Module system 

System instructions and prompt management.

Prompt variants share one canonical base contract plus thin mode deltas and compact runtime addenda. Project-specific behavior comes from dynamically loaded instruction maps (AGENTS.md/CLAUDE.md), dynamic tool guidance, skill metadata, and runtime notices.

Re-exports

pub use crate::prompts::static_prompts::agent_identity_label;

pub use crate::prompts::static_prompts::default_lightweight_prompt;

pub use crate::prompts::static_prompts::default_system_prompt;

pub use crate::prompts::static_prompts::lightweight_instruction_text;

pub use crate::prompts::static_prompts::minimal_instruction_text;

pub use crate::prompts::static_prompts::minimal_system_prompt;

pub use crate::prompts::static_prompts::specialized_instruction_text;

pub use crate::prompts::static_prompts::specialized_system_prompt;

pub use crate::prompts::static_prompts::static_profile_prompt;

Structs

SystemPromptConfig

System instruction configuration

SystemPromptReport

Result of measuring a composed system prompt against the configured token budget (agent.max_system_prompt_tokens).

Enums

SectionKind

Identifies which layer of the system prompt a PromptSection belongs to.

Constants

CONTRACT_HEADER

DEFAULT_OPERATING_PROFILE_DELTA

DEFAULT_SPECIFIC_LINES

Default/Lightweight/Specialized mode: extended working style beyond the universal runtime guidance (instruction map, communication detail, corrections, delegation, code style, and test design). Each rule has exactly one home; do not restate runtime-guidance rules here.

LIGHTWEIGHT_OPERATING_PROFILE_DELTA

MINIMAL_OPERATING_PROFILE_DELTA

MINIMAL_SPECIFIC_LINES

Minimal mode has no additional contract lines; universal behavior lives in the compiled runtime-guidance section shared by every profile.

OPERATING_TASK_TRACKER

Shared operating-profile sentences reused across modes.

PLANNING_WORKFLOW_EXIT_INSTRUCTION_LINE

Shared Planning workflow instruction line for transitioning to implementation.

PLANNING_WORKFLOW_HINT

Hint shown when planning workflow is active.

PLANNING_WORKFLOW_IMPLEMENTATION_PROMPT

Shared Planning workflow task-tracking line clarifying availability and aliasing. Implementation prompt used when transitioning from planning to execution.

PLANNING_WORKFLOW_IMPLEMENT_REMINDER

Shared reminder appended when presenting plans while still in Planning workflow.

PLANNING_WORKFLOW_INTERVIEW_POLICY_LINE

PLANNING_WORKFLOW_NO_AUTO_EXIT_LINE

Shared Planning workflow guard line requiring explicit transition from planning to execution.

PLANNING_WORKFLOW_NO_REQUEST_USER_INPUT_POLICY_LINE

PLANNING_WORKFLOW_PLAN_PERSISTENCE_POLICY_LINE

Canonical contract for model-authored plan output and runtime-owned persistence.

PLANNING_WORKFLOW_PLAN_POLICY_LINE

Shared Planning workflow policy line directing context-aware read-only research and plain-text question resolution.

PLANNING_WORKFLOW_PLAN_QUALITY_LINE

Compact, spec-like plan quality line. The previous wording (“summary, steps, test cases, assumptions”) let the model emit verbosely large plans that blew the generation token budget and were cut off mid-<proposed_plan> — which previously re-triggered the recovery loop forever. This mandates a tight spec that fits a small token budget and prefers file:symbol references over prose. It also forbids wrapping those references in markdown link syntax or editor/IDE URI schemes (e.g. vscode-file://, file://) — plans are read in terminals and other non-hyperlink surfaces, and a bare path/to/file.rs:42 reference is portable while a broken pseudo-link pointing at the editor binary itself is not. The canonical one-line step format, mirrored from tools::handlers::planning_workflow::artifacts::CANONICAL_STEP_FORMAT (const contexts cannot concat!, so the sync is enforced by the plan_quality_line_shows_canonical_step_format test below). Showing the exact shape up front matters: the repair directive prints it only after a rejection, and turn_912/913 showed planners repeatedly failing “step lacks a concrete target or verification” without ever seeing an example. The optional ## Expected Outcomes / ## Dependencies and Prerequisites sections are requested only “when material” so plans carry outcomes and prerequisites without inflating every plan past the token budget.

PLANNING_WORKFLOW_READ_ONLY_HEADER

Shared Planning workflow header used by both static and incremental prompt builders.

PLANNING_WORKFLOW_READ_ONLY_NOTICE_LINE

Shared Planning workflow notice line describing strict read-only enforcement.

PLANNING_WORKFLOW_RESEARCH_SCOPE_LINE

Scale research effort to the request instead of always exhaustively enumerating the repository. Checkpoint turn_647 showed a “make a simple plan to improve launch time” request burn 70+ tool calls across dozens of files until the turn’s tool wall-clock budget was exhausted with no plan delivered — the model had no signal to stop researching and draft. This line gives it a concrete budget to self-regulate against.

PLANNING_WORKFLOW_TASK_TRACKER_LINE

PROMPT_INTRO

Identity line. apply_agent_identity replaces the first VT Code in this line with the active agent label, so the label must appear exactly once.

PROMPT_ROLE_PARAGRAPH

Natural-language role framing inserted between the identity line and the runtime guidance for the Default and Specialized profiles. It sets the working posture and effort calibration; the concrete rules live in the sections below it. Omitted from Minimal and Lightweight modes to respect their compact budgets (see the parent-ratio guard in subagents/config.rs).

PROMPT_TITLE

SHARED_CONTRACT_LINES

Contract rules shared across all prompt modes that are not universal user-facing runtime guidance: state the harness must carry across compaction. Grounding, honesty, and scope rules live in runtime_guidance::RUNTIME_GUIDANCE_SECTION so they also survive a workspace system.md override.

SPECIALIZED_OPERATING_PROFILE_DELTA

Functions

apply_coordinator_role_guidance

Align cached base guidance with the selected role, including switches away from the coordinator.

apply_output_style

Apply output style to a generated system instruction

compose_system_instruction_text

Compose the base system instruction plus compact tool/skill/environment addenda.

compose_system_instruction_with_report

Compose the system instruction and return the token-budget report alongside it. See SystemPromptReport and SectionKind::trim_priority for the budget/trim behavior driven by agent.max_system_prompt_tokens, agent.system_prompt_budget_warning, and agent.trim_system_prompt.

estimate_token_count

Estimate prompt tokens through the shared workspace tokenizer.

generate_lightweight_instruction

Generate a lightweight system instruction for simple operations

generate_minimal_instruction

Generate a minimal system instruction (pi-inspired, <1K tokens)

generate_specialized_instruction

Generate a specialized system instruction for advanced operations

generate_system_instruction_with_config

Generate the stable base system instruction with configuration-aware sections.

generate_system_instruction_with_config_and_report

Same as generate_system_instruction_with_config but also returns the SystemPromptReport for the composed prompt, whether served from cache or freshly built.

generate_system_instruction_with_context_and_report

Generate a system instruction using a context-aware cache identity.

measure_system_prompt_size

Measure the system prompt size without applying budget trimming or warnings.