SelfieK

SelfieK (selfiek) is a fast, evidence-friendly CLI for the K/Kate selfie pipeline.

It keeps the proven ChatGPT/CDP image-generation path and moves the orchestration into:

• Rust core for fast deterministic status, dice validation, sampling, prompt-card compilation, stock consumption, cleanup, and cdper invocation.
• TypeScript npm CLI for ergonomic distribution through npm install -g selfiek.

SelfieK intentionally does not replace cdper-gpt-image. cdper remains the image-generation adapter; SelfieK is the stock kitchen manager.

Boundary: SelfieK manages the fridge; cdper handles ad-hoc cooking. SelfieK can call cdper to produce inventory, but ad-hoc/one-off image requests should go directly to cdper instead of adding an instant lane to SelfieK.

Current defaults

SelfieK defaults to the existing local K pipeline paths:

/home/agent/K-original
/home/agent/k-selfie-new
/home/agent/k-selfie-used
/home/agent/.hermes/scripts/k-selfie-generator/dice_config.json
/home/agent/obsidian-vault/raw/selfie-prompts
/home/agent/.local/bin/cdper-gpt-image

Override paths with flags or environment variables (SELFIEK_K_ORIGINAL, SELFIEK_NEW_DIR, SELFIEK_USED_DIR, SELFIEK_DICE_CONFIG, SELFIEK_PROMPT_LIB, SELFIEK_CDPER_BIN). For stock-only product-line generalization, use --profile-file / SELFIEK_PROFILE_FILE plus --profile and --stock-limit.

Profile manifest example:

schema_version: selfiek.profile.v1
id: k-selfie
paths:
  dice_config: dice_config.json
  reference_dir: /home/agent/K-original
  new_dir: /home/agent/k-selfie-new
  used_dir: /home/agent/k-selfie-used
  prompt_lib: /home/agent/obsidian-vault/raw/selfie-prompts
  runtime_dir: /home/agent/.hermes/scripts/k-selfie-generator
  cdper_bin: /home/agent/.local/bin/cdper-gpt-image
stock:
  new_limit: 100

Profile paths may be relative to the manifest file. Precedence is: built-in defaults < profile file < environment variables < CLI flags.

CLI

selfiek --version
selfiek status --json
selfiek profile show --json
selfiek profile validate --file ./profile.yml --json
selfiek validate-config --json
selfiek library lint --json
selfiek library report --json
selfiek library optimize --dry-run --json
selfiek feedback rate --image /path/to/used.png --score 2 --reason "natural" --like scene.concert,outfit.casual --dislike face_likeness --json
selfiek preference compile --json
selfiek preference report --json
selfiek preference evolve --dry-run --json
selfiek compile --json
selfiek compile --use-orderk --json
selfiek draw --json --use-templates --explain
selfiek generate --json --dry-run --use-templates  # manual stock generation only; writes to stock new_dir
selfiek produce --json --use-templates --dry-run
selfiek produce --json --use-templates --quiet
selfiek next --json --use-templates               # consumes stock only; does not generate on empty stock
selfiek cleanup-used --json

Version model

• runtime_version: latest release
• contract_version: latest release

SelfieK 3.7 adds the Preference Engine: feedback rate records immutable feedback events linked to image sidecars and prompt cards; preference compile aggregates those events into preference_model.json, preference_report.json, and preference-aware weights.json; preference report explains positive/negative element trends; preference evolve --dry-run proposes candidate boosts/penalties without rewriting prompts. The hot path remains deterministic and offline: no LLM optimizer, no UI, no daemon, no MCP loop, and no cdper rewrite.

SelfieK 3.6 upgraded prompt-library absorption: library lint validates the Obsidian recipe book, library ingest creates source/template drafts, library optimize --dry-run produces an offline no-write cleanup plan, library report includes coverage and inventory-quality sections, compile emits template_index.json, fragment_index.json, prompt_cards.jsonl, library_report.json, and actionable weights.json v2, and draw/generate --use-templates attach prompt_card v2 for feedback attribution. orderk is compile/report-only; production cron still consumes compiled artifacts and cdper remains the image-generation adapter.

The prompt-library layer also includes lightweight guardrails borrowed from prompt-optimizer-style tooling without adopting its product shape: raw prompts are stored as JSON evidence references instead of copied wholesale, risky role/injection-like source text is reported by library lint, {{placeholder}} variables are tracked through prompt cards, structured JSON sources expose top-level key preservation warnings, feedback templates are nudged toward visible facts, and library report emits machine-readable quality_signals. There is still no LLM prompt optimizer, UI, daemon, MCP optimize tool, or cdper rewrite step.