@tangleai/gmpl

GMPL supplies immutable prompt, role, domain and pattern content for @tangleai/mas. MAS owns validation, planning, agents, concurrency, bounded loops, durable interactions and replay. The host supplies clients, context, store, clock and any permitted capabilities. Importing GMPL performs no I/O.

The fourteen source packs compile with Jaren JOSL (parseToml) and JTLT (compileJtltStylesheet). Installed consumers use artifacts/catalog.json; no checkout prompt directory or archived module is needed. Outcome reflection and trading/research domain content are outside this catalog.

import {
  gmplArtifacts, createGmplCatalog, renderGmplPrompt,
} from '@tangleai/gmpl';

const loaded = await createGmplCatalog(gmplArtifacts);
if (!loaded.valid) throw new Error(JSON.stringify(loaded.issues));
const prompt = loaded.value.prompt('analysis-analyst');
if (!prompt) throw new Error('Missing analyst artifact');
const rendered = renderGmplPrompt(prompt, {
  query: 'Compare these records.', evidence: [],
  context: { participant: 'analyst-1' },
});

compileGmplPromptPack(sourceText, { variables, outputSchema }) accepts TOML text and explicit variable contracts. It returns a schema-bound artifact or coded issues. Named variables and balanced {{#if name}} blocks are the only source syntax. Variables outside optional blocks are required. Unknown names, traversal, helpers and malformed blocks refuse. Conditional presence means true, a nonempty string/container or a nonzero finite number. Substituted text is never parsed again; JSON values render in stable property order. System instructions are static. Source-byte, parsed-pack, role, schema, compilation policy and compiled-template changes move identity.

createGmplDomainBinding and createGmplRecipe create revisioned data. gmplCatalogDocument seals a document containing prompts, domains and recipes; createGmplCatalog checks those identities, references and schemas and returns a detached immutable catalog. Domain data names prompt bindings and pure answer projector ids/versions; it cannot contain clients, functions, credentials or filesystem paths.

materializeGmplTemplate(recipe, domain, hostSnapshot, catalog) fixes the participant topology and produces a MAS template, registry snapshot and identity receipt. The host snapshot contains { registry, config, profile }, using the MAS public registry/catalog constructors. Child workflows inherit registry and CONFIG pins; the final registry exists before a top-level template pins it. The top-level template is never inserted into the registry it hashes.

instantiateGmplPattern(materialized, { caps? }, hostSnapshot, catalog) uses MAS specialization, full validation and planning. Parameters can only lower declared caps. Changing participant counts or round policy requires a newly materialized recipe; there is no node-array patch. createGmplHostBindings returns pure task handlers and exact message adapters for MAS compilation. Neither function executes agents. Parallel analysis executes independent analysts and an ordered, evidence-preserving synthesis. Peer review, red team and structured debate use bounded MAS loops, explicit early termination and no-consensus at the round cap. The catalog contains all six default recipes. Peer/red single-round recipes remain available with scope: 'round'. Debate synthesizes once after terminal judgment and preserves dissent. Execution coverage is reported by the benchmark.

Clarification inspects intent, asks one to three questions (q1 through qN), and waits through MAS. The host supplies { answers: { q1: "..." } } with exactly the requested keys. Nonblank string answers are bounded to 4,096 characters. Accepted responses become labelled host evidence; they clarify intent and do not independently verify external facts. Each turn has a distinct durable interaction path. Resolved intent permits one answer stage; unresolved turn exhaustion returns needs-information with outstandingQuestions and makes no answer call.

Delphi polls independently and exposes a closed projection of prior rounds to peers: round-local panel/finding aliases, answer keys, estimates, confidence, citation ids and statistics. Private finding ids, origins, reasons and invocation paths remain in host traces. This protects structured attribution from peers; external evidence ids remain visible, and answer keys deliberately containing an identity are not anonymized free text. Confidence is diagnostic. Text domains require unanimous normalized keys; numeric domains use Jaren sample standard deviation on their declared scale. Unresolved supported contradictions block agreement. The final synthesizer receives the full retained finding ledger.

createGmplHostBindings(materialized, catalog, { answerProjector? }) checks a pure custom projector's id/version against the domain before execution. Built-in text-answer and estimate-answer projectors use version 1. A custom binding returns a GmplOutcome<{ key, estimate }>; functions never enter artifact JSON.

| Pattern | Default | Participant range | Round/turn range | |---|---|---|---| | Parallel analysis | 2 analysts, 1 synthesis | 1–8 | one pass | | Peer review | 2 critics, 3 cycles, acceptance >=0.7 | 1–8 | 1–10 | | Red team | 1 attacker/defender, 3 rounds, resilience >=0.7 | 1–8 | 1–10 | | Structured debate | 2 participants, 3 rounds, final synthesis | 2–8 | 1–10 | | Clarification | at most 5 human turns | 1–3 questions/turn | 1–10 | | Delphi | 5 panelists, 3 rounds, sample deviation <=0.2 | 2–20 | 1–10 |

Claims cite visible evidence ids and exact digests. The shared evidence gate rejects foreign citations and silent loss or reattribution of supported findings. Changing a finding disposition requires a new evidenced reason. These checks establish citation scope and ledger retention; they do not prove entailment or that an answer is true. Pure post-agent validation preserves a content failure as a failed MAS attempt after the already incurred agent/normalization cost.

Public content issues use TGMPL1001 (shape), 1002 (identity), 1003 (unknown id), 1004 (prompt), 1005 (evidence/findings), 1006 (round/disposition), and 1007 (host capability). MAS/Jaren errors keep their owning codes or causes.

Development gates: npm run emit:gmpl, npm run gmpl:artifacts, npm run test:gmpl, and npm run benchmark:gmpl -- --require complete. The benchmark is keyless scripted conformance. It is not evidence of live pattern quality improvement, and it supplies no implicit paid-provider approval.

The public host walkthrough is examples/gmpl.ts. npm run gmpl:smoke (or bun scripts/gmpl-consumer-smoke.ts) runs the document-review text domain, the estimate-panel domain on [0,1], and a two-turn scripted clarification. Domain replacement changes immutable payload restrictions, prompt roles, evidence and host profile identities while retaining the same participant topology and round policy. Every domain payload is checked before any model request. The host registers versions, runs the existing MAS worker, reopens SQLite, reads the result/finding ledger and projects the saved plan with projectMasPlan. A second reconciliation creates no new effect.

npm run benchmark:gmpl -- --require comparison publishes the 288 primary scripted executions and seven separately registered diagnostic ablations. All pairs share question/evidence bytes, model/profile/config, output contract, scorer, total caps and no incremental human information. Clarification waits on nine cases and scores 0.625 versus its control's 1.000; the other five pairs tie at 1.000. These are protocol/scorer fixtures, not measured model improvements.

npm run benchmark:gmpl:locomo -- --plan freezes 64 seeded category 1–4 questions and their evidence slices from the local LoCoMo submodule. It never loads credentials. --replay PATH accepts an explicitly bound, content-addressed bundle using the existing model wire replay keys; misses fail closed and partial rows are ineligible. Reports separate original purchases from zero incremental replay transport. Test-created responses remain scripted. See the plan and replay report for identities, costs and limits. Live LoCoMo pattern-versus-single-agent quality remains unmeasured.