Overview

Following Previous Post: #12, this is the 13th cycle. 39 commits across backend, frontend, and infrastructure — the largest single cycle yet. Three threads weave together: product (multi-tone reactions, angle/lens presets, custom retry prompts), data layer (alembic migrations for new reaction shapes), and production cutover (EC2 + Nginx + Terraform).

Three Threads in One Cycle

graph TD
    A[Cycle #13] --> B[Product features]
    A --> C[Data layer]
    A --> D[Production deploy]
    B --> B1[Multi-tone reactions]
    B --> B2[Angle/lens presets]
    B --> B3[Model auto-injection]
    B --> B4[Custom retry prompts]
    C --> C1[Reaction unique constraint]
    C --> C2[Multi-tone schema]
    C --> C3[Injected model column]
    D --> D1[Terraform main.tf]
    D --> D2[Nginx config]
    D --> D3[setup-ec2.sh]
    D --> D4[ecosystem.config.js]

Multi-Tone Reactions

The reaction system gained tone awareness — instead of a single 👍/👎 per image, users can react with multiple tone+angle combinations. Three alembic migrations land it:

• add_multi_tone_angle_text_reactions_*.py — schema for the new reaction shape
• add_unique_constraint_to_image_reactions.py — prevents duplicate reactions per (user, image, tone)
• add_injected_model_filename_to_*.py — tracks which model produced the reacted-to variant

Frontend changes (ReactionButtons.tsx, LikesTab.tsx, FeedbackModal.tsx) surface the tone picker.

Angle and Lens Presets

A library of named photographic presets — angle (“eye-level”, “low-angle”, “dutch tilt”) and lens (“35mm”, “85mm portrait”, “fisheye”) — that get injected into generation prompts. Backend tests under backend/tests/test_angle_presets.py and test_lens_presets.py verify the prompt construction. Frontend AnglePicker.tsx provides the visual selector.

Model Auto-Injection for Person-Intent Prompts

When the user’s prompt is detected as person-focused, the generation pipeline auto-injects a person-tuned model checkpoint. The injection logic lives in backend/src/generation/injection.py and is wired through prompt.py and service.py. The migration add_injected_model_filename_to_*.py records which model was injected for each generation, so the UI can show provenance.

Production Cutover

The biggest infra delta. Previous cycles ran on a developer EC2 with manual deployment. This cycle:

• terraform/main.tf — defines the prod VPC, EC2 instance, and security groups
• terraform/keys/prod.pub — production SSH key
• infra/nginx/diffs-image-agent.conf — Nginx reverse proxy config (TLS termination, route splitting between frontend and backend)
• scripts/setup-ec2.sh — provisioning script (uv, node, postgres client, pm2)
• ecosystem.config.js — pm2 process definitions, with a fix to remove APP_ENVIRONMENT (was conflicting with the .env loader)

After this cycle, the app lives at a real domain behind Nginx with auto-restart via pm2.

Commit Log (Highlights — 39 total)

인사이트 (Insights)

Three things stand out from this 39-commit cycle. First, the alembic migration count tracks product velocity — three migrations in one cycle means the data model is genuinely evolving, not just being patched. Second, landing prod deploy in the same cycle as new features is risky but fast — historically these are split across cycles, but bundling them means the new features get tested under realistic conditions immediately. Third, the angle/lens preset pattern (named presets injected into prompts) is the same pattern as the model auto-injection — both are forms of prompt enrichment based on user signal. That’s the right abstraction to formalize next cycle: a unified prompt enrichment pipeline where presets, model selection, and persona injection all flow through the same hook.