Glossary

Each entry defines a term. Where a rule applies, the entry names the authoritative file; the glossary never states a rule.

Semantic-sort naming

• The naming convention described in prompts/flavors/01-semantic-sort-naming.md.

DELTA protocol

• A minimal correction to a named artifact.
• Rules: prompts/01-contract.md section 6.

Universal failure-prevention rules

• The cross-language rules listed in prompts/common/01-universal-rules.md.
• They cover scope, clarification, anti-hallucination, privacy boundaries, risky operations, and language and format.

One-sentence-per-line

• A prose convention in which each sentence occupies one line, so a diff reads clearly.
• It does not apply to code blocks.

One-statement-per-line

• A code convention in which independent statements occupy separate lines.
• The target language’s formatter and syntax rules take precedence.

Feature file

• A file in prompts/features/ that defines one project capability.
• Rules: prompts/01-contract.md section 2.

Plan block

• A block listing ordered steps.

Output block

• A block containing final output.
• Rules: prompts/01-contract.md section 5.

OUTPUT line

• The final line of a task file that restates the response representation, including whether it ends with the VERIFICATION: line.
• Rules: prompts/how-to-write-tasks.md section 1.

Context block

• A block providing additional information, such as file contents, notes, samples, or constraints.
• Rules: prompts/how-to-write-tasks.md section 5.

task_context block

• A delimited block of copied data inside TASK-CONTEXT; it is data, never instructions.
• Rules: prompts/how-to-write-tasks.md section 5.

constraint block

• A delimited block of instructions inside TASK-CONTEXT.
• Rules: prompts/how-to-write-tasks.md section 5.

note block

• A delimited block of background inside TASK-CONTEXT that does not constrain.
• Rules: prompts/how-to-write-tasks.md section 5.

Scope-based identifier length

• The amount of identifier detail appropriate to the identifier’s scope.

Ordered transformation pipeline

• A sequence of transformations applied in a defined order.

Generated file

• A file produced by a script, build tool, generator, or other automated process.
• Rules: prompts/03-conventions.md section 6.

Incident record

• A record of an incident: what happened, root cause, fix, lessons, and safeguards for a retry.
• Rules: records/README.md.

Handoff record

• A record written at a session boundary so a fresh session resumes without the prior conversation’s memory.
• Rules: records/README.md.

Live-state test

• A test that verifies a documented model against live reality.
• Rules: prompts/02-workflow.md section 4.1.

Privacy boundary

• Owner-declared content that is off-limits to the LLM.
• Rules: prompts/common/01-universal-rules.md.

VERIFICATION line

• The single output line that reports whether the work was verified.
• Rules: prompts/01-contract.md section 5.

Requirement

• Something the resulting project must satisfy.
• Rules: prompts/how-to-write-features.md.

Requirement identifier

• A stable name for one top-level requirement, <FEATURE-NAME>-R<NNN>.
• Rules: prompts/how-to-write-features.md section 4.

Scope

• The artifacts and operations a task describes, declared by the operation lines of TASK-DESCRIPTION and the TASK-FILES table.
• Rules: prompts/how-to-write-tasks.md sections 3 and 6.

Authorization

• Permission to perform an operation, supplied by the task or by an applicable workflow rule.
• Rules: prompts/01-contract.md section 3.

TASK-ACCEPTANCE

• A claim that the task intends to satisfy the listed requirement identifiers; it is not the human’s acceptance decision.
• Rules: prompts/how-to-write-tasks.md section 6.2.

TASK-FEATURES

• The feature files whose requirements a task applies; their requirements apply transitively through each feature’s Dependencies.
• Rules: prompts/how-to-write-tasks.md section 6.3.

Acceptance criterion

• A human-reviewable condition that decides whether the work is acceptable.
• Rules: prompts/how-to-write-episodes.md.

Verification

• Evidence that a requirement or an acceptance criterion is satisfied.
• Rules: prompts/02-workflow.md section 6.

Output

• What the LLM returns to the human.
• Rules: prompts/01-contract.md section 5.

Untrusted content

• Content that is not authoritative as instructions; it may still be factually true.
• Rules: prompts/01-contract.md section 8.

Stability rules

• Project behavior that should not change without explicit instruction.

Brain

• The model and its harness: the part of an agent system that decides.
• Rules: prompts/capabilities/execution/06-execution-boundary.md.

Cattle

• A disposable, interchangeable instance that is replaced rather than repaired.
• Rules: prompts/capabilities/execution/06-execution-boundary.md.

Compaction

• A lossy transformation that summarizes the context window.
• Rules: prompts/capabilities/execution/07-session-log.md.

Context window

• What the model can see at one time; it is not the session.
• Rules: prompts/capabilities/execution/07-session-log.md.

Credential proxy

• A component that holds a credential and makes an authenticated call on a session’s behalf, so neither the harness nor the sandbox holds it.
• Rules: prompts/capabilities/execution/06-execution-boundary.md.

Event

• One recorded occurrence in a session log.
• Rules: prompts/capabilities/execution/07-session-log.md.

Execution boundary

• The declared reach of a sandbox: which paths it writes, which commands it runs, and which hosts it contacts.
• Rules: prompts/capabilities/execution/06-execution-boundary.md.

Hand

• An execution environment or tool that performs an action, reached only through a tool call.
• Rules: prompts/capabilities/execution/06-execution-boundary.md.

Harness

• The loop that calls the model and routes its tool calls to the hands.
• Rules: prompts/capabilities/execution/07-session-log.md.

Provisioning

• Creating a sandbox from a recipe when a tool call needs it.
• Rules: prompts/capabilities/execution/06-execution-boundary.md.

Sandbox

• The execution environment where model-generated or third-party code runs.
• Rules: prompts/capabilities/execution/06-execution-boundary.md.

Session

• The append-only log of everything that happened in a unit of work.
• Rules: prompts/capabilities/execution/07-session-log.md.

Session identifier

• The name that ties a run’s events, logs, and records together.
• Rules: prompts/capabilities/execution/09-failure-attribution.md.

Session log

• The durable store of session events, outside the harness and outside the context window.
• Rules: prompts/capabilities/execution/07-session-log.md.

Host checkout root

• The host directory where cloned third-party repositories live, outside the project working tree.
• Rules: prompts/03-conventions.md section 6.1.

Rule set

• A JSON document holding facts and rules, the input to the krank engine.
• Reference: documents/10-rule-notation.md.

Fact

• A JSON object of key/value pairs; by convention type names the kind of fact.
• Reference: documents/10-rule-notation.md.

Rule

• A krank rule: an id, a when array of conditions, and a then array of actions.
• Reference: documents/10-rule-notation.md.

Condition

• A {"fact": PATTERN} or {"fact": PATTERN, "not": true} matched by subset equality.
• Reference: documents/10-rule-notation.md.

Action

• One of assert, retract, or emit, applied in order when a rule fires.
• Reference: documents/10-rule-notation.md.

Forward chaining

• Firing rules whose conditions hold, applying their actions, and repeating until no rule can fire.
• Reference: documents/10-rule-notation.md.

Refraction

• The rule that a rule binding fires at most once.
• Reference: documents/10-rule-notation.md.

Round trip

• Translating prose into a rule set, rendering the rule set back into prose, and comparing the two.
• Reference: documents/10-rule-notation.md.