@besales/ops-framework 0.1.23 → 0.1.25

package/CHANGELOG.md

@@ -1,5 +1,18 @@
 # Changelog
 
+## 0.1.25
+
+- Stabilized Verify LLM `task-manifest.json` input by stripping volatile Check telemetry, timestamps and verbose loop history from verifier packs.
+- Kept previous `verify.md` context compact in external Verify packs so verifier runs do not re-ingest large internal verify excerpts.
+- Raised `standard_plus` context cap to keep ~27k compact Verify packs out of slow strict mode.
+- Narrowed external Verify policy: local scratch DB migrations, fixture imports and bounded local-only backfills can close with internal Verify when evidence covers required gates; external CLI remains for production/real-data/destructive/security/financial/broad-risk work or explicit human request.
+
+## 0.1.24
+
+- Raised `standard_plus` Check/Verify context cap so near-cap plans stay in compact mode instead of jumping to slow strict context.
+- Stabilized Check LLM `taskManifest` input by removing volatile `llmInput`, `lastCheckResult`, timestamps and verbose loop history from the prompt payload.
+- Reduced duplicate external Check reruns after `task-manifest.json` refreshes by keeping prompt input stable when plan, memory and risk inputs have not changed.
+
 ## 0.1.23
 
 - Added deterministic import/ingestion planning gates for representative real fixtures, raw downstream metadata extraction and explicit duplicate-import behavior.

package/bin/lib/llm-input-pack-utils.mjs

@@ -12,8 +12,8 @@ export const LLM_CONTEXT_MODES = ['fast', 'standard', 'standard_plus', 'strict']
 export const LLM_CONTEXT_CAPS = {
   fast: 8000,
   standard: 20000,
-  standard_plus: 26000,
-  strict: 45000,
+  standard_plus: 34000,
+  strict: 50000,
 };
 
 const TOKEN_ESTIMATE_CHARS_PER_TOKEN = 1.8;
@@ -225,7 +225,7 @@ export function buildCheckerLlmInputPack({
     relevantPlaybooks: selectedMode === 'strict'
       ? renderRelevantPlaybooks(readRelevantPlaybooks(checkContext.riskTriggers || []), { mode: 'strict' })
       : renderRelevantPlaybooks(readRelevantPlaybooks(checkContext.riskTriggers || []), { mode: 'compact' }),
-    taskManifest,
+    taskManifest: stableTaskManifestForCheck(taskManifest),
     projectMemory: compactProjectMemory(projectMemory, selectedMode),
     taskArtifacts: artifacts,
     outputContract: {
@@ -249,6 +249,61 @@ export function buildCheckerLlmInputPack({
   return withPackMetadata(input, selectedMode);
 }
 
+export function stableTaskManifestForCheck(taskManifest) {
+  let parsed = null;
+  if (typeof taskManifest === 'string') {
+    try {
+      parsed = JSON.parse(taskManifest);
+    } catch {
+      return taskManifest;
+    }
+  } else if (taskManifest && typeof taskManifest === 'object' && !Array.isArray(taskManifest)) {
+    parsed = taskManifest;
+  }
+  if (!parsed || typeof parsed !== 'object' || Array.isArray(parsed)) {
+    return taskManifest || '';
+  }
+
+  const stable = {
+    schemaVersion: parsed.schemaVersion,
+    taskId: parsed.taskId,
+    mode: parsed.mode,
+    phase: parsed.phase,
+    gates: parsed.gates,
+    context: {
+      planSha: parsed.context?.planSha,
+      planFingerprintVersion: parsed.context?.planFingerprintVersion,
+      memorySha: parsed.context?.memorySha,
+      riskProfile: parsed.context?.riskProfile,
+      riskTriggers: parsed.context?.riskTriggers,
+      riskWarnings: parsed.context?.riskWarnings,
+      checkContextCurrent: parsed.context?.checkContextCurrent,
+    },
+    requiredEvidenceIssues: parsed.requiredEvidenceIssues,
+    qualitySignals: parsed.qualitySignals,
+    loopDetector: {
+      threshold: parsed.loopDetector?.threshold,
+      requiresConsolidatedRemediation: parsed.loopDetector?.requiresConsolidatedRemediation,
+      repeatedReasons: parsed.loopDetector?.repeatedReasons,
+    },
+    consolidatedRemediationAccepted: parsed.consolidatedRemediationAccepted,
+    consolidatedRemediationArtifact: parsed.consolidatedRemediationArtifact,
+  };
+  return JSON.stringify(pruneUndefined(stable), null, 2);
+}
+
+function pruneUndefined(value) {
+  if (Array.isArray(value)) {
+    return value.map(pruneUndefined);
+  }
+  if (!value || typeof value !== 'object') {
+    return value;
+  }
+  return Object.fromEntries(Object.entries(value)
+    .filter(([, item]) => item !== undefined)
+    .map(([key, item]) => [key, pruneUndefined(item)]));
+}
+
 export function buildVerifierLlmInputPack({
   taskDir,
   taskId,
@@ -263,7 +318,7 @@ export function buildVerifierLlmInputPack({
         'brief.md': readTaskFile(taskDir, 'brief.md'),
         'research.md': readTaskFile(taskDir, 'research.md'),
         'plan.md': readTaskFile(taskDir, 'plan.md'),
-        'task-manifest.json': readTaskFile(taskDir, 'task-manifest.json'),
+        'task-manifest.json': stableTaskManifestForCheck(readTaskFile(taskDir, 'task-manifest.json')),
         'check.result.json': readTaskFile(taskDir, 'check.result.json'),
         'check.md': compactCheckMarkdown({
           checkMarkdown: readTaskFile(taskDir, 'check.md'),
@@ -274,7 +329,7 @@ export function buildVerifierLlmInputPack({
         'human-gate-summary.md': truncateMiddle(readTaskFile(taskDir, 'human-gate-summary.md'), 3500),
         'execution.md': readTaskFile(taskDir, 'execution.md'),
         'execution-ledger.json': compactExecutionLedger(readOptionalJson(taskDir, 'execution-ledger.json'), 'strict'),
-        'verify.md': compactVerifierMarkdown(readTaskFile(taskDir, 'verify.md'), readOptionalJson(taskDir, 'verify.result.json'), 'standard'),
+        'verify.md': compactVerifierMarkdown(readTaskFile(taskDir, 'verify.md'), readOptionalJson(taskDir, 'verify.result.json'), 'fast'),
         'status.md': compactStatus(readTaskFile(taskDir, 'status.md')),
         'feedback.md': compactArtifact(taskDir, 'feedback.md', 'standard', ['feedback event', 'classification', 'supervisor decision']),
         'execution-feedback.md': compactArtifact(taskDir, 'execution-feedback.md', 'standard', ['feedback event', 'classification', 'supervisor decision']),
@@ -284,7 +339,7 @@ export function buildVerifierLlmInputPack({
         'brief.md': compactArtifact(taskDir, 'brief.md', selectedMode, ['goal', 'scope', 'success criteria']),
         'research.md': compactArtifact(taskDir, 'research.md', selectedMode, ['findings', 'evidence', 'repo']),
         'plan.md': compactArtifact(taskDir, 'plan.md', selectedMode, VERIFY_PLAN_SECTIONS),
-        'task-manifest.json': readTaskFile(taskDir, 'task-manifest.json'),
+        'task-manifest.json': stableTaskManifestForCheck(readTaskFile(taskDir, 'task-manifest.json')),
         'check.result.json': readTaskFile(taskDir, 'check.result.json'),
         'check.md': compactCheckMarkdown({
           checkMarkdown: readTaskFile(taskDir, 'check.md'),
@@ -295,7 +350,7 @@ export function buildVerifierLlmInputPack({
         'human-gate-summary.md': truncateMiddle(readTaskFile(taskDir, 'human-gate-summary.md'), charLimitForMode(selectedMode, 1200, 2500)),
         'execution.md': compactArtifact(taskDir, 'execution.md', selectedMode, VERIFY_EXECUTION_SECTIONS),
         'execution-ledger.json': compactExecutionLedger(readOptionalJson(taskDir, 'execution-ledger.json'), selectedMode),
-        'verify.md': compactVerifierMarkdown(readTaskFile(taskDir, 'verify.md'), readOptionalJson(taskDir, 'verify.result.json'), selectedMode),
+        'verify.md': compactVerifierMarkdown(readTaskFile(taskDir, 'verify.md'), readOptionalJson(taskDir, 'verify.result.json'), 'fast'),
         'status.md': compactStatus(readTaskFile(taskDir, 'status.md')),
         'feedback.md': compactArtifact(taskDir, 'feedback.md', selectedMode, ['feedback event', 'classification', 'supervisor decision']),
         'orchestration-log.md': compactOrchestrationLog(readTaskFile(taskDir, 'orchestration-log.md'), selectedMode),

package/bin/lib/llm-input-pack-utils.test.mjs

@@ -112,12 +112,61 @@ describe('llm input pack utilities', () => {
     });
 
     expect(pack.meta.mode).toBe('standard_plus');
-    expect(pack.meta.capTokens).toBe(26000);
+    expect(pack.meta.capTokens).toBe(34000);
     expect(pack.input.llmInputPolicy.contextInsufficientFallback).toBe('rerun_strict');
     expect(pack.input.taskArtifacts['plan.md']).toContain('<!-- compacted:plan.md');
     expect(pack.meta.compactedArtifacts).toContain('plan.md');
   });
 
+  it('stabilizes checker task manifest by excluding volatile check telemetry', () => {
+    const taskDir = createTask();
+    const pack = buildCheckerLlmInputPack({
+      taskDir,
+      taskId: 'TASK-999-token-pack',
+      checkerPromptSha: 'sha256:test',
+      cacheKey: { test: true },
+      checkContext: {
+        planSha: 'sha256:plan',
+        memorySha: 'sha256:memory',
+        riskProfile: 'high',
+        riskTriggers: ['source-sync-provider'],
+      },
+      checkEvidence: '# Evidence\n\nok',
+      checkerContextPack: '# Checker Context Pack\n\nok',
+      taskManifest: JSON.stringify({
+        schemaVersion: 1,
+        taskId: 'TASK-999-token-pack',
+        mode: 'standard',
+        phase: 'check',
+        gates: { sourceSyncProvider: { required: true, planComplete: true } },
+        context: {
+          planSha: 'sha256:plan',
+          memorySha: 'sha256:memory',
+          riskProfile: 'high',
+          riskTriggers: ['source-sync-provider'],
+          checkContextCurrent: true,
+        },
+        llmInput: { check: { updatedAt: 'volatile', attempts: [{ mode: 'strict' }] } },
+        lastCheckResult: { verdict: 'ready_for_human_gate', createdAt: 'volatile' },
+        timestamps: { updatedAt: 'volatile' },
+        loopDetector: {
+          threshold: 2,
+          requiresConsolidatedRemediation: true,
+          repeatedReasons: [{ normalizedReason: 'context_overflow', count: 3 }],
+          reasons: { noisy: { normalizedReason: 'verbose history' } },
+        },
+      }, null, 2),
+      projectMemory: [],
+      mode: 'standard',
+    });
+
+    expect(pack.input.taskManifest).toContain('"requiresConsolidatedRemediation": true');
+    expect(pack.input.taskManifest).not.toContain('lastCheckResult');
+    expect(pack.input.taskManifest).not.toContain('llmInput');
+    expect(pack.input.taskManifest).not.toContain('timestamps');
+    expect(pack.input.taskManifest).not.toContain('verbose history');
+  });
+
   it('preserves protected verification sections when compacting long plans', () => {
     const taskDir = createTask();
     const longPlan = [

package/bin/run-verify.mjs

@@ -503,7 +503,8 @@ function writeInternalSupervisorVerify({
     '## residual risks',
     '',
     '- This is not an independent fresh-context verifier run.',
-    '- Use `--verify-mode external_cli` for production-readiness, R4/R5, material migrations/backfills, destructive/security/financial work, broad ambiguous refactors or explicit human request.',
+    '- Use `--verify-mode external_cli` for production-readiness, R4/R5, production or real-user-data migrations/backfills, destructive/security/financial work, broad ambiguous refactors or explicit human request.',
+    '- Local scratch DB migrations, fixture imports and bounded local-only backfills can close with internal Verify when evidence covers the plan gates.',
     '',
     '## latest status excerpt',
     '',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@besales/ops-framework",
-  "version": "0.1.23",
+  "version": "0.1.25",
   "type": "module",
   "bin": {
     "ops-agent": "bin/ops-agent.mjs"

package/prompts/supervisor.md

@@ -57,7 +57,7 @@ Supervisor является code-level orchestrator по контракту: rou
 23. Любой user feedback, вопрос, correction, review note или новое наблюдение на любом этапе сначала записывается в `feedback.md` через `ops-agent intake-feedback` и классифицируется. Feedback не является инструкцией к изменению implementation, пока не классифицирован.
 24. После `Execute` задача не может перейти в `Retrospective`, `Human Closeout Gate`, `Closed`, `Accepted` или task switch без `Verify` и structured `verify.result.json`.
 25. `verify.result.json` должен сверять `plan.md` с фактическим `execution.md`, diff/files/tests и явным execution evidence. Self-reported executor checks без verifier verdict не являются достаточным Verify.
-26. `verify.result.json.verdict = pass | pass_with_notes` допустим при `verificationMode = internal_supervisor` для обычных `R0-R3` local engineering slices. Это cost-saving режим без независимого CLI/model verifier и он является default, если shared defaults или project agents override задают `verifier.mode = internal_supervisor`. `external_cli` обязателен только для R4/R5, production-readiness, destructive/security/financial/broad operational actions, material Prisma/data migrations/backfills, broad ambiguous refactors или explicit human request.
+26. `verify.result.json.verdict = pass | pass_with_notes` допустим при `verificationMode = internal_supervisor` для обычных `R0-R3` local engineering slices. Это cost-saving режим без независимого CLI/model verifier и он является default, если shared defaults или project agents override задают `verifier.mode = internal_supervisor`. `external_cli` обязателен только для R4/R5, production-readiness, destructive/security/financial/broad operational actions, production or real-user-data Prisma/data migrations/backfills, broad ambiguous refactors или explicit human request. Local scratch DB migrations, fixture imports and bounded local-only backfills can close with internal Verify when execution evidence covers the plan gates.
 27. Если external verifier/checker/browser tooling начинает тратить непропорционально много времени или блокируется окружением, Supervisor обязан остановить loop и вынести human decision: принять internal verify/evidence, запустить external escalation вручную или изменить scope.
 
 ## Hard Gate: Material Scope Expansion -> Brief Reset

package/prompts/verifier.md

@@ -69,7 +69,7 @@
 
 Для обычных `R0-R3` local engineering slices verdict `pass` или `pass_with_notes` может быть выдан в режиме `internal_supervisor`, если verifier сверил `plan.md`, `execution.md`, diff/files/tests and explicit evidence.
 
-External CLI verifier обязателен только для escalation triggers: `R4/R5`, production-readiness, destructive/security/financial/broad operational actions, material Prisma/data migrations/backfills, broad ambiguous refactors или explicit human request.
+External CLI verifier обязателен только для escalation triggers: `R4/R5`, production-readiness, destructive/security/financial/broad operational actions, production or real-user-data Prisma/data migrations/backfills, broad ambiguous refactors или explicit human request. Local scratch DB migrations, fixture imports, docs-only changes and bounded R0-R3 local engineering slices can close with `internal_supervisor` when execution evidence includes the required tests/build/lint/smoke/apply proof.
 
 Минимальный shape `verify.result.json`:
 

package/templates/verify.md

@@ -18,7 +18,7 @@
 
 `internal_supervisor | external_cli`
 
-Use `internal_supervisor` for ordinary local R0-R3 slices when focused tests/build/lint/smoke evidence is sufficient. Use `external_cli` for R4/R5, production-readiness, destructive/security/financial/broad operational work, material migrations/backfills, broad ambiguous refactors or explicit human request.
+Use `internal_supervisor` for ordinary local R0-R3 slices when focused tests/build/lint/smoke evidence is sufficient. Use `external_cli` for R4/R5, production-readiness, destructive/security/financial/broad operational work, production or real-user-data migrations/backfills, broad ambiguous refactors or explicit human request. Local scratch DB migrations, fixture imports and bounded local-only backfills can close with internal Verify when evidence covers the plan gates.
 
 ## Verification ladder coverage