Skip to content

Core Concepts

Boundline is easiest to understand as a runtime that keeps AI-assisted delivery bounded, inspectable, and recoverable.

Session Lifecycle

Bounded Sessions

A bounded session is the active container for one delivery goal. It records the goal, context, plan, execution state, findings, traces, checkpoints, and next action.

Start one explicitly:

bash
boundline init --workspace .
boundline goal --workspace . --goal "Add validation for imported customer emails"
boundline plan --workspace .

The session prevents broad, uncontrolled work by making the current boundary visible.

Runtime State

Runtime state lives in the workspace, normally under .boundline/.

Important surfaces:

  • .boundline/session.json: active session state
  • .boundline/config.toml: workspace configuration
  • .boundline/traces/: persisted decision and execution traces
  • .boundline/checkpoints/: rollback manifests for mutating runs

Assistant chat history is not runtime state.

Context Packs

A context pack is the bounded evidence set Boundline uses to plan or continue. It can include:

  • the captured goal
  • authored briefs
  • relevant files
  • failing validation output
  • recent changed files
  • previous traces
  • guidance sources
  • Canon project memory or governed artifacts when configured

If context is insufficient, Boundline should stop and ask for clarification or better evidence.

Expert Packs

Expert packs provide reusable expertise for languages, frameworks, domains, review profiles, and delivery roles. Boundline can select packs from built-in sources, bundled assistant packs, workspace overrides, or compatible governed inputs.

See Shared Expert Packs.

Guidance

Guidance shapes work before and during action. It can describe architecture expectations, testing discipline, language practices, security rules, migration policy, or domain conventions.

Guidance answers: "What should the worker know before acting?"

See Guidance And Guardians.

Guardians

Guardians check work after action or at quality boundaries. They can be deterministic, LLM-based, or hybrid. They emit structured findings such as observations, warnings, risks, blockers, or errors.

Guardians answer: "What should be checked before continuing?"

Structured Findings

Findings are not vague comments. They should preserve severity, source, scope, evidence, and blocking outcome when available. Use status and inspect to read them:

bash
boundline status --workspace . --json
boundline inspect --workspace . --json

Delivery Councils & Authority Zones

Boundline can assemble a Delivery Council at critical delivery boundaries (e.g., planning, implementation, verification) based on the risk and authority zone of the work.

  • Authority Zones: Control classes mapped from Canon's authority-governance-v1 metadata (green, yellow, red, restricted).
  • Council Profiles: Depending on the authority zone, Boundline activates specific council profiles (e.g., yellow_pair, red_five) that dictate the number and type of independent reviewers required before the session can proceed.

Adjudication & Producer Responses

When a Council Guardian emits a concern or blocking finding, it is not just informational. The producer must explicitly respond (accepted, rejected, deferred) with rationale.

  • Accepting a blocking finding automatically generates remediation work or follow-up tasks within the session.
  • Unresolved blocking findings result in a hard_stop or adjudication_required state, preventing the work from bypassing governance.

Reasoning Profiles

Reasoning profiles are bounded runtime challenge patterns that activate when governance posture, stage policy, or operator policy requires stronger challenge than a single linear execution step.

In 0.63.0, the same runtime now carries that reasoning story through the assistant-facing follow-through surfaces: status, inspect, why, risk, evidence, next-best, challenge, and explain-plan can all disclose the active profile, selection reason, contribution, and fallback wording.

The current shipped reasoning-profile surface is intentionally small:

  • bounded_self_consistency
  • independent_pair_review
  • heterogeneous_security_review
  • bounded_reflexion
  • debate as bounded substrate rather than a standalone shipped profile
  • adjudication as a shared primitive rather than a standalone shipped profile

Control Graduation & Adaptive Governance

Starting from version 0.63.0, Boundline keeps governance explicit through rollout profiles and runtime states rather than introducing a second orchestration system or a hidden learning layer.

  • Rollout Profiles: Workspaces progress through maturity levels: minimal, guided, governed, and strict.
  • Runtime States: Governance operates in distinct states (advisory, catch, rule, hook). New or uncalibrated surfaces start in advisory mode.
  • Canon Inputs: Canon can supply required authority-governance-v1 metadata and optional additive adaptive-governance-v1 metadata, but Boundline still resolves the effective runtime posture locally.
  • Degradation & Escalation: If evidence is weak, trust is low, or reviewers are unavailable, Boundline safely degrades (e.g., smaller council, advisory fallback) or escalates explicitly, maintaining traceability rather than silently weakening governance.

Traces

Traces explain what happened and why. They preserve decisions, selected context, route choices, executed commands, skipped checks, findings, blocking states, and next actions.

The 0.63.0 line also projects human-facing inspect closures from the same trace authority: inspect_context, inspect_council, and inspect_timeline.

Delight Feedback Signals

Delight feedback signals are lightweight session-scoped usefulness measures. They are not a second telemetry runtime. Current examples include time to first useful answer, explanation attribution rate, next-action acceptance rate, and the latest next-action outcome.

See Traces And Inspectability.

Recovery

Recovery starts from runtime state, not guesswork:

bash
boundline status --workspace .
boundline next --workspace .
boundline inspect --workspace .

When a checkpoint exists, prefer the reported restore command over a generic reset.

Released under the MIT License.