by nothing for its entire life. Its own first line reads *"A feature is NOT
+complete until it is verified working in production with real data. 'Deployed' and 'working' are two
+completely different things."*
+
+- **A new gate that deletion cannot satisfy.** The previous check asked only "does every reference
+  file have a link?" — so deleting the spec files made it pass. The new one asserts, per subsystem,
+  that the file EXISTS **and** is LINKED **and** is INVOKED by name. Proven by attacking all three.
+
+### Fixed
+- `massu-pattern-scanner.sh` Check 40 used a relative path and silently failed from any directory but
+  the repo root — including `packages/core/`, which is where npm runs `prepublishOnly`. The
+  pre-publish check that runs it had therefore never worked.
+- `diff-commands-vs-docs.sh` derived its repo root via `git rev-parse || pwd`; outside a git tree the
+  fallback turned "I cannot find the repo" into a confident wrong answer.
+
 ## [1.15.4] - 2026-07-13
 
 ### Security

package/commands/massu-golden-path/references/phase-3.5-security-audit.md

@@ -0,0 +1,108 @@
+# Phase 3.5: Deep Security Audit
+
+> Reference doc for `/massu-golden-path`. Return to main file for overview.
+
+```
+[GOLDEN PATH -- PHASE 3.5: DEEP SECURITY AUDIT]
+```
+
+## Purpose
+
+Run a full adversarial security audit loop against ALL files changed in this golden path run. This is a deep, iterative audit with parallel red-team agents that converges to zero findings. It runs AFTER simplification (Phase 3) so the audit targets the final, cleaned-up code -- and BEFORE pre-commit verification (Phase 4) so all security fixes are included in the verification gates.
+
+**This phase is NEVER skipped.** Security is non-negotiable regardless of change size, type, or scope.
+
+---
+
+## 3.5.1 Determine Audit Scope
+
+Collect ALL files changed during this golden path run:
+
+```bash
+git diff --name-only HEAD
+```
+
+If files were already committed in earlier phases, also include:
+```bash
+git diff --name-only main...HEAD
+```
+
+The audit scope is the union of all changed files. Do NOT narrow scope -- every changed file gets audited.
+
+**Output:**
+```
+SECURITY AUDIT SCOPE:
+  Files: [N]
+  [list of files]
+```
+
+---
+
+## 3.5.2 Execute Deep Security Audit
+
+Run the full security audit protocol against the scoped files:
+
+1. **Launch 2-4 parallel adversarial reviewer agents** adapted to the codebase area:
+   - Backend/API code: 4 agents (Injection, Network/Leakage, DoS/Resources, Red Team Bypass)
+   - Frontend code: 3 agents (XSS/Injection, Auth/Data Exposure, Input Validation/Logic)
+   - Infrastructure/config: 2 agents (Secrets/Config, Dependencies/Supply Chain)
+
+2. **Consolidate findings** -- deduplicate across agents, take higher severity on disagreements
+
+3. **Fix ALL findings** -- CRITICAL first, then HIGH, MEDIUM, LOW. INFO documented only.
+
+4. **Verify fixes** -- import checks, input validation tests, functionality preserved
+
+5. **Loop until zero findings** -- max 5 iterations, escalate to user if still failing after 5
+
+---
+
+## 3.5.3 Attack Vector Coverage
+
+Every audit iteration MUST verify the complete attack vector checklist:
+
+### Universal
+- Hardcoded secrets / API keys / credentials
+- Error messages leaking internal details
+- Dependency vulnerabilities
+- Input validation on ALL external boundaries
+
+### Backend / API
+- SQL injection, command injection, path traversal
+- SSRF, authentication bypass, authorization bypass
+- DoS via unbounded inputs, memory leaks, race conditions
+- Response validation, type confusion
+
+### Frontend
+- XSS, open redirects, sensitive data in client state
+- CSRF, client-side auth bypass
+
+### LLM / AI Specific
+- Prompt injection, model output trust
+- Tool argument injection, vision/multimodal injection
+
+---
+
+## 3.5.4 Completion Gate
+
+The phase completes ONLY when the audit loop achieves a clean pass with zero findings.
+
+```
+SECURITY_AUDIT_GATE: PASS
+  Iterations: [N]
+  Total findings fixed: [N]
+  Breakdown: [X] CRITICAL, [X] HIGH, [X] MEDIUM, [X] LOW fixed
+  Clean pass: Iteration [N]
+```
+
+**Do NOT proceed to Phase 4 until SECURITY_AUDIT_GATE = PASS.**
+
+---
+
+## Rules
+
+1. **NEVER skip this phase** -- not for small changes, not for docs, not for config
+2. **NEVER proceed with findings unfixed** -- zero means zero
+3. **ALL severity levels get fixed** -- CRITICAL through LOW
+4. **No commit prompt** -- unlike standalone security audit commands, do NOT offer to commit here (Phase 4 handles commits)
+5. **Findings feed Phase 4** -- security fixes are verified by Phase 4's type check, build, lint, and secrets gates automatically

package/commands/massu-golden-path.md

@@ -106,15 +106,25 @@ Read `references/phase-1-plan-creation.md` for full details.
 
 Read `references/phase-2-implementation.md` for full details.
 
-**Summary**: Seven sub-phases:
+**Summary**: Eight sub-phases:
 - **2A**: Extract plan items into tracking table, initialize session state
+- **2A.5**: **Sprint contracts** — negotiate definition-of-done per plan item BEFORE implementing it (scope boundary, acceptance criteria, VR-* mapping). See `references/sprint-contract-protocol.md`
 - **2B**: Implementation loop (pre-check, execute, guardrail, verify, update per item)
-- **2C**: Multi-perspective review (3 parallel agents: security, architecture, quality)
-- **2D**: Verification audit loop (subagent, circuit breaker CR-37, max 10 iterations)
+- **2C**: Multi-perspective review (3 parallel agents: security, architecture, quality) + **QA evaluator** (conditional, UI plans only — adversarial Playwright-based acceptance testing **against the sprint contracts agreed in 2A.5**). See `references/qa-evaluator-spec.md`
+- **2D**: Verification audit loop (subagent, circuit breaker CR-37, sprint-contract verification, max 10 iterations)
 - **2E**: Post-build reflection + memory persist (CR-38)
 - **2F**: Documentation sync (if user-facing features)
 - **2G**: Browser verification & fix loop (auto-triggers if UI files changed, Playwright MCP)
 
+> **2A.5 and 2C are load-bearing, and they were missing.** Commit `8fff05d` — whose message
+> announced only ADDITIONS ("CR-45 zero-tolerance mandate") — silently deleted the sprint
+> contracts and the QA evaluator from this file. Every golden-path run since has implemented
+> against no agreed definition of done, and accepted its own work without an adversarial check.
+> The spec files never stopped shipping; nothing invoked them. Restored 2026-07-14.
+>
+> The order matters: **2A.5 agrees what "done" means BEFORE 2B builds it, so 2C has something
+> to judge against.** A QA evaluator with no contract is grading an exam with no answer key.
+
 **Gate**: APPROVAL POINT #2: NEW PATTERN (if needed, any sub-phase)
 
 ---
@@ -123,7 +133,12 @@ Read `references/phase-2-implementation.md` for full details.
 
 Read `references/phase-2.5-gap-analyzer.md` for full details.
 
-**Summary**: After implementation completes, run a continuous gap and enhancement analysis loop. A subagent analyzes all changed files across 6 categories (functional gaps, UX gaps, data integrity, security, pattern compliance, enhancements). Every gap/enhancement found is fixed immediately. The loop re-runs until a full pass discovers ZERO gaps. Max 10 iterations. Skippable only for documentation-only changes or explicit user request.
+**Summary**: After implementation completes, run a continuous gap and enhancement analysis loop. A subagent analyzes all changed files across **7 categories** (functional gaps, UX gaps, data integrity, security, pattern compliance, enhancements, **sprint-contract compliance**). **VR-VISUAL uses weighted 4-dimension scoring (threshold >= 3.0)** — see `references/vr-visual-calibration.md` for the 5/3/1 calibration examples. Every gap/enhancement found is fixed immediately. The loop re-runs until a full pass discovers ZERO gaps. Max 10 iterations. Skippable only for documentation-only changes or explicit user request.
+
+> **The 7th category and VR-VISUAL's scoring were deleted by `8fff05d` and restored 2026-07-14.**
+> Without the sprint-contract category, the loop could report ZERO gaps while the work missed the
+> definition of done it was built against. Without the weighted scoring, VR-VISUAL was a
+> pass/fail vibe check with a calibration file nobody read.
 
 ---
 
@@ -135,6 +150,35 @@ Read `references/phase-3-simplify.md` for full details.
 
 ---
 
+## PHASE 3.5: DEEP SECURITY AUDIT
+
+Read `references/phase-3.5-security-audit.md` for full details.
+
+**Summary**: A full adversarial security audit loop against EVERY file changed in this run —
+parallel red-team agents, iterating until the audit converges to ZERO findings. It runs AFTER
+simplification (so it audits the final, cleaned-up code) and BEFORE pre-commit verification (so
+every security fix is caught by the Phase 4 gates).
+
+**THIS PHASE IS NEVER SKIPPED.** Security is non-negotiable regardless of the size, type, or scope
+of the change. Do not narrow the audit scope: every changed file gets audited.
+
+> **This phase SHIPPED in v0.6.1 and had vanished by v1.0.0.** It exists in no commit in this
+> repository's history — the only surviving copy on the machine was in a consumer repo that had
+> committed its installed commands back in April. So for every version from v1.0.0 to v1.15.5,
+> every user's golden path ran with **no security audit phase at all**, and nothing anywhere said
+> so.
+>
+> It was found by accident, on 2026-07-14, when a `git status` showed the file being deleted from
+> that consumer during a reinstall. **The reference-link guard could never have found it:** that
+> guard asks "does every reference file have a link?", and a file that has been deleted outright
+> has no link to be missing. A check that only sees *broken* things is blind to *absent* ones —
+> which is the same failure as a scanner that reports CLEAN because it scanned zero files.
+>
+> Restored and wired 2026-07-14. Guarded by `golden-path-quality-subsystems-wired.test.ts`, which
+> asserts the file EXISTS, is LINKED, and is INVOKED — a conjunction no deletion can satisfy.
+
+---
+
 ## PHASE 4: PRE-COMMIT VERIFICATION
 
 Read `references/phase-4-commit.md` for full details.
@@ -155,6 +199,30 @@ Read `references/phase-5-push.md` for full details.
 
 ---
 
+## PHASE 5.5: PRODUCTION VERIFICATION
+
+Read `references/phase-5.5-production-verify.md` for full details.
+
+**Summary**: After push and CI success, verify the deployment actually LANDED and the feature
+actually WORKS — against production, with real data. Confirm the deploy reached READY, read the
+runtime logs for startup errors, then exercise the feature itself and read back what the system
+returns. Skippable ONLY when the change has no deployed surface (docs-only, tooling-only).
+
+**Core principle**: *A feature is NOT complete until it is verified working in production with real
+data. "Deployed" and "working" are two completely different things.*
+
+> **This phase was written, shipped in v0.6.0 — and NEVER WIRED IN.** It has been sitting in
+> `references/` for its entire life with nothing invoking it: a `git log -S` over this file finds
+> no commit that ever linked it. It was not deleted; it was never enabled. That is its own failure
+> class ("hunt the built-and-never-enabled"), and it is the more embarrassing one, because the
+> phase that would have caught "we shipped it but never checked it worked" was itself shipped and
+> never checked. Wired 2026-07-14.
+>
+> It goes AFTER the push (you cannot verify a deployment that has not happened) and BEFORE
+> completion (you may not call it done until the world agrees it is).
+
+---
+
 ## PHASE 6: COMPLETION
 
 Read `references/phase-6-completion.md` for full details.
@@ -171,12 +239,17 @@ This skill is a folder. The following files are available for reference:
 |------|---------|-----------|
 | `references/phase-0-requirements.md` | Requirements interview, ambiguity detection, 10-dimension coverage map | Starting a new implementation from a task description |
 | `references/phase-1-plan-creation.md` | Config/schema reality check, blast radius analysis, plan generation, audit loop | Writing or auditing a plan |
-| `references/phase-2-implementation.md` | Item loop, multi-perspective review, verification audit, browser testing | Executing implementation; any Phase 2 sub-phase |
-| `references/phase-2.5-gap-analyzer.md` | Gap/enhancement analysis loop, 6 categories, fix-and-repass until zero | After implementation, before simplification |
+| `references/phase-2-implementation.md` | Item loop, sprint contracts, multi-perspective review, QA evaluator, verification audit, browser testing | Executing implementation; any Phase 2 sub-phase |
+| `references/sprint-contract-protocol.md` | Sprint contract template, quality bar, negotiation rules, skip conditions | Phase 2A.5 — agreeing definition-of-done before implementing |
+| `references/qa-evaluator-spec.md` | Adversarial QA evaluator: 4 dimensions, anti-leniency rules, known failure patterns | Phase 2C — QA evaluation (UI plans only) |
+| `references/vr-visual-calibration.md` | Score 5/3/1 calibration examples for VR-VISUAL weighted dimensions | Calibrating VR-VISUAL scoring in Phase 2.5 |
+| `references/phase-2.5-gap-analyzer.md` | Gap/enhancement analysis loop, 7 categories (incl. sprint-contract compliance), fix-and-repass until zero | After implementation, before simplification |
 | `references/phase-3-simplify.md` | Pattern scanner fast gate, dead code detection, parallel semantic review agents | Running simplification after implementation |
+| `references/phase-3.5-security-audit.md` | Adversarial red-team security audit loop over every changed file, iterating to ZERO findings | Phase 3.5 — after simplification, before pre-commit. NEVER skipped |
 | `references/phase-4-commit.md` | Auto-verification gates, quality scoring, commit format | Preparing a commit |
 | `references/phase-5-push.md` | Pre-flight, 4-tier push verification, regression detection | Preparing to push to remote |
-| `references/phase-6-completion.md` | Final report, plan status update, auto-learning | After push; completing the golden path |
+| `references/phase-5.5-production-verify.md` | Deployment landed, runtime logs clean, feature exercised against production with real data | Phase 5.5 — after push, before calling it done |
+| `references/phase-6-completion.md` | Final report, plan status update, auto-learning | After production verification; completing the golden path |
 | `references/approval-points.md` | Exact format and options for all 4 approval points (5 with --competitive: Plan, New Pattern, Winner Selection, Commit, Push) | Presenting any approval gate to the user |
 | `references/competitive-mode.md` | Competitive mode protocol: agent spawning, scoring, winner selection | Using --competitive flag |
 | `references/error-handling.md` | Abort handling, non-recoverable errors, post-compaction re-verification, competitive mode errors | On user abort, blocker error, or after context compaction |

package/dist/cli.js

@@ -4357,9 +4357,9 @@ function syncDirectory(sourceDir, targetDir, framework, manifest, manifestKeyPre
         continue;
       }
       if (lastInstalledHash === void 0) {
-        manifest.entries[manifestKey] = existingHash;
         process.stderr.write(
-          `First-install heuristic: keeping existing ${targetPath} (differs from upstream).
+          `First-install: keeping existing ${targetPath} (differs from upstream, and massu has no record of writing it).
+  It will be kept on every future install until you resolve it explicitly.
   To accept upstream: rm ${targetPath} && npx massu install-commands
 `
         );

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@massu/core",
-  "version": "1.15.4",
+  "version": "1.15.6",
   "type": "module",
   "description": "AI Engineering Governance MCP Server - Session memory, knowledge system, feature registry, code intelligence, rule enforcement, tiered tooling (12 free / 73 total), 56 workflow commands, 11 agents, 20+ patterns",
   "main": "./dist/cli.js",