Claude of Alexandria · Claude Code mod

README

Claude of Alexandria Plugin

Rigorous analytical skills for biblical study and teaching preparation.

Model policy

Sonnet is the default for retrieval, smoke checks, and deep-study stages. Haiku is not supported for any study or evaluation. A caller may explicitly select Opus when a study warrants stronger reasoning; existing Opus choices stay intact. Context-preserving skills retain a Sonnet or Opus caller. If the caller is on Haiku, switch to Sonnet or Opus before starting the study. Do not delegate to a built-in agent that would select Haiku.

The deep-study request accepts optional model: "sonnet" | "opus"; omission means sonnet, and other values are rejected before any stage launches. The selected model applies to every stage, including repair and re-verification. For ordinary agent calls, use the explicit model: opus override when escalation is justified.

This policy governs plugin-controlled selections. The plugin also ships one small guard (hooks/guard.ts, Claude Code 2.1.287 or later) for the two ways a session could still put its agents on Haiku:

a coa: agent spawned with model: haiku runs on Sonnet instead;
/coa:deep-study refuses to start while CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL names Haiku, and says to unset it or set it to sonnet or opus.

Other forced model settings or organizational substitutions can still override plugin defaults. Evaluation traces reject observed Haiku execution rather than certifying such a run. Installed copies receive changes through plugin updates.

What runs on your machine

The guard is the plugin's only code. It runs inside Claude Code as a mod, reads one environment variable (CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL), and touches no files, network or processes. claude plugin validate plugins/claude-of-alexandria lists exactly what it does. Where Anthropic or an organisation switches mods off, the guard does not load and the skills work as before. Optional extras that run more code live in the separate coa-desk plugin.

Architecture

The MCP server runs on Cloudflare Workers + D1 (edge SQLite). Skills call MCP tools automatically to retrieve linguistic data — morphology, discourse features, vocabulary frequencies, paragraph markers — with no local installation required.

COA uses separate execution paths for separate jobs:

Path	Use
Context-aware skills	Preserve the user's current request and await specialist results
`deep-study` workflow	Repeatable, bounded orchestration with structured stage contracts
`study-team`	Optional attended discussion between three in-process teammates

Some skills delegate specialized work to sub-agents:

study-evaluator (Sonnet)
    └── biblical-scholar (Sonnet)
            └── data-retriever (Sonnet)

The data-retriever agent (Sonnet) handles all MCP data gathering and compression. Higher-level agents and skills spawn it automatically, keeping token usage low and data handling centralized.

MCP Server

The reference server exposes 27 tools. Skills call these automatically. V4 requires native JSON arrays, explicit modes for variant tools, and cursor iteration through page.next_cursor.

Core linguistic tools:

Tool	Queries	Coverage
`query_morphology`	Word-level morphological parsing	Both
`query_discourse_features`	Levinsohn NT discourse features	NT
`query_paragraph_breaks`	Masoretic petuchah/setumah markers	OT
`query_vocabulary`	Lemma frequencies, thematic keywords, clustering	Both
`query_ot_quotes`	OT quotations and allusions in the NT	NT
`query_themes_for_lemmas`	Resolve morphology lemmas to vocabulary theme names	Both
`query_lemmas`	Cross-book lemma distribution	Both
`query_theme_distribution`	Cross-book distribution of a thematic keyword group	Both
`query_syntax`	OpenText clause-level semantic role annotations	NT
`query_variants`	Textual variant edition comparison across 9 editions	NT
`query_lexicon`	Strong's-based word definitions (TBESH/TBESG)	Both
`list_books`	Available books, metadata, and thematic keyword groups	Both

Entity and context tools:

Tool	Queries	Coverage
`query_cross_references`	Editorial tradition cross-references between verses	Both
`trace_cross_reference_path`	Bounded path search through the cross-reference graph	Both
`query_places`	Geographic locations with coordinates	Both
`query_people`	Named individuals with cross-canonical appearances	Both
`query_events`	Timeline events with participants and locations	Both
`query_person_network`	Family relationships and co-appearances	Both
`query_speakers`	Speaker attribution with quotation type	Both
`check_versification`	Hebrew-English verse numbering differences	OT
`query_ot_structure`	Verse-edge syntax, participant, and speech boundary context	OT

Bible text and commentary tools:

Tool	Queries	Coverage
`bible_lookup`	Verse text in 6 translations	Both
`commentary_lookup`	Commentary entries from 6 commentaries	Both
`parallel_text`	Compare verse text across multiple translations	Both
`confessional_lookup`	Creeds and confessions (4 query modes)	—
`liturgical_lookup`	Church-year season → passages + themes, passage → season(s) lookup	—
`query_controversies`	Contested topics by topic or passage	—

Tech stack: TypeScript, Cloudflare Workers, D1 (edge SQLite), MCP SDK (HTTP transport). No local runtime needed.

Available Skills

exegetical-notes

Produces structured, context-neutral exegetical analysis of a biblical passage. Data-grounded. Always English. Saves to file as a reusable reference document.

Runs pericope check first - Verifies boundaries before analysis
Every claim sourced - Bundled data or web-verified scholarly sources
Training knowledge supplements - Never substitutes for data

Usage:

/coa:exegetical-notes Romans 8:1-11

consult-biblical-scholar

Answers questions about a biblical passage's meaning, validates analogies or ideas against the text, and provides cross-references with scholarly evidence. Works with or without an explicit passage anchor.

Confidence tiering - Every answer declares its evidence basis
MCP data before answering - No training-data-only verdicts
Formal verdict for analogy questions - Structured evaluation, not vibes

Usage:

/coa:consult-biblical-scholar Is "rest" in Hebrews 4 about salvation or sanctification?

argument-flow

Maps the logical argument of a biblical passage using discourse markers and morphological data. Produces a connective-anchored proposition chain showing how clauses relate to each other.

Conjunction-first analysis - MCP morphology data shapes the argument, not training memory
Genre-aware - Different structural methods for epistles, narrative, poetry
Numbered proposition chain - Every response includes one, no exceptions

Usage:

/coa:argument-flow Phil 2:1-4

pericope-delimitation

Validates whether a biblical passage constitutes a coherent discourse unit. Recommends boundary extensions or contractions based on linguistic evidence from bundled data.

Data-grounded boundaries - Never from commentary tradition alone
Inline response - Quick consultation before deeper study
Checks actual discourse features - Levinsohn markers, not familiarity with "famous passages"

Usage:

/coa:pericope-delimitation Romans 8:1-17

biblical-segmentation

Divides biblical books into coherent teaching units with integrity safeguards.

Refuses impossible divisions - Won't pretend Philemon can be 12 sessions
Presents multiple frameworks - Because interpretation isn't a dictatorship
Validates against ancient markers - Masoretic markers, Levinsohn discourse features
Handles contested books - Isaiah's unity debate gets frameworks, not false consensus

Coverage: All 66 canonical books

Usage:

/coa:biblical-segmentation Divide Romans into 12 sessions for a sermon series.

passage-glossary

Builds a passage reader with a deduplicated, MCP-grounded lemma glossary.

/coa:passage-glossary Philippians 2:5-11

study-team

Runs an optional three-perspective review of a completed deep-study report. It requires an attended interactive session with agent teams already enabled and supports in-process teammates only. It does not change settings or fall back to ordinary sub-agents.

/coa:study-team

Available Workflow

deep-study

Runs the fixed evidence → boundary → parallel analysis → verification → optional one-time repair → synthesis graph. Input must include the passage, research question, output language, context, and constraints. At most two calls run concurrently and eight run in total. The workflow binds each call to one of four leaf agent types. Retrieval and verification can use only the plugin's read-only MCP tools; analysis and synthesis are tool-free. None can spawn a descendant.

/coa:deep-study passage=Philippians 2:5-11; research question=How does the passage present Christ's humiliation and exaltation?; output language=English; context=Teaching outline; constraints=Distinguish evidence from interpretation

Available Agents

Sub-agents are selected by skills or workflows automatically. They are not invoked directly.

Agent	Model	Role
`data-retriever`	Sonnet	Fetches MCP data, compresses into structured summaries with testament-aware routing
`biblical-scholar`	Sonnet	Scholarly analysis (ANALYZE, VALIDATE, TRACE modes) with confidence tiers
`study-evaluator`	Sonnet	Evaluates study materials against exegetical standards with drift classification
`pericope-delimitation`	Sonnet	Boundary validation with structured verdicts
`argument-flow`	Sonnet	Logical structure mapping with proposition chains
`smoke-test`	Sonnet	Pipeline verification
`deep-study-retrieval`	Sonnet	Read-only MCP evidence retrieval and provenance capture
`deep-study-analysis`	Sonnet	Tool-free boundary, discourse, interpretation, and repair analysis
`deep-study-verification`	Sonnet	Read-only MCP claim and reference verification
`deep-study-synthesis`	Sonnet	Tool-free final report rendering
`study-team-textual-evidence`	Sonnet	Leaf teammate for textual structure and evidence integrity
`study-team-interpretation`	Sonnet	Leaf teammate for the evidence-to-interpretation chain
`study-team-critique`	Sonnet	Leaf teammate for adversarial confidence and synthesis review

Development

This plugin is built using Test-Driven Development. Model-free CI validates every shipped definition, reference, generated workflow, and canonical RED/GREEN configuration. Maintainer-run promptfoo acceptance uses the Claude Agent SDK with live MCP data; paid model evaluations do not run in PR CI.

See CLAUDE.md for development guidelines and the Librarian's instructions.

Hermeneutical Framework

All skills follow historical-grammatical method with theological guardrails:

Anti-moralism mandate - No "try harder" applications without gospel
Christ-centeredness - Traces redemptive-historical arc
Context primacy - Respects discourse units, pericopes, books
Genre governance - Right method for the text type
Covenantal awareness - Attends to progressive revelation

License

GNU General Public License v3.0 - See repository root for details.

Source 1 files

hooks/guard.ts 27 lines

1// The only code the coa plugin runs: a guard that keeps coa agents off Haiku,
2// which this project does not support (docs/adr/0002, 2026-09-30 amendment;
3// docs/adr/0006). Everything else in coa is Markdown and a remote MCP server.
4//
5// It reads one environment variable and logs one line. It reads no files,
6// makes no network requests and starts no processes.
7
8const HAIKU = /haiku/i
9
10export function register(on) {
11  // An Agent-tool spawn can be moved to Sonnet. A workflow agent's model cannot
12  // be changed from a hook, so a Haiku workflow override is refused instead.
13  on('agent.spawn', { subagentType: /^coa:/ }, async ($, e, next) => {
14    if (e.workflow) {
15      const forced = await $.env.get('CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL')
16      if (!HAIKU.test(String(forced ?? '')) && !HAIKU.test(String(e.model ?? ''))) return next(e)
17      return {
18        deny: 'coa workflow agents do not run on Haiku. Unset CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL or set it to sonnet or opus, then run the workflow again.',
19      }
20    }
21    if (!HAIKU.test(String(e.model ?? ''))) return next(e)
22    await $.ui.log(`${e.subagentType} asked for ${e.model}; coa agents run on Sonnet instead.`)
23    return next({ ...e, model: 'sonnet' })
24    // Fail open: a fault in this guard must never stop coa agents from starting.
25  }).catch(async ($, e, next) => next(e))
26}
27

Install

$claude plugin marketplace add davebream/claude-of-alexandria

$claude plugin install coa@coa

Entry name may differ; see the README.

$claude plugin marketplace add https://www.slopshopper.com/community/marketplace.json

$claude plugin install coa@slopshopper-community

Points at the author's repo. Nothing re-hosted.

$git clone https://github.com/davebream/claude-of-alexandria

$claude --plugin-dir ./claude-of-alexandria/plugins/claude-of-alexandria

One session, nothing installed.

In the preview

Nothing visible. Works behind the scenes, or needs what the sandbox lacks.

Hooks

agent.spawn{subagentType=/"^coa:"/}

Reaches

$.env.get$.ui.log

env: CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL

Checks

validate: 1 error
Path is a file; skills entries must be directories containing SKILL.md: ./skills

Tests ✓ · no state contract · first seen 2026-10-17