Skip to content

Garry's Opinionated OpenClaw/Hermes Agent Brain

TypeScript

fix(skillopt): success-reflect prompt states the edit schema (#7257)

garrytanb1384cfe8cae
diff --git a/src/core/skillopt/reflect.ts b/src/core/skillopt/reflect.ts
index 08b095564e..e807c52db1 100644
--- a/src/core/skillopt/reflect.ts
+++ b/src/core/skillopt/reflect.ts
@@ -79,15 +79,18 @@ export function describeJudges(tasks: ReadonlyArray<{ judge: Judge }>): string {
return blocks.join('\n');
}
-const FAILURE_REFLECT_SYSTEM = `You are SkillOpt's optimizer. You analyze AGENT FAILURE TRAJECTORIES and propose specific edits to a SKILL document so the agent does better next time.
-
-Output ONLY a single JSON object on one or more lines:
+/** The output format both reflect modes send (#7257): parseEdits accepts exactly these op and field names. */
+const EDIT_SCHEMA = `Output ONLY a single JSON object on one or more lines:
{"edits": [{"op": "add|replace|delete", ...}, ...]}
Edit ops:
add: {"op": "add", "anchor": "<exact heading text>", "content": "<new markdown>", "reason": "<one sentence>"}
replace: {"op": "replace", "target": "<exact text to find>", "replacement": "<new text>", "reason": "<one sentence>"}
- delete: {"op": "delete", "target": "<exact text to remove>", "reason": "<one sentence>"}
+ delete: {"op": "delete", "target": "<exact text to remove>", "reason": "<one sentence>"}`;
+
+const FAILURE_REFLECT_SYSTEM = `You are SkillOpt's optimizer. You analyze AGENT FAILURE TRAJECTORIES and propose specific edits to a SKILL document so the agent does better next time.
+
+${EDIT_SCHEMA}
Rules:
- Each edit MUST address a SPECIFIC failure pattern you observed.
@@ -100,7 +103,9 @@ Rules:
const SUCCESS_REFLECT_SYSTEM = `You are SkillOpt's optimizer. You analyze AGENT SUCCESS TRAJECTORIES and propose specific edits to a SKILL document so the agent CONSISTENTLY does what worked here.
-Output format and rules are identical to the failure-reflect mode — same {edits: [...]} shape.
+${EDIT_SCHEMA}
+
+anchor / target MUST be uniquely identifiable in the skill body (exact match). Output at MOST 8 edits.
When successes are present, look for: which rules were FOLLOWED to produce success, which rules could be MADE EXPLICIT (not yet stated, but exemplified), which anti-patterns the agent successfully AVOIDED that should be stated.
diff --git a/test/skillopt/reflect-success-schema-7257.test.ts b/test/skillopt/reflect-success-schema-7257.test.ts
new file mode 100644
index 0000000000..15689a0088
--- /dev/null
+++ b/test/skillopt/reflect-success-schema-7257.test.ts
@@ -0,0 +1,65 @@
+/**
+ * #7257 — the success-reflect call states the same edit schema as the
+ * failure-reflect call. Its system prompt used to say only that the format is
+ * "identical to the failure-reflect mode", a prompt that call never sends, so
+ * the optimizer guessed field names (find/replace) and parseEdits dropped
+ * every success-only edit.
+ */
+import { describe, expect, test } from 'bun:test';
+import { runReflect } from '../../src/core/skillopt/reflect.ts';
+import type { ChatOpts, ChatResult } from '../../src/core/ai/gateway.ts';
+import type { ScoredRollout } from '../../src/core/skillopt/types.ts';
+
+const MODEL = 'anthropic:claude-opus-4-7';
+type ChatFn = NonNullable<Parameters<typeof runReflect>[0]['chatFn']>;
+
+function scored(id: string, score: number): ScoredRollout {
+ return {
+ trajectory: {
+ task_id: id, task: `Task ${id}`, final_text: `output ${id}`, tool_calls: [],
+ usage: { input_tokens: 0, output_tokens: 0, cache_read_tokens: 0, cache_creation_tokens: 0 },
+ turns: 1, stop_reason: 'end', duration_ms: 1,
+ },
+ score,
+ };
+}
+
+function reply(text: string): ChatResult {
+ return { text, blocks: [{ type: 'text', text }], stopReason: 'end',
+ usage: { input_tokens: 1, output_tokens: 1, cache_read_tokens: 0, cache_creation_tokens: 0 }, model: MODEL, providerId: 'anthropic' };
+}
+
+async function systemsFor(successes: ScoredRollout[], failures: ScoredRollout[]): Promise<string[]> {
+ const systems: string[] = [];
+ const chatFn = (async (opts: ChatOpts) => { systems.push(String(opts.system)); return reply('{"edits":[]}'); }) as ChatFn;
+ await runReflect({ skillBodyText: '# S\n## People\n', successes, failures, rejected: [], optimizerModel: MODEL, chatFn });
+ return systems;
+}
+
+const EDIT_OPS = /Edit ops:[\s\S]*?delete:[^\n]*\n/;
+
+describe('#7257: both reflect modes send the edit schema', () => {
+ test('a success-only batch sends the op and field names parseEdits accepts', async () => {
+ const systems = await systemsFor([scored('s', 1)], []);
+ expect(systems).toHaveLength(1);
+ for (const field of ['"op"', '"anchor"', '"content"', '"target"', '"replacement"', '{"edits": [']) {
+ expect(systems[0]).toContain(field);
+ }
+ });
+
+ test('the success and failure prompts carry the identical output format and edit ops', async () => {
+ const systems = await systemsFor([scored('s', 1)], [scored('f', 0)]);
+ expect(systems).toHaveLength(2);
+ const [failure, success] = systems.map(p => EDIT_OPS.exec(p)?.[0]);
+ expect(failure).toBeDefined();
+ expect(success).toBe(failure);
+ });
+
+ test('a success-only reply written in that schema parses into edits', async () => {
+ const text = JSON.stringify({ edits: [{ op: 'replace', target: '## People', replacement: '## People and roles', reason: 'clearer' }] });
+ const chatFn = (async () => reply(text)) as ChatFn;
+ const res = await runReflect({ skillBodyText: '# S\n## People\n', successes: [scored('s', 1)], failures: [], rejected: [], optimizerModel: MODEL, chatFn });
+ expect(res.successEdits).toHaveLength(1);
+ expect(res.errors).toEqual([]);
+ });
+});