WEBDIENSTE/ECC
1
0
Fork 0
mirror of https://github.com/affaan-m/ECC.git synced 2026-09-28 20:45:11 +02:00
Code Issues Packages Projects Releases Wiki Activity

Integrate reviewed contributor PR2768

Browse Source
This commit is contained in:
affaan-m
2026-09-28 01:50:12 -04:00
parent 0d8070b32e e8cba155a8
commit ccb4c51737
1 changed files with 60 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files
commands/santa-loop.md
+60 -27
View File
@@ -70,38 +70,71 @@ Launch an Agent (subagent_type: `code-reviewer`, model: `opus`) with the full ru
- "You are an independent quality reviewer. You have NOT seen any other review. Your job is to find problems, not to approve."
- Return the structured JSON verdict above
#### Reviewer B: External Model (Claude fallback only if no external CLI installed)
#### Reviewer B: External Model (Claude fallback if no external reviewer is ready)
First, detect which CLIs are available:
```bash
command -v codex >/dev/null 2>&1 && echo "codex" || true
command -v gemini >/dev/null 2>&1 && echo "gemini" || true
```
Antigravity is optional and requires the external `ccg-workflow` runtime; it is not included in the base ECC install. An existing installation must provide both an executable `~/.claude/bin/codeagent-wrapper` and a readable regular reviewer role file at `~/.claude/.ccg/prompts/antigravity/reviewer.md`. Otherwise, retain the Claude fallback. Use the wrapper's maintained model-selection contract; do not add an Antigravity model pin.
Run detection, prompt creation, and the selected external review together in this single Bash subshell. Replace the prompt placeholder with the same rubric, file contents, and review-only instructions given to Reviewer A. The order remains Codex, Gemini, Antigravity, then Claude. The Claude fallback creates no prompt file.
Build the reviewer prompt (identical rubric + instructions as Reviewer A) and write it to a unique temp file:
```bash
PROMPT_FILE=$(mktemp /tmp/santa-reviewer-b-XXXXXX.txt)
cat > "$PROMPT_FILE" << 'EOF'
(
set -e
set -o pipefail
REVIEWER_WRAPPER="$HOME/.claude/bin/codeagent-wrapper"
REVIEWER_ROLE="$HOME/.claude/.ccg/prompts/antigravity/reviewer.md"
if command -v codex >/dev/null 2>&1; then
REVIEWER_BACKEND=codex
elif command -v gemini >/dev/null 2>&1; then
REVIEWER_BACKEND=gemini
elif [ -x "$REVIEWER_WRAPPER" ] && [ -f "$REVIEWER_ROLE" ] && [ -r "$REVIEWER_ROLE" ]; then
REVIEWER_BACKEND=antigravity
else
printf '%s\n' 'CLAUDE_FALLBACK'
exit 0
fi
cleanup_reviewer_prompt() {
reviewer_status=$?
trap - EXIT
if ! rm -f -- "$PROMPT_FILE"; then
printf '%s\n' 'Could not remove reviewer prompt; remove the private temp file before continuing.' >&2
if [ "$reviewer_status" -eq 0 ]; then reviewer_status=1; fi
fi
exit "$reviewer_status"
}
umask 077
PROMPT_FILE=$(mktemp "${TMPDIR:-/tmp}/santa-reviewer-b.XXXXXX")
trap cleanup_reviewer_prompt EXIT
trap 'exit 129' HUP
trap 'exit 130' INT
trap 'exit 143' TERM
cat > "$PROMPT_FILE" << 'EOF'
... full rubric + file contents + reviewer instructions ...
EOF
case "$REVIEWER_BACKEND" in
codex)
codex exec --sandbox read-only -m gpt-5.4 -C "$(pwd)" - < "$PROMPT_FILE"
;;
gemini)
REVIEWER_PROMPT=$(cat "$PROMPT_FILE")
gemini -p "$REVIEWER_PROMPT" -m gemini-2.5-pro
;;
antigravity)
{
printf 'ROLE_FILE: %s\n' "$REVIEWER_ROLE"
cat "$PROMPT_FILE"
} | "$REVIEWER_WRAPPER" --backend antigravity - "$PWD"
;;
esac
)
```
Use the first available CLI:
The subshell removes its prompt on success, backend or prompt-write failure, and handled HUP/INT/TERM signals, preserving the original failure status. A cleanup-only failure also returns nonzero. Forced termination such as SIGKILL cannot run cleanup; inspect private temp files after an interrupted process. A failed external invocation is not an approval and must not silently become a fallback or proceed to the verdict gate.
**Codex CLI** (if installed)
```bash
codex exec --sandbox read-only -m gpt-5.4 -C "$(pwd)" - < "$PROMPT_FILE"
rm -f "$PROMPT_FILE"
```
**Gemini CLI** (if installed and codex is not)
```bash
gemini -p "$(cat "$PROMPT_FILE")" -m gemini-2.5-pro
rm -f "$PROMPT_FILE"
```
**Claude Agent fallback** (only if neither `codex` nor `gemini` is installed)
Launch a second Claude Agent (subagent_type: `code-reviewer`, model: `opus`). Log a warning that both reviewers share the same model family — true model diversity was not achieved but context isolation is still enforced.
**Claude Agent fallback:** If the subshell prints `CLAUDE_FALLBACK`, launch a second Claude Agent (subagent_type: `code-reviewer`, model: `opus`) with the rubric and file contents directly. The marker is a dispatch instruction, not a review verdict. Log that both reviewers share the same model family; keep their contexts separate.
In all cases, the reviewer must return the same structured JSON verdict as Reviewer A.
@@ -166,9 +199,9 @@ Result: [PUSHED / ESCALATED TO USER]
## Notes
- Reviewer A (Claude Opus) always runs — guarantees at least one strong reviewer regardless of tooling.
- Model diversity is the goal for Reviewer B. GPT-5.4 or Gemini 2.5 Pro gives true independence — different training data, different biases, different blind spots. The Claude-only fallback still provides value via context isolation but loses model diversity.
- Strongest available models are used: Opus for Reviewer A, GPT-5.4 or Gemini 2.5 Pro for Reviewer B.
- External reviewers run with `--sandbox read-only` (Codex) to prevent repo mutation during review.
- Model diversity is the goal for Reviewer B. Record the actual model/provider reported by the selected backend; the Antigravity wrapper name alone does not establish a different model family. The Claude-only fallback provides context isolation but loses model diversity.
- Use each backend's maintained model-selection contract. Do not pin a transient Antigravity model ID in this workflow.
- Request review-only output from every reviewer. Codex uses `--sandbox read-only`; Gemini and Antigravity permissions depend on their installed runtime configuration. The CCG wrapper invocation alone does not guarantee a read-only sandbox.
- Fresh reviewers each round prevents anchoring bias from prior findings.
- The rubric is the most important input. Tighten it if reviewers rubber-stamp or flag subjective style issues.
- Commits happen on NAUGHTY rounds so fixes are preserved even if the loop is interrupted.