1. Input Parsing

1.1 Required Parameters

target_file

.md

.yaml

.yml

.json

.toml

.txt

1.2 Optional Parameters

• target_score=N

  (default 90) — Minimum Overall score required to exit the loop (percentage, 0-100). If the score is lower than this value, the iteration continues.

• max_rounds=N

  (default 10) — Upper limit of maximum iteration rounds. If the target is not reached when the limit is hit, stop (no error thrown, output a closing report indicating the target is not met).

• review_focus="..."

  (default empty) — Focus of the review. Examples:

  review_focus="schema instruction"

  or

  review_focus="OpenCode skill body"

  . If left empty, the oracle uses a general baseline.

2. Global Rules

2.1 Fast Fail Rule

task()

tool()

skill()

1. Immediately stop the current round
2. Report "🔴 [Round Name] Interrupted: [Call Name] call failed. [Error Message]"
3. Wait for user intervention

2.2 Scoring System (Percentage)

2.3 target_score Mode

• target_score >= 95

  is considered "Strict Mode" (each refinement requires higher ROI)

• target_score < 80

  is considered "Lenient Mode" (oracle will focus more on blocking issues)

• target_score < 0

  or

  > 100

  is considered invalid, stop and prompt the user

3. R0: Baseline Assessment

task(subagent_type="oracle", prompt="
You are the Oracle Review Agent, responsible for running a 5-dimensional quality assessment (baseline assessment) on <target_file>.

**Task**:

1. Read this skill body (§2.2 5-dimensional definitions + §5 Output Format + §2.3 target_score mode — all are authoritative sources from the skill body)
2. Read <target_file> (the file to be reviewed)
3. Output 5-dimensional scores + 3 types of findings + Quick wins according to §5 Output Format

**Target Score** (percentage): <target_score>, < 80 means lenient / ≥95 means strict / default 90.
   ")

4. R1+: Refinement-Review Loop

4.1 Loop Entry

• 🔴 Must fix (this round)
• 🟡 Fix based on ROI (this round, e.g., changes ≤ 3 lines)
• ⚪ Select 1-2 items with the highest ROI to fix (this round)

4.2 Apply Refinements

target_file

1. Read to confirm the current line numbers before modification
2. Use Edit to replace (exact match, no fuzzy replacement)
3. After modification, run verification such as

   bun test

   /

   openspec schema validate

   via

   bash

   (if the target file is associated with a project that has these tests). If no test suite is available (e.g., reviewing README/standalone .md), skip this step and only rely on oracle re-review as the quality gate.

4.3 Re-extract + Re-review

1. (No need to re-read target_file — the oracle automatically re-reads it every round; if there are changes, directly proceed to the next round of review)
2. Re-run the oracle (same task() call as §3, reuse session_id to maintain cross-round context). Specific mechanism: Extract the

   session_id

   field (OpenCode standard) from the return value of R0's

   task()

   , and pass it as a parameter in subsequent

   task(subagent_type="oracle", session_id=<extracted value>, prompt=...)

   calls. Do not omit session_id — omitting it will cause the oracle to lose known cross-round findings during each re-review, leading to R(N) score fluctuations.
3. Parse the new score, record it as R(N)

4.4 Loop Exit Conditions (Check before each round of refinement)

• Overall × 20 ≥ target_score → Exit the loop, proceed to §5 Closing
• Round count ≥ max_rounds → Force stop, proceed to §5 Closing (indicate target not met + remaining findings)
• Otherwise → Proceed to §4.1 for the next round

5. Closing Report

## score-prompt Closing Report

**Target File**: <target_file>
**Target Score**: <target_score>
**Actual Score**: <R(N) Overall × 20>/100
**Iteration Rounds**: R0 → R(N)
**Time Spent**: <estimated minutes>

### Score Evolution

| Round | Overall | Δ   | Refinement Summary |
| ----- | ------- | --- | ------------------ |
| R0    | 80/100  | —   | Baseline assessment |
| R1    | 88/100  | +8  | Fixed 3 🟡 + 2 ⚪ |
| R2    | 92/100  | +4  | Fixed remaining 1 🟡 |
| ...   |         |     |                    |

### Target Status

✅ Target score reached (≥ 90) / ❌ Target not reached (N remaining findings)

### Remaining Findings (if any)

- 🟡 L45: ...
- ⚪ L78: ...

### Follow-up Recommendations

- To achieve a higher score, re-run this skill with target_score=95
- Items marked [Low ROI] in remaining findings can be handled in future reviews

6. Error Handling

• Target file does not exist → Immediately report "🔴 target_file does not exist: <path>" and stop
• Unsupported file type → Report "🔴 Unsupported file type: <ext>, this skill supports .md/.yaml/.yml/.json/.toml/.txt" and stop
• Oracle agent unavailable → Report "🔴 Oracle agent is unavailable, this skill relies heavily on oracle for quality review" and stop
• Target not reached within max_rounds → No error thrown, output closing report (mark ❌ target not met), return control
• Invalid target_score → Report "🔴 target_score must be between 0 and 100" and stop ",