Skip to content

System components

This document identifies the parts required to make the Conductor vision usable. It defines responsibilities and boundaries. Ownership vocabulary lives in the authority model. Core versus organization layering lives in the organization profile. Detailed missing contracts remain in the vision gap register.

Primary flow

text Requesting agent │ ├── preparation queries │ └── policy + state + estate + precedent │ └── proposed plan / document bundle │ ▼ Conductor supervisory agent │ ├── asks only unresolved questions ├── reviews against current context └── authorizes an exact plan │ ▼ Requesting agent executes │ ├── updates Notion state ├── updates production evidence └── submits reusable learning when warranted

Conductor reads organizational and production state but does not write routine project state. The executing agent owns the Notion and production updates required by the authorized plan.

Component boundaries

Each component below is classified as core (reusable Conductor), organization (profile or org policy), adapter (environment wiring), state (live records in SoR), or derived (index/cache).

Requesting agent

Aspect Definition
Purpose Propose work, obtain authorization, execute the authorized plan, persist state and evidence
Inputs User intent; authorized pack; permitted production tools
Outputs Production artifacts; Notion updates; optional experience or proposal submissions
Reads Authorized pack; tools granted by role
Writes Notion state and production artifacts; feedback submissions
Depends on MCP gateway; authoritative production systems
Non-responsibilities Issuing authorization; changing governed corpus; acting as Initiative SoR
Classification External actor (not Conductor core)

MCP gateway

Aspect Definition
Purpose Portable authenticated interface to Conductor tool families
Inputs Authenticated client requests
Outputs Versioned tool responses
Reads Auth configuration; routes to tool layers
Writes Nothing in project state; may emit audit via authorization audit and observability components
Depends on Client identity and authentication; access-policy enforcement
Non-responsibilities Compilation policy; Notion writes; production mutations
Classification Core interface; endpoint and tokens are adapter/org

Preparation tool family

Aspect Definition
Purpose Deterministic discovery and optional fast-path validate
Inputs Line, stage, bindings, Initiative/workstream identifiers, query parameters
Outputs Lists, context summaries, validation findings (validated or gap findings)
Reads Corpus; organization profile; Notion/GitHub read adapters; estate index
Writes Preparation results only
Depends on Corpus; read adapters; estate index
Non-responsibilities Authorization; policy activation; mandatory gate before conversation
Classification Core

Conversation tool family

Aspect Definition
Purpose Plan review, refinement, and authorize of an exact plan
Inputs Messages; proposed plan or document bundle; thread context
Outputs Questions; findings; refusals; revision requests; escalations; authorized pack when issued
Reads Preparation results; corpus; organization profile; state snapshots; proposed-plan store
Writes Conductor review artifacts and authorization records only — not proposed plans, not Notion project state
Depends on Supervisory agent; compiler and selector
Non-responsibilities Executing work; Hector approval for exceptions; routine writeback
Classification Core

Feedback and proposal tool family

Aspect Definition
Purpose Accept reusable experiences, improvement candidates, and governed policy proposals (propose)
Inputs Structured experience or proposal payloads
Outputs Stored experience/candidate references; proposal issue or PR linkage
Reads Linked evidence metadata; corpus allowlists
Writes Experience store; candidate store; proposal records — not live Initiative state
Depends on Governance layer; experience store
Non-responsibilities Activating policy; completing Initiatives
Classification Core

Conductor supervisory agent

Aspect Definition
Purpose Review, question, reconcile, and authorize ordinary execution plans within delegated policy
Inputs Proposed bundle; preparation context; corpus; organization profile
Outputs Authorization or findings; review artifacts; escalation when human approval is required
Reads All preparation and policy inputs; proposed plan from requesting agent
Writes Conductor review artifacts and authorization records only — not proposed plans; not Notion project state
Depends on Compiler; corpus; organization profile; read adapters
Non-responsibilities Writing routine Notion/project state; replacing Hector for policy/exceptions; multi-role packs
Classification Core

Compiler and selector

Aspect Definition
Purpose Select the smallest relevant policy and context set for one stage and role
Inputs Initiative/workstream snapshot; stage; role; bindings; corpus version
Outputs Versioned execution pack content (with authorization metadata when issued)
Reads Corpus; organization profile; state snapshot
Writes Pack artifacts only
Depends on Corpus; organization profile
Non-responsibilities Discovering estate alone; mutating SoR; inventing missing bindings
Classification Core

Corpus

Aspect Definition
Purpose Store governed workflows, inventories, standards, roles, patterns, and templates
Inputs Approved PRs from governance
Outputs Units consumed by compiler and preparation
Reads Approved policy sources
Writes Only through proposal → Hector approval → PR
Depends on Governance layer
Non-responsibilities Holding live Initiative state; replacing organization profile for host-specific IDs
Classification Core shapes plus organization policy units (see organization profile)

Organization profile

Aspect Definition
Purpose Bind reusable Conductor concepts to this organization's roles, tools, providers, vocabulary, and approval model
Inputs Hector's organization decisions
Outputs Configuration consumed by runtime and compiler
Reads Organization decisions
Writes Governed configuration only
Depends on Authority model layering
Non-responsibilities Redefining core semantics; storing secret values; acting as Asset registry
Classification Organization

Notion read adapter

Aspect Definition
Purpose Retrieve Initiative, workstream, Product, Asset, decision, and event state
Inputs Stable IDs; query filters; role permission context
Outputs Versioned state snapshots with provenance and freshness
Reads Notion
Writes Nothing
Depends on Adapter credentials; organization Notion location references
Non-responsibilities Creating or updating Notion pages; owning completion writeback
Classification Core adapter contract; wiring is environment adapter

GitHub read adapter

Aspect Definition
Purpose Retrieve repositories and software delivery evidence
Inputs Repo identity; PR/check/release queries
Outputs Evidence metadata with timestamps and provenance
Reads GitHub
Writes Nothing
Depends on Adapter credentials; organization-allowed orgs/repos
Non-responsibilities Creating issues/PRs; owning Initiative state; replacing Asset registry
Classification Core adapter contract; wiring is environment adapter

Asset registry

Aspect Definition
Purpose Authoritative descriptions of enduring systems (applications, services, sites, tools, prototypes)
Inputs Human-reviewed Asset records; linked provenance
Outputs Capability and relationship data for reuse decisions
Reads Via Notion read adapter
Writes Agents/humans with Asset permission — not the Conductor compiler
Depends on Notion as organizational state layer
Non-responsibilities Being a second GitHub; storing full source trees
Classification State (schema designed in M1)

Estate index

Aspect Definition
Purpose Normalize searchable metadata across Notion Assets, GitHub, and approved sources for retrieval speed
Inputs Registry and adapter reads
Outputs Ranked search hits with relevance explanation and source pointers
Reads Notion registry, GitHub, other approved sources
Writes Derived cache or index only
Depends on Asset registry; read adapters
Non-responsibilities System of record; silent merge of duplicates; overriding reviewed Asset text
Classification Derived

Experience store and scheduled reviewer

Aspect Definition
Purpose Persist reusable experiences; periodically create non-binding improvement candidates from new evidence
Inputs Feedback submissions; review cursor
Outputs Experience records; candidate improvements
Reads New structured experiences only within privacy rules
Writes Experience and candidate stores — never live corpus activation
Depends on Feedback layer; privacy classes from organization profile
Non-responsibilities Editing governed assets; authorizing execution; writing Notion Initiative state
Classification Core stores; cadence defaults may be organization

Governance layer

Aspect Definition
Purpose Protect behavior-bearing changes through proposal, Hector approval, and PR
Inputs Proposals and candidate evidence
Outputs Approved merges; rejection or duplicate closure
Reads Proposal and candidate evidence
Writes Approved PR content only; no automatic activation without merge
Depends on Approver identities from organization profile
Non-responsibilities Ordinary plan authorization; project management
Classification Core process; approver list is organization

Conductor identity allocator

Aspect Definition
Purpose Mint stable identifiers for Initiatives, workstreams, Products, Assets, Decision/Events, and authorization records
Inputs Allocation request type and minimal context needed for uniqueness
Outputs Fresh initiative_id, workstream_id, product_id, asset_id, event ID, or authorization ID
Reads Existing ID registries / uniqueness stores as required
Writes Allocator registry entries only — never Notion Initiative/Product/Asset project fields
Depends on Runtime identity of the requesting client
Non-responsibilities Persisting IDs into Notion or production SoR; deciding Initiative content; issuing plan authorization
Classification Core

The identity allocator is the sole minting component for every ID above, including authorization IDs. The Conductor supervisory agent may request an authorization ID and then emit it inside the authorization record; it does not co-allocate. The executing agent acting for the responsible decision owner persists Initiative, workstream, Product, Asset, and Decision/Event IDs into the authoritative store. Minting alone is not project-state writeback.

Client identity and authentication

Aspect Definition
Purpose Establish who the calling client is (authenticate)
Inputs Credentials or tokens presented by the client
Outputs Authenticated client identity or rejection
Reads Auth configuration; secret references for token verification
Writes Authentication events to the runtime authN store
Depends on Secret resolution; organization endpoint configuration
Non-responsibilities Plan authorize; Initiative outcomes; privacy redaction of pack content
Classification Core controls; credential wiring is adapter/org

Access-policy enforcement

Aspect Definition
Purpose Decide whether an authenticated client may invoke a tool or see a resource (enforce access)
Inputs Authenticated identity; requested operation; resource metadata
Outputs Allow or deny with reason
Reads Access policy configuration; organization profile permission bindings
Writes Access-decision logs
Depends on Client identity and authentication
Non-responsibilities Plan authorize; corpus edits; Notion writeback
Classification Core controls

Secret resolution

Aspect Definition
Purpose Resolve configured secret names to secret values at runtime with least privilege
Inputs Secret name / binding from organization profile or adapter config
Outputs Ephemeral secret value to the caller that needs it; or explicit missing-secret failure
Reads Secret provider
Writes Nothing in project state; may write resolution failure telemetry
Depends on Secret provider; organization secret-name bindings
Non-responsibilities Choosing which secret names exist (profile); storing values in packs, corpus, docs, or state
Classification Core controls; provider account is adapter

Privacy filtering

Aspect Definition
Purpose Omit or redact context the active role or client must not see
Inputs Candidate context; privacy class; active role / client identity
Outputs Filtered context or explicit redaction markers
Reads Privacy class assignments from organization profile; state and adapter payloads
Writes Nothing in authoritative SoR
Depends on Access-policy enforcement; organization privacy classes
Non-responsibilities Deciding Initiative outcomes; plan authorize
Classification Core controls; class assignments are organization

Authorization audit

Aspect Definition
Purpose Persist attributable records of plan authorization, refusal, and escalation
Inputs Authorization or escalation events from the supervisory agent
Outputs Immutable audit entries (authorization ID, digest, versions, outcome, time)
Reads Supervisory agent outputs
Writes Authorization audit store only
Depends on Conductor supervisory agent; identity allocator for authorization IDs
Non-responsibilities Notion Decision/Event content; client authenticate; mutating packs after issue
Classification Core controls

Operational observability

Aspect Definition
Purpose Emit health, latency, error, and operational telemetry for the Conductor runtime
Inputs Runtime metrics and logs
Outputs Telemetry streams and dashboards
Reads Runtime process signals
Writes Observability store
Depends on Runtime hosts
Non-responsibilities Initiative state; estate truth; plan authorization decisions
Classification Core controls; hosts are adapter/org

These six runtime responsibilities may share one deployment later. Architecture treats them as separate inputs, outputs, writes, dependencies, and non-responsibilities. None of them performs MCP authorize.

Knowledge layers

Conductor must keep different kinds of knowledge separate because they have different authority and lifecycles.

Layer Example Authority
Core model Workstream, stage, inventory, authorization Conductor architecture
Organization policy Secret management, deployment defaults, approval roles Hector-approved corpus/profile
Initiative state Outcome, active workstreams, next action, evidence Notion
Software delivery state Repository, PR, checks, release, deployment GitHub
Estate description Existing app purpose, capabilities, lifecycle, links Notion asset registry with source provenance
Descriptive history What happened and why Notion events plus authoritative artifact links
Experience Blocker, workaround, failure, repeated question Conductor feedback store
Candidate policy Possible new or changed pattern Non-binding until governed approval

State model

Notion

Notion is the organizational state layer. The minimum proposed model is:

Record Purpose
Initiative Why work exists, outcome, definition of done, priority, relationships
Workstream One line's stage, role, status, bindings, evidence, and next action
Product Product evaluation, scorecards, gates, and ship signals
Asset Enduring system, application, service, site, tool, or prototype that already exists
Decision/Event Durable evidence of transitions, decisions, exceptions, and outcomes

Whether Workstreams and Decisions are separate databases or structured children of an Initiative is decided for M1 in state-model.md: five separate databases (Gate A draft). Assets must remain distinct from Initiatives: an Initiative is work with an outcome; an Asset persists and may be changed by many Initiatives.

GitHub

GitHub is authoritative for software production evidence:

  • Repository identity and source.
  • Branches, commits, and pull requests.
  • Required checks and test outcomes.
  • Releases and deployment evidence.
  • Optional repository-local issues.

GitHub does not contain the canonical Initiative or estate description. A README may be evidence, but it is not guaranteed to answer why an asset exists, which need it serves, or whether it overlaps another asset.

Asset registry fields

The Asset registry enables the central checkpoint: before authorizing new work, Conductor must determine whether an existing system can satisfy the need, should be extended, or is materially different.

Minimum Asset metadata:

Field Purpose
asset_id Stable identity independent of repository or machine
name and aliases Discovery across renamed or duplicate projects
purpose Why the asset exists
users and needs Who it serves and what problem it addresses
capabilities What it currently does, in comparable terms
lifecycle status Proposed, prototype, active, maintained, superseded, archived
estate Organization-specific classification such as Prototype, Lab, Portfolio-org
repositories GitHub and any exceptional local-only source references
environments and URLs Where verified instances run
stack and deployment shape Technical comparison and reuse constraints
related Initiatives and Assets Provenance, replacements, forks, dependencies, and overlap
privacy and access class What Conductor and requesting agents may see
last verified and evidence Freshness and source links

Capability descriptions matter more than names or stacks. The three real-estate applications may use different technologies while still solving substantially the same user need.

Read access and indexing

Conductor requires read access to Notion and GitHub wherever it runs. Secret values remain only in the configured secret provider and are never included in a pack. Secret names and provider bindings come from the organization profile or adapter configuration.

The read path must provide:

  • Organization and user identity.
  • Permission-filtered queries.
  • Source timestamps or versions.
  • Stable identifiers and provenance.
  • Explicit stale, unavailable, and partial results.
  • No assumption that the local machine contains the full estate.

The estate index is derived for retrieval speed. It is not a new system of record. Search results must point back to Notion, GitHub, or another authoritative source and state why each result is relevant.

Initial prepopulation

Conductor cannot prevent duplicate work until the current estate is represented. Bootstrap requires a one-time, reviewable inventory process (M1):

  1. Enumerate repositories from the approved GitHub users and organizations.
  2. Enumerate known local-only prototypes, workspace pointers, and deployed services not represented by GitHub.
  3. Create candidate Asset records with source provenance.
  4. Extract candidate purpose, users, capabilities, stack, deployments, and relationships.
  5. Group likely duplicates and related systems for human review.
  6. Confirm, merge, or distinguish records and add missing capability descriptions.
  7. Link Assets to historical and active Initiatives where known.
  8. Record a verified baseline date and unresolved inventory gaps.

Automated extraction may prepare candidate records; it must not silently decide that two systems are equivalent or overwrite reviewed descriptions.

Reuse decision

Before authorizing a new Asset or substantial implementation, Conductor should return one of:

Result Meaning
reuse Existing capability already satisfies the need
extend An existing Asset is the best base and the delta is justified
new No existing Asset sufficiently fits and a new one is justified
insufficient-context Estate or requirement evidence is too weak to decide

The result must cite compared Assets, overlap, material differences, freshness, and the evidence used. new is not valid merely because the requesting agent did not know another project existed.

Contracts still required

The components above depend on contracts already tracked in the gap register:

  • Authority and adapter boundaries: G1 (M0 verified; authority-model.md, organization-profile.md).
  • Plan bundle and authorization: G2.
  • Agent-owned Notion writeback: G3 (state-model.md draft).
  • Compact pack selection: G4.
  • Complete stage behavior: G5.
  • Corpus alignment: G6.
  • Estate and precedent discovery: G7 (asset-registry.md draft).
  • History and experience: G8-G10.
  • Reconciliation and MCP tools: G11-G12.
  • Organization portability: G13-G15 (read-access-contract.md draft for G14 adapters).

This document must not restate those acceptance criteria. The gap register remains their source.