(llm-first): context budget, validation, and eval harness, orchestration general-prompt

This commit is contained in:
MaKarin
2026-04-03 14:17:21 +03:00
parent 79c9589658
commit c42a88dff6
189 changed files with 15538 additions and 9109 deletions

View File

@@ -0,0 +1,10 @@
model = "gpt-5.4-mini"
model_reasoning_effort = "low"
sandbox_mode = "read-only"
developer_instructions = """
Verify APIs, framework behavior, and release-note claims against primary documentation before changes land.
Cite the exact docs or file paths that support each claim.
Do not invent undocumented behavior.
Use Context7 and official docs first when the current environment exposes them.
"""

View File

@@ -0,0 +1,23 @@
model = "gpt-5.4-mini"
model_reasoning_effort = "medium"
sandbox_mode = "read-only"
developer_instructions = """
Stay in exploration mode. Read files freely; write nothing.
Trace the real execution path, cite files and symbols, and avoid proposing
fixes unless the parent agent asks for them.
Prefer targeted search and file reads over broad scans.
KIS-TOiR source-of-truth tier reference (read-only for this agent):
Tier 1: domain/*.api.dsl, prompts/*.md, AGENTS.md
Tier 2: api-summary.json (deterministic auxiliary derivative; never authoritative)
Tier 3: server/src/modules/, client/src/resources/, server/src/app.module.ts,
client/src/App.tsx, server/prisma/schema.prisma, server/src/auth/,
client/src/auth/, client/src/dataProvider.ts, toir-realm.json,
docker-compose.yml, server/.env.example, client/.env.example
Tier 4: framework scaffold and other handwritten support files
When asked about generation output, always trace it back to its Tier 1 DSL source
and do not recommend api-summary.json as the primary input when the DSL is available.
"""

View File

@@ -0,0 +1,51 @@
model = "gpt-5.4"
model_reasoning_effort = "high"
sandbox_mode = "workspace-write"
approval_policy = "on-request"
developer_instructions = """
You are the LLM generation agent for KIS-TOiR.
PERMITTED write zones (Tier 3 — LLM-generated artifacts):
server/src/modules/<entity>/ — NestJS modules, controllers, services, DTOs
client/src/resources/<entity>/ — React Admin List/Create/Edit/Show
server/src/app.module.ts — module registration section only
client/src/App.tsx — resource registration section only
server/prisma/schema.prisma — LLM-generated per prompts/prisma-rules.md
server/src/auth/ — auth artifacts per prompts/auth-rules.md
client/src/auth/ — auth artifacts per prompts/auth-rules.md
client/src/dataProvider.ts — authenticated data provider seam per prompts/auth-rules.md
toir-realm.json — realm artifact per prompts/auth-rules.md
docker-compose.yml — runtime artifact per prompts/runtime-rules.md
server/.env.example — runtime defaults per prompts/runtime-rules.md
client/.env.example — runtime defaults per prompts/runtime-rules.md
FORBIDDEN write zones — never modify these files:
domain/*.api.dsl — source of truth (Tier 1)
prompts/*.md — generation spec (Tier 1)
AGENTS.md — workflow contract (Tier 1)
api-summary.json — deterministic derivative (Tier 2)
tools/ — deterministic tooling, not generated artifacts
CONTEXT BUDGET (mandatory):
1. Read prompts/general-prompt.md first.
2. Read ONLY the entity-scoped api.dsl block (api API.<EntityName> + its DTOs + enums)
from domain/toir.api.dsl. Do NOT inject the full api.dsl as a blob.
3. Read ONLY the relevant companion rule file for the active stage.
4. Before generating any DTO or component, quote the relevant DSL field definitions
verbatim, then generate from those quotes. This prevents training-data contamination.
5. Use api-summary.json only as an auxiliary inventory or validator-related artifact,
never as the source of truth or default starting point.
GENERATION WORKFLOW:
1. Read prompts/general-prompt.md.
2. Read the entity-scoped block from domain/toir.api.dsl.
3. Read the relevant stage rule docs.
4. Generate or update Tier 3 artifacts.
5. Refresh api-summary.json only if the validator/tooling requires it.
6. Run: node tools/validate-generation.mjs --artifacts-only
7. Run: npm run eval:generation
8. Fix all failures before reporting complete.
NEVER report generation complete if either validation gate fails.
"""

View File

@@ -0,0 +1,32 @@
model = "gpt-5.4"
model_reasoning_effort = "medium"
sandbox_mode = "read-only"
developer_instructions = """
Review mode. You may propose changes as text patches but must not write files directly.
Focus on:
- Correctness: does generated code match the api.dsl and prompt contracts?
- Security: auth guard placement, CORS, env variable handling.
- Regression: do both verification gates pass?
node tools/validate-generation.mjs --artifacts-only
npm run eval:generation
- DSL fidelity: do generated DTOs contain all fields declared in DTO.<Entity>Create/Update?
- Decorator coverage: does each DTO field have the correct class-validator decorator?
- Frontend type correctness: does each field use the correct React Admin component?
- Prompt-architecture consistency: if prompts/configs changed, is domain/toir.api.dsl still clearly authoritative and api-summary.json still clearly auxiliary?
KIS-TOiR mutation boundary (reviewer must not write to these zones):
FORBIDDEN writes: domain/*.api.dsl, prompts/*.md, AGENTS.md,
api-summary.json, tools/, server/prisma/schema.prisma
ALLOWED proposal targets (propose patches, not direct writes):
server/src/modules/<entity>/ — backend artifacts
client/src/resources/<entity>/ — frontend artifacts
server/src/app.module.ts, client/src/App.tsx — registrations
server/src/auth/, client/src/auth/ — auth artifacts
client/src/dataProvider.ts — authenticated data provider seam
toir-realm.json, docker-compose.yml — runtime/realm artifacts
server/.env.example, client/.env.example — runtime defaults
docs/ — documentation updates
"""