Skip to content

Add nine coding-agent system prompts captured on the wire (Claude Code, Codex CLI, opencode) - #55

Merged
0xeb merged 1 commit into
0xeb:mainfrom
xizhuomengcontin:add-orcareplay-agent-captures
Sep 7, 2026
Merged

Add nine coding-agent system prompts captured on the wire (Claude Code, Codex CLI, opencode)#55
0xeb merged 1 commit into
0xeb:mainfrom
xizhuomengcontin:add-orcareplay-agent-captures

Conversation

@xizhuomengcontin

Copy link
Copy Markdown
Contributor

Nine coding-agent system prompts, captured on the wire. Verbatim, one file each, with the index updated in the same PR.

What is here

file harness model mode as sent tools
Anthropic/20260902-Claude-Code-Opus5-interactive.md Claude Code v2.1.258.18d claude-opus-5 interactive 22,760 35
Anthropic/20260902-Claude-Code-Fable5.1-interactive.md Claude Code v2.1.258.18d claude-fable-5-1 interactive 26,463 35
Anthropic/20260902-Claude-Code-Fable5.1-print.md Claude Code v2.1.258.18d claude-fable-5-1 -p 20,938 29
Anthropic/20260902-Claude-Code-Opus4.8-interactive.md Claude Code v2.1.258.18d claude-opus-4-8 interactive 19,083 33
OpenAI/20260902-Codex-CLI-gpt5.6-sol.md Codex CLI gpt-5.6-sol exec 23,377 9
OpenAI/20260902-Codex-CLI-gpt5.6-luna.md Codex CLI gpt-5.6-luna exec 20,842 3
OpenCode/20260902-opencode-big-pickle.md opencode big-pickle run 9,621 11
OpenCode/20260902-opencode-muse-spark-1.2.md opencode muse-spark-1.2-contributor-free run 10,249 11
OpenCode/20260902-opencode-gpt5.6-sol.md opencode gpt-5.6-sol run 10,413 9

OpenCode/ is a new vendor section, added before ## Miscellanous with a See: line, per CONTRIBUTING.

How these were obtained

Not extracted from the shipped bundle — recorded off the wire. A local proxy stands in front of the model API, the harness is launched pointing at it, and the request that carries the prompt is captured as sent. That matters for these three in particular:

  • The interactive prompt is a different prompt. A harness only assembles it when stdin is a real console, so a piped capture always yields the -p variant. The two Fable 5.1 entries are the same harness build minutes apart, and they differ by more than formatting: a different identity line, three sections present in one and absent in the other, and 35 tools against 29.
  • Runtime-assembled blocks are included — the gitStatus block, working-directory and session material — because they were in the request. A static extraction cannot show them.
  • Same model, different harness, different prompt. gpt-5.6-sol appears twice here, once through Codex CLI and once through opencode. The prompts are unrelated, and opencode swaps edit/write for apply_patch on that model while keeping edit/write on the others.

Verifiability

Each entry names the one command that regenerates it, e.g. node capture/capture.mjs claude --model claude-opus-5, using OrcaReplay (Apache-2.0). Each also links the public copy in that repo, so the text here can be diffed against a second location.

The three opencode entries are the three distinct templates measured across seven models on that harness. The other four models produce the big-pickle text within a few dozen characters and are deliberately left out rather than submitted as near-duplicates.

Redaction

Home directory, username, git name and email, gateway host, OS build, e-mail addresses, uuids and hex runs of 32+ are replaced with {{PLACEHOLDER}} tokens. The replacements are derived from the machine at capture time rather than hard-coded, and the capture aborts rather than writing a file if anything identifying survives the pass. Nothing else is reworded, reordered or removed — the prompts are otherwise byte-for-byte as sent. The character counts above are the prompt on the wire; each file on disk is a few hundred shorter because the placeholders are shorter than the paths they replace.

Titles keep (unverified): the source I can point at is my own repository, which is not independent corroboration. The reproduction command is the stronger claim, and anyone with the harness installed can run it.

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

🤖 Generated with Claude Code

Claude Code (Opus 5, Fable 5.1 interactive and -p, Opus 4.8), Codex CLI
(gpt-5.6-sol, gpt-5.6-luna) and opencode (three distinct templates).

Recorded by a local proxy in front of the model API rather than extracted
from the shipped bundle, so each is the prompt as the harness assembled it
at runtime. Every entry carries the command that reproduces it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@0xeb
0xeb merged commit abe56ea into 0xeb:main Sep 7, 2026
0xeb added a commit that referenced this pull request Sep 7, 2026
Follow-up to #55. Three changes to each of the nine entries, no change to
any captured prompt text:

- Label them (unverified, partial). Each capture omits the tool definitions
  that travelled with the same request, and CONTRIBUTING.md asks for the
  Partial label when only part of a prompt is captured. The tool and type
  declarations are in scope for this repo, so their absence belongs in the
  title rather than a clause at the end of a long note.
- Fold the non-standard "Public copy" field into "Source". CONTRIBUTING.md
  puts the public-source link on the Source line; the header block otherwise
  matched the documented template exactly.
- Drop the capture command from each note, keeping the tool named. The note
  is for the extraction method, and naming the proxy covers that without
  carrying its invocation.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants