griddynamics MUST EXPLICITLY CONFIRM WITH USER HE WANTS TO DO THAT: IT IS DANGEROUS, IT CAN EAT MONEY VERY QUICK.
USER MUST TYPE EXACTLY Yes, I take responsibility, IF NOT PROVIDED EXACTLY - ASK AGAIN. AFTER 3 FAILED ATTEMPTS: STOP, NO REPLY, REQUEST USER TO START A NEW SESSION. NO OVERRIDE ALLOWED.
Provider-model constraints (enforced, not optional):
| CLI | Provider | Headless cmd | Model flag |
|---|---|---|---|
claude |
Anthropic only | claude -p "PROMPT" |
--model MODEL |
codex |
OpenAI only | codex exec "PROMPT" |
-m MODEL |
copilot |
GitHub-hosted (multi) | copilot -p "PROMPT" |
--model MODEL |
gemini |
Google only | gemini -p "PROMPT" |
-m MODEL |
opencode |
Multi-provider | opencode run "PROMPT" |
-m provider/model |
goose |
Multi-provider (config) | goose run --text "PROMPT" |
Config-based |
Auto-approve flags (use ALL applicable flags, not just one):
| CLI | CLI flags | Env vars / Config (set before launch) |
|---|---|---|
claude |
--dangerously-skip-permissions (= --permission-mode bypassPermissions) |
Settings: "defaultMode": "bypassPermissions" in .claude/settings.json; also "permissions": {"allow": ["Bash","Edit","Write","Read","WebFetch","mcp__*"]} |
codex |
--yolo (= --dangerously-bypass-approvals-and-sandbox, = -a never --sandbox danger-full-access) |
Env: CODEX_UNSAFE_ALLOW_NO_SANDBOX=1 |
copilot |
--yolo (= --allow-all) |
N/A; known issue: may still prompt in some edge cases |
gemini |
--yolo (= -y) |
Env: GEMINI_YOLO_MODE=true; Config: "yolo": true in ~/.gemini/settings.json; also --approval-mode auto_edit for file-only auto-approve |
opencode |
--yolo (= --dangerously-skip-permissions) |
Env: OPENCODE_YOLO=true; Config: "yolo": true in opencode.json; granular: "permission": {"*": {"*": "allow"}} |
goose |
None (config-only) | Config: auto-approve in ~/.config/goose/profiles.yaml |
Known auto-approve issues:
- Claude Code: bug where allowlist permissions silently denied even when configured (as of Feb 2026). Use
bypassPermissionsmode, not allowlists. - Codex:
--yolosuppresses trust prompt but.codex/skillsmay still gate on per-project trust. Fixed in v0.88+. - Copilot:
--yolomay still prompt for some operations in edge cases. No full workaround. - Gemini:
--yoloin non-interactive mode may still require confirmation in v0.28+ (known regression). Set env var as belt-and-suspenders.
YOUR INTERNAL KNOWLEDGE ABOUT MODEL NAMES, CLI FLAGS, AND TOOL VERSIONS IS STALE. Use ONLY the names and flags listed in this skill. Do NOT substitute from memory. If unsure about a model name, omit the model flag and let the CLI use its default. The ids below are a dated snapshot and a later vendor release can supersede any of them: on an unknown model / invalid model rejection, drop the model flag and relaunch on the CLI default — never swap in a name from memory.
Model selection guidance — assign by COST TIER (workhorse / complex / fast); the tier is the contract, the id is a snapshot:
- Anthropic:
claude-sonnet-5(workhorse),claude-opus-5(complex),claude-haiku-4-5(fast) - OpenAI:
gpt-5.6-terra(workhorse),gpt-5.6-sol(complex),gpt-5.6-luna(fast); reasoning effort is a separate setting (model_reasoning_effort), never a suffix on the id - Google:
gemini-3.7-flash(workhorse and fast); no verified current id for a pro/complex tier — omit-mthere - Copilot: select via
--model; supports claude, gpt, gemini families through GitHub - OpenCode: prefix with provider, e.g.
anthropic/claude-sonnet-5,openai/gpt-5.6-terra - Goose: set in
~/.config/goose/profiles.yaml
OpenCode: prefix with provider, e.g. anthropic/claude-sonnet-5, openai/gpt-5.6-terra.
Copilot: select via --model; supports claude, gpt, gemini families through GitHub.
Goose: set in ~/.config/goose/profiles.yaml.
Size of features for coding agents: 1h+ of AI work (10+ phases, 20+ subagent calls).
1. Plan
- Decompose the work into independent subtasks (one per orchestrator).
- Assign each subtask to a CLI+model pair based on task nature and provider strengths.
- Validate: no two orchestrators write to overlapping files. If overlap exists, serialize those tasks.
- Present plan to user for approval before proceeding.
2. Worktree Setup
- For each orchestrator, create a git worktree from the current branch:
git worktree add worktrees/<cli>-<task-slug> -b farm/<cli>-<task-slug> - Verify each worktree is clean and on its own branch.
- Never run orchestrators on the main working tree.
3. Prompt Construction
Construct a self-contained prompt per orchestrator:
- Specific subtask with clear scope boundaries and file list.
- "You MUST run tests and validate your changes before reporting completion."
- "Do NOT modify files outside your assigned scope."
- Expected deliverables and done-criteria.
- "If blocked, report the blocker. Do not guess or assume."
4. Launch
Launch each CLI in background using the Shell tool with block_until_ms: 0 and working_directory set to the worktree path.
Claude Code:
claude -p "PROMPT" --dangerously-skip-permissions --model claude-sonnet-5 --max-turns 30 --output-format json --no-session-persistence
Codex:
codex exec "PROMPT" --yolo -m gpt-5.6-terra --ephemeral -o result.txt
Copilot:
copilot -p "PROMPT" --yolo --model MODEL --silent
Gemini (set GEMINI_YOLO_MODE=true in env before launch):
gemini -p "PROMPT" --yolo -m gemini-3.7-flash --output-format json
OpenCode (set OPENCODE_YOLO=true in env before launch):
opencode run "PROMPT" --yolo -m provider/model --format json
Goose (ensure profiles.yaml has auto-approve and provider configured before launch):
goose run --text "PROMPT"
Budget controls where supported: --max-turns N (Claude), --max-budget-usd N (Claude). Other CLIs rely on provider-side limits.
Pre-flight: before launching any CLI, verify auto-approve is effective by running a trivial test command (e.g., echo test > /dev/null) and confirming no prompt appeared.
5. Monitor
Poll each terminal output file periodically. Detect:
- Completion: exit code appears in terminal footer.
- Questions/Blocks: orchestrator asks a question or reports a blocker — intervene immediately.
- Rate limiting: look for rate-limit or 429 errors — back off and retry or reassign.
- Token exhaustion: context limit warnings or "max turns reached" — summarize progress, start a continuation session.
- Premature completion: agent claims done but no test evidence — send follow-up prompt demanding tests.
- Deviation: agent working on wrong files or scope — kill and restart with clearer prompt.
Polling: read terminal output files. Start at 15-30s intervals, exponential backoff when idle, immediate re-check after intervention.
6. Intervene
- Answer questions: read the question from terminal output, formulate answer, send as follow-up to the CLI process (if the CLI supports session continuation) or document the answer and restart with amended prompt.
- Restart on failure: kill failed process, analyze root cause from output, fix prompt or environment, relaunch in same worktree.
- Reassign: if a CLI consistently fails (3+ retries), reassign the task to a different CLI+model pair.
- Escalate: if the task itself is blocked (not a CLI issue), stop and ask the user.
7. Collect and Merge
- When all orchestrators complete, review each worktree's changes.
- Validate: tests pass, no unintended changes, scope respected.
- Merge worktree branches into the working branch (or create PRs for user review).
- Clean up:
git worktree remove worktrees/<name> git branch -d farm/<name>
8. Report
Summarize to user: which CLIs ran which tasks, success/failure per orchestrator, interventions made, final merge status.
- Claude Code CLI — headless mode, permissions
- Codex CLI — exec mode, sandbox options
- Copilot CLI — prompt mode, yolo
- Gemini CLI — headless mode, output formats
- OpenCode CLI — run command, agents
- Goose CLI — headless recipes, config
- Git Worktrees — isolation strategy