Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion .claude-plugin/marketplace.base.json
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@
"plugins": [],
"metadata": {
"description": "Agent collaboration plugin marketplace",
"version": "4.9.0",
"version": "4.9.1",
"repository": "https://github.com/sumitake/agent-collab"
}
}
4 changes: 2 additions & 2 deletions .claude-plugin/marketplace.json
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@
{
"name": "agent-collab",
"description": "Unified dynamic-host collaboration package. Centralized skills and async coordination work without legacy packages; every model-execution route requires the verified signed plugin artifact.",
"version": "4.9.0",
"version": "4.9.1",
"author": {
"name": "John Osumi"
},
Expand All @@ -32,7 +32,7 @@
],
"metadata": {
"description": "Agent collaboration plugin marketplace",
"version": "4.9.0",
"version": "4.9.1",
"repository": "https://github.com/sumitake/agent-collab"
}
}
894 changes: 216 additions & 678 deletions README.md

Large diffs are not rendered by default.

15 changes: 15 additions & 0 deletions changelog.d/2026-08-06-public-architecture-handbook.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,15 @@
### Changed

- agent-collab 4.9.1 adds a sanitized, indexed public architecture handbook
covering system context, capabilities, governance and authority, lifecycle
operations, repository/release architecture, and lifecycle-state evidence.
- Refocus the root README on general users and link maintainers to the existing
low-level package reference instead of duplicating its protocol detail.

### Fixed

- Correct stale public statements about the package version, committed
activation artifact, reviewed signing anchor, and current runtime-manifest
schema while removing private-producer provenance from active documentation.
- Add a design-evidence registry so cited design-of-record sections,
superseded drafts, and historical adversarial reviews are not conflated.
108 changes: 108 additions & 0 deletions docs/architecture/README.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,108 @@
# agent-collab architecture handbook

This handbook explains the public architecture of `agent-collab` for users,
contributors, reviewers, and release maintainers. It describes stable
boundaries and public contracts. It does not reproduce private runtime
implementation, build credentials, operator-specific paths, or machine state.

The handbook is descriptive. Source, tests, manifests, release evidence, and
the contribution contract remain authoritative for their respective claims.
When two generations differ, use the status vocabulary below instead of
silently treating them as the same thing.

## Read this handbook in order

1. [System context](system-context.md) explains what the project is, what it is
not, its actors, and its public/private boundary.
2. [Capabilities and workflows](capabilities-and-workflows.md) maps the skill
surface to common user jobs and explains availability limits.
3. [Governance and authority](governance-and-authority.md) explains family
independence, sealed authority, review evidence, and operator control.
4. [Lifecycle and operations](lifecycle-and-operations.md) covers installation,
verification, use, updates, troubleshooting, safe mode, rollback, and
removal.
5. [Repository and release architecture](repository-and-release.md) maps the
public package, generated sources, validation, and release flow.
6. [Status and evidence](status-and-evidence.md) defines the binding lifecycle
labels and records the dated evidence snapshot used by this documentation.

Historical and design material is indexed separately in
[`docs/design/README.md`](../design/README.md). It is not automatically the
current runtime contract.

## Status vocabulary

| Label | Meaning | Appropriate evidence |
| --- | --- | --- |
| **current** | The behavior or contract is present in the checked-out public repository baseline and covered by current source, manifests, or tests. | Merged source plus focused tests or generated-manifest checks. |
| **repository-only** | The implementation or artifact exists in the repository, but no cited release and host observation proves it is installed and selected. | Repository source or artifact identity only. |
| **staged** | Material is prepared for a later compilation, publication, activation, or selection step. | A changelog fragment, candidate artifact, tag input, or generated release input. |
| **installed/active** | A specific version has been positively observed as installed, selected, and ready on a specific host. | Host package inventory plus provider-free readiness evidence. |
| **proposed** | A design or change has not become the current merged contract. | Draft, branch, issue, or unmerged design. |
| **historical** | Retained evidence explains earlier decisions but does not define current behavior. | Changelog, superseded design, or past review record. |
| **retired** | The public source explicitly removes or blocks the old surface. | Migration policy, absence tests, and clean-package checks. |

These labels are deliberately narrower than words such as “released” or
“available.” A manifest can advertise a route while a host still reports it
unavailable. A tag can exist without a GitHub release. A repository version can
be newer than both. See [Status and evidence](status-and-evidence.md).

## Architectural invariants

The public repository and package preserve these invariants:

1. There is one installable package, `agent-collab`, not one plugin per host or
provider.
2. Callers select a skill or logical collaboration job. They do not receive a
raw provider-execution escape hatch.
3. Governance review requires a model family independent of the active primary
and the reviewed artifact's known author family.
4. Route authorities are closed. Read-only, output-only, and unavailable
actions do not promote themselves because another route failed.
5. Provider output is an artifact for the trusted primary to inspect. It does
not merge, deploy, or change policy by returning successfully.
6. Native execution is optional at the package level and fail-closed. A listed
skill or advertised contract does not prove current host readiness.
7. The public package may contain only the final reviewed native bundle and
closed manifest metadata. Native implementation and build/sign credentials
remain outside this repository.
8. Retired packages are migration evidence, not rollback targets.

## Sanitization contract

Architecture documentation is public only when it records portable boundaries
rather than one operator's environment. These pages therefore:

- use repository-relative paths and generic roles;
- omit personal account names, host names, network details, tokens,
credentials, local cache paths, and private repository locations;
- describe managed provider roles without publishing raw provider invocation
recipes or discovery commands;
- distinguish public artifact-verification requirements from the private
systems that produce those artifacts; and
- label uncertainty instead of inferring installed or active state.

Do not add copied diagnostic output, private build identifiers, notarization
submission identifiers, environment dumps, or secret-bearing examples to this
handbook. Suspected exposure follows [the security policy](../../SECURITY.md),
not a public issue or pull request.

## Source map

| Question | Start here | Authoritative public evidence |
| --- | --- | --- |
| What does the package install? | [System context](system-context.md) | Host manifests, generated marketplaces, package tree, and distribution tests. |
| Which workflows exist? | [Capabilities and workflows](capabilities-and-workflows.md) | `skill-specs/`, generated skills, package reference, and skill-contract tests. |
| Who may review, write, or merge? | [Governance and authority](governance-and-authority.md) | Host policy, coordinator, public governance contract, PR template, and compliance checks. |
| Is a route usable now? | [Status and evidence](status-and-evidence.md) | Installed version plus provider-free readiness on that host. Repository presence alone is insufficient. |
| How do I install or recover? | [Lifecycle and operations](lifecycle-and-operations.md) | Current host CLI, migration doctor, runtime-management surface, and migration policy. |
| How is a release produced? | [Repository and release architecture](repository-and-release.md) | Release scripts, workflows, signed-tag contract, archive checks, and release evidence. |

## Maintenance rule

Update the relevant architecture page in the same change when a public
component, authority boundary, host surface, route contract, lifecycle state,
or release boundary changes. Keep each claim tied to repository-relative source
and focused evidence. If delivery evidence lags source, keep both observations
and label them; do not “fix” the discrepancy by declaring one generation
active everywhere.
148 changes: 148 additions & 0 deletions docs/architecture/capabilities-and-workflows.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,148 @@
# Capabilities and workflows

< [Architecture handbook index](README.md)

The unified package exposes 50 generated skills under the
`agent-collab` namespace. A skill is a workflow contract: it defines when to
use the capability, what evidence to gather, which authority is acceptable,
and when to stop. It is not a promise that every managed route is currently
available on every host.

The definitive low-level inventory remains the
[package reference](../../plugins/agent-collab/README.md#skills); this page
organizes that inventory by user outcome.

## Capability map

| User job | Skills | Typical outcome |
| --- | --- | --- |
| Identity, routing, and readiness | `agent-readiness`, `agent-runtime-status`, `migration-doctor`, `route`, `start-inbox-monitor`, `teamwork` | Establish which host, family, route, or async target is eligible before relying on it. |
| Planning and architecture | `architect`, `brainstorm`, `compose-skills`, `intent-check`, `second-opinion` | Clarify a design, widen options, select workflows, or obtain an independent read. |
| Governance and assurance | `autonomy-readiness`, `code-review`, `governance-review`, `logic-check`, `qa-verify`, `red-team`, `untrusted-audit` | Test correctness, independence, security, provenance, and completion evidence. |
| Deliberation and stakeholder lenses | `debate`, `simulate-user` | Expose conflicting arguments or test a proposal against a persona. |
| Delegation and implementation | `delegate`, `dev-delegate`, `worker` | Return bounded research or output for the primary to integrate. |
| Large-context and knowledge work | `knowledge-compile`, `long-context` | Extract or synthesize a bounded corpus with explicit provenance. |
| Reproducible workflows | `chain`, `chain-configurator`, `orchestrate` | Define and execute repeatable multi-step coordination. |
| Integration and conflict handling | `merge-resolve` | Analyze and resolve a bounded merge conflict while preserving intent. |
| Visual guidance | `ui-to-code`, `visual-review` | Guide primary-only visual work when typed image transport is absent; never invent a managed attachment path. |
| Language expertise | `rust-engineer`, `go-engineer`, `elixir-engineer`, `sql-engineer` | Apply domain-specific engineering practices. |
| Infrastructure and reliability | `kubernetes-specialist`, `terraform-engineer`, `sre-engineer`, `incident-responder` | Design, review, or troubleshoot operational systems. |
| Data and AI | `mlops-engineer`, `llm-architect`, `postgres-engineer`, `data-engineer` | Apply data, model, database, and pipeline expertise. |
| Evaluation and writing quality | `eval-engineer`, `prompt-regression-tester`, `hallucination-investigator`, `ai-writing-auditor` | Evaluate model-backed behavior, prompt drift, factual failure, or prose quality. |
| Engineering process | `decision-map`, `prototype`, `architecture-review` | Create decision tickets, answer one design question with a throwaway prototype, or sweep module boundaries. |

## Execution shapes

### Managed synchronous routes

Routed review, context, planning, governance, and worker workflows submit a
bounded request through the installed package's public coordinator. Policy
selects an eligible managed route and preserves one of the authorities defined
in [Governance and authority](governance-and-authority.md).

Current repository route contracts cover:

- Gemini advisory, governance, and long-context work;
- Codex advisory and governance work;
- OpenCode planning and output-only build work;
- Grok read-only architecture, governance, and huge-context work; and
- Composer-compatible output-only code generation.

Codex build is recognized but typed unavailable in this repository generation.
The route list is a **current repository contract**, not installed/active
evidence. Readiness is resolved immediately before use.

### Primary-executed workflows

Some skills guide the active primary directly rather than selecting a managed
provider. The engineering-process pack is explicitly self-executed. Visual
skills also remain primary-only where the current protocol has no typed image
or binary-media transport. These workflows can still require local tools,
tests, or user approval; “primary-executed” is not “unchecked.”

### Async coordination

Claude and Antigravity participation can use host-owned asynchronous inbox
transport after the exact target identity, family, session, and current
readiness are observed. The public coordinator exposes readiness only. It does
not send, does not create a synchronous Claude route, and does not treat an
async reply as independent governance merely because it arrived.

### Reproducible composition

- `chain` executes a versioned YAML-defined sequence of skill invocations.
- `chain-configurator` helps create a chain definition interactively.
- `orchestrate` coordinates a dependency graph with bounded tasks and explicit
gates.
- `teamwork` coordinates role-based milestones and stop conditions.

Composition does not erase the authority of individual steps. A read-only
review inside a chain remains read-only, and a worker artifact still returns to
the primary for integration.

## Common workflows

### Independent review

1. Identify the exact artifact and its author lineage.
2. Invoke `second-opinion`, `code-review`, `governance-review`, `red-team`, or
`qa-verify` according to the risk.
3. Policy excludes the primary and artifact-author families where independence
is required.
4. Preserve the raw finding and typed status.
5. The primary adjudicates, changes the artifact if needed, and re-verifies the
exact head.

### Bounded delegation

1. The primary keeps objective interpretation and integration ownership.
2. A worker receives a bounded scope, authority, evidence contract, and stop
condition.
3. Read-heavy work can use a cheaper capable tier; judgment and landing stay
with the primary.
4. Output-only code is returned as an artifact from a private temporary
workspace; it is not written into the caller's checkout.
5. The primary reviews and tests all delegated output before it becomes source.

### Architecture and planning

Use `brainstorm` to widen the option space, `architect` for an independent
read-only architecture consultation, `architecture-review` for a primary-led
codebase sweep, `intent-check` to validate task interpretation, and
`decision-map` when the effort is too large for one session.

### Large-context work

Use `knowledge-compile` when multiple sources must become a durable cited
dossier. Use `long-context` when a managed large-context route is needed. The
input remains bounded and source-specific; long-context transport does not
grant governance or mutation authority.

## Host and package support

| Surface | Public package evidence | User expectation |
| --- | --- | --- |
| Claude Code | Claude-compatible plugin manifest and marketplace metadata. | Native package install and `/agent-collab:*` skills. Claude model participation remains async-only. |
| Codex CLI/app | Codex-native manifest and generated Codex marketplace. | Native package install and the same skill namespace. Start a new task after install/update. |
| Antigravity | Dynamic host policy and async target model; no separate package. | Use only through a compatible package host and observed async readiness. |
| OpenCode and ZCode | Dynamic host/model policy and managed OpenCode routes; no separate package. | A compatible host/plugin surface is required. OpenCode is a transport; the selected model supplies family lineage. |
| Custom host | Explicit primary identity fields and the closed package contract. | If the host cannot load the package safely, it is unsupported; do not recreate provider-specific shims. |

## Availability rules

A capability is usable only when all applicable gates pass:

- the unified package is installed and selected;
- no active retired package blocks migration;
- primary identity is complete enough for the requested authority;
- the requested family is eligible and independent where required;
- the manifest advertises the exact route/action contract;
- the native boundary and provider-free readiness checks pass; and
- provider authentication, quota, and request execution succeed.

Failing one gate produces a typed result such as `unavailable`,
`same_family_blocked`, `unknown_family`, or a runtime error category. Do not
turn that result into a raw-provider fallback or broader authority.

For installation and recovery, continue to
[Lifecycle and operations](lifecycle-and-operations.md).
Loading