Mcs External Positioning

mcs-external-positioning

Overview

This repo is pre-1.0 (issue #3, open, 0 of 22 boxes checked as of 2026-07-05, re-verify with gh issue view 3 --json body -q .body | grep -c '\- \[ \]') and is actively mutating under a live ship run. Public claims must be evidence-backed today, not aspirational, and must not overstate what a fresh clone (no owner-only private repo, no owner's machine) actually gets. Core principle: rank claims by demonstrated evidence, not by design intent; anything without committed proof is "candidate," not "shipped."

When to use, and when NOT to

Use for: drafting/reviewing public-facing claims (README pitch, launch checklist sign-off, "what makes this different"), v1.0 readiness verdicts, license/attribution audits before release, and structuring (never resolving) the privacy/history-exposure decision.

Do NOT use for:

• Fixing README/ARCHITECTURE drift itself (the numbers, the file paths): that's mcs-docs-of-record.
• Judging whether a technique is a genuine research contribution: that's mcs-research-frontier.
• Executing the actual git-history scrub, force-push, or BFG pass: no skill does this; it is an owner-only action, never automated.

Positioning inventory (ranked by evidence strength, verified 2026-07-02)

1. Acceptance-contract-gated autonomous execution. hooks/acceptance-gate.sh, wired PreToolUse on TaskUpdate (lines 69/75 as of 2026-07-05 via grep -n '"PreToolUse"\|acceptance-gate' settings.json; approximate only, re-grep before citing, since settings.json is under active rewrite), denies (exit 2) completion while unverified criteria remain; selects the contract by mtime, skips stale ones. Took 3 fail-open bugs to reach this state (299a264 → 6bc7c1a → 4bb8bf2 → 605112d). Live proof: today's session verified AC-001..AC-006 with real pytest runs before merging PR #67/#68. Gap: no observed case of it catching a real regression, only "fail-open bug" or "criteria passed". Claim "gates completion," not "catches bugs."
2. Shared-engine multi-theme council on native Workflow substrate. commands/_council-engine.md (1,753 lines) is parameterized by 14 $THEME_ID-style variables; two themes instantiate it: commands/council.md (public) and commands/ece.md (private-repo symlink, undocumented in README/ARCHITECTURE). Backend fallback workflow → teams → sequential is real and recorded per session. Gap: the substrate is Anthropic's, not this repo's differentiator (claim the degradation design, not the tools); say "proven by two themes, one public," not "multiple public themes."
3. Enforced skill-governance pipeline with 3-tier evals. Pre-commit Tier 1 (static) is real and hard-blocking (skill-frontmatter, skill-references, skill-isolation, skill-context-load, skill-commit-msg; skill-token-budget/skill-prose-check warn-only, per .pre-commit-config.yaml). Spec defines Tier 2 (E2E via claude -p) and Tier 3 (LLM-as-judge) at pipeline/specs/SKILL-GOVERNANCE-SPEC.md §7.6, with working scripts (run-evals.sh, judge-skill-quality.sh). Gap: Tiers 2-3 are not wired into CI or pre-commit, zero committed run output. Claim "Tier 1 static enforcement is continuous; Tiers 2-3 are unexercised tooling," not "3-tier enforcement."

Commodity, don't lead with these: the skills/commands/agents collection itself (many public repos do this); the symlink-based installer; project scaffolding commands (/new-python, /new-terraform). Competent execution, not novelty.

Cross-cutting gap for all three ranked claims: zero committed cost or quality telemetry for council sessions. skills/council/registry.json (v2.1) tracks per-skill usage counts only (uses, last_used, sessions), no token cost, latency, or quality score. (grep -n cost skills/council/registry.json matches only a skill named cost-analysis, not a telemetry field.) Any "council sessions are cheap/fast/high-quality" claim is unmeasured: label it "candidate, no telemetry."

v1.0 readiness ledger

Full issue-#3 checklist with live per-item verification: references/v1-readiness-ledger.md. Headline blockers as of 2026-07-02:

• Docs accuracy: README's agent/skill/command counts are drifted from reality. Fix routes to mcs-docs-of-record; this skill only flags it as a v1.0 gate.
• Third-party notices: THIRD_PARTY_NOTICES.md lists 2 dead directories (skills/terraform-skill/, skills/tdd/) and misses 2 real adaptations (agentic-council v1.2.0, steipete-derived skill-audit tooling). See Licensing below.
• CI/security gates: validate.yml (ubuntu+macos) + secret-scan (gitleaks) are required status checks, but enforce_admins: false and required_approving_review_count: 0; CODEOWNERS exists but isn't enforced. Vulnerability alerts and Dependabot security updates are enabled; Dependabot version updates cover only GitHub Actions, not mcp/openrouter's Python deps.
• Onboarding validation: issue #3 requires 3+ onboarding feedback reports via the issue template; zero exist.
• Private-repo coupling (unstated blocker): 7 symlinks point into $HOME/Development/my-claude-setup-private. install.sh's existence guards silently skip broken symlinks, so a fresh clone's install is functionally fine; gap is cosmetic. Tracked by open issue #65 (US-007, queued).
• Machine-local paths in settings.json (unstated blocker): 2 hardcoded /Users/danielsong/... paths remain (OpenRouter MCP command, research-toolkit-local marketplace), opt-in only but non-functional for anyone else. No issue owns this gap yet.
• README prerequisites incomplete: python3 is a hard install.sh dependency, absent from README's Prerequisites list.
• Newly resolved, don't carry forward: prior "fresh clone gets broken telemetry hooks" is fixed as of PR #67 (2026-07-02): hooks/telemetry-dispatch.sh is fail-soft, wired to all 5 events, zero private-repo strings remain in settings.json.

Privacy and history exposure: decision framing only, no action

Real personal content is exposed on two different surfaces (corrected 2026-07-06): SOC/security and ECE content IS reachable from pushed remote refs (6 branches); the named individual's resume was never on main and survives only as one orphaned commit object on a fork, a much smaller exposure. Full options ledger, exact refs, and consequences: references/privacy-history-exposure.md. This skill states options; it never recommends or executes one. Deleting remote refs or rewriting pushed history (BFG/force-push) is irreversible-in-practice and owner-only. Do not perform it, do not suggest a default, and do not treat commit 69cdb56's "scrubbing history is a separate decision" note as consent.

Licensing and attribution

• Root license: MIT, copyright 2025 Daniel Song (LICENSE:1-3).
• THIRD_PARTY_NOTICES.md has 3 entries: skills/terraform-skill/ (stale, directory deleted), skills/vercel-react-best-practices/ (accurate stub), skills/tdd/ (stale, directory deleted; only commands/tdd.md redirect stub remains).
• Missing attribution: the agentic-council v1.2.0 orchestration adoption (commit 9cee199) and the steipete/agent-scripts skill-cleaner-derived audit tooling. Both real, neither in THIRD_PARTY_NOTICES.md or docs/upstream-sources.md (the latter is documented only in prose inside pipeline/scripts/README.md's "Origin" section, uncross-referenced).
• A maintenance procedure exists (docs/upstream-sources.md) but nothing enforces it: no hook checks notice entries against actual directory existence, which is why it drifted silently twice.

Reproducibility standard for any public claim

A public claim needs a fresh-clone quickstart that actually works, or it doesn't ship.

• ./install.sh --dry-run --preset skills runs clean; private-repo symlinks are silently skipped by design ([[ -d ]] on a broken symlink is false), verified by reading install.sh:83-149, not by an executed absent-private-repo CI run. Label this "verified by code inspection," not "verified by CI."
• No CI job actually simulates the absent-private-repo state; the ubuntu/macos matrix runs inside this same checkout.
• Benchmark/token numbers: never present an estimate as a measurement. The only committed telemetry today is skills/council/registry.json usage counts (18 total uses as of 2026-07-02). No cost, latency, or token data is committed anywhere. Any number not traceable to a git show-able file is an estimate and must be labeled one.

Gotchas

• WRONG: "First system with acceptance-contract-gated autonomous execution." RIGHT: state what's verified and stop; novelty claims belong to mcs-research-frontier.
• WRONG: assuming THIRD_PARTY_NOTICES.md is complete because it's well-formatted. Check both directions: every listed path exists, every real adaptation has an entry.
• WRONG: treating registry.json's usage counts as quality or cost evidence: no cost/quality/outcome field exists.
• WRONG: assuming PR review is required for main. enforce_admins: false means the owner can and does push directly.
• WRONG: reading "install.sh is stable" from CI alone: CI smoke-tests skills/core only, never full.
• WRONG: quoting README's agent/skill counts in a launch post: pull live counts at write time (mcs-docs-of-record owns the fix).
• SILENT FAILURE: settings.json's 2 hardcoded personal paths only surface as a broken MCP entry for a --with-settings opt-in, no install-time error.
• Everything here drifts fast under the live ship run. Re-run Provenance commands before repeating any number publicly.

Provenance and maintenance

Last verified: 2026-07-05, against a live, actively-mutating repo (PRs #67-#70 merged since original authoring, US-002 model tier alias / US-003 permissions rewrite / US-004 OpenRouter model ID refresh / settings schema guard work). Re-verify every fact below before reusing it.