diff --git a/docs/reference/trajectory-assessment.md b/docs/reference/trajectory-assessment.md index b51c254986..7d618ecee9 100644 --- a/docs/reference/trajectory-assessment.md +++ b/docs/reference/trajectory-assessment.md @@ -304,14 +304,18 @@ leading signs, precision and percent suffixes stay exact, while spelled-out unit are explained in the output language. An adjacent topic cannot resolve an ambiguous referent without supporting field-bound evidence. This remains model guidance; the exact-literal qualifier and original-text fallback are unchanged. -The curator selects it in a new versioned input through the existing analysis -action. Earlier encodings retain their exact prompts, identities and artifacts; -reading or refreshing them never starts or upgrades analysis. +New CLI preparations and default pack registrations select it when a captured +display snapshot has no explicit encoding. The normalized input is saved with +that choice before analysis. Explicit encodings/configurations and frozen inputs +retain their exact prompts, identities and artifacts; reading or refreshing them +never starts or upgrades analysis. Translation quality belongs to the existing +analysis generation, with original text available for comparison. 显式数值编码 v3 增加生成时的逐项忠实性核对：正负号、精度和百分号按允许的原始写法 保留，文字单位用目标语言解释；不能仅凭邻近主题补全含糊所指。这仍是模型指导， -既有字面校验和原文回退不放宽。由原输入整理者通过既有分析入口明确生成新版本， -旧编码的提示、身份和产物保持，阅读与刷新不升级或执行分析。 +既有字面校验和原文回退不放宽。新CLI准备和默认pack登记在快照未指定编码时采用v3， +先将选择保存到版本化输入，再由既有分析入口生成。显式编码/配置及历史输入的提示、 +身份和产物保持，阅读与刷新不升级或执行分析。译文质量由默认生成承担，原文仍可核对。 Translated prose is applied to decoded narrative fields; it is never serialized and reparsed as an Explore decision. Owner-format labels or list markers inside diff --git a/examples/personal-workspace-browser/research-history.mjs b/examples/personal-workspace-browser/research-history.mjs index df02cf6186..600f619373 100644 --- a/examples/personal-workspace-browser/research-history.mjs +++ b/examples/personal-workspace-browser/research-history.mjs @@ -719,6 +719,8 @@ export const researchHistoryScenario = { await unavailablePanel.getByRole('status').filter({hasText: '分析来源未就绪'}).waitFor(); assert.equal(await unavailablePanel.locator('header p').count(), 0, 'An empty reader does not repeat analysis instructions'); await unavailablePanel.press('Escape'); + await page.waitForFunction(element => element === document.activeElement, + await history.getByRole('button', {name: '深度分析', exact: true}).elementHandle()); assert(await history.getByRole('button', {name: '深度分析', exact: true}).evaluate(e => e === document.activeElement)); assert.equal(analysisWrites, 0, 'Unavailable analysis, help and recovery are reads'); await page.setViewportSize({width: 1280, height: 900}); diff --git a/loopx/capabilities/benchmark_toolkit/trajectory_assessment.py b/loopx/capabilities/benchmark_toolkit/trajectory_assessment.py index c31ea17b04..69ad903a00 100644 --- a/loopx/capabilities/benchmark_toolkit/trajectory_assessment.py +++ b/loopx/capabilities/benchmark_toolkit/trajectory_assessment.py @@ -49,6 +49,7 @@ def prepare( "benchmark.trajectory.prepare", { "input": json.loads(input_path.read_text(encoding="utf-8")), + "generation_defaults": output_schema_revision == 2, "config": {"model": model, "effort": effort, "service_tier": service_tier, **({"output_schema_revision": output_schema_revision} if output_schema_revision is not None else {}), **({"research_synthesis": True, "research_synthesis_revision": 2} if research_synthesis else {})}, diff --git a/loopx/capabilities/benchmark_toolkit/trajectory_assessment_store.py b/loopx/capabilities/benchmark_toolkit/trajectory_assessment_store.py index 218254c848..3abf690a69 100644 --- a/loopx/capabilities/benchmark_toolkit/trajectory_assessment_store.py +++ b/loopx/capabilities/benchmark_toolkit/trajectory_assessment_store.py @@ -351,6 +351,7 @@ def register_pack( { "input": json.loads(input_path.read_text(encoding="utf-8")), "config": config or DEFAULT_CONFIG, + "generation_defaults": not config, }, ) record = { @@ -571,7 +572,7 @@ def _continuation(runtime_root, owner_goal_id, record, prepared, result, languag if base_id is None: return None, None try: - _, previous = _resolve(runtime_root, owner_goal_id, base_id) + owner, previous = _resolve(runtime_root, owner_goal_id, base_id) base = inspect_pack(runtime_root=runtime_root, owner_goal_id=owner_goal_id, pack_id=base_id, output_language=language, include_continuation=False) old = _prepared(previous, language) if language is not None else _prepared(previous) @@ -589,9 +590,24 @@ def _continuation(runtime_root, owner_goal_id, record, prepared, result, languag except (OSError, ValueError): readable = False usable = readable and base["result"] is not None and base["reason_code"] != "method_changed" - prior = {**old, **({"output": base["result"]} if usable else {})} + # Readback has restored display literals. Requalify the exact successful + # attempt's provider response, including when a later retry failed. + output = None + if usable: + directory = _job(owner, old, old["analysis_key"]) + if base.get("result_job"): + directory = directory / f"attempt-{base['result_job']['attempt']}" + try: + output = json.loads((directory / "model-output.json").read_text(encoding="utf-8")) + except (OSError, ValueError): + usable = False + prior = {**old, **({"output": output} if usable else {})} + # Compare analytical rows without resending the restored display sidecar; + # the raw response above supplies the versioned display data for reuse. + comparison_output = {key: result[key] for key in ("assessments", "limitations", "episode_questions", "synthesis") + if key in result} if result is not None else None projection = effect_runtime_result("benchmark.trajectory.continuation", { - **prepared, "previous": prior, **({"output": result} if result is not None else {}), + **prepared, "previous": prior, **({"output": comparison_output} if result is not None else {}), }) return { **projection, diff --git a/loopx/control_plane/capabilities/trajectory_assessment.ts b/loopx/control_plane/capabilities/trajectory_assessment.ts index 10b27f326c..78b3c8bd43 100644 --- a/loopx/control_plane/capabilities/trajectory_assessment.ts +++ b/loopx/control_plane/capabilities/trajectory_assessment.ts @@ -33,7 +33,7 @@ function metric(v: unknown): TrajectoryMetric { if (!TRAJECTORY_METRICS.includes(v as TrajectoryMetric)) throw new EffectRuntimeRequestError('unsupported metric'); return v as TrajectoryMetric; } -export function parseTrajectoryInput(raw: unknown): TrajectoryInput { +export function parseTrajectoryInput(raw: unknown, generationDefaults = false): TrajectoryInput { const x = object(raw, 'input'); if (x.schema_version !== 'trajectory_assessment_input_v1') throw new EffectRuntimeRequestError('unsupported input schema'); const s = object(x.scope, 'scope'), selection = object(x.selection, 'selection'); @@ -67,7 +67,9 @@ export function parseTrajectoryInput(raw: unknown): TrajectoryInput { if (selection.exhaustive && population !== episodes.length) throw new EffectRuntimeRequestError('exhaustive requires a complete known population'); const display = x.display_snapshot === undefined ? undefined : object(x.display_snapshot, 'display_snapshot'); if (display && display.schema_version !== 'research_display_snapshot_v1') throw new EffectRuntimeRequestError('unsupported display snapshot'); - const encoding = display?.translation_encoding; + const requestedEncoding = display?.translation_encoding; + const encoding = display && generationDefaults && requestedEncoding === undefined + ? 'source_text_aliases_with_literal_values_v3' : requestedEncoding; if (encoding !== undefined && encoding !== 'source_text_aliases_v1' && encoding !== 'source_text_aliases_with_literals_v1' && encoding !== 'source_text_aliases_with_literal_values_v1' && encoding !== 'source_text_aliases_with_literal_values_v2' && encoding !== 'source_text_aliases_with_literal_values_v3') { @@ -242,7 +244,9 @@ const synthesisSchema = {type: 'object', additionalProperties: false, }}; export function prepareTrajectoryAssessment(params: Record) { - const input = parseTrajectoryInput(params.input); + // Only new writer requests select defaults. Frozen inputs are re-prepared on + // every read and must retain their original prompt and analysis identity. + const input = parseTrajectoryInput(params.input, params.generation_defaults === true); const config = object(params.config, 'config'); if (config.research_synthesis !== undefined && typeof config.research_synthesis !== 'boolean') { throw new EffectRuntimeRequestError('research_synthesis must be boolean'); diff --git a/tests/capabilities/test_trajectory_assessment_store.py b/tests/capabilities/test_trajectory_assessment_store.py index b542afd06f..78a00af6b0 100644 --- a/tests/capabilities/test_trajectory_assessment_store.py +++ b/tests/capabilities/test_trajectory_assessment_store.py @@ -583,10 +583,12 @@ def test_missing_source_blocks_new_execution_and_leaves_registration_visible(pac execute_pack(**operation, codex="must-not-run") -@pytest.mark.parametrize("encoding", [None, "source_text_aliases_v1", "source_text_aliases_with_literals_v1", "source_text_aliases_with_literal_values_v1", "source_text_aliases_with_literal_values_v2", "source_text_aliases_with_literal_values_v3"]) +@pytest.mark.parametrize("encoding", ["default", None, "source_text_aliases_v1", "source_text_aliases_with_literals_v1", "source_text_aliases_with_literal_values_v1", "source_text_aliases_with_literal_values_v2", "source_text_aliases_with_literal_values_v3"]) @pytest.mark.parametrize("context", [False, True]) def test_full_history_display_uses_the_same_job_and_preserves_original_versions(pack, tmp_path, encoding, context): - scope, _, _, input_path, sources, _ = pack + from loopx.capabilities.benchmark_toolkit.trajectory_assessment import prepare + + scope, _, initial, input_path, sources, _ = pack data = json.loads(input_path.read_text()) history = {"ok": True, "goal_id": "solver", "total": 2, "next_cursor": None, "items": [ dict(schema_version="loopx_explore_result_event_v0", event_id="question-old", result_id="question", goal_id="solver", event_kind="node", @@ -594,19 +596,25 @@ def test_full_history_display_uses_the_same_job_and_preserves_original_versions( dict(schema_version="loopx_explore_result_event_v0", event_id="question-new", result_id="question", goal_id="solver", event_kind="node", node_kind="question", recorded_at="2026-01-02T00:00:00Z", title="Does V19 help?", summary="Still unknown.", status="blocked"), ]} - if encoding in {"source_text_aliases_with_literal_values_v1", "source_text_aliases_with_literal_values_v2", "source_text_aliases_with_literal_values_v3"}: + if encoding in {"default", "source_text_aliases_with_literal_values_v1", "source_text_aliases_with_literal_values_v2", "source_text_aliases_with_literal_values_v3"}: history["items"][0]["summary"] = "Trial4failed2checks" text = json.dumps(history, ensure_ascii=False) (sources / "history.json").write_text(text) data["sources"].append(dict(id="history", artifact="history.json", artifact_sha256=hashlib.sha256(text.encode()).hexdigest(), first_line=1, text=text)) data["display_snapshot"] = dict(schema_version="research_display_snapshot_v1", source_id="history", source_namespace="source-a") - if encoding is not None: + if encoding is not None and encoding != "default": data["display_snapshot"]["translation_encoding"] = encoding context_citation = dict(source_id="source", first_line=2, last_line=2) if context: data["display_snapshot"]["contexts"] = [dict(event_id="question-old", field="summary", citations=[context_citation])] input_path.write_text(json.dumps(data)) - registered = register_pack(**scope, input_path=input_path, source_root=sources, title="Whole history · synthetic") + registered = register_pack(**scope, input_path=input_path, source_root=sources, title="Whole history · synthetic", + config=None if encoding == "default" else initial["config"]) + if encoding == "default": + prepared = prepare(input_path, model="gpt-6.1-sol", effort="max", service_tier="fast") + assert prepared["input"]["display_snapshot"]["translation_encoding"] == "source_text_aliases_with_literal_values_v3" + assert prepared["analysis_key"] == registered["analysis_key"] + assert json.loads(input_path.read_text()) == data current = inspect_pack(**scope, pack_id=registered["pack_id"], output_language="zh") assert current["display_request"]["eligible"] == 4 and current["job"] is None provider = tmp_path / "display-provider" @@ -640,7 +648,7 @@ def translation(field): read_citation(**operation, source_id="history", first_line=1, last_line=1) else: assert not any("context_citations" in field for field in result["result"]["display"]["fields"]) - if encoding in {"source_text_aliases_with_literal_values_v1", "source_text_aliases_with_literal_values_v2", "source_text_aliases_with_literal_values_v3"}: + if encoding in {"default", "source_text_aliases_with_literal_values_v1", "source_text_aliases_with_literal_values_v2", "source_text_aliases_with_literal_values_v3"}: assert any(field["text"] == "第4次尝试有2项检查未通过（Trial4failed2checks）。" for field in result["result"]["display"]["fields"]) assert result["job"]["usage"]["input_tokens"] == 19 # one shared attempt, not four new jobs @@ -657,6 +665,25 @@ def translation(field): assert prior["job"]["status"] == "failed" and prior["job"]["attempt"] == 2 assert prior["result_attempt"] == 1 and prior["result_job"]["usage"]["input_tokens"] == 19 assert prior["result"] == result["result"] + if encoding == "default": + # A failed retry keeps the prior successful response reusable. Readback + # contains restored literals, so it cannot substitute for raw output. + successor_input = json.loads(input_path.read_text()) + successor_input["scope"]["revision"] = "successor-capture" + successor_path = tmp_path / "successor.json" + successor_path.write_text(json.dumps(successor_input)) + successor = register_pack(**scope, input_path=successor_path, source_root=sources, + title="Same evidence, new capture", base_pack_id=registered["pack_id"]) + successor_view = inspect_pack(**scope, pack_id=successor["pack_id"], output_language="zh") + assert successor_view["continuation"]["base_available"] and successor_view["continuation"]["can_reuse"] + assert successor_view["job"] is None + reuse = execute_pack(**scope, pack_id=successor["pack_id"], analysis_key=successor_view["analysis_key"], + output_language="zh", codex="must-not-run") + assert reuse["status"] == "succeeded" and reuse["reused"] + reused = inspect_pack(**scope, pack_id=successor["pack_id"], output_language="zh") + assert reused["result"]["display"] == prior["result"]["display"] + assert reused["result"]["metrics"] == prior["result"]["metrics"] + assert reused["continuation"]["verdict_changes"] == [] provider.write_text(successful_provider) assert execute_pack(**operation, codex=str(provider), retry=True)["status"] == "succeeded" result = inspect_pack(**scope, pack_id=registered["pack_id"], output_language="zh") diff --git a/tests/control_plane_ts/trajectory_assessment.test.ts b/tests/control_plane_ts/trajectory_assessment.test.ts index d06fa179d0..1a7651bcf4 100644 --- a/tests/control_plane_ts/trajectory_assessment.test.ts +++ b/tests/control_plane_ts/trajectory_assessment.test.ts @@ -51,6 +51,34 @@ const contextInput = () => { const translatedOutput = (p = displayPrepared()) => ({...output(), episode_questions: [{episode_id: 'e1', question: 'A 是否有帮助？'}, {episode_id: 'e2', question: 'B 是否有帮助？'}], display_translations: JSON.parse(p.prompt.split('\n\ndisplay_fields:\n')[1]!).map((row: {field_id: string; text: string}) => ({field_id: row.field_id, text: `【译文】${row.text}`}))}); +test('new generation selects faithful display defaults without upgrading frozen or explicit inputs', () => { + for (const language of ['zh', 'en']) { + const original = contextInput(), configForLanguage = {...config, output_language: language}; + delete original.display_snapshot!.translation_encoding; + const before = structuredClone(original); + const frozen = prepareTrajectoryAssessment({input: original, config: configForLanguage}); + const next = prepareTrajectoryAssessment({input: original, config: configForLanguage, generation_defaults: true}); + assert.equal(next.input.display_snapshot!.translation_encoding, 'source_text_aliases_with_literal_values_v3'); + assert.notEqual(next.analysis_key, frozen.analysis_key); + assert.equal(next.config.output_language, language); + assert.deepEqual(next.input.sources, original.sources); + assert.deepEqual(next.input.episodes, original.episodes); + assert.deepEqual(original, before, 'Writer normalization does not modify the captured input'); + assert.deepEqual(prepareTrajectoryAssessment({input: next.input, config: next.config}), next); + assert.deepEqual(prepareTrajectoryAssessment({input: original, config: configForLanguage}), frozen); + for (const encoding of ['source_text_aliases_v1', 'source_text_aliases_with_literal_values_v2'] as const) { + const explicit = structuredClone(original); + explicit.display_snapshot!.translation_encoding = encoding; + assert.deepEqual(prepareTrajectoryAssessment({input: explicit, config: configForLanguage, generation_defaults: true}), + prepareTrajectoryAssessment({input: explicit, config: configForLanguage})); + } + const invalid = {...original, display_snapshot: {...original.display_snapshot, translation_encoding: null}}; + assert.throws(() => prepareTrajectoryAssessment({input: invalid, config: configForLanguage, generation_defaults: true}), + /unsupported display translation encoding/); + } + assert.deepEqual(prepareTrajectoryAssessment({input: input(), config, generation_defaults: true}), prepared()); +}); + test('field-bound evidence disambiguates identical notes without sharing translations or changing source authority', () => { const input = contextInput(), p = prepareTrajectoryAssessment({input, config: {...config, output_language: 'zh'}}); const rows = JSON.parse(p.prompt.split('\n\ndisplay_fields:\n')[1]!);