Change Intake

Purpose

Use Change Intake after a codebase exists or after WF8/WF9 when the operator brings a new request, idea, config change, or code delta. The goal is to route the change without silently rerunning WF1-WF12 or making broad edits.

References

Read these first:

• ../../../AGENTS.md
• ../../../CLAUDE.md
• ../../../.agents/references/context-layering-policy.md
• ../../../.agents/references/contract-gating-rule.md
• ../../../.agents/references/evidence-chain-rule.md
• ../../../.agents/references/project-map-rule.md
• ../../../.agents/references/language-policy.md
• ../../../.agents/references/ubiquitous-language.md
• ../../../.agents/references/research-supervision-patterns.md
• ../../../.agents/references/research-supervision/idea-evaluation.md

Classifier Inputs

Read current context when present:

• PROJECT_STATE.json
• project_map.json
• docs/20_facts/Codebase_Map.md
• docs/10_contract/Project_Contract.md
• docs/10_contract/Evaluation_Contract.md
• docs/10_contract/Baseline_Contract.md
• docs/10_contract/Claim_Boundary.md
• latest relevant iteration_log.json entry

Routes

• bugfix -> $build
• experiment_delta -> $iterate
• stable_code_delta -> $build
• architecture_delta -> delta grill, $build
• evaluation_delta -> evaluation delta route plus Gate ledger and pre_eval_commit requirement before metric-bearing eval
• claim_boundary_delta -> Claim Delta Evidence plus owning claim/release or writing route
• new_research_direction -> new Research Intent Draft branch
• harness_guardrail_delta -> $harness-maintenance
• unknown -> STEER

Use idea-evaluation.md when the request changes the research direction, claim boundary, evaluation target, dominant improvement axis, or feasibility story. It provides pressure-test context only; it does not approve the change.

Recorded route

Use the visible typed Harness route when the operator wants a recorded change:

python tooling/harness.py change --json

The typed change Recipe records its Action Result under .harness_runtime/ and routes through the Campaign when attached. It does not silently invoke other Operations.

Fail-Closed Rules

• confidence=low routes to STEER, not code edits.
• Evaluation, Baseline, or Claim Boundary impact routes to the owning workflow with a Gate ledger, Claim Delta Evidence when claims change, and a commit checkpoint requirement before train/eval. It does not pause for approval when the change stays inside the active Automation Policy.
• Public interface, config schema, data flow, or primary metric changes are not plain bugfixes.
• Harness guardrail changes never route to $code-debug.

Exit Condition

Produce a Change Request matching schemas/change_request.schema.json, or pause with one concrete steering question and a Gate ledger.