Starter products: agent writers and context profiles

Non-normative starter-product guidance (RFC 0078). Not part of the standard.

This page moved from UW Protocol §IX under RFC 0078 (D3). Section numbers are kept so existing references still find the text. The requirement keywords below (MUST, SHOULD, MAY) describe how the reference starter products behave and what a tool must do to stay compatible with them. They are not conformance requirements of the standard, and a conforming Reader or Verifier owes none of them.

The reference toolkit hosts no agents. @uwmd/core and the uwmd CLI run no model, ship no provider adapter, tool schema or layer registry, and have no uwmd run or uwmd layers command (RFC 0112 removed the layer vocabulary). What remains is the context profiles (buildContext) and the generic writers (writeAgentBlock, writeErrorEntry, buildMeta). Any host that runs its own agents writes their records through those writers or through the edit API (applyEdit, uwmd edit --agent-id/--agent-version); the format's provenance, supersede and pipeline-log rules, and its producer obligations (Format §3.6), apply to it exactly as to any other producer.

IX. AI Host Contract

§IX.1 (layer DAG), §IX.2 (context bundling) and §IX.8 (layer-declared profile consumption) described the removed layer registry and are retired with it (RFC 0112).

IX.4 _meta substitution

Agents may include _meta and _notes inside section_data. The host MUST strip both keys from the agent payload and substitute the canonical _meta constructed via buildMeta(). This prevents agents from forging provenance. The host passes that _meta to writeAgentBlock(), which also drops any _role or _overrides the payload carries.

IX.5 Pipeline log

Every successful agent write MUST append a pipeline_log entry with timestamp, agent_or_actor, event_type, and status. Failures MUST also be logged, with status: "failed" and an error string. writeAgentBlock() appends the success entry and writeErrorEntry() the failure entry.

IX.7 Context Profiles

A consumer of a .uwx.md file SHOULD declare which context profile it consumes. The profile determines which sections are included, whether superseded blocks appear, whether prose is included, and the approximate token budget the producer targets when constructing the context payload.

The five profiles are:

Profile	Audience	Includes	Notes
summary	Chat agents, orchestration heuristics	Frontmatter slim subset (including status and deal_stage), quick_metrics, gaps head	Soft target ≤ 600 tokens (chars/4 estimate).
live	Calc-aware editors	All non-superseded sections, full prose	Default for human-facing tools.
compact	LLMs at scale	All non-superseded sections, JSON only, minified, ordered stable→volatile	Targets ≤ 55% of live token count on the canonical Parkview fixture. _meta stripped by default.
full	Archival / forensic	Every byte, prose included, no compaction	Round-trips the input file byte-for-byte when no truncation.
relevant	Callers that need a subset	Caller-supplied section list, live-style payload	The caller passes the sections to include.

A producer that satisfies a request for profile X MUST NOT return a context strictly larger than X allows; it MAY return a smaller context (e.g. truncated to fit maxTokens) and MUST set truncated: true in the result when it does.

The compact profile MUST order sections stable-first so prompt- cache prefixes survive frequent edits to volatile downstream sections. The canonical order is:

1. frontmatter
2. property
3. ownership
4. borrower_sponsor
5. debt_structure
6. sources_uses
7. valuation
8. rent_roll
9. operating_statement
10. noi_model
11. market_analysis
12. compliance
13. risk_assessment
14. dcf
15. stress_tests
16. custom_calculations
17. gaps
18. extensions

Sections not on this list trail the canonical prefix in Object.keys order.

Token estimates use the chars/4 approximation; producers MUST document the approximation and SHOULD validate periodically against the true tokenizer cost (typical drift ±5% on .uwx.md content).