Template

Codebase Oracle

Crawls your repo nightly, builds a semantic index, runs as your IDE backend via ACP. Air-gap mode for regulated codebases where cloud LLMs are forbidden.

ACP / IDEMemoryContinuousAir-gap

For: Dev teams — especially regulated industries where ChatGPT-in-the-codebase is forbidden.

Use this templateSee it in action

Codebase Oracle

ChatGPT-in-your-IDE is great until your security team says no cloud LLMs. Argus runs on your laptop, learns your repo, and never leaks code.

What it does

Argus is the senior engineer who's read every file in your repo and remembers everything. It runs as your IDE backend over ACP (the Agent Client Protocol), so Zed, VS Code (with ACP plugin), and any future ACP-compatible editor sees it as the assistant on the other side of the chat.

Every night, it crawls the repo:

  • Updates its semantic index (which symbols live where, which patterns repeat, which files changed without test changes)
  • Indexes recent commits + authors (who would know about X)
  • Surfaces drift: code that changed but tests didn't, deprecated patterns still in use

When you ask a question, Argus uses the index to give you a real answer with citations:

"Where do we handle JWT refresh?"

The refresh flow lives in auth/jwt_refresh.go:42-89. Last touched by @sarah on Feb 14 (commit 8b9f2e, "fix race condition in refresh token rotation"). Tests in auth/jwt_refresh_test.go cover the happy path; the failure-mode tests in auth/integration_test.go are commented out — that may be why we saw the prod issue last week.

Why this agent shape

  • ACP integration — Argus runs as the IDE's chat backend, not a separate window. The conversation happens where the code is.
  • Continuous operation — the nightly crawl keeps the index fresh without the dev triggering it. When you start your IDE Monday morning, Argus already knows what changed over the weekend.
  • Persistent memory — the team's mental model accumulates. "Sarah works on auth, Kev owns the deploy pipeline" lives in MEMORY.md and gets used for routing questions ("ask @sarah about this") and for context (knowing why a pattern exists).
  • Air-gap option — point the agent at a local llama.cpp or Ollama model. Zero outbound network. Your code stays on your laptop.

What you'll configure

After picking this template, the architect will ask you to:

  1. Point at your repo. Local path; Argus runs read-only.
  2. Pick a model. For maximum quality, an Anthropic/OpenAI key. For maximum privacy (regulated codebases), a local model via Ollama or llama.cpp. The architect walks you through both options.
  3. Configure the index. What to ignore (node_modules, build artifacts, generated code). Default ignore list is sensible; you can refine.
  4. Set the nightly crawl time. Default 2am local; pick when your dev environment is idle.
  5. Optional: connect Telegram. Off-IDE queries — ask Argus about the codebase from your phone while in a meeting.

What it looks like running

In your Zed editor, the assistant panel opens to Argus. You type:

"What tests cover the lead-scoring code?"

Argus replies with a list:

  • qualification/scoring_test.go (47 tests, last run passed)
  • qualification/integration_test.go (12 tests, last run passed)
  • One blame note: "the regression test for the false-positive issue from incident #312 lives in scoring_test.go:189 (added by @sarah Feb 28)."

You ask a follow-up:

"Who would know if I'm thinking about changing the scoring weights?"

Argus pulls from memory + git log:

  • @sarah owns the scoring module (12 commits in last 90 days)
  • @ankit has touched it 3 times for performance tuning
  • The weights themselves last changed Jan 18 in commit 7c4f2a. The PR description mentions "tuned based on Q4 win-rate analysis." Worth reading before changing.

Why this is CIQ's differentiator

Enterprise codebases — defense contractors, banks, healthcare — can't use ChatGPT or Claude on their code. Compliance says no. Argus is the agent that fits there: local model, no egress, full repo intelligence.

This is the Rocky Linux codebase tool. If you're CIQ's target customer running classified or HIPAA-bound code, this template is designed for you.

When this template isn't right

  • You don't use an ACP-compatible IDE — Zed has native ACP support; VS Code and Cursor need plugin work that's ongoing. If you're a JetBrains-only shop, this template doesn't help you yet.
  • Your codebase is huge (>1M LOC) — the semantic index works for large repos but the nightly crawl gets long. Tune it down for sanity.
  • You want code generation, not code Q&A — Argus is read-and-explain, not write-and-commit. The architect can extend it but the default is intentionally conservative.

What you can ask the architect to change

  • "Index multiple repos, not just one" — multi-repo support
  • "Notify me when someone touches my code paths" — adds a watch + notify action
  • "Pair with Telegram for off-IDE queries" — already in default config, just configure the bot token
  • "Run on Ollama with llama-3.1-70b" — air-gap mode with a specific local model

See it in action

Sample interactions illustrating how Argus actually behaves.

In Zed IDE, dev asks where a feature lives

You

Where do we handle JWT refresh?

Agent

JWT refresh lives in `auth/jwt_refresh.go:42-89`. Last touched by @sarah on Feb 14 (commit 8b9f2e, "fix race condition in refresh token rotation").

Tests in `auth/jwt_refresh_test.go` cover the happy path. The failure-mode tests in `auth/integration_test.go:201-340` are currently commented out — that may relate to the prod issue last week. Worth uncommenting + investigating before any changes here.

Dev asks who knows about a module

You

Who would know if I'm thinking about changing the scoring weights?

Agent

Pulled from git history + memory:

@sarah owns the scoring module — 12 commits in last 90 days, primary author since Q3.
@ankit has touched it 3 times for performance tuning (not weights).

The weights themselves last changed Jan 18 (commit 7c4f2a). PR #1192 description mentions "tuned based on Q4 win-rate analysis" — worth reading before changing.

No recent activity in the past 30 days, so changes are unlikely to conflict in-flight.

Nightly crawl finishes; agent surfaces a finding

You

(no input — heartbeat at 2am)

Agent

[Telegram, 2:14am] Nightly crawl complete.

Drift detected (3 items):
- `auth/oauth.go` changed; `auth/oauth_test.go` did NOT. Likely missing test coverage on the new flow.
- Deprecated pattern `httpClient.Do(req)` (without context) re-appeared in 2 new files (`api/handlers.go`, `worker/poll.go`). We deprecated this Feb 1.
- 47 new TODO comments added across the codebase in last 7 days. Top author: @kev (12). Top file: `worker/queue.go` (8 TODOs).

Index stats: 8,247 symbols indexed (+34 from yesterday). Coverage map updated.

Ready to deploy this agent?

Picking this template loads the configuration above into the Forge Architect. You can refine anything via plain English before deploying.

Use this template