Post-Plan Enrichment Swarm

Author: Larry the Laptop Lobster
Date: 2026-02-20
Status: Proposal
Audience: PlanExe Contributors, OpenClaw Agent Architects

Pitch

When PlanExe finishes generating a plan, it drops a rich set of structured artifacts onto disk: document TODOs, a full WBS CSV, a Gantt schedule, assumptions, team roles, and physical location data. Today those artifacts sit idle.

This proposal defines a post-plan enrichment swarm — five specialised agents that fire automatically when PlanExe completes, read those artifacts as grounded inputs, and commit real enrichment layers (fetched documents, a populated GitHub project board, validated assumptions, sourced candidates, and jurisdiction-specific compliance requirements) back to the plan repository.

This is not a modification to PlanExe. The planning pipeline is untouched. The swarm runs entirely outside PlanExe, triggered by a single completion signal.

What This Is NOT

Concern	Answer
Does this change PlanExe's planning pipeline?	No. Zero changes to `planexe/` code or APIs.
Does this overlap with proposal #41 (Autonomous Execution)?	No. #41 builds an execution engine for running tasks after a plan is approved. This proposal enriches the plan artifacts themselves — fetching documents, setting up the project board, validating assumptions — before any execution begins.
Does this overlap with proposal #03 (Distributed Execution)?	No. #03 parallelises PlanExe's internal plan-generation workers. This is a separate post-generation layer.
Does this overlap with proposal #47 (OpenClaw Skill)?	Partially reuses it. #47 packages PlanExe as an OpenClaw skill so agents can call PlanExe. This proposal builds the response to PlanExe completing — the outbound enrichment pass.

Problem

A PlanExe run for even a simple plan (e.g. a small Connecticut egg farm) produces:

017-5-identified_documents_to_find.json — a structured list of real documents that need to be located (zoning ordinances, predator population data, health codes, etc.)
017-6-identified_documents_to_create.json — a structured list of internal documents to draft (Project Charter, Risk Register, Communication Plan, etc.)
023-4-wbs_project_level1_and_level2_and_level3.csv — a full Level 1–4 WBS with task UUIDs
026-3-schedule_gantt_machai.csv — project schedule with start/end dates per task
021-task_dependencies_raw.json — dependency graph
003-11-consolidate_assumptions_short.md — key assumptions in plain Markdown
013-team.md — role roster with contract types
005-2-project_plan.md — master plan with resources, permits, and budget narrative
002-21-physical_locations.md — jurisdiction(s) where the project will run

All of these are machine-readable and semantically rich. Without enrichment they just sit in a folder.

Solution: Enrichment Swarm Architecture

Trigger

PlanExe writes 999-pipeline_complete.txt to the run directory when the pipeline finishes. This file is the trigger.

/run/29131a8e-95d1-4f43-9891-920fae2b90ef/999-pipeline_complete.txt

An OpenClaw file-watch hook (or a simple inotifywait wrapper on the run root) detects the file and fires the enrichment swarm:

# Example hook: watch for completion signal
inotifywait -m -r /run --include '999-pipeline_complete\.txt' -e create \
  | while read dir event file; do
      planexe-enrich "$dir"
    done

planexe-enrich is a thin shell wrapper that reads the run directory and invokes the Lobster pipeline below.

Orchestration: Lobster Pipeline

The enrichment swarm is orchestrated as a single Lobster workflow file. Lobster provides:

Deterministic stage ordering — stages run in sequence, not ad-hoc
Approval gates — human review between stages before side effects commit
Resumable tokens — a crashed pipeline resumes from the last completed stage
JSON piping — each stage emits structured JSON consumed by the next

See: https://docs.openclaw.ai/tools/lobster.md

State Machine: Git as Ground Truth

Each enrichment agent commits its outputs to the plan repository before the next stage begins. Idempotency rule: if the output path already exists in git, the agent skips that step. This makes every stage safe to re-run.

plan-repo/
  docs/                          ← Agent 1 output
    project-charter.md
    risk-register.md
    CT-zoning-litchfield.pdf     (fetched external doc)
  validation/                    ← Agent 3 output
    assumptions-check.md
  sourcing/                      ← Agent 4 output
    team-leads.md
    vendors.md
  compliance/                    ← Agent 5 output
    requirements.md
  .github/                       ← Agent 2 sets up milestones + issues

The `.lobster` Workflow File

File: skills/planexe-enrich/enrich.lobster

name: planexe-post-plan-enrichment
args:
  run_dir:
    description: "Absolute path to the PlanExe run directory"
    required: true
  plan_repo:
    description: "Absolute path to the git plan repository"
    required: true
  github_repo:
    description: "GitHub repo slug (owner/repo) for project board setup"
    required: true

steps:
  # ── Stage 1: Document Executor ─────────────────────────────────────────
  - id: document_executor
    command: >
      planexe-enrich-agent document-executor
        --run-dir "$run_dir"
        --plan-repo "$plan_repo"
        --find-list "017-5-identified_documents_to_find.json"
        --create-list "017-6-identified_documents_to_create.json"
        --output-dir "docs/"
        --json
    condition: "! git -C $plan_repo ls-files --error-unmatch docs/project-charter.md 2>/dev/null"

  - id: approve_documents
    command: >
      planexe-enrich-agent summarize-stage
        --stage document_executor
        --stdin
        --json
    stdin: $document_executor.stdout
    approval: required
    approval_prompt: "Agent 1 has fetched/drafted documents into docs/. Review and approve to commit."

  - id: commit_documents
    command: >
      git -C "$plan_repo" add docs/ &&
      git -C "$plan_repo" commit -m "enrich(docs): document executor pass [planexe-swarm]"
    condition: $approve_documents.approved

  # ── Stage 2: Project Board Setup ───────────────────────────────────────
  - id: project_board_setup
    command: >
      planexe-enrich-agent project-board
        --run-dir "$run_dir"
        --github-repo "$github_repo"
        --wbs-csv "023-4-wbs_project_level1_and_level2_and_level3.csv"
        --gantt-csv "026-3-schedule_gantt_machai.csv"
        --deps-json "021-task_dependencies_raw.json"
        --json
    condition: $commit_documents.exitcode == 0

  - id: approve_board
    command: >
      planexe-enrich-agent summarize-stage
        --stage project_board_setup
        --stdin
        --json
    stdin: $project_board_setup.stdout
    approval: required
    approval_prompt: "Agent 2 will create GitHub milestones and issues. Review the plan and approve."

  - id: apply_board
    command: >
      planexe-enrich-agent project-board
        --run-dir "$run_dir"
        --github-repo "$github_repo"
        --wbs-csv "023-4-wbs_project_level1_and_level2_and_level3.csv"
        --gantt-csv "026-3-schedule_gantt_machai.csv"
        --deps-json "021-task_dependencies_raw.json"
        --apply
        --json
    condition: $approve_board.approved

  # ── Stage 3: Assumption Validator ──────────────────────────────────────
  - id: assumption_validator
    command: >
      planexe-enrich-agent assumption-validator
        --run-dir "$run_dir"
        --plan-repo "$plan_repo"
        --assumptions-md "003-11-consolidate_assumptions_short.md"
        --output "validation/assumptions-check.md"
        --json
    condition: "! git -C $plan_repo ls-files --error-unmatch validation/assumptions-check.md 2>/dev/null"

  - id: approve_validation
    command: >
      planexe-enrich-agent summarize-stage
        --stage assumption_validator
        --stdin
        --json
    stdin: $assumption_validator.stdout
    approval: required
    approval_prompt: "Agent 3 validated assumptions against live data. Discrepancies flagged. Approve to commit."

  - id: commit_validation
    command: >
      git -C "$plan_repo" add validation/ &&
      git -C "$plan_repo" commit -m "enrich(validation): assumption validator pass [planexe-swarm]"
    condition: $approve_validation.approved

  # ── Stage 4: Team & Resource Sourcer ───────────────────────────────────
  - id: team_sourcer
    command: >
      planexe-enrich-agent team-sourcer
        --run-dir "$run_dir"
        --plan-repo "$plan_repo"
        --team-md "013-team.md"
        --project-plan-md "005-2-project_plan.md"
        --output-dir "sourcing/"
        --json
    condition: "! git -C $plan_repo ls-files --error-unmatch sourcing/team-leads.md 2>/dev/null"

  - id: approve_sourcing
    command: >
      planexe-enrich-agent summarize-stage
        --stage team_sourcer
        --stdin
        --json
    stdin: $team_sourcer.stdout
    approval: required
    approval_prompt: "Agent 4 found real candidates and vendors. Review contacts before committing."

  - id: commit_sourcing
    command: >
      git -C "$plan_repo" add sourcing/ &&
      git -C "$plan_repo" commit -m "enrich(sourcing): team and vendor sourcer pass [planexe-swarm]"
    condition: $approve_sourcing.approved

  # ── Stage 5: Compliance Researcher ─────────────────────────────────────
  - id: compliance_researcher
    command: >
      planexe-enrich-agent compliance-researcher
        --run-dir "$run_dir"
        --plan-repo "$plan_repo"
        --project-plan-md "005-2-project_plan.md"
        --locations-md "002-21-physical_locations.md"
        --output "compliance/requirements.md"
        --json
    condition: "! git -C $plan_repo ls-files --error-unmatch compliance/requirements.md 2>/dev/null"

  - id: approve_compliance
    command: >
      planexe-enrich-agent summarize-stage
        --stage compliance_researcher
        --stdin
        --json
    stdin: $compliance_researcher.stdout
    approval: required
    approval_prompt: "Agent 5 found real permit requirements for the jurisdiction. Approve to commit."

  - id: commit_compliance
    command: >
      git -C "$plan_repo" add compliance/ &&
      git -C "$plan_repo" commit -m "enrich(compliance): compliance researcher pass [planexe-swarm]"
    condition: $approve_compliance.approved

Invoking the Pipeline

# Run the enrichment swarm for a completed PlanExe run
lobster run skills/planexe-enrich/enrich.lobster \
  --arg run_dir=/run/29131a8e-95d1-4f43-9891-920fae2b90ef \
  --arg plan_repo=/repos/egg-farm-ct \
  --arg github_repo=acme/egg-farm-ct

# Resume after approving an approval gate (token returned by previous call)
lobster resume <resumeToken> --approve

The Five Enrichment Agents

Agent 1 — Document Executor

Purpose: Turn PlanExe's document TODO lists into real artifacts.

Inputs:

{run_dir}/017-5-identified_documents_to_find.json
{run_dir}/017-6-identified_documents_to_create.json

Example input item (find-list):

{
  "id": "46451098-6d0f-4efd-8367-0ee5247ed4b7",
  "document_name": "Connecticut Local Zoning Regulations for Poultry Farming",
  "description": "Regulations regarding chicken farming and egg sales in Litchfield, Tolland, and New London Counties",
  "recency_requirement": "Current regulations",
  "responsible_role_type": "Regulatory Compliance Assistant",
  "steps_to_find": [
    "Contact the zoning department in each county.",
    "Search the county websites for zoning ordinances.",
    "Consult with a local land use attorney."
  ],
  "access_difficulty": "Medium"
}

Example input item (create-list):

{
  "id": "e4b5c157-3e2f-4fc5-93f4-6fa4f09aeb12",
  "document_name": "Project Charter",
  "document_template_primary": "PMI Project Charter Template",
  "steps_to_create": [
    "Define project objectives and scope based on the goal statement.",
    "Identify key stakeholders and their roles.",
    "..."
  ],
  "approval_authorities": "Project Sponsor"
}

Actions: 1. For each item in the find-list: run a targeted web search using document_name + steps_to_find[0] as query; save result to docs/{slug}.md (or PDF if direct link found). 2. For each item in the create-list: call llm-task (Lobster JSON-only LLM step) with the document_template_primary as prompt scaffold and PlanExe plan context as input; save draft to docs/{slug}.md. 3. Skip any item whose output file already exists in git (idempotent).

Outputs committed to plan repo:

docs/
  CT-zoning-litchfield-poultry.md      # fetched from county website
  CT-health-codes-egg-sales.md         # fetched from DEEP / health dept
  CT-predator-population-data.md       # fetched from DEEP wildlife survey
  project-charter.md                   # LLM-drafted from PMI template
  risk-register.md                     # LLM-drafted
  communication-plan.md                # LLM-drafted
  [... all items from both JSON lists]

CLI contract:

planexe-enrich-agent document-executor \
  --run-dir /run/UUID \
  --plan-repo /repos/my-plan \
  --find-list 017-5-identified_documents_to_find.json \
  --create-list 017-6-identified_documents_to_create.json \
  --output-dir docs/ \
  --json
# stdout: { "created": 12, "fetched": 8, "skipped": 0, "errors": [] }

Agent 2 — Project Board Setup

Purpose: Populate a GitHub project board from the WBS and Gantt CSVs.

Inputs:

{run_dir}/023-4-wbs_project_level1_and_level2_and_level3.csv
{run_dir}/026-3-schedule_gantt_machai.csv
{run_dir}/021-task_dependencies_raw.json

Example WBS row:

Level 1;     Level 2;                        Level 3;    Level 4;                         Task ID
Egg Operation;Project Initiation & Planning;;            ;9394ca93-...
              ;                              ;Define Project Scope and Objectives;;50744ef4-...
              ;                              ;            ;Identify Stakeholders and Their Needs;56e85987-...

Example Gantt row:

project_key,project_name,project_start_date,project_end_date,...
9394ca93-...,Project Initiation & Planning,2/21/2026,3/26/2026,...

Actions: 1. Parse WBS CSV: Level 2 rows → GitHub milestones (with start/end from Gantt CSV). 2. Parse WBS CSV: Level 3 rows → GitHub issues (assigned to milestone, labelled with Level 1). 3. Parse WBS CSV: Level 4 rows → GitHub sub-issues (linked to parent Level 3 issue). 4. Parse 021-task_dependencies_raw.json → add "Depends on: #N" lines to issue bodies. 5. In dry-run mode (--json only): emit the plan as JSON without creating anything. 6. In apply mode (--apply): call gh api to create milestones and issues.

Outputs:

GitHub project: acme/egg-farm-ct
  Milestones:
    "Project Initiation & Planning"  (due: 2026-03-26)
    "Site Preparation & Construction" (due: ...)
    [one per WBS Level 2]
  Issues:
    #1  Define Project Scope and Objectives
    #2  Develop Detailed Budget
    [one per WBS Level 3]
  Sub-issues / task links:
    #1 → children: Identify Stakeholders, Define Egg Production Goals, ...

CLI contract:

# Dry run (returns JSON plan, no GitHub writes)
planexe-enrich-agent project-board \
  --run-dir /run/UUID \
  --github-repo owner/repo \
  --wbs-csv 023-4-wbs_project_level1_and_level2_and_level3.csv \
  --gantt-csv 026-3-schedule_gantt_machai.csv \
  --deps-json 021-task_dependencies_raw.json \
  --json
# stdout: { "milestones": 6, "issues": 42, "sub_issues": 127, "plan": [...] }

# Apply (creates milestones and issues on GitHub)
planexe-enrich-agent project-board ... --apply --json
# stdout: { "milestones_created": 6, "issues_created": 42, "errors": [] }

Agent 3 — Assumption Validator

Purpose: Check each plan assumption against current real-world data.

Input:

{run_dir}/003-11-consolidate_assumptions_short.md

Example assumption (from actual egg-farm run):

## Assumptions
- Demand exists for locally sourced eggs.
- Startup capital is available.
- Land is suitable for chickens.

Actions: 1. Parse each bullet or numbered assumption from the Markdown. 2. For each assumption: run a web search with the assumption text + plan jurisdiction as query. 3. Score each assumption: CONFIRMED, UNCERTAIN, or CONTRADICTED based on search results. 4. Produce a structured Markdown report with evidence citations.

Output committed to plan repo:

validation/assumptions-check.md

Example output format:

# Assumption Validation Report
Generated: 2026-02-20 | Plan: Egg Operation, Litchfield County CT

## Assumption 1: Demand exists for locally sourced eggs
Status: ✅ CONFIRMED
Evidence: CT farm direct sales up 18% YoY (CT Dept of Agriculture, 2025).
Source: https://portal.ct.gov/doag/...

## Assumption 2: Land is suitable for chickens
Status: ⚠️ UNCERTAIN
Evidence: Litchfield County has active zoning restrictions on flock size.
Recommend: Verify specific parcel zoning before committing.
Source: https://litchfieldct.gov/zoning/...

## Assumption 3: Startup capital is available
Status: ❓ UNVERIFIABLE
Notes: Plan-specific; no external data source applies.

CLI contract:

planexe-enrich-agent assumption-validator \
  --run-dir /run/UUID \
  --plan-repo /repos/my-plan \
  --assumptions-md 003-11-consolidate_assumptions_short.md \
  --output validation/assumptions-check.md \
  --json
# stdout: { "confirmed": 1, "uncertain": 1, "contradicted": 0, "unverifiable": 1 }

Agent 4 — Team & Resource Sourcer

Purpose: Find real local candidates and vendors for each required role.

Inputs:

{run_dir}/013-team.md         (role roster with contract types)
{run_dir}/005-2-project_plan.md  (resources section)

Example input (from actual egg-farm run):

# Roles
## 1. Poultry Husbandry Advisor
...
## 2. Coop Construction Specialist
...

Actions: 1. Parse role names and contract types from 013-team.md. 2. Extract location from 002-21-physical_locations.md (e.g. "Litchfield County, CT"). 3. For each role: run targeted web searches for real professionals/vendors in that location. 4. For each resource mentioned in 005-2-project_plan.md: find actual local suppliers. 5. Compile findings with names, contact info, and notes.

Outputs committed to plan repo:

sourcing/team-leads.md
sourcing/vendors.md

Example team-leads.md:

# Team Sourcing — Egg Operation (Litchfield County CT)

## Poultry Husbandry Advisor
- CT Poultry Association: https://ctpoultry.org/  +1-860-...
- UConn Extension Poultry Program: https://extension.uconn.edu/...
  Contact: Dr. [Name], poultry@uconn.edu

## Coop Construction Specialist
- [Local contractor], Litchfield CT — specialises in farm structures
  Phone: +1-860-... | Website: ...

CLI contract:

planexe-enrich-agent team-sourcer \
  --run-dir /run/UUID \
  --plan-repo /repos/my-plan \
  --team-md 013-team.md \
  --project-plan-md 005-2-project_plan.md \
  --output-dir sourcing/ \
  --json
# stdout: { "roles_sourced": 8, "vendors_found": 5, "errors": [] }

Agent 5 — Compliance Researcher

Purpose: Find real regulatory requirements for the plan's jurisdiction and domain.

Inputs:

{run_dir}/005-2-project_plan.md      (permits mentioned)
{run_dir}/002-21-physical_locations.md  (jurisdiction)

Example location (actual run):

## Location 1
Connecticut, Litchfield County
A small farm in Litchfield County, CT

Actions: 1. Extract jurisdiction (state, county, municipality) from 002-21-physical_locations.md. 2. Extract permit types and regulatory domains mentioned in 005-2-project_plan.md. 3. For each regulatory requirement: search for actual permit names, forms, fees, and filing procedures. 4. Compile into a structured compliance report.

Output committed to plan repo:

compliance/requirements.md

Example output:

# Compliance Requirements — Egg Operation, Litchfield County CT

## Connecticut Poultry Registration
- Requirement: Flocks of 50+ birds must register with CT DOAG
- Form: PR-1 (Poultry Registration)
- Fee: No fee for small flocks
- Filing: https://portal.ct.gov/doag/registration/poultry
- Authority: CT Dept of Agriculture, Animal Industry Division

## Litchfield County Zoning — Agricultural Use
- Requirement: A-1 Agricultural zoning required for poultry operations
- Setback: Coops must be 50ft from property lines (Litchfield Town Code §8-2)
- Permit: Zoning Certificate of Compliance
- Fee: $75 (2025 fee schedule)
- Filing: Litchfield Town Hall, Zoning Office

## CT Cottage Food Law / Egg Sales
- Requirement: Direct farm sales of own eggs exempt from dealer licence if <500 dozen/month
- Labelling: Producer name, address, grade required on carton
- Authority: CT Public Health Code §19-13-B42

CLI contract:

planexe-enrich-agent compliance-researcher \
  --run-dir /run/UUID \
  --plan-repo /repos/my-plan \
  --project-plan-md 005-2-project_plan.md \
  --locations-md 002-21-physical_locations.md \
  --output compliance/requirements.md \
  --json
# stdout: { "requirements_found": 8, "jurisdictions_searched": 3, "errors": [] }

Crash Recovery

If the pipeline crashes mid-run, resume from the last approved + committed stage:

# Lobster returns a resumeToken when paused at an approval gate
lobster resume <resumeToken> --approve

# If the process crashes entirely, re-invoke — idempotency conditions
# (git ls-files checks) skip already-committed stages automatically
lobster run skills/planexe-enrich/enrich.lobster \
  --arg run_dir=/run/UUID \
  --arg plan_repo=/repos/my-plan \
  --arg github_repo=owner/repo

File Layout (Skill Package)

skills/
  planexe-enrich/
    enrich.lobster                    # Lobster workflow (this proposal)
    SKILL.md                          # Skill documentation
    agents/
      document_executor.py            # Agent 1
      project_board_setup.py          # Agent 2
      assumption_validator.py         # Agent 3
      team_sourcer.py                 # Agent 4
      compliance_researcher.py        # Agent 5
    lib/
      planexe_artifacts.py            # Shared: parse WBS CSV, Gantt CSV, etc.
      git_state.py                    # Shared: idempotency checks, commit helpers
    tests/
      fixtures/
        017-5-identified_documents_to_find.json   # Sample from real run
        023-4-wbs_project_level1_and_level2_and_level3.csv
      test_document_executor.py
      test_project_board_setup.py
      test_assumption_validator.py

What Gets Committed to the Plan Repo

plan-repo/
  docs/
    CT-zoning-litchfield-poultry.md
    CT-health-codes-egg-sales.md
    project-charter.md
    risk-register.md
    communication-plan.md
    [all 017-5 and 017-6 items]
  validation/
    assumptions-check.md
  sourcing/
    team-leads.md
    vendors.md
  compliance/
    requirements.md

Each file is committed by the enrichment agent in a dedicated git commit with message prefix enrich(...), making the enrichment layer auditable and separately revertable from the original plan artifacts.

Success Metrics

Metric	Target
Documents found/drafted per run	≥80% of items in `017-5` + `017-6`
GitHub issues created	100% of WBS Level 3 tasks
Assumptions validated with citation	≥75% (remainder marked UNVERIFIABLE)
Team leads identified per role	≥1 real contact per role
Compliance requirements found	≥90% of permit types mentioned in plan
Time from trigger to pipeline complete (excluding approval wait)	<15 minutes

Risks

Risk	Mitigation
Web search returns stale/wrong content	Agent cites source + date; human review at each approval gate
GitHub API rate limits	Batch issue creation; use conditional requests
LLM-drafted documents are generic	Use plan context as system prompt; flag AI-drafted docs with frontmatter `ai_generated: true`
Compliance information is jurisdiction-specific and may change	Output includes source URLs and retrieval date; not a substitute for legal advice
Approval gates slow the workflow	Gates are optional per deployment; can be disabled for trusted environments

Staged Rollout

Phase 1 — Agent 3 (Assumption Validator) only

Lowest risk, highest signal value. Validates the plan's core assumptions against live data. No external side effects (no GitHub writes, no document commits beyond the plan repo).

Phase 2 — Agents 1 + 3 (Documents + Validation)

Adds document fetching and drafting. All output stays inside the plan repo.

Phase 3 — Agent 5 (Compliance Researcher)

Adds jurisdiction-specific regulatory research. Still repo-only output.

Phase 4 — Agent 2 (Project Board Setup)

Adds GitHub writes. Requires github_repo arg and gh CLI auth.

Phase 5 — Agent 4 (Team & Resource Sourcer)

Adds external candidate/vendor data. Review carefully before sharing output externally.

Relationship to Existing Proposals

Proposal	Relationship
#03 Distributed Plan Execution	Orthogonal. #03 parallelises PlanExe's internal plan-generation workers. This swarm fires after generation is complete.
#41 Autonomous Execution of a Plan	Sequential, not competing. #41 builds the execution engine for running tasks after a plan is approved. This swarm enriches the plan before execution begins — it's the prep layer.
#43 Assumption Drift Monitor	Agent 3 (Assumption Validator) is a one-shot point-in-time check at plan completion. #43 is a continuous monitoring loop during execution. Both are needed.
#47 OpenClaw Skill Integration	#47 enables agents to call PlanExe. This proposal defines what happens after PlanExe responds. Complementary.
#49 Distributed Physical Task Dispatch	Agent 4 (Team Sourcer) provides the initial vendor/candidate list that #49's dispatch protocol would later use to route physical tasks. Feeds into #49.

References

Lobster workflow runtime: https://docs.openclaw.ai/tools/lobster.md
PlanExe run artifact schema: see expected_filenames1.json in any completed run directory
Sample run used to ground this proposal: 29131a8e-95d1-4f43-9891-920fae2b90ef (Egg Operation, Litchfield County CT)

Post-Plan Enrichment Swarm

Pitch

What This Is NOT

Problem

Solution: Enrichment Swarm Architecture

Trigger

Orchestration: Lobster Pipeline

State Machine: Git as Ground Truth

The .lobster Workflow File

Invoking the Pipeline

The Five Enrichment Agents

Agent 1 — Document Executor

Agent 2 — Project Board Setup

Agent 3 — Assumption Validator

Agent 4 — Team & Resource Sourcer

Agent 5 — Compliance Researcher

Crash Recovery

File Layout (Skill Package)

What Gets Committed to the Plan Repo

Success Metrics

Risks

Staged Rollout

Phase 1 — Agent 3 (Assumption Validator) only

Phase 2 — Agents 1 + 3 (Documents + Validation)

Phase 3 — Agent 5 (Compliance Researcher)

Phase 4 — Agent 2 (Project Board Setup)

Phase 5 — Agent 4 (Team & Resource Sourcer)

Relationship to Existing Proposals

References

The `.lobster` Workflow File