Add six 2026 coding-agent captures: Claude Code (incl. -p mode), Codex CLI, opencode - #162

Six coding-agent system prompts, captured 2 September 2026. New OPENCODE/ vendor directory.

ANTHROPIC/Claude-Code-Opus-5_Sep-02-2026.md          22,760 chars  35 tools  interactive
ANTHROPIC/Claude-Code-Fable-5.1_Sep-02-2026.md       26,463 chars  35 tools  interactive
ANTHROPIC/Claude-Code-Print-Mode_Sep-02-2026.md      20,938 chars  29 tools  -p / print
OPENAI/Codex-CLI-gpt-5.6-sol_Sep-02-2026.md          23,377 chars   9 tools  exec
OPENCODE/opencode-chat-completions-template_...md     9,621 chars  11 tools  run
OPENCODE/opencode-apply-patch-template_...md         10,413 chars   9 tools  run

The Claude Code entry currently in ANTHROPIC/ is Claude_Code_03-04-24.md. These are the v2.1.258.18d harness on three current models.

Captured off the wire, not out of the bundle

A local proxy sits in front of the model API, the harness is launched pointing at it, and the request that carries the prompt is recorded exactly as sent. Every file's header names the one command that regenerates it — node capture/capture.mjs claude --model claude-opus-5 and so on — using OrcaReplay (Apache-2.0), and links a public copy so the text can be diffed against a second location.

That method reaches two things a bundle extraction cannot:

1. -p mode is a different prompt. Every Claude Code prompt in circulation is the interactive one. --print is the non-interactive mode — one turn in, answer to stdout — and it is what a script, a CI job, or the Agent SDK actually receives:

• "You are a Claude agent, built on Anthropic's Claude Agent SDK." — not "You are Claude Code, Anthropic's official CLI for Claude."
• billing header cc_entrypoint=sdk-cli, not cli
• gone: the ! <command> hint, the Scratchpad Directory section, the gitStatus block
• 29 tools instead of 35 — Artifact, AskUserQuestion, EnterPlanMode, ExitPlanMode are never offered when nobody is at the terminal
• the injected role:"system" turn drops from 11,529 to 7,585 characters

Same harness build as the interactive Fable 5.1 file in this PR, captured minutes apart, so the two diff cleanly.

2. opencode has three prompts, not one. Measured across seven models on the same harness:

Two of the three are in this PR (the first and the third — the widest-shared and the outlier). The prompt, the wire dialect and the tool set are each chosen per model. The apply_patch one is worth reading beside OPENAI/Codex-CLI-gpt-5.6-sol_Sep-02-2026.md: same model, different harness, unrelated prompt.

Scrubbing

Home directory, username, git name and email, gateway host, OS build, e-mail addresses, uuids and hex runs of 32+ become {{PLACEHOLDER}} tokens — derived from the machine at capture time, not hard-coded, and the capture aborts rather than write a file if anything identifying survives the pass. Nothing else is reworded, reordered or removed. Tool definitions travelled in the same requests and are not included in these files.

Disclosure: I work on OrcaReplay, the tool used to capture these.

🤖 Generated with Claude Code