SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsclaude-code

claude-code

Agents & models

View repository

Updated 11-07 17:53

About this skill

Drive Claude Code CLI

Claude Code is a separate agent with its own tools, permissions, project memory (CLAUDE.md) and session store. Headless mode (-p) runs one turn to completion and exits. You talk to it through files: a brief on stdin, a JSON result on stdout, and a session id for follow-ups.

Check the installed version before relying on a flag: claude --help.

Start a task

D=<scratch dir outside the repo>
cd /path/to/repo && claude -p --model opus --effort high \
  --permission-mode acceptEdits --output-format json \
  < "$D/brief.md" > "$D/result.json" 2> "$D/log.txt"
jq -r .result "$D/result.json" > "$D/report.md"
jq -r .session_id "$D/result.json"
Part Why
cd <dir> The working directory is the project root: it decides which CLAUDE.md, settings and skills load. Use --add-dir for extra directories.
--model <alias or id> Aliases fable, opus, sonnet (and haiku) select the latest of each; a full id such as claude-opus-5-5 pins one.
--effort <level> low, medium, high, xhigh, max. Use high for design and review.
--permission-mode <mode> plan (no edits), acceptEdits, auto, dontAsk, bypassPermissions. Headless runs cannot answer prompts; pick a mode that will not need one.
--allowedTools / --disallowedTools Narrow what it may run, e.g. --disallowedTools "Bash(git commit *)" "Bash(git push *)".
--output-format json One JSON object: result, session_id, is_error, subtype, stop_reason, permission_denials, safety_stops, total_cost_usd, num_turns.
stdin brief Long briefs in an argument break on quoting.

Optional: --max-budget-usd caps spend, --json-schema constrains the final answer, --session-id <uuid> sets the id up front (handy when you launch in the background and want to record it before it ends), --append-system-prompt adds standing rules, --no-session-persistence for throwaway runs.

Run long tasks in the background and wait for the process to exit. Do not poll in a loop. --bg starts a managed background session instead (claude agents, claude logs <id>, claude stop <id>).

Continue a session

cd /path/to/same/repo && claude -p --resume "$id" --model opus --effort high \
  --permission-mode acceptEdits --output-format json \
  < "$D/followup.md" > "$D/result-2.json" 2> "$D/log-2.txt"
  • Resume from the same directory; sessions are stored per project path.
  • Pass model, effort and permission mode again.
  • --fork-session with --resume branches a copy instead of extending it.
  • Resume the worker that did the work for follow-ups; start fresh for an independent review.

Check the outcome

  1. is_error false and subtype success. Otherwise read log.txt.
  2. permission_denials non-empty: it wanted something the mode did not allow. Decide whether to widen the mode or rescope the task.
  3. safety_stops non-empty, or a result saying the model's safeguards flagged the message: the turn was stopped, not answered. Do not resend the same brief. Rephrase around the actual goal, split out the part that tripped it, or do that part yourself (see collaborate-with-claude-5).
  4. Read git status / git diff in its working tree; the report describes intent, the diff is what happened. Run the checks yourself.

Run other models

Claude Code also hosts non-Claude models, such as GLM-5 and Qwen 3, through Anthropic-compatible endpoints. Set the endpoint and map the model tiers in the environment of that run only, so your own sessions are untouched:

ANTHROPIC_BASE_URL=<vendor endpoint> ANTHROPIC_AUTH_TOKEN="$VENDOR_KEY" \
ANTHROPIC_DEFAULT_OPUS_MODEL=<model> ANTHROPIC_DEFAULT_SONNET_MODEL=<model> \
ANTHROPIC_DEFAULT_HAIKU_MODEL=<small model> API_TIMEOUT_MS=3000000 \
  claude -p --model opus --output-format json < "$D/brief.md" > "$D/result.json"
  • Map all three tiers explicitly; Claude Code picks tiers internally (for subagents and summaries) and a missing mapping sends those calls to a model name the endpoint does not know.
  • --effort is passed through and mapped by the vendor; check its docs for what each level becomes.
  • Claude-specific features (safeguard fields, some beta tools) may be absent or behave differently. Check is_error and the diff, not only result.
  • The key is the user's; ask before using it.
Model family ANTHROPIC_BASE_URL Tier models (opus / sonnet / haiku)
GLM-5 (Z.ai) https://api.z.ai/api/anthropic (China: https://open.bigmodel.cn/api/anthropic) glm-5.3[1m] / glm-5.3[1m] / glm-5.3-flash[1m]
Qwen 3 (Alibaba) Token Plan: https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic; Coding Plan: https://coding-intl.dashscope.aliyuncs.com/apps/anthropic qwen3.8-max / qwen3.8-flash / qwen3.6-flash

The Qwen URL ends in /apps/anthropic with no /v1, and a key from a different plan returns 401. Neither endpoint lists models, so name them explicitly. Which member to pick and how to work with it is in collaborate-with-glm-5 and collaborate-with-qwen-3.

Parallel workers and isolation

The same rules as for any agent that edits files: one scope per worker, never two workers on the same files, no commits from workers. When your own checkout has uncommitted work, give Claude a separate detached worktree seeded with that work (the codex skill ships scripts/isolate.sh for this) and run it with cd into that worktree.

Failure modes

Symptom Cause / fix
Hangs or stops early with denials Permission mode needs a prompt nobody can answer. Choose a mode up front.
Resume says no such session Different working directory than the original run.
Ignores project conventions Started outside the repo root, so CLAUDE.md did not load.
Empty result, safety_stops set Safeguard stop. Rephrase or reroute that part; do not retry verbatim.
Cost or turns run away Add --max-budget-usd; narrow the scope in the brief.