How to Backtrack and Recover from AI Agent Codebase Malfunctions
makedream2026년 7월 30일
0
Computing/Software
Comments (0)
Log in to leave a comment
No posts yet
makedreammakedreamLog in to leave a comment
No posts yet
When you integrate AI agents like Claude Code or Codex into a backend repository with over 100,000 lines of code, you're initially blown away by the speed. Then, at some point, the moment a few lines of bad prompt go in, the AI breaks through its guardrails and starts tampering with core utilities or environment configuration files. Session logs inflate by megabytes, builds fail, and after spending an hour scanning Git Diff just to find the culprit file, you find yourself thinking it would have been better to just write the code yourself from scratch.
To grant agents autonomy while preventing unnecessary file corruption, you need a strict hard-isolation mechanism and a visual trace tracking system.
Natural language instruction files like CLAUDE.md or .aiignore are easily bypassed by agents as the context window grows long. Requests at the prompt level are merely recommendations, so the moment context is lost, the agent's tool execution ruthlessly crosses the fence.
To physically block the agent's reach, you should directly place .claude/settings.json—a deterministic permission control engine—in the project root.
json { "permissions": { "deny": [ "Edit(src/core/config/**)", "Edit(src/shared/utils/**)", "Read(./.env*)", "Bash(rm -rf *)" ], "ask": [ "Edit(src/api/v1/legacy/**)" ] } }
After establishing explicit block rules, it is safer to lock things down once more at the file system level. For production build scripts or security key files, taking away write permissions at the OS level is the cleanest approach.
.claude/settings.json in the project root.permissions.deny array.chmod 444 .env*.By controlling permissions this way, you can eliminate a significant portion of build errors caused by AI randomly modifying shared modules.
If session logs (.jsonl) exist only on an individual developer's local machine, it becomes difficult to determine the root cause when side effects hit later on. Set up a GitHub Actions workflow to compress and validate session traces at the point a PR is opened.
`yaml
name: AI Agent Session Audit & Risk Analysis
on:
pull_request:
types: [opened, synchronize]
jobs:
audit-session:
runs-on: ubuntu-latest
steps:
- name: Checkout Code
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Setup Mindwalk Telemetry Extractor
run: |
docker run --rm -v ${{ github.workspace }}:/workspace \
cosmtrek/mindwalk:v0.1.0 parse --log-dir=./.claude/sessions --out=telemetry.json
- name: Run Deterministic Repo Audit
id: repo_audit
run: |
curl -sSL https://github.com/aletheore/releases/download/v1.0/aletheore -o aletheore
chmod +x aletheore
./aletheore audit --telemetry=telemetry.json --max-allowed-depth=3
- name: Upload Mindwalk 3D Visual Artifact
uses: actions/upload-artifact@v4
with:
name: mindwalk-3d-trace-${{ github.event.pull_request.number }}
path: telemetry.json
`
By pairing a Rust-based analysis engine or Tree-sitter to check whether the agent has strayed out of scope, you can immediately fail the CI build if parent modules outside the allowed boundary are modified.
It is useful to calculate the risk score based on node connectivity rather than simple lines of code changed.
Here, is the number of modified files, and is the average 3D connectivity of the modified nodes. Setting it up to send a Slack notification with a 3D visualization map link when the threshold is exceeded allows reviewers to grasp the risk level right away.
When an agent creates a mess, instead of discarding the entire commit, you should trace the timeline and selectively recover only the corrupted files. Once you extract the specific time range where the malfunction started, use jq to pull the corresponding prompt and reasoning path from the JSONL log.
bash jq -c 'select(.timestamp >= "2026-07-11T14:20:00Z" and .timestamp <= "2026-07-11T14:25:00Z") | {timestamp: .timestamp, prompt: .payload.prompt, tool_use: .payload.tool_use, thinking: .payload.thinking}' \ ./.claude/sessions/session_abc123.jsonl
Once you've confirmed the cause, proceed with the recovery process in the following 3 steps:
git checkout <PRE_AI_COMMIT_SHA> -- src/legacy/broken_module.ts)git clean -fd -- src/unwanted_generated_dir/)Using this approach, you can cleanly pick up and restore only damaged files without the misfortune of wiping out working code.
In a monorepo environment, when nodes swell into the tens of thousands, the visualization tool itself can crash. If the browser slows down, debugging becomes a chore. To trim unnecessary directories from the rendering target, you need to clearly set up the mindwalk.config.json exclusion filter.
json { "visualization": { "excludePatterns": [ "**/node_modules/<strong>", "</strong>/vendor/<strong>", "</strong>/.git/<strong>", "</strong>/dist/<strong>", "</strong>/coverage/<strong>", "</strong>/*.log", "**/*.pb.go" ], "maxDepth": 5, "groupingStrategy": "directory-segmented" } }
Finally, when launching visualization pipelines or analysis tools in a background Docker container, resource allocations must be capped so your local IDE doesn't freeze.
mindwalk.config.json to exclude dependency directory nodes from the visualization target.--cpus="1.5" and --memory="2g" options when running the container.--nice=19 flag to the end of the background process execution command to lower the CPU priority to the absolute lowest.