Programmer model · Online

Read the codebase first,
then write the code.

DeepCodex does more than answer the current question. It understands the repository layout, team conventions and past decisions first, then carries on like a programmer who has worked on this project before.

Let it read my codebase How to connect

1,284: Files indexed
4: Relevant memories
12/12: Tests passed

payment-service / live

Forming the change

goal: Stop duplicate webhooks from crediting twice

Read the repository
1,284 files indexed
Recall experience
4 memories matched
Start coding
payment.ts +18 −6
Verify
12 tests passed

src/payments/webhook.ts+18 −6

export async function handleWebhook(event) {

+ const key = event.idempotency_key;

+ const existing = await ledger.find(key);

+ if (existing) return existing.result;

await ledger.credit(event.amount);

+ await ledger.commit(key);

}

pnpm test payments12 passed

Side by side

A generic model vs. a programmer model with codebase context

Same model; the difference is whether it knows your codebase.

Aspect

Generic model, direct

DeepCodex programmer model

Knowledge of the codebase

Generic model, directOnly the few files you just sent

✓Has read the layout, conventions and past decisions

Cost of re-explaining

Generic model, directEvery new session starts with the backstory

✓The backstory is in the profile; start on the task

Pitfalls

Generic model, directThe same trap can be hit again

✓Failed approaches are recorded and avoided

Switching terminals

Generic model, directWhat you told Codex CLI, Claude Code does not know

✓One project's knowledge, shared by every terminal

Style of changes

Generic model, directCode that runs but does not look like your team's

✓Follows existing naming, layout and test style

Verifiable results

Generic model, directOnly the chat log

✓Which entries were injected and how often they hit

Context architecture

Layered codebase context, not an indiscriminate pile

More is not more accurate. Stable facts, layout, conventions, lessons and current progress are kept apart and enter a request by relevance.

Stack

Framework, runtime, dependencies and build

Structure

Module boundaries, entry points and key files

Conventions

Naming, testing, commit and code-organisation rules

Pitfalls

Failed approaches, edge cases and fixes

Stage summary

Current goal, finished work and upcoming risks

Memory lifecycle

Memories have a source and a lifecycle

What the assistant extracts starts as a candidate; once confirmed it is injected, it can be disabled when stale, and a deleted entry keeps an auditable state.

Extracted from real tasks
Long-term only once confirmed
Hit counts reflect real use

payment-service / context4 hits this request

structureWebhook handling lives in src/payments/webhooksconfirmed
conventionEvery write must carry an idempotency keyconfirmed
pitfallNever infer payment completion from terminal stateconfirmed
taskAdding duplicate-webhook testscandidate

Request assembly

How one request gains codebase context

Every step has an explicit input and result, traceable in the task replay.

Detect intent
Locate the project scope involved
Retrieve profiles
Drop low-relevance and disabled entries
Assemble context
Respect model and allowance limits
Record hits
Write back task and memory statistics

Engineering search

Recover the evidence, then let the model continue

Search tasks, requests, profiles and memories across projects. Every result keeps its source, state and hit count; fuzzy similarity is never presented as fact.

MemoryEvery wallet write must carry an idempotency keyconfirmed · 42 hits
TaskFix duplicate credit from payment webhookL3 · 18 turns
Requestreq_8f2a9c4.2K tokens · 12 s ago

Keep project knowledge in the project

Profiles accumulate from the first API key bound to a project. Every entry can be viewed, disabled and deleted.

Start free Read the docs

Loading page

Read the codebase first,then write the code.

A generic model vs. a programmer model with codebase context

Layered codebase context, not an indiscriminate pile

Stack

Structure

Conventions

Pitfalls

Stage summary

Memories have a source and a lifecycle

How one request gains codebase context

Detect intent

Retrieve profiles

Assemble context

Record hits

Recover the evidence, then let the model continue

Keep project knowledge in the project

Read the codebase first,
then write the code.