Parent
Part of #46.
Relationships
Depends on #57, #58, #59, #60 and #61.
Goal
Stop treating incidental English wording in system prompts as an API. Replace brittle phrase-presence tests with tests of the runtime contracts and agent behaviours the wording was intended to protect.
Problem
The current prompt/agent test suite contains many assertions of the form:
expect(prompt).toContain("...")
Those tests make prompt sediment difficult to remove because changing wording can fail CI even when behaviour is unchanged or better.
Scope
Audit prompt and agent tests and classify assertions into:
Keep exact/textual assertions only when wording is a real contract
Examples may include:
- required machine-readable terminal markers;
- exact handoff tokens that a parser consumes;
- required include/loading invariants;
- intentionally stable tool/role identifiers.
Replace prose-presence tests with behaviour/structure tests
Examples:
- Todo updates happen automatically rather than the prompt saying they should;
- invalid assignments are rejected rather than the prompt warning about overlap;
- lifecycle gates agree with status rather than the prompt explaining transition rules;
- dispatch payload contains assigned task/finding context rather than role prose repeating it;
- Reviewer output preserves finding semantics rather than checking for one sentence explaining them.
Testing strategy
Prefer, in order:
- pure deterministic unit tests;
- host/tool integration tests;
- dispatch contract tests;
- focused prompt structure/size tests;
- model-driven behavioural scenarios where semantic behaviour cannot be tested deterministically.
Do not replace brittle string tests with equally brittle snapshots of entire prompts.
Acceptance criteria
Out of scope
- Final prompt budget values.
- Full end-to-end benchmark suite.
- Weakening machine-consumed handoff/review contracts.
Parent
Part of #46.
Relationships
Depends on #57, #58, #59, #60 and #61.
Goal
Stop treating incidental English wording in system prompts as an API. Replace brittle phrase-presence tests with tests of the runtime contracts and agent behaviours the wording was intended to protect.
Problem
The current prompt/agent test suite contains many assertions of the form:
Those tests make prompt sediment difficult to remove because changing wording can fail CI even when behaviour is unchanged or better.
Scope
Audit prompt and agent tests and classify assertions into:
Keep exact/textual assertions only when wording is a real contract
Examples may include:
Replace prose-presence tests with behaviour/structure tests
Examples:
Testing strategy
Prefer, in order:
Do not replace brittle string tests with equally brittle snapshots of entire prompts.
Acceptance criteria
Out of scope