Cambium

English | 简体中文

Cambium is a governance standard and reference toolset for knowledge repositories maintained with LLM agents.

It helps an operator answer five practical questions:

1. What rules apply to this repository?
2. What work is required, and who may change shared state?
3. What evidence must exist before work can close?
4. How can an interrupted task resume without guessing?
5. Which decisions belong to the operator rather than the agent?

Cambium is not a knowledge base, a RAG engine, an agent scheduler, or a default domain policy. It governs work; it does not supply the corpus or decide its meaning.

Start Here

• To understand the model, read The Mental Model.
• To adopt Cambium for a repository, follow Adopt Cambium.
• To resume existing work, run the command in Start Or Resume A Task before writing anything.
• To connect an agent host, see Use Cambium From An Agent Host.
• For every tool and its exact arguments, see Tools/README.md.
• For what is complete, in progress, or only conditional, see ROADMAP.md.

The Mental Model

effective governance
  = Cambium kernel
  + exactly one selected profile
  + adopter-owned runtime state

The diagram shows how these layers connect to runtime routes, deterministic tools, and agent execution contexts.

| Layer | What it owns | |---|---| | kernel/ | Cross-domain governance semantics, invariants, state meanings, and extension points | | Card/ | Curated, non-authoritative flight checklists for an already selected task route or phase | | Read Set/ | Machine-resolvable declarations of what canonical material an already selected route or phase must load | | Selected profile | One repository's scope, language, architecture, sources, priorities, roles, scans, and allowed extensions | | .cambium/ | The adopter's current governance identity, task state, Queue, plans, deltas, receipts, and recovery evidence | | Tools/ | Stable public commands and Area/Domain implementations for deterministic checks, controlled writes, schemas, and generated projections |

The kernel is normative. A profile can fill or tighten an extension point, but cannot disable a kernel rule. Tools execute declared rules; they do not make the final semantic judgment.

Cards are short, curated checklists, not routes or a second copy of the standard. Read Sets own the static loading boundary. When a Card is insufficient or disputed, its read-back hook resolves through the paired Read Set to the canonical owner.

This repository is intentionally uninstantiated. It contains one candidate Profile template and non-authoritative examples, but selects no adopter profile and creates no fabricated task state.

What Ships Today

Cambium currently provides:

• one empty TOML Profile candidate, Agent-assisted interviews, safe creation and snapshot-bound editing tools, read-only review/status views, and CUE-backed Profile checks;
• persistent Coverage, Required Queue, and Progress state for resumable work;
• deterministic task and batch transitions, controlled Amendments, active-task Standards adoption, interruption recovery, and build or maintenance closure;
• append-only receipts and Terminal Proof bindings;
• explicit Global Map, Capability Matrix, and Gap Register validation;
• deterministic page, structure, vocabulary, link, boundary, freshness, and residual-content checks;
• a generated host-neutral interface: each tool's own CLI declaration and the closed agent-interface capability policy compile into the agent-facing MCP projection and per-host configuration; every active caller-visible path is retained as a descriptor capability through subprocess consumption;
• a typed Task Runtime Runner that advances registered deterministic tools to the next Agent, user, Host, repair, or terminal boundary;
• Card-first activation and progressive Read Set delivery primitives.

The generated MCP surface exposes both leaf calls and the bounded Runner. The Runner is not a scheduler or governance engine: it derives one identity-bound next action from current runtime state, invokes only registered capabilities, reads the result back, and stops at every semantic boundary. Each underlying Tool still decides whether its operation is valid and whether its evidence counts. For every active typed path, the transport retains the admitted file or parent-directory descriptor through subprocess consumption; an unsupported platform fails server initialization instead of claiming this assurance.

What Does Not Ship Yet

Cambium does not currently bundle:

• agent dispatch or scheduling;
• isolated worker workspaces;
• a complete single-writer integrator loop;
• durable Assignment lifecycle management;
• authenticated actor or reviewer identity;
• protected whole-workspace execution against arbitrary concurrent mutation;
• automatic corpus-wide dependency propagation;
• an independent evaluator that re-derives the complete expected corpus;
• an installable OpenAI Plugin package, Hooks, UI, or marketplace entry.

These boundaries are intentional. A host may add capabilities, but it must not claim evidence for a capability it cannot prove. See ROADMAP.md for the delivery order.

The Three Runtime Ledgers

Long-running work uses three state objects with different owners:

| State object | What it answers | |---|---| | Coverage Ledger | Which knowledge objects exist, what disposition they have, and which batch currently owns unfinished work? | | Required Queue | Which batches exist, what are their manifests and dependencies, and what lifecycle state is each batch in? | | Progress Ledger | What is the task contract, whole-task state, checkpoint, Standards identity, and accepted Queue fingerprint? |

They must agree, but they are not interchangeable task lists.

The adopter-owned namespace contains six lifecycle classes:

.cambium/
├── <canonical current state>
├── <bound operational inputs>
├── <evidence and history>
├── <recovery state>
├── <transient workspace>
└── <derived projections>

Do not edit canonical state by hand. Use the owning writer so revisions, hashes, receipts, and recovery evidence move together. Tools/execution/task_runtime/runtime_paths.py is the single machine owner of the current physical path spellings and object classifications; this README does not maintain a second directory contract.

Adopt Cambium

Adoption creates and approves one profile for one repository. Copying a template or example does not select it.

Run the authoring workflow from a Cambium source checkout after the Agent-driven Host preparation. With terminal access and installation authorization, the Agent prepares and verifies the required toolchain; users do not choose dependency versions or fill in local paths. The user supplies and confirms repository decisions, without manually copying template files or writing TOML.

1. Create a candidate profile

python3 Tools/scaffold_profile.py . --profile-id my-profile
python3 Tools/scaffold_profile.py . --profile-id my-profile --apply

The first command is a dry run. The second creates profiles/my-profile/profile.toml with the confirmed identity and empty slots, copies only the declared supporting files, and refuses to overwrite an existing candidate. It makes no policy choice and performs no adoption.

2. Answer the open decisions and validate

The assisting agent uses profiles/interview.yaml to discuss the repository's needs and Tools/profile_candidate.py to read, preview, edit, and render the candidate. User answers live once in profile.toml; independently referenced policy bodies retain their own owner. profiles/README.md describes the exact workflow and snapshot preconditions.

The Kernel owns slot meaning and legal values through K00/19 and domain-owned contracts. Tools own TOML encoding—including the root version, slots packaging, and draft-validation entry point—plus file layout, evaluation, and presentation. Existing domain YAML contracts remain sole owners where other consumers need them; their CUE projections are generated and checked, not parallel handwritten rules.

python3 Tools/profile_onboarding_status.py . --profile-id my-profile --json
python3 Tools/check_profile.py profiles/my-profile

Unanswered draft fields remain unanswered: omission is not agreement to disable an option or inherit a default. Existing legal defaults still apply where the completed contract permits them, but do not prove user confirmation. Mechanical validity, user confirmation, and adoption are separate; a rendered view or successful check never selects the Profile.

3. Approve the profile through R09

Prepare a plan from Tools/schemas/profile_adoption_plan.template.yaml, then dry-run and apply it:

python3 Tools/apply_profile_adoption.py . --plan <plan>.yaml \
  --upstream-root <local-cambium-repository> --upstream-ref <git-ref>
python3 Tools/apply_profile_adoption.py . --plan <plan>.yaml \
  --upstream-root <local-cambium-repository> --upstream-ref <git-ref> --apply

The transaction resolves the upstream ref to its full Git commit SHA and records that SHA as the sole Standards identity in upstream_revision_id. It binds the selected Profile and resulting adopter-owned contracts/evidence, and restores the previous control plane if any step fails. It never restamps or rewrites the adopter's upstream Card bytes. Adoption remains an explicit CLI maintenance operation: its external upstream repository input is never exposed as an unrestricted MCP argument.

An empty corpus follows the same adoption contract. First perform bounded founding work to create real canonical owners and the residual-scan witness — one page may serve as both owner and witness when that is semantically natural, but pages are never merged only to save files. Then a second R09 revision configures the Corpus Planning slot before large-scale work begins. The candidate and adoption boundary is documented in profiles/README.md.

Start Or Resume A Task

Always check for existing runtime state before writing:

python3 Tools/check_queue.py . --resume-status

If .cambium/state/ exists, this command reports the recorded task, locks, holds, in-flight batches, recovery state, and exact next_action. Do not initialize over it.

Bounded work does not need persistent state. For long-running, resumable, or multi-batch work, first copy and complete the single Task Plan:

cp Tools/schemas/task_plan.template.yaml \
  .cambium/deltas/task-plans/TP-001.yaml

python3 Tools/init_state.py . \
  --plan .cambium/deltas/task-plans/TP-001.yaml

python3 Tools/init_state.py . \
  --plan .cambium/deltas/task-plans/TP-001.yaml --apply

# Run the exact compile_queue command printed by init_state.py; it already carries the Queue revision and SHA bound to the published Task Plan.
python3 Tools/compile_queue.py . --apply --actor-role integrator \
  --expected-queue-revision REVISION \
  --expected-sha256 SHA256

python3 Tools/check_queue.py .
python3 Tools/render_queue.py .

init_state.py has no parallel flags for task identity, objective, scope, Standards, Profile, completion model, or concurrency. Those confirmed values have one owner: the Task Plan. The command atomically publishes the empty Queue, complete Task Contract, planning-only Coverage, and the Receipt retained by Progress; compile_queue.py remains the sole Queue materializer