@besales/ops-framework 0.1.24 → 0.1.26

package/CHANGELOG.md

@@ -1,5 +1,16 @@
 # Changelog
 
+## 0.1.26
+
+- Raised `standard_plus` context cap to 36k so compact Verify packs around 29-31k estimated tokens stay out of strict mode after safety margin.
+
+## 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.

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

@@ -12,7 +12,7 @@ export const LLM_CONTEXT_MODES = ['fast', 'standard', 'standard_plus', 'strict']
 export const LLM_CONTEXT_CAPS = {
   fast: 8000,
   standard: 20000,
-  standard_plus: 30000,
+  standard_plus: 36000,
   strict: 50000,
 };
 
@@ -318,7 +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'),
@@ -329,7 +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']),
@@ -339,7 +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'),
@@ -350,7 +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

@@ -112,7 +112,7 @@ describe('llm input pack utilities', () => {
     });
 
     expect(pack.meta.mode).toBe('standard_plus');
-    expect(pack.meta.capTokens).toBe(30000);
+    expect(pack.meta.capTokens).toBe(36000);
     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');

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": "0.1.24",
+  "version": "0.1.26",
   "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