Skip to main content

FORGE Deep Research Pipeline

The FORGE Deep Research (DR) pipeline is a two-phase autonomous system that scans a GitHub organization, produces a structured state audit, and generates an execution plan that runs overnight without human intervention.

Pipeline Architecture

Phase 1: Organization State Audit

The State Audit Generator (FORGE-DR-001) scans everything:
Scan TargetWhat It Checks
RepositoriesClassification (active, empty, archived, legacy), default branch, visibility
BranchesNon-standard defaults, divergence, orphaned feature branches
IssuesOpen/closed counts, orphans, duplicates, missing labels
Pull RequestsConflicts, stale PRs, duplicate PRs
Local WorkspaceClone coverage, branch drift, uncommitted changes
Bundles/EpicsCompletion validation, hollow completions
Cross-RepoContract integrity, dangling references, event catalogue consistency
StageEvidence-backed maturity assessment (PLANNING through PRODUCTION)

Output Files

  • ORGANIZATION_STATE_AUDIT.md — 12-section comprehensive report (15-25 pages)
  • state_inventory.yaml — Machine-readable state for Phase 2 consumption
  • risk_register.yaml — Prioritized risks with severity, likelihood, mitigation

Phase 2: Autonomous Execution Planner

The Execution Planner (FORGE-EX-001) consumes the audit and produces:

Epic-to-Taskset Hierarchy

A2A Handoff Protocol

Every taskset boundary has a formal contract:
FieldDescription
handoff_idUnique: HO---
artifactsFiles/PRs produced by source taskset
preconditionsWhat must hold before target starts
quality_gateAutomated: lint, type-check, test, build
status_codesOK, RETRY_REQUEST, STEP_FAILED, QUALITY_GATE_FAILURE, PIPELINE_COMPLETE
retry_policyExponential backoff: base=30s, max=15min, 5 retries

Overnight Execution

4-Stream Verification

Both phases employ four parallel verification streams (FORGE Code 06 pattern):

Stream 1: Primary

Execute the methodology as specified. Produces the primary output.

Stream 2: Alternative

Independent approach using a different analytical framework.

Stream 3: Critical Review

Identify gaps, errors, unstated assumptions, blind spots.

Stream 4: Factual Verify

Re-verify claims against live data, flag discrepancies.
Streams are synthesized with confidence ratings: HIGH (4/4 agree), MEDIUM (3/4), LOW (2/4 or less).

Traceability

Trace IDRequirementRationaleVerification
DR-T-001Zero human gatingEliminate overnight blocking on manual approvalexecution_manifest.yaml consumed by n8n without intervention
DR-T-0024-stream verificationReduce single-pass blind spots per FORGE Code 06Synthesis section in all outputs
DR-T-003Epic-to-taskset traceabilityAudit finding → action mappingTraceability matrix in execution plan
DR-T-004A2A handoff contractsFormal agent-to-agent interface specificationhandoff_contracts.yaml validatable by JSON Schema
DR-T-005Exponential backoffRespect API rate limits, handle transient failuresRetry policy in handoff contracts

FORGE Pipeline Stages

FORGE Codes 01-06: ConOps through Churnery

Veritas Integration

How Veritas prompt library integrates with FORGE