| name | wf-agents-docs |
| description | Source-backed CLI invocation guide for Claude Code, Codex, and OpenCode automation. Use when invoking peer CLIs, writing batch tests, collecting cache telemetry, debugging command-line flags, or documenting cross-runtime agent usage for Harness workflows. |
WF Agents Docs
Use this skill before shelling out to claude, codex, or opencode from Harness workflows, peer review, cache tests, or automation scripts.
Source Order
- Prefer installed help:
claude --help, codex exec --help, opencode run --help.
- Check official docs for flags that affect cost, auth, JSON, resume, tools/MCP, or telemetry.
- When adding automation, record command, source, stdout/stderr shape, and failed patterns.
Claude Code CLI
- Interactive:
claude.
- Non-interactive JSON: pipe ASCII or UTF-8-safe stdin into
claude -p --output-format json.
- Stream JSON requires verbose mode:
claude -p --output-format stream-json --verbose.
- Continue/resume:
claude -c -p "..." or claude -p --resume <session-id> "..."; for PowerShell automation, prefer stdin and validate non-empty JSON before parsing.
- Use
--max-budget-usd <amount> in scripted probes.
- Use
--strict-mcp-config without --mcp-config to ignore configured MCP servers for a run. Use --safe-mode only to disable project customizations. Use --bare only for minimal CLI probes, not for Harness/cache attribution, because it skips CLAUDE.md, skills, plugins, MCP, hooks, and auto memory.
- Use
--tools "Read,Grep,Glob" or explicit --allowedTools/--disallowedTools for read-only probes.
- Prompt-cache telemetry appears in JSON
usage.cache_read_input_tokens / usage.cache_creation_input_tokens, and in statusline context_window.current_usage.*.
Codex CLI
- Interactive:
codex.
- Non-interactive:
codex exec "task".
- Read stdin as the full prompt:
cat prompt.txt | codex exec -.
- Prompt plus stdin context:
some-command | codex exec "summarize this output".
- Machine output:
codex exec --json "task" emits JSONL events; parse turn.completed.usage, including cached_input_tokens when present.
- Resume:
codex exec resume --last "..." or codex exec resume <SESSION_ID> "...".
- Permissions: default is read-only; set
--sandbox workspace-write only when edits are required. Use --ignore-user-config / --ignore-rules for controlled automation.
OpenCode CLI
- Interactive:
opencode.
- Non-interactive:
opencode run [message..].
- JSON events:
opencode run --format json "task".
- Resume:
opencode run --continue "..." or opencode run --session <id> "...".
- Peer role:
opencode run --agent reviewer --dir . "review prompt".
- Reuse a server to avoid MCP cold boot:
opencode serve, then opencode run --attach http://localhost:4096 "task".
- On Windows, first verify
opencode exists before writing automation around it.
PowerShell Automation Rules
- Prefer stdin over trailing prompt args for
claude -p in PowerShell.
- Use ASCII prompts or explicitly UTF-8-safe input for automated probes.
- Do not trust exit code alone. Fail on empty/non-JSON stdout or error/budget terminal fields.
- Avoid naming function parameters
$Args; PowerShell treats $Args specially.
- Store telemetry outside the repo, e.g.
$HOME/.claude/cache-telemetry/*.json, so git status does not perturb prefixes.
Evidence-Packet Review Pattern
For peer review, route smokes, cache analysis, and audits, gather evidence
first; the peer judges only the bounded packet.
- Gather paths, line snippets, command names, exits, and invariants with
rg,
node scripts, validators, or small reads.
- Send only that packet. Exclude full docs, raw logs, timestamps, session IDs,
and screenshots unless they are the evidence.
- Prefer no tools for judgment-only review; otherwise allow only read-only
tools and name the exact read set.
- Controller accepts, rejects, or escalates findings. Peers do not own scope.
No Scratch-File Rule
- Do not write CLI probe output under
%TEMP%, $env:TEMP, /tmp, or other
system temp directories.
- Prefer stdout, JSON/JSONL streaming, or in-memory parsing.
- Persistent repo evidence goes under
Harness/tasks/<task-id>/evidence/.
- Cache telemetry may live under
$HOME/.claude/cache-telemetry/ to avoid repo
prompt-cache churn.
- Do not create prompt temp files. Use stdin.
Subagent Output Contract
Require bounded structured returns:
Agent: <claude|codex|opencode|role name>
Probe: <what was tested or reviewed>
Mode: <read-only|review|telemetry|implementation>
Files examined: <exact paths or none>
Evidence: <commands, exit codes, paths, line refs>
Passes: <confirmed invariants>
Findings: <severity, file/path, reason, suggested fix>
Risks: <residual uncertainty or none>
Tool/CLI issues: <auth, timeout, budget, JSON parse, MCP, or none>
Verdict: PASS | FAIL | BLOCKED
Next: <smallest next controller action>
For JSON, use the same keys. Do not return transcripts, full file bodies, decorative logs, or speculation.
Cache Discipline
Follow Harness/specs/runtime/context-loading.md#Cache-First Context Contract: stable instructions first, volatile output in the dynamic suffix, and no provider cache claims without telemetry. Claude Code L2 uses cache_read_input_tokens; Codex JSONL may emit cached_input_tokens.
Batch-Test Pattern
- Probe command availability with
Get-Command claude,codex,opencode -ErrorAction SilentlyContinue.
- Build a compact evidence packet before invoking peer agents; use the peer
only for judgment unless the test explicitly requires live agent discovery.
- Run a cold turn and capture session id.
- Resume that session for two warm turns.
- For each turn record input, cache creation, cache read, ratio, cost, model/session id, and exact flags.
- Compare against a control mode. Do not attribute a provider-wide cache feature to Harness unless the Harness-shaped run improves or stabilizes cache behavior against a comparable baseline.
Official References
