Feat/add knowleage base and agent

[anuragk16]

target issue : Knowledge Base for Agentic Systems #11

Testing in IAGES:

• After adding the knowledge base and the agent i gave this prompt :

• Then this :

• So it created me these :

• So for testing if the knowledge base is used by the agents or not, I give it a new requirement for the similar task. (
  02-architecture/adr-002-vcp-annual-revenue-lock.md | #597 — Revenue locked after provisional payment )

input

output

Conclusion

• It is using the knowledge base , also as the third primary source of truth (first is claude.md , and second is memory.md which is agent specific )

NOTE : Found some issue in the prompt and the business flow index file which is not creating a new tree sturcture instead it is extenting the indix file in the business flow for the all uses ( issue is file get to long in very short time and Ai will be reading the whole file which will make the token consuption more)

• fixed this by adding a more generic prompt at the business flow index file.

Test 2

• Without calling any specific agent , validating if the Claude is using the knowledge base or not for the information.

Conclusion : Yes , it is seeing the knowledge base as the first source then going to explore the codebase.

[anuragk16]

Knowledge Base Adoption Testing Template

1) Project Info

• Project Name: IAGES
• Repository URL: https://github.com/coloredcow-admin/iages-backend
• Branch / Commit: N/A (manual validation run)
• Test Owner: @anuragk16
• Test Date: 2026-03-03
• Knowledge Base Mode: Using KB

2) Test Scope (Baseline Only)

• Scope Type: Baseline
• Goal: Confirm knowledge-base integration is safe to adopt and does not break normal agent workflows.
• Out of Scope: Deep workflow-specific validations (to be refined later).

3) Baseline Gate Checklist (Mandatory)

Mark each item: PASS / FAIL / BLOCKED / NOT APPLICABLE

ID	Check	Status	Evidence	Notes
B1	Agent workflows run without hard failure	PASS	IAGES run notes	Knowledge base + agent setup completed and produced outputs.
B2	User prompt instructions still take precedence over KB context	PASS	IAGES run notes	Validation responses were KB-aware and aligned to prompt context.
B3	No unintended file changes outside expected scope	BLOCKED	IAGES run notes	Not explicitly validated in this baseline run.
B4	If KB is missing/disabled, agents still run in fallback mode	BLOCKED	N/A	Negative-path fallback test not executed yet.
B5	CI/check pipeline status is known and recorded	PASS	PR checks	Status is visible/recorded; baseline interpretation pending full test matrix.

4) IAGES Testing Snapshot

• Initial setup prompts were executed and KB artifacts were created.
• Validation prompt used a similar requirement context:
  • 02-architecture/adr-002-vcp-annual-revenue-lock.md
  • #597 — Revenue locked after provisional payment
• Observed outcome from the run:
  • Agent responses referenced and used KB context.
  • KB acted as an additional source of truth alongside CLAUDE.md and agent memory.

Evidence thread: Testing in IAGES

5) New Findings Update (Latest)

• Found an issue in the prompt + business flow index behavior:
  • It was extending 01-business-flows/00-index.md instead of using tree-structured separate flow files.
  • This can make the index grow too fast and increase token usage because agents read larger files repeatedly.
• Applied fix in IAGES test setup:
  • Updated with a more generic prompt direction in the business-flow index context to enforce tree structure behavior.

Test 2 (No specific agent invocation)

• Test objective: Validate whether Claude uses the knowledge base even without explicitly calling a specific agent.
• Result: PASS
• Conclusion:
  • Claude checked KB first for domain/business-flow context, then explored codebase for implementation details.
  • This confirms KB is being treated as an early source of context in normal prompting.
• Evidence: Screenshot shared in this PR thread (latest testing update).

6) Result Summary

• Overall Result: GO WITH RISKS
• Blocking Issues: B3 and B4 are still BLOCKED (not yet validated in this baseline run).
• Known Risks Accepted for This Stage: Proceeding with baseline confidence only; deeper sync/fallback coverage pending.
• Next Refinement Target (single next step): Run explicit B3 and B4 checks and then finalize GO/NO-GO.

7) Decision Rule (Simple)

Use this simple rule until process maturity increases:

• GO only if all mandatory baseline checks (B1 to B5) are PASS or NOT APPLICABLE
• NO-GO if any mandatory baseline check is FAIL
• GO WITH RISKS if no FAIL, but one or more mandatory checks are BLOCKED