A self-improving RLM agent for coding workflows and long-running autonomous tasks.
Rust
fix(coding-agent): derive refine output budgets from the model (#598)
Seth Karten2c651cee7e7f
diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.mdindex fae19b0624..884f59eca5 100644--- a/packages/coding-agent/CHANGELOG.md+++ b/packages/coding-agent/CHANGELOG.md@@ -2,6 +2,10 @@## [Unreleased]+### Fixed++- Fixed `/refine` failing with an opaque JSON parse error when the refiner exceeded a fixed 4096-token output cap; output budgets now derive from the selected model, and a truncated reply reports the exhausted budget directly.+## [0.5.0] - 2026-08-03### Breaking Changesdiff --git a/packages/coding-agent/src/core/refinement/refinement.ts b/packages/coding-agent/src/core/refinement/refinement.tsindex 589d5259c8..c7ce97ee6e 100644--- a/packages/coding-agent/src/core/refinement/refinement.ts+++ b/packages/coding-agent/src/core/refinement/refinement.ts@@ -184,6 +184,26 @@ Return JSON only:"instructions": "optional concise instructions for /refine if shouldRefine is true"}`;+/**+ * Output budgets are derived from the selected model instead of fixed literals.+ * /refine input scales with harness size (entry overview, refinement history, and+ * the trajectory slice), so a constant output cap silently truncates exactly the+ * large multi-edit proposals that matter most. Math.min keeps small models honest.+ */+const REFINEMENT_MAX_OUTPUT_TOKENS = 32_000;+const AUTO_REFINE_REVIEW_MAX_OUTPUT_TOKENS = 4_096;++const TRUNCATED_JSON_ERROR =+ "the model stopped before completing its JSON object. This usually means the output budget was exhausted; retry with a smaller request.";++function refinementMaxOutputTokens(model: Model<any>): number {+ return Math.min(model.maxTokens, REFINEMENT_MAX_OUTPUT_TOKENS);+}++function autoRefineReviewMaxOutputTokens(model: Model<any>): number {+ return Math.min(model.maxTokens, AUTO_REFINE_REVIEW_MAX_OUTPUT_TOKENS);+}+function now(): string {return new Date().toISOString();}@@ -541,19 +561,71 @@ function historyForPrompt(history: RefinementResult[]): string {.join("\n\n");}+/**+ * Whether a JSON candidate ends mid-value: an unterminated string, or unclosed+ * objects/arrays. A reply cut off by an exhausted output budget is incomplete in+ * this sense, while a complete-but-malformed reply is balanced. Brace slicing can+ * also produce a balanced fragment, so callers treat "balanced" as malformed.+ */+function isIncompleteJson(candidate: string): boolean {+ let depth = 0;+ let inString = false;+ let escaped = false;+ for (const char of candidate) {+ if (escaped) {+ escaped = false;+ continue;+ }+ if (inString) {+ if (char === "\\") escaped = true;+ else if (char === '"') inString = false;+ continue;+ }+ if (char === '"') inString = true;+ else if (char === "{" || char === "[") depth++;+ else if (char === "}" || char === "]") depth--;+ }+ return inString || depth > 0;+}++function parseJsonCandidate(candidate: string): unknown {+ try {+ return JSON.parse(candidate);+ } catch (error) {+ // A truncated reply and a malformed one both fail here, and JSON.parse+ // describes the fragment rather than the cause. Name the cause instead.+ if (isIncompleteJson(candidate)) {+ throw new Error(TRUNCATED_JSON_ERROR);+ }+ throw new Error(`the model did not return valid JSON: ${error instanceof Error ? error.message : String(error)}`);+ }+}+function extractJsonObject(text: string): unknown {const trimmed = text.trim();if (trimmed.startsWith("{") && trimmed.endsWith("}")) {- return JSON.parse(trimmed);+ // A reply truncated after a nested closing brace still looks well-formed+ // here, so this path needs the same diagnosis as the slicing fallback.+ return parseJsonCandidate(trimmed);}const fenced = trimmed.match(/```(?:json)?\s*([\s\S]*?)```/);if (fenced) {- return JSON.parse(fenced[1].trim());+ return parseJsonCandidate(fenced[1].trim());}+ // Brace slicing recovers JSON wrapped in prose. On a reply truncated inside the+ // edits array it slices to an earlier edit's closing brace, so a failure here+ // is diagnosed against the original text rather than the balanced fragment.const start = trimmed.indexOf("{");const end = trimmed.lastIndexOf("}");if (start !== -1 && end > start) {- return JSON.parse(trimmed.slice(start, end + 1));+ try {+ return JSON.parse(trimmed.slice(start, end + 1));+ } catch {+ return parseJsonCandidate(trimmed.slice(start));+ }+ }+ if (isIncompleteJson(trimmed)) {+ throw new Error(TRUNCATED_JSON_ERROR);}throw new Error("Refiner did not return a JSON object");}@@ -844,12 +916,15 @@ export async function planRefinement(systemPrompt: REFINEMENT_SYSTEM_PROMPT,messages: [{ role: "user", content: [{ type: "text", text: userPrompt }], timestamp: Date.now() }],},- { maxTokens: 4096, signal, apiKey, headers },+ { maxTokens: refinementMaxOutputTokens(model), signal, apiKey, headers },);if (response.stopReason === "error") {throw new Error(`Refinement failed: ${response.errorMessage || "Unknown error"}`);}+ if (response.stopReason === "length") {+ throw new Error(`Refinement failed: ${TRUNCATED_JSON_ERROR}`);+ }const text = response.content.filter((content): content is { type: "text"; text: string } => content.type === "text")@@ -907,11 +982,14 @@ ${conversationText}systemPrompt: AUTO_REFINE_REVIEW_SYSTEM_PROMPT,messages: [{ role: "user", content: [{ type: "text", text: userPrompt }], timestamp: Date.now() }],},- { maxTokens: 1024, signal, apiKey, headers },+ { maxTokens: autoRefineReviewMaxOutputTokens(model), signal, apiKey, headers },);if (response.stopReason === "error") {throw new Error(`Auto-refine review failed: ${response.errorMessage || "Unknown error"}`);}+ if (response.stopReason === "length") {+ throw new Error(`Auto-refine review failed: ${TRUNCATED_JSON_ERROR}`);+ }const text = response.content.filter((content): content is { type: "text"; text: string } => content.type === "text").map((content) => content.text)diff --git a/packages/coding-agent/test/refinement.test.ts b/packages/coding-agent/test/refinement.test.tsindex 79966a475c..3b2f7037cf 100644--- a/packages/coding-agent/test/refinement.test.ts+++ b/packages/coding-agent/test/refinement.test.ts@@ -1030,8 +1030,9 @@ describe("harness refinement", () => {"During a local refinement, global entries are read-only context: never propose update or delete edits for them",),});+ // Budget is derived from the model (8192) rather than a fixed literal.expect(completeSimpleMock.mock.calls[0][2]).toMatchObject({- maxTokens: 4096,+ maxTokens: 8192,apiKey: "api-key",headers: { "x-test-header": "1" },});@@ -1047,6 +1048,77 @@ describe("harness refinement", () => {);});+ it("caps the refinement output budget by the model's own maxTokens", async () => {+ const state = loadHarnessState(makeTempDir());+ completeSimpleMock.mockResolvedValueOnce(+ assistantText(JSON.stringify({ summary: "s", rationale: "r", expectedOutcome: "o", edits: [] })),+ );+ // A large model must receive the policy ceiling, not its full output width.+ const wideModel = { ...createRefineModel(false), maxTokens: 128_000 };++ await refineHarness([], state, [], wideModel, "api-key", {});++ expect(completeSimpleMock.mock.calls[0][2]).toMatchObject({ maxTokens: 32_000 });+ });++ it("reports an exhausted output budget instead of a JSON parse error", async () => {+ const state = loadHarnessState(makeTempDir());+ // A reply truncated inside the edits array: brace slicing would otherwise+ // recover a fragment and surface an opaque JSON.parse message.+ const truncated = `{+ "summary": "s",+ "rationale": "r",+ "expectedOutcome": "o",+ "edits": [+ { "action": "create", "kind": "memory", "id": "a", "title": "t", "content": "first" },+ { "action": "create", "kind": "memory", "id": "b", "title": "t2", "content": "second`;+ completeSimpleMock.mockResolvedValueOnce({ ...assistantText(truncated), stopReason: "length" });++ await expect(refineHarness([], state, [], createRefineModel(false), "api-key", {})).rejects.toThrow(+ /output budget was exhausted/,+ );+ });++ it("rejects a truncated proposal that never reports a length stop reason", async () => {+ const state = loadHarnessState(makeTempDir());+ const truncated = `{+ "summary": "s",+ "edits": [+ { "action": "create", "kind": "memory", "id": "a", "title": "t", "content": "first" },+ { "action": "create", "kind": "memory", "id": "b", "title": "t2", "content": "second`;+ completeSimpleMock.mockResolvedValueOnce(assistantText(truncated));++ await expect(refineHarness([], state, [], createRefineModel(false), "api-key", {})).rejects.toThrow(+ /stopped before completing its JSON object/,+ );+ });++ it("reports truncation when a JSON-only reply is cut after a nested closing brace", async () => {+ const state = loadHarnessState(makeTempDir());+ // Ends on "}" so it takes the startsWith/endsWith fast path rather than+ // the brace-slicing fallback, but is still an incomplete object.+ const truncated = `{+ "summary": "s",+ "edits": [+ { "action": "create", "kind": "memory", "id": "a", "title": "t", "content": "first" }`;+ completeSimpleMock.mockResolvedValueOnce(assistantText(truncated));++ await expect(refineHarness([], state, [], createRefineModel(false), "api-key", {})).rejects.toThrow(+ /stopped before completing its JSON object/,+ );+ });++ it("reports malformed JSON as invalid rather than as an exhausted budget", async () => {+ const state = loadHarnessState(makeTempDir());+ // Complete and balanced, but not valid JSON: this is a model formatting+ // failure, not a truncation, and must not blame the output budget.+ completeSimpleMock.mockResolvedValueOnce(assistantText('Here is the result: {"edits": [oops]}'));++ await expect(refineHarness([], state, [], createRefineModel(false), "api-key", {})).rejects.toThrow(+ /did not return valid JSON/,+ );+ });+it("rolls back created, updated, and deleted entries from refinement history", async () => {const state = loadHarnessState(makeTempDir());seedEntry(state, "memory", "kept_memory");