foreach-agent

Run Liquid templates in batch across AI coding agent CLIs (claude, cursor-agent, gemini, codex), with:

• interactive TUI mode
• headless CLI mode
• programmatic API usage

Storage

All data is persisted under ~/.foreach-agent/:

• templates/: editable templates
• runs/<runId>/run.json: consolidated run state
• runs/<runId>/prompts/*.md: rendered prompt per task
• runs/<runId>/transcripts/*.jsonl: raw transcript per task
• exports/: run exports (JSON/CSV)

Run

Interactive mode (default):

cd apps/foreach-agent
pnpm start

or:

pnpm start tui

Build:

pnpm --filter @c3-oss/foreach-agent build

Headless CLI

Main commands:

pnpm start run --template my-template.liquid --entries '[{"name":"world"}]'
pnpm start templates list --output-format table
pnpm start runs --query 'status == "failed"'
pnpm start tasks --run-id <runId> --query 'provider == "codex" and hasError'
pnpm start configs --query 'concurrency >= 10 and autoApproval'
pnpm start export --run-id <runId> --format csv --output /tmp/run.csv

--output-format (tfplan-explorer style)

Supported values:

• interactive
• table
• json
• csv

Defaults:

• templates list, runs, tasks, configs: interactive
• run, templates create/read/update/delete: table

Note:

• --query requires non-interactive output (table, json, or csv)

Examples:

pnpm start runs --output-format interactive
pnpm start runs --output-format table
pnpm start runs --output-format json --query 'failedTasks > 0'
pnpm start templates list --output-format csv

Template CRUD subcommands

Create:

pnpm start templates create --name my-template --content 'Hello {{ name }}'

Read:

pnpm start templates read --id my-template.liquid --output-format json

Update:

pnpm start templates update --id my-template.liquid --content 'Hi {{ name }}'

Delete:

pnpm start templates delete --id my-template.liquid

Content input for create/update:

• --content (inline)
• --content-path (file path)
• --content-stdin (stdin)

Run without TUI

pnpm start run \
  --template my-template.liquid \
  --entries-path /abs/path/entries.yaml \
  --providers 'claude,codex' \
  --codex-model gpt-5.3-codex \
  --concurrency 10 \
  --retries 3 \
  --timeout-seconds infinite \
  --cwd /abs/path/project \
  --auto-approval true

Entries input options:

• --entries (inline JSON/YAML)
• --entries-path (JSON/YAML file)
• --entries-stdin (stdin)

--query (Filtrex)

--query follows the same model used in mzi-tfplan-explorer via filtrex.

Examples:

pnpm start runs --query 'failedTasks > 0'
pnpm start templates list --query 'variablesCount >= 3 and content ~= "TODO"'
pnpm start tasks --run-id <runId> --query 'provider == "codex" and status != "success"'
pnpm start configs --query 'timeoutSeconds == -1 and autoApproval'

Available query fields:

• templates: id, name, filePath, updatedAt, updatedAtEpoch, content, contentLength, variablesCount
• runs: id, status, templateName, templateId, providers, providerCount, entryCount, taskCount, successTasks, failedTasks, concurrency, retries, timeoutSeconds, autoApproval, cwd, createdAt, startedAt, finishedAt
• tasks: id, runId, status, provider, model, entryIndex, attempt, maxAttempts, durationMs, hasError, hasOutput, errorMessage, outputText, templateName, concurrency, retries, autoApproval, cwd
• configs: runId, runStatus, templateName, templateId, providerCount, providers, entryCount, concurrency, retries, timeoutSeconds, autoApproval, cwd, createdAt

Programmatic usage

import { executeHeadlessRun, filterWithQuery } from '@c3-oss/foreach-agent'

const run = await executeHeadlessRun({
  template: 'my-template.liquid',
  entriesText: '[{"name":"world"}]',
  providers: ['codex'],
})

const failedTasks = filterWithQuery(
  run.tasks,
  'status != "success"',
  (task) => ({ status: task.status, provider: task.provider }),
  (task) => task.id,
)

Main TUI shortcuts

Templates screen:

• j/k: navigate
• n: create template
• e: edit template
• d: delete template
• r: open run setup
• u: open runs history
• q: quit

Run setup screen:

• j/k: navigate fields
• space: toggle provider/boolean/source
• enter: edit selected field
• enter on start run: start execution

Run monitor screen:

• j/k: navigate tasks
• enter: open pretty logs
• s: status filter
• p: provider filter
• /: text filter
• x: export JSON
• X: export CSV

Multi-line editor:

• Ctrl+S: save
• Ctrl+L: clear
• Esc: cancel

Notes

• Template engine: liquidjs
• Entries accept JSON/YAML inline or file path
• Default concurrency: 10
• Default retries: 3 attempts per task
• Default timeout: infinite (minimum configurable value: 20s)
• Default auto-approval: enabled
• Default codex model: gpt-5.3-codex
  • gpt-codex-5.3 was tested and did not work in this environment