feat(demo): add interview quality audit

2026-07-09 11:18:49 +08:00
parent a6c2d4459c
commit 9c9a0024d4
37 changed files with 1162 additions and 86 deletions
@@ -0,0 +1 @@
+ready
@@ -0,0 +1 @@
+committed
@@ -0,0 +1,117 @@
+# Design: Interview Demo Quality Audit
+
+## Overview
+
+The change adds auditability and demo readiness without changing Agent routing.
+
+```text
+ChatService prompt resources
+  -> PromptAuditService
+  -> verifier_evaluation.prompt_audit
+  -> Trace API
+  -> DiagnosisTraceEvaluator
+  -> baseline report
+
+mvp/demo/scripts/run-interview-demo-check.ps1
+  -> health/readiness check
+  -> payment timeout chat
+  -> trace fetch
+  -> feedback
+  -> demo output bundle
+```
+
+## Prompt Audit
+
+Add a compact `prompt_audit` object under:
+
+```text
+diagnosis_session.self_evaluation.verifier_evaluation.prompt_audit
+```
+
+Shape:
+
+```json
+{
+  "version": "chat-prompts-v1",
+  "prompts": [
+    {
+      "name": "chat_planner",
+      "version": "chat-planner-v1",
+      "resource": "prompts/chat-planner-prompt.md"
+    }
+  ]
+}
+```
+
+Design choices:
+
+- Use explicit local metadata, not full prompt hashes, because the interview goal is explainable version audit rather than cryptographic integrity.
+- Keep this metadata in code or a small resource catalog near prompt loading.
+- Persist prompt audit with every Chat verifier evaluation, including fallback/degraded paths.
+- Do not include full prompt text in trace.
+
+## Eval Expansion
+
+Extend `DiagnosisEvalCase` with optional fields:
+
+- `requirePromptAudit`
+- `expectedPromptAuditVersion`
+- `expectedPromptVersions`
+- `requireGatekeeperRules`
+
+Evaluator behavior:
+
+- If `requirePromptAudit=true`, `verifier_evaluation.prompt_audit.version` must exist.
+- If `expectedPromptAuditVersion` is set, it must match.
+- If `expectedPromptVersions` is set, each listed prompt name/version pair must exist.
+- If `requireGatekeeperRules=true`, `gatekeeper_result.rules` must be a non-empty list and each item must include `id`, `enabled`, and `default_severity`.
+
+Add at least two fixture-backed cases:
+
+- A positive audit closure case that requires prompt audit + Gatekeeper rules.
+- A metadata-gap negative case represented as `LOW_CONFID`/safe final answer, used to prove the evaluator catches missing audit metadata when configured.
+
+The baseline must remain fully passing after fixtures are updated.
+
+## Demo Stabilization
+
+Add a PowerShell script:
+
+```text
+mvp/demo/scripts/run-interview-demo-check.ps1
+```
+
+Responsibilities:
+
+- Accept base URL and session id parameters.
+- Check that the service is reachable.
+- Run the existing payment-timeout chat request.
+- Fetch trace for the same session id.
+- Submit useful feedback.
+- Write outputs under `mvp/demo/output/`.
+- Emit a concise summary with session id, verdict, Gatekeeper rule version, prompt audit version, and output paths.
+
+The script should fail fast with actionable messages when the service is unavailable.
+
+## Documentation
+
+Add/update:
+
+- `mvp/demo/README.md`: mention the preflight script.
+- `mvp/demo/ten-minute-interview-demo.md`: use the preflight script as the recommended path.
+- `mvp/demo/interview-q-and-a.md`: concise interview answers for Agent engineering tradeoffs.
+- `mvp/architecture/harness-quality-gates.md`: record prompt audit as part of the quality gate.
+
+## Verification
+
+Required:
+
+- Targeted unit/eval tests for prompt audit persistence and evaluator checks.
+- Regenerated baseline JSON/Markdown reports.
+- OpenSpec validation.
+
+E2E:
+
+- If local dependencies are available, run Spring Boot with `mvp-demo` profile and execute the new preflight script.
+- If unavailable, record the reason and rely on deterministic unit/eval evidence.
+
@@ -0,0 +1,58 @@
+# Change: Interview Demo Quality Audit
+
+## Problem
+
+SuperBizAgent MVP is now strong enough to demonstrate traceable Agent engineering, but the interview path still has three gaps:
+
+- The live demo has a run script, but no preflight command that checks service readiness and produces a concise interview evidence bundle.
+- The diagnosis eval baseline covers the V2 evidence pipeline, but it does not yet assert prompt-version audit data and has limited coverage for audit metadata gaps.
+- Gatekeeper already exposes `rule_set_version`, but prompt versions are not persisted with the verifier evaluation, making prompt changes harder to explain, compare, and roll back in an interview.
+
+This change stabilizes the MVP as an interview artifact rather than adding a new diagnosis architecture.
+
+## Proposed Solution
+
+Implement a small internal quality/audit increment:
+
+1. Add prompt version audit metadata to Chat verifier evaluation.
+2. Extend deterministic diagnosis eval cases/fixtures to assert prompt audit metadata and audit-metadata failures.
+3. Add an interview demo preflight script and documentation that can be run before or during a demo to verify service readiness, execute the payment timeout path, fetch trace, and record key audit fields.
+4. Add/update MVP documentation for interview Q&A and the new audit/preflight workflow.
+
+## Scope
+
+In scope:
+
+- Internal `diagnosis_session.self_evaluation.verifier_evaluation` audit JSON.
+- Diagnosis eval case schema, evaluator checks, fixtures, and baseline reports.
+- MVP demo scripts/docs.
+- Architecture/demo documentation for prompt and Gatekeeper version audit.
+
+Out of scope:
+
+- Public HTTP API changes.
+- Database schema changes.
+- New Agent roles, MCP tool server migration, process isolation, or AIOps LLM Verifier.
+- Replacing existing `Planner -> Executor -> Gatekeeper -> Verifier -> Composer` orchestration.
+- Guaranteeing live LLM `PASS` for every demo run. Live demo compatibility and deterministic fixture regression are separate acceptance paths.
+
+## Context Constraints From devflow
+
+- Evidence Tools produce incident facts and must be recorded in `tool_invocation`.
+- Chat quality gates are layered: Gatekeeper verifies evidence references, Verifier judges derivability, Composer controls expression.
+- `diagnosis eval` is deterministic and fixture-backed; no LLM-as-judge.
+- Demo assets should be runnable, but interview safety should not depend solely on live LLM behavior.
+- Gatekeeper rule metadata is metadata-only; dynamic rule execution is out of scope.
+
+## Interface Impact
+
+Level: L2 internal contract change.
+
+Reason: `verifier_evaluation` gains a compact `prompt_audit` object. Existing public API shape remains the same, and the value is exposed only through already-existing trace/self-evaluation JSON.
+
+## Risks
+
+- Baseline report churn is expected when adding cases; JSON and Markdown reports must be regenerated together.
+- Prompt audit must be deterministic and stable enough for eval fixtures; avoid hashing full prompt text with environment-specific content.
+- Demo preflight must not hardcode secrets and must tolerate local service unavailability with clear failure messages.
+
@@ -0,0 +1,16 @@
+## MODIFIED Requirements
+
+### Requirement: Verifier SHALL be observable
+The Verifier's verdict SHALL be persisted for observability.
+
+#### Scenario: prompt audit written to verifier evaluation
+- **WHEN** the Chat verifier evaluation is persisted
+- **THEN** the system SHALL include a `prompt_audit` object under `diagnosis_session.self_evaluation.verifier_evaluation`
+- **AND** `prompt_audit.version` SHALL identify the Chat prompt audit catalog version
+- **AND** `prompt_audit.prompts` SHALL include the planner, executor, verifier, and composer prompt names and versions
+- **AND** full prompt text SHALL NOT be persisted in `prompt_audit`
+
+#### Scenario: prompt audit available on fallback paths
+- **WHEN** Chat verifier parsing fails, Composer parsing fails, or Chat produces a degraded answer
+- **THEN** the persisted verifier evaluation SHALL still include `prompt_audit`
+
@@ -0,0 +1,21 @@
+## MODIFIED Requirements
+
+### Requirement: Diagnosis eval SHALL validate trace fixtures deterministically
+The diagnosis eval harness SHALL evaluate saved trace fixtures without invoking an LLM judge.
+
+#### Scenario: prompt audit assertions are enforced
+- **WHEN** an eval case sets `requirePromptAudit=true`
+- **THEN** the evaluator SHALL require `verifier_evaluation.prompt_audit.version`
+- **AND** when `expectedPromptAuditVersion` is configured, it SHALL match exactly
+- **AND** when `expectedPromptVersions` is configured, each configured prompt name SHALL appear with the expected version
+
+#### Scenario: Gatekeeper rule metadata assertions are enforced
+- **WHEN** an eval case sets `requireGatekeeperRules=true`
+- **THEN** the evaluator SHALL require `verifier_evaluation.gatekeeper_result.rules` to be non-empty
+- **AND** each rule item SHALL include `id`, `enabled`, and `default_severity`
+
+#### Scenario: expanded baseline remains passing
+- **WHEN** the committed fixture set is evaluated
+- **THEN** every case SHALL pass
+- **AND** baseline JSON and Markdown reports SHALL reflect the expanded case count and verdict distribution
+
@@ -0,0 +1,22 @@
+## MODIFIED Requirements
+
+### Requirement: MVP demo SHALL be reproducible for interviews
+The MVP demo SHALL provide a repeatable way to show a diagnosis answer, trace, verifier evaluation, and feedback.
+
+#### Scenario: interview demo check script records an evidence bundle
+- **WHEN** the user runs the interview demo check script against a running `mvp-demo` service
+- **THEN** the script SHALL submit a fixed Chat diagnosis request
+- **AND** it SHALL fetch the trace for the same session id
+- **AND** it SHALL submit useful feedback for that session
+- **AND** it SHALL write chat, trace, feedback, and summary outputs under `mvp/demo/output/`
+
+#### Scenario: interview demo check fails with actionable readiness output
+- **WHEN** the target service is not reachable
+- **THEN** the script SHALL fail before issuing diagnosis requests
+- **AND** the failure message SHALL name the base URL and the expected startup profile
+
+#### Scenario: interview documentation explains audit fields
+- **WHEN** an interviewer asks how prompt or Gatekeeper changes are audited
+- **THEN** the demo documentation SHALL point to `prompt_audit.version` and `gatekeeper_result.rule_set_version`
+- **AND** it SHALL explain that deterministic eval fixtures are the regression source of truth
+
@@ -0,0 +1,32 @@
+## 1. OpenSpec And devflow
+
+- [x] 1.1 Create OpenSpec proposal/design/spec/tasks for `interview-demo-quality-audit`.
+- [x] 1.2 Record context, question pool, interface impact, audit, and verification plan in devflow decisions.
+- [x] 1.3 Pass OpenSpec validation and create `.committed`.
+
+## 2. Prompt/Gatekeeper Version Audit
+
+- [x] 2.1 Add compact Chat prompt audit metadata for planner, executor, verifier, and composer prompts.
+- [x] 2.2 Persist `prompt_audit` under `verifier_evaluation` for Chat verifier/composer outcomes.
+- [x] 2.3 Add focused tests proving prompt audit appears in persisted verifier evaluation.
+- [x] 2.4 Extend eval checks for prompt audit and Gatekeeper rule metadata.
+
+## 3. Eval Expansion
+
+- [x] 3.1 Extend diagnosis eval case/result schema for prompt audit fields.
+- [x] 3.2 Add fixture-backed cases for audit closure coverage.
+- [x] 3.3 Regenerate baseline JSON and Markdown reports.
+- [x] 3.4 Update eval docs/schema.
+
+## 4. Interview Demo Stabilization
+
+- [x] 4.1 Add `run-interview-demo-check.ps1` with service preflight, chat, trace, feedback, and summary output.
+- [x] 4.2 Update demo README and 10-minute script to use the preflight path.
+- [x] 4.3 Add interview Q&A documentation focused on Agent engineering tradeoffs.
+
+## 5. Verification And Archive
+
+- [x] 5.1 Run targeted tests for ChatService/prompt audit and diagnosis eval.
+- [x] 5.2 Run relevant broader regression tests.
+- [x] 5.3 Run E2E demo check with `mvp-demo` profile if dependencies are available; otherwise record the blocker.
+- [x] 5.4 Archive the OpenSpec change, update devflow artifacts, and commit implementation + archive.