How it works

Code decides, AI proposes, people approve. A character is a package on a fixed, shared core; every model step runs through the same harness; every action waits for the approval its risk requires.

Four layers, from the one changed most often to the one changed least

Runtime settings

Database, no deploy

Cut-offs, approver groups, templates, autonomy within ceilings - bounded by each package's settings.schema.json.

Character packages

Git, reviewed deploy

characters/<slug>/: character.yaml, workflow.py, its own schema and migrations, api.py and its workspace screens.

Shared blocks

Versioned, pinned

blocks/connectors, tools and skills, each with a manifest: what it does, what it may change, how risky that is.

Platform core

Platform release only

Identity, Cedar-style policies, the gateway, the harness, the engine, approvals, the ledger, cases, the queue.

One model step, as the harness runs it for every character

Step 1Assemble context
Only the step's declared inputs, after the access check; untrusted text wrapped and labelled.
Step 2Scan inputs
The injection scan; a high score goes to a person, not to the model.
Step 3Model call
Local or hosted, through one entry point, with the output schema enforced.
Step 4Validate output
Schema, invariants, quotes against the source, numbers against cited paragraphs.
Step 5Retry or abstain
One repair with the errors spelled out; then the case goes to a person with the reason.
Step 6Save and trace
Output with its provenance; the engine - not the model - picks the next step.

Where things live

core/                platform: harness, gateway, engine, approvals, ledger, cases, packages loader
blocks/connectors/   google, microsoft, telegram, whatsapp, erp, geocoder, files, internal ...
blocks/tools/        document_text, pii_redact, injection_scan, rubric_rank, cite_sources, check_numbers ...
blocks/skills/       screen_cv, draft_jd, draft_brd, draft_process_map, classify_request, research_plan, ...
characters/<slug>/   character.yaml, workflow.py, data/models.py + migrations, api.py, ui/ screens
ui/kit/              the design system every screen is built from
ui/web/              this shell; workspaces are mounted from each package's ui/manifest.json