Architecture

Boundline is a bounded delivery runtime for AI-assisted engineering work. It keeps software delivery in inspectable, governed, and recoverable runtime state instead of treating chat history as the source of truth.

High-Level Architecture

Component Drill-Down

Primary Runtime Model

The normal operator path is goal-first:

init → goal → plan → run → status → next → inspect

Two optional preflight surfaces can sit in front of that loop:

• boundline models auth ... for user-scoped provider credentials
• boundline probe for a read-only readiness answer

Read-side commands project persisted runtime state from .boundline/ instead of inventing a new story from logs or chat memory.

Repo-Visible Document Boundary

Boundline keeps runtime state separate from repo-visible delivery knowledge:

• .boundline/ owns session state, traces, checkpoints, and transient governance artifacts.
• .boundline/context-intelligence/ owns the derived retrieval DB, manifest.json, and SQLite WAL/SHM sidecars used by local semantic retrieval.
• .canon/ owns raw Canon packets and Canon runtime payloads.
• docs/project/ owns stable reusable project memory.
• docs/evidence/ owns durable feature outputs and evidence bundles.

See [[Project Memory Structure|Project-Memory-Structure]].

Core Runtime Decisions

• Session runtime is the authoritative control plane for lifecycle transitions, goal state, next-action projection, and recovery posture.
• Planning is evidence-driven: goals, briefs, workspace evidence, recent traces, and compatible Canon artifacts shape one bounded context pack.
• Execution stays bounded: reads, edits, validation, and follow-through happen as explicit actions, with checkpoints available before mutation.
• Guidance, guardians, traces, and checkpoints make the path inspectable rather than conversationally implied.
• Canon remains an external governed-knowledge boundary, not Boundline's orchestrator.

The current Boundline 0.66.0 line documents Canon 0.63.0 support for canon governance start|refresh|capabilities --json v1.

The same runtime now owns the derived semantic index lifecycle under .boundline/context-intelligence/, including the retrieval DB, manifest.json, explicit boundline index ... commands, and lightweight stale-mark hook behavior when the operator opts in.

Framework Adapter Boundary

Framework adapters extend the runtime without replacing it.

• Boundline remains the orchestrator and the default execution path.
• One workspace may select one adapter or none.
• The host owns capability validation, config persistence, routing decisions, and operator-visible status or inspect output.
• Adapters only own the stages they explicitly declare and successfully claim.

The V1 wire contract stays bounded to one-shot trusted local subprocess commands, UTF-8 JSON over stdin/stdout, one standard success or error envelope on stdout, optional structured stderr copied into traces only as enrichment, and no graceful shutdown or persistent transport lifecycle.

Current public repositories on this line:

• boundline-framework-template: reference template for building a compatible framework adapter.
• boundline-adapter-speckit: concrete Speckit adapter implementation used to exercise the claimed plan and run handoff path.

Runtime Families

Boundline currently ships two runtime-owned algorithm families:

• review-council assembly, independence guarding, vote resolution, and bounded adjudication projection
• reasoning-profile activation, independence assessment, bounded profile outcomes, and confidence handoff

The currently surfaced reasoning profiles are bounded_self_consistency, independent_pair_review, heterogeneous_security_review, and bounded_reflexion.

Runtime Data Flow

1. Observe: assemble bounded context from workspace evidence, authored input, traces, and compatible Canon artifacts.
2. Decide: resolve the next bounded action and any required governance, council, or reasoning behavior.
3. Act: execute the action and persist session plus trace updates.
4. Verify: use validation, guardians, and any required follow-through to decide whether the session may continue, degrade, or stop.