Look

Converts user-provided images into validated, decision-ready action drafts. Routes images by inferred intent across domains: events from flyers, macros from meal photos, places to save, products to price, receipts to log, documents to file, civic issues to report. Supports reverse image search via Yandex and Google. NOT for generic OCR, computer vision research, or surveillance.

Interactive Menu

When invoked interactively, present a two-level menu. See references/interactive-menu.md for the full menu structure.

When to Use

• Image analysis and visual content extraction
• Reverse image search and source identification
• Visual verification of web content
• Screenshot analysis and OCR
• When any skill needs image understanding
• A flyer or poster → calendar event or ticket purchase draft
• A meal photo → macro estimation
• A storefront or sign → save to try-list
• A product photo → pricing comparison or order draft
• A civic issue photo → 311 report draft
• A receipt → expense entry
• A document → searchable PDF and filing draft

When NOT to Use

• Image generation (use Imagine)
• Video processing or analysis
• Bulk image processing at scale
• Real-time camera feed analysis
• Generic OCR utility
• Background surveillance or tracking
• Generic automation framework
• Images where the user has no clear intent in mind

Workflow

The image-to-draft workflow routes user images through intent inference, domain routing, and draft generation. This workflow exists because raw images alone are not actionable — the agent must extract structured meaning before proposing next steps.

1. Ingest — Accept user-provided image (file, URL, or paste)
2. Classify intent — Determine domain (event, meal, place, product, receipt, document, civic)
3. Extract — OCR, vision analysis, reverse image search as needed
4. Generate draft — Produce actionable output (calendar entry, macro estimate, expense log, etc.)
5. Confirm — Present draft to user for approval before acting

Example: a user provides a flyer image → agent classifies as "event" → extracts date/venue → generates calendar draft → confirms with user.

Responsibility boundary

Look owns image-to-action conversion: ingest, context inference, domain routing, draft generation, and execution with confirmation.

Look does not own: web research (Sift), preference persistence (Taste), communications (Dispatch).

Ontology types

Look works with these types from spec-ocas-ontology.md:

• Entity/Person — people identified in images (public figures, named contacts).
• Place — venues, locations, and scenes extracted from image context.
• Concept/Event — events visible in images (gatherings, occasions, dated scenes).
• Thing/DigitalArtifact — the source image itself.

Look includes entity signals in journal payloads using the Signal schema from spec-ocas-shared-schemas.md. The payload.type field must be set to the ontology type of the primary extracted entity (Person, Place, Event, or DigitalArtifact). source_journal_type is "Observation".

Supported domains

Events, food macros, places, products, civic issues, receipts, documents.

Read references/domain_playbooks.md for detailed per-domain behavior.

Core workflow

See references/workflow.md for the full step-by-step procedure, error handling table, and clarification rules.

Commands

• look.ingest.image — ingest image(s) with optional EXIF and device pre-parse
• look.propose.actions — generate ActionDrafts with DecisionRecords
• look.execute.action — execute a confirmed draft (requires explicit approval)
• look.rollback.action — attempt rollback for reversible actions
• look.status — last ingest, pending drafts, items awaiting confirmation
• look.config.set — update configuration
• look.journal — write journal for the current run; called at end of every run
• look.update — pull latest from GitHub source; preserves journals and data
• lookup.reverse_search — perform Google reverse image search on an image URL or local file. Returns matching pages, titles, and similar image URLs.

Confirmation and rollback rules

• Draft-first always. No execution without explicit confirmation.
• High-risk actions (purchases, 311 submission, health writes): require per-draft confirmation token.
• Reversible actions (calendar, maps): expose rollback information.

Permission discipline

Default deny. Request minimally. Drafting continues even without execution permissions. Blocked execution reported, not silently skipped.

Boundaries

• Never invent OCR text, barcodes, prices, or license plates
• EXIF capture location is not the event venue
• iOS relay pre-parse is optional evidence, not truth
• The skill must work without relay upload

Error handling

See references/workflow.md for the full error handling table.

Reverse Image Search

Look uses reverse image search during step 5 (Research and validate externally). Primary method: Yandex Images via browser (works from cloud IPs). Fallback: google-image-source-search PyPI package (residential IPs only).

Quick reference:

• Local file → upload to Imgur first → public URL → Yandex search
• Already a public URL → use directly with Yandex
• Google fallback only on residential IP — never retry more than once from cloud

Detailed instructions: references/google-image-source-search.md (Google fallback), references/credential-files.md (Imgur client ID).

Gotchas (also in Gotchas section):

• Google blocks cloud IPs completely — use Yandex
• TinEye returns JS-rendered results — use browser
• Yandex results may contain Cyrillic tags — URLs still usable

Storage layout

text

{agent_root}/commons/data/ocas-look/
  config.json
  state.json
  events.jsonl
  decisions.jsonl
  reports/
  artifacts/

{agent_root}/commons/journals/ocas-look/
  YYYY-MM-DD/
    {run_id}.json

Default config: see references/default_config.md.

OKRs

See references/okrs.md for the full OKR definitions. Universal OKRs from spec-ocas-journal.md apply to all runs.

Optional skill cooperation

• Sift — web research for validation during draft generation (via SearchX)
• Chronicle — entity signals emitted via journal payloads after draft generation

Journal outputs

Observation Journal — all image ingestion and draft generation runs.

Initialization

On first invocation of any Look command, run look.init:

1. Create {agent_root}/commons/data/ocas-look/ and subdirectories (reports/, artifacts/)
2. Write default config.json and state.json if absent
3. Create empty JSONL files: events.jsonl, decisions.jsonl
4. Create {agent_root}/commons/journals/ocas-look/
5. Ensure journal payload fields (see interfaces specification) exists (create if missing)
6. Register cron job look:update if not already present (check the platform scheduling registry first)
7. Log initialization as a DecisionRecord in decisions.jsonl
8. Reverse image search setup (run once):
   • Read references/pip-venv.md for the full setup pattern (install google-image-source-search, pin requests, verify import)

Background tasks

Job name	Mechanism	Schedule	Command
look:update	cron	0 0 * * * (midnight daily)	look.update

text

# Task declared in SKILL.md frontmatter metadata.{platform}.cron

Self-update

look.update pulls the latest package from the source: URL in this file's frontmatter. Runs silently — no output unless the version changed or an error occurred.

See references/self_update.md for the full self-update procedure.

Visibility

public

Gotchas

• Google reverse image search is blocked from cloud IPs — Use Yandex Images via browser instead. Do not retry Google more than once.
• Vision model config: vision_analyze uses the vision.model and vision.provider fields in config.yaml, NOT the auxiliary model. Set to a vision-capable model like nvidia/nemotron-nano-12b-v2-vl:free (provider: openrouter). The agent cannot patch config.yaml directly — use hermes config set vision.model <model> or edit manually.
• Reasoning model incompatibility: Some free vision models (nex-agi/nex-n2-pro:free, nvidia/nemotron-3-nano-omni) are reasoning-only models that return content: null with the answer in message.reasoning. The vision_analyze tool doesn't handle this. Use nvidia/nemotron-nano-12b-v2-vl:free for reliable vision analysis.
• Wikimedia image URLs blocked: OpenRouter proxy returns 400 when fetching wikimedia.org image URLs. Use base64-encoded images or alternative image hosts.
• LLM7.io has no free vision model: GLM-4.6V-Flash is paywalled. Use OpenRouter free vision models instead.
• EXIF location ≠ event venue — The capture location embedded in a photo's EXIF data is where the photo was taken, not necessarily the event venue.
• Never invent OCR text, barcodes, prices, or license plates — If the image content is ambiguous, state uncertainty rather than guessing.
• TinEye returns JS-rendered results — Use the browser if TinEye is needed; it can't be scraped with curl.
• Reverse image search requires a public URL — Upload local files to Imgur first, then use the returned URL for Yandex reverse search.

Support File Map

File	When to read
references/schemas.md	Before creating evidence, drafts, or receipt records; when validating data structures
references/domain_playbooks.md	Before domain routing or draft generation; when checking per-domain behavior
references/decision_policy.md	Before risk assessment or confirmation decisions; when classifying action risk level
references/command_reference.md	Before any command execution; when checking command syntax or parameters
references/storage_and_config.md	Before config changes or storage operations; when modifying skill state
references/journal.md	Before calling look.journal; at end of every run
references/pip-venv.md	Before installing any pip package for a skill; when setting up Python dependencies
references/google-image-source-search.md	Before using Google reverse image search fallback; when on residential IP and Yandex is unavailable
references/credential-files.md	Before reverse image search with local files; when Imgur client ID is needed
references/default_config.md	Before initialization or config changes; when referencing default config values
references/okrs.md	Before writing or evaluating OKRs; when checking skill-level targets
references/workflow.md	Before executing the core workflow; when handling errors during processing
references/self_update.md	Before running look.update; when debugging self-update failures

Update command

This skill self-updates every 24 hours via:

bash

look.update

This pulls the latest version from GitHub and restarts the skill's background tasks if applicable.