— SCOPING & DISCOVERY · A CODEDROVERS FIELD UNIT —
An Agent That Reasons Like a Senior Implementation Architect.
The Scout close-reads a client request, introspects the live workspace read-only, and returns a structured discovery agenda — the questions worth asking before a single use case is scoped. He proposes. The architect decides.
0WRITE PATHS IN THE RUNTIME
4TYPED GAP CLASSES
25ITERATION SAFETY CAP
1ARCHITECT WHO OWNS THE OUTPUT
— THE PREMISE · THE AGENDA IS THE DIFF —
Local knowledge
What should be true
A curated use-case library: the events, attributes, and integrations each use case requires when built correctly.
Live workspace, via MCP
What is true
Read-only introspection of the marketing platform: event schemas, taxonomy mappings, customer property schemas, workspace scale.
The deliverable
The discovery agenda
The diff between the two, classified into typed gaps — grounded in workspace evidence, not client assertion. Ready for the next client conversation.
The cross-reference is the architectural move that makes the Scout an agent rather than a lookup tool: he queries what should be true and probes what is true in the same reasoning step. A single request triggers an iterative loop — batched tool calls, each round informed by the last — that ends when the agenda is complete or at the 25-iteration cap, whichever comes first. One vague-input session ran more than 20 tool calls before the Scout was satisfied.
— FOUR TYPED GAPS · AS CLASSIFIED IN THE FIELD —
Event-family gap
demo: double_opt_in not present
The required event type does not exist in the workspace at all.
Attribute gap
demo: consent_list absent on consent event
The event exists but lacks a field the use case depends on.
Integration gap
demo: SMS provider unconfirmed in schema
The use case needs an integration the workspace can't confirm is wired up.
Mapping ambiguity
demo: 12 taxonomy slots null, 27 event types flowing
Events flow under custom names never wired to the standard taxonomy.
Field finding · the load-bearing condition
Manual inventory recorded 21 event types and flagged view_item as absent. The live probe returned 27, with view_item carrying 89,717 events — and every one of the 12 standard taxonomy slots mapped to null. Rich behavioral data flowing; none of it wired to the taxonomy.
A finding manual inspection missed and live probing surfaced. It reframed the discovery agenda before a single use case was scoped.
— CODE OF CONDUCT · RESPONSIBLE DESIGN —
Every claim below is one the Scout can back with receipts. The ones he can't make are listed too.
Read-Only by Design
No write path exists in the runtime. Every tool is a get, list, search, or find. The Scout does not write data, modify configuration, create segments, or send messages. This is the property the rest of the design depends on.
Recommendation, Not Execution
The output is a discovery agenda for architect review — never a scope, a quote, or anything client-facing without a human signature. The reasoning trace is fully visible, so every classification can be interrogated. The human owns the output.
Nothing Persists
Session state lives in memory and is discarded on reset. No workspace data, customer records, or response payloads are persisted beyond the active session. Trace export exists — but it is architect-initiated and visible, never background.
Everything Is Logged
Every tool call emits a structured access line: timestamp, tool, source, parameters, size, status, latency. Payloads never reach the audit tier. The access log is always on — it is the trail the whole posture depends on.
Bounded Reasoning
The loop is capped at 25 tool iterations per session; at the cap, the Scout surfaces what he has rather than riding on. Rate limits are respected by pacing calls under the platform's cap by design, not by aggressive retry.
No Credentials on Paper
Keys live in the environment, never in source or logs; error messages name the variable, not the value. Token-bearing responses are explicitly excluded from logging.
What is not claimed
"Propose, don't decide" is a workflow property, enforced in the system prompt and the review step — not a database-level write block beyond the read-only design above. And no claim is made that a malicious operator with code access couldn't rewire the agent; the posture is honest about how the system is deployed and reviewed, not armored against adversaries no agent with credentials could survive. Identified for pre-production: managed audit-log retention, broader anomaly detection, a general redaction layer.
THE SCOUT
— A CODEDROVERS FIELD UNIT —
Practice note:
"The map before the ride."
— The Scout, agenda attached
"The map before the ride."
— The Scout, agenda attached
THE PRACTICE
© Indifferencer LLC · The Scout is a scoping & discovery unit of the CodeDrovers practice. · Demonstrated on reference workspaces; no client data lives here. · Architect review required for all client deliverables.