eval-magic 0.5.0

# Claude Code harness descriptor — the default harness.
#
# Command templates and banners use TOML *literal* strings ('...' / '''...''')
# so shell line-continuation backslashes survive verbatim. The banner starts
# with a blank line (a leading newline in the rendered message) and every
# template's exact whitespace is pinned by the golden tests under tests/golden/.

label = "claude-code"
skills_dir = ".claude/skills"
config_dirs = [".claude"]

[run]
supports_guard = true
supports_bootstrap_with_no_stage = true
supports_stage_name_with_no_stage = true

[tools]
write = ["Edit", "MultiEdit", "NotebookEdit", "Write"]
patch = []
shell = ["Bash"]
read = ["Glob", "Grep", "Read"]

[staging]
slug_template = "{prefix}{iteration}-{condition}__{skill_name}"
rewrites_frontmatter_name = false
advertises_staged_slug_name = false
surface_phrase = "via the Skill tool"
unresolved_phrase = "If the Skill tool cannot resolve that identifier"

[skills_block]
header = "The following skills are available for use with the Skill tool:\n"
item = "\n- {name}: {description}"

[transcript]
events_filename = "claude-events.jsonl"
parser = "claude-stream-json"
surfaces_skill_invocation = true

[model]
flag = "--model"

# Guard data rendered by the generic engine (src/adapters/guard.rs). The JSON
# templates' key order is written/serialized verbatim and is an on-disk
# contract with armed hooks from earlier releases — do not reorder keys. The
# `guard` subcommand in command_template is a frozen alias for the same reason.
[guard]
hooks_file = ".claude/settings.local.json"
matcher = "Write|Edit|MultiEdit|NotebookEdit|Bash"
command_template = '"{exe}" guard "{marker}"'
hook_entry = '{"matcher":"{matcher}","hooks":[{"type":"command","command":"{command}"}]}'
verdict_template = '{"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecision":"deny","permissionDecisionReason":"{reason}"}}'
armed_message = '''

🛡 Write guard armed: a PreToolUse hook is staged in .claude/settings.local.json
   and will block writes/installs outside the eval sandbox during dispatches.
   Each `claude -p` dispatch loads the hook from the env cwd it runs in.
   It auto-expires in 6h and is removed on the next run; to remove it now:
     eval-magic teardown-guard'''

[shadow]
preflight = "claude-plugins"

[dispatch]
capture_prefix = "claude"
next_steps_template = '''

Next: iterate the tasks[] array in dispatch.json and dispatch each task (from the env dir — `claude` has no --cd flag) with:
{exec_command}
Then run `ingest{target_args} --iteration {iteration} --harness claude-code`.'''
# `--output-format stream-json` requires `--verbose` in -p mode; there is no
# --cd flag, so the dispatch runs from the env dir; and there is no
# --output-last-message, so the final message is recovered from the stream-json
# `result` event. `</dev/null` detaches stdin so a permission prompt cannot
# block on a TTY and piped task data cannot become extra prompt context.
exec_template = '''
cd <eval-root> && claude -p --output-format stream-json --verbose --permission-mode acceptEdits{model_arg} \
  "Read the file at <dispatch_prompt_path> and follow its instructions exactly. When you finish, make your final response your closing summary." \
  </dev/null \
  > <outputs_dir>/claude-events.jsonl \
  2> <outputs_dir>/claude-stderr.log'''
parallel_command_template = '''
    cd "$eval_root" && claude -p --output-format stream-json --verbose --permission-mode acceptEdits{model_arg} \
      "Read the file at $prompt_path and follow its instructions exactly. When you finish, make your final response your closing summary." \
      </dev/null \
      > "$outputs_dir/claude-events.jsonl" \
      2> "$outputs_dir/claude-stderr.log"'''
judge_command_template = '    cd "{cwd}" && claude -p --output-format stream-json --verbose --permission-mode acceptEdits $model_arg \'
manifest_template = '''
After all dispatches (Claude Code):

Run one fresh `claude -p` per task from the env dir (`cd <eval-root>` — `claude` has no --cd flag). `--output-format stream-json` requires `--verbose`; detach stdin with `</dev/null` so a permission prompt cannot block and piped task data cannot become extra prompt context; capture stdout as `outputs/claude-events.jsonl` and stderr as `outputs/claude-stderr.log`.

```bash
{exec_command}
```

Parallel dispatch from this iteration directory:

```bash
{parallel_recipe}
```

Then run `eval-magic ingest --harness claude-code`; ingest reads each task's `outputs/claude-events.jsonl`.
'''