Skip to content

Latest commit

 

History

History
846 lines (693 loc) · 44 KB

plugins.md

File metadata and controls

846 lines (693 loc) · 44 KB

Plugin System

The AI Code Assistant ships an extensible plugin architecture so that Stellar, Soroban, GitHub, AI-provider, and developer-tooling features can be added without modifying the core application.

Status: Phase 8 + Phase 9 security foundations. Manifest validation, the registry, the capability model, the event system with fail-closed dispatch-time capability enforcement, a workspace-scoped plugin management API

  • UI, publisher trust (trusted keys, embedded and detached Ed25519 signatures), catalog-pinned remote installation, the plugin lock-entry model, a member-readable marketplace browser (install/update/rollback for owners), transitive catalog dependency validation, and the published marketplace JSON schema are implemented and tested. Artifact download/execution remains out of scope — installs register validated metadata only.

Concepts

Concept Description
Manifest manifest.json describing a plugin (id, version, capabilities…).
Plugin A runtime plugin instance built from a manifest.
Registry Central registration, discovery, lookup, enable/disable.
Capability A named permission a plugin may request (never granted by default).
Capability grant An explicit per-workspace grant row; nothing is granted implicitly.
Event / hook A plugin can subscribe to supported events and react.
Event dispatcher Dispatches events to subscribers; handler failures are isolated.

Plugin manifest

A plugin lives in its own directory and is described by manifest.json. The JSON Schema is at plugins/plugin.schema.json.

{
  "id": "stellar-tools",
  "name": "Stellar Developer Tools",
  "version": "0.1.0",
  "description": "Soroban / Stellar analysis helpers.",
  "author": "AI Code Assistant Team",
  "entry_point": "plugins.stellar_tools.plugin:StellarPlugin",
  "compatibility": ">=0.8.0",
  "capabilities": ["PROJECT_READ", "STELLAR_READ", "AI_ACCESS"],
  "permissions": ["read:project:files"],
  "dependencies": [],
  "configuration": {
    "networks": ["testnet", "mainnet"]
  }
}

Manifest validation

app/services/plugins.py::PluginManifest.from_dict rejects manifests that:

  • Miss a required field (id, name, version, description, author, entry_point, capabilities).
  • Use an invalid id (must match ^[a-z][a-z0-9_-]*$).
  • Use an invalid semantic version.
  • Use an invalid entry_point (must be module.path:ClassName).
  • Declare an unknown capability.
  • Declare an empty capability list.
  • Declare an invalid compatibility specifier (when present, it must be valid PEP 440 syntax such as >=0.8.0 or >=0.1.0,<0.5.0).

Validation errors raise ManifestValidationError; unreadable or malformed manifest files raise PluginError.

Compatibility (PEP 440)

The optional compatibility field declares the application versions a plugin supports (==, !=, <=, >=, <, >, ~=, wildcards, and comma/space separated ranges). Enforcement lives in app/services/plugin_compat.py and is compared against the running app version (installed package metadata, falling back to pyproject.toml):

  • Omitted / empty → the plugin supports any application version.
  • Invalid specifier → the manifest is rejected during validation.
  • Incompatible range → installation is refused with a clear message naming the required range and the current app version.
  • Updates are gated too: POST .../plugins/<plugin_id>/update refuses a manifest whose range excludes the running version before writing anything, so an incompatible release can never replace a working installation.
  • Compatible plugins register normally, and the management API exposes a compatibility badge in plugin metadata (compatibility, compatible_with_app, and app_version).

Registry

PluginRegistry (in app/services/plugins.py) provides:

  • register(plugin) — duplicate plugin ids raise PluginRegistrationError.
  • discover(dir) — scan a directory for manifest.json files; invalid manifests are skipped (with a warning), never fatal.
  • get(plugin_id) / list_all() / list_enabled().
  • enable(plugin_id) / disable(plugin_id).
  • validate_capability(plugin_id, capability).

Capabilities

Capabilities are defined in app/services/capabilities.py and include PROJECT_READ, PROJECT_WRITE, GITHUB_READ, GITHUB_WRITE, AI_ACCESS, AI_ANALYSIS, WORKSPACE_READ, WORKSPACE_WRITE, NOTIFICATION_CREATE, STELLAR_READ, STELLAR_WRITE, STELLAR_ANALYSIS, REVIEW_READ, REVIEW_CREATE, PROJECT_DELETE.

Capabilities are never granted automatically. A plugin receives a capability only when an explicit CapabilityGrant row exists for that plugin in a workspace (CapabilityStore.grant). The workspace role model maps each role to the capabilities it may exercise (ROLE_CAPABILITY_MAPPING), and validate_plugin_capability fails closed.

The CapabilityGrant model lives at app/models/plugin.py (table plugin_capability_grants) and is created by the Phase 8 migration.

Events / hooks

app/services/events.py provides an EventDispatcher and a global instance via get_dispatcher(). Supported event types include:

  • project.created, project.updated, project.deleted
  • review.created, review.completed, review.finding.added
  • workspace.created, workspace.member_added, workspace.member_removed
  • github.connected, github.disconnected
  • ai.analysis.completed
  • stellar.analysis.completed, stellar.network.detected

Plugins subscribe with dispatcher.subscribe(event_type, handler, plugin_id). Subscribing to an unsupported event type raises EventError. When an event is dispatched, one failing handler never prevents other handlers from running, and the request itself is never crashed (routes use emit_event, which swallows and logs failures).

Event capability enforcement (fail closed)

Before a plugin handler is invoked, the dispatcher verifies (in app/services/events.py):

  1. the event type is supported and has a required capability mapped;
  2. the plugin exists (Plugin row) and is enabled;
  3. for workspace-scoped events, the event carries a valid workspace_id, the plugin is installed (PluginInstallation) and enabled in that workspace, and the emitting user is authorized for the workspace;
  4. the plugin holds the capability required by EVENT_CAPABILITY_MAP through an explicit CapabilityGrant for the event's workspace.

If any condition cannot be proven, the handler is denied (recorded in the dispatch result as a denial) and never invoked. EVENT_CAPABILITY_MAP associates each event with the capability required to receive it, for example:

  • project.created / project.updated / project.deleted → PROJECT_READ
  • workspace.member_added / workspace.member_removed → WORKSPACE_READ
  • ai.analysis.completed → AI_ACCESS
  • stellar.analysis.completed / stellar.network.detected → STELLAR_READ
  • github.connected / github.disconnected → GITHUB_READ (global events, delivered to enabled plugins without a per-workspace grant, since grants are workspace-scoped)

Subscribers registered without a plugin_id are internal handlers (trusted application code) and are exempt from plugin capability enforcement. See docs/security.md for the full security model.

The application already emits events from real flows: project import/delete, workspace member add/remove, AI analysis completion, Stellar analysis completion, GitHub connection, and plugin lifecycle.

Lifecycle hooks (on_enable / on_disable / on_uninstall)

A loaded plugin class may implement optional lifecycle methods (app/services/plugins.py):

class MyPlugin:
    def __init__(self, app=None, manifest=None):
        ...

    def on_enable(self): ...      # after the registry enables the plugin
    def on_disable(self): ...     # after the registry disables the plugin
    def on_uninstall(self): ...   # before the registry removes the plugin

PluginRegistry.enable / disable / uninstall invoke the matching hook for plugins already loaded in-process. Hooks are isolated and non-fatal: an absent hook is a no-op, and a raising hook is logged without preventing the state change — enabling/disabling always lands in the new state and uninstall always removes the plugin, so the registry never ends inconsistent. The management API never loads or executes plugin code on its own.

Per-workspace lifecycle is also observable through the dispatcher events plugin.enabled, plugin.disabled, and plugin.uninstalled (mapped to WORKSPACE_READ; emitted only on real transitions). The API exposes enable/disable and an owner-only uninstall action that revokes every workspace capability grant and removes the workspace installation.

Execution timeout (#408)

Plugin event handlers and lifecycle hooks run with a bounded execution budget (app/services/plugin_execution.py). The budget is PLUGIN_EXECUTION_TIMEOUT seconds (default 5; 0 or negative disables it):

  • the callback runs in a short-lived daemon thread while the request waits;
  • if it overruns, the dispatcher stops waiting, records one bounded error report (plugin_error_reports, operation dispatch:<event> or hook:<hook>), and continues with the next handler or lifecycle step;
  • state changes still land: enabling/disabling always completes, and dispatch keeps running the remaining subscribers.

Python cannot interrupt a thread, so an over-budget callback is abandoned rather than killed. It may still finish in the background, which is why plugin code must be idempotent and must not depend on being stopped.

Plugin management (API + UI)

Operator CLI

app/services/plugins_cli.py registers a flask plugins command group that acts as the operator, reading/writing the persisted Plugin rows (the same store the management API and dispatch-time authorization use) through app/services/plugin_ops.py:

flask plugins list                      # id, version, enabled state
flask plugins inspect <plugin_id>       # full manifest metadata + state
flask plugins enable <plugin_id>        # enable (operator scope)
flask plugins disable <plugin_id>       # disable (operator scope)
flask plugins install <local-path>      # register from a local manifest dir
flask plugins publishers                # list trusted publisher ids (never keys)

Every command supports --json for stable, parseable output. Exit codes distinguish success (0), unknown plugin (1), and validation/service errors (2). Security rules: installs accept only local filesystem paths with a validated manifest.json (URLs are refused), the CLI never grants capabilities (they remain explicit and per-workspace), and it never loads or executes plugin code.

The user-facing API + UI is workspace-scoped: the plugins blueprint (app/plugins/) provides the management layer. Plugin state is workspace-scoped: a plugin exists globally (a Plugin row) while having a separate PluginInstallation per workspace, each with its own enabled state and capability grants.

Authorization reuses the Phase 7 workspace role model: any workspace member may view; only the workspace owner may manage (the manage_plugins capability). Non-members get 404; members who are not owners get 403 for management operations.

Method Path Purpose Role
GET /plugins/ Management page (workspace selector) any member
GET /plugins/api/workspaces Workspaces the user can access any member
GET /plugins/api/workspaces/<id>/plugins List plugins + installation state any member
GET /plugins/api/workspaces/<id>/plugins/<plugin_id> Inspect one plugin any member
GET /plugins/api/workspaces/<id>/catalog (alias /marketplace) Browse catalog metadata, trust, dependencies and install/update/rollback state any member
GET /plugins/api/workspaces/<id>/lockfile Read the resolved lockfile pin set any member
GET /plugins/api/workspaces/<id>/plugins/updates Installed plugins with a newer, compatible catalog version any member
GET /plugins/api/trusted-publishers List configured trusted publisher ids (never keys) and the active policy authenticated
POST /plugins/api/workspaces/<id>/plugins/install Install a trusted/local manifest owner
POST /plugins/api/workspaces/<id>/plugins/install-from-catalog Install a catalog-pinned plugin (trust policy + dependency graph enforced) owner
POST /plugins/api/workspaces/<id>/plugins/<plugin_id>/update Update a registered plugin from a manifest (compatibility-gated) owner
POST /plugins/api/workspaces/<id>/plugins/<plugin_id>/update-from-catalog Atomically update from a catalog version owner
POST /plugins/api/workspaces/<id>/plugins/<plugin_id>/rollback Roll back to the lockfile/catalog target owner
POST /plugins/api/workspaces/<id>/plugins/<plugin_id>/enable Enable the workspace installation owner
POST /plugins/api/workspaces/<id>/plugins/<plugin_id>/disable Disable the workspace installation owner
POST /plugins/api/workspaces/<id>/plugins/<plugin_id>/capabilities Explicitly grant/revoke capabilities ({"grant": [...], "revoke": [...]}) owner

Installation (local manifests or signed catalog entries)

Local manifests are the default. Remote installation is supported only through the operator-controlled catalog described in Remote install from the validated catalog; the API never accepts a client-supplied artifact URL.

POST .../plugins/install accepts a manifest object ({"manifest": {...}}). The manifest is fully validated (PluginManifest.from_dict); malformed manifests are rejected with 400. The plugin id is taken only from the validated manifest — a client cannot override it (identity binding). Installing:

  • creates or reuses the global Plugin row (a conflicting entry point for an existing id returns 409);
  • creates a PluginInstallation for the workspace (409 if already installed there);
  • never loads or executes plugin code;
  • never creates capability grants.

Capabilities granted through the API must be declared in the plugin's manifest (declared vs granted are kept distinct). The manage_plugins role gate is the only way to grant/revoke through the UI; grants are never implicit.

Publisher trust: keys and detached signatures (#385, #386)

Operators list trusted publishers in PLUGIN_TRUSTED_KEYS, a JSON object that maps a publisher id (lowercase [a-z0-9._-], at most 64 characters) to an Ed25519 public key. Keys may be base64-encoded raw 32-byte keys or PEM SubjectPublicKeyInfo text:

export PLUGIN_TRUSTED_KEYS='{"acme-labs": "<base64 or PEM Ed25519 public key>"}'
export PLUGIN_TRUST_POLICY=required   # or if-present (default)

Key material is validated when the application starts (app.services.plugin_trust.validate_trust_config): a malformed key, a non-canonical publisher id, or an unknown policy aborts startup with a clear message instead of silently treating every signature as invalid. Keys are never returned by the API or the CLI — inspect the configured ids with:

flask plugins publishers             # human-readable ids + policy
flask plugins publishers --json      # {"publishers": [...], "count": N, "policy": ...}

GET /plugins/api/trusted-publishers exposes the same read-only view (any authenticated user): publisher ids, a count, and the active policy, never key bytes or fingerprints.

A publisher signs the canonical manifest body: compact, key-sorted JSON with signature, trust_state, and trust_publisher removed. The signature is a base64 Ed25519 signature over those bytes and may be:

  • embedded in the manifest ("signature": {"publisher", "algorithm": "ed25519", "value"}), or
  • detached, delivered next to the manifest (a .sig file or the catalog entry's signature field).

app/services/plugin_signatures.py is the single verification path for both forms. Policy handling is deliberately strict:

Situation if-present required
Valid signature from a configured publisher accepted (Trusted) accepted (Trusted)
No signature accepted (Unverified) refused
Forged/malformed signature, or unknown publisher refused refused

Tolerating absence must never mean tolerating forgeries, so an invalid signature fails even in if-present mode. Local-path installs keep working under the default policy; set required to demand signatures everywhere.

Remote install from the validated catalog (#389)

POST /plugins/api/workspaces/<id>/plugins/install-from-catalog installs a plugin that an operator has already published in the catalog. The request only names the entry — a client can never supply a manifest, an artifact URL, or a key:

curl -X POST https://<host>/plugins/api/workspaces/7/plugins/install-from-catalog \
  -H "Content-Type: application/json" \
  -d '{"plugin_id": "hello-world", "version": "0.1.0"}'

version is optional; without it the newest catalogued version wins (PEP 440 ordering). PLUGIN_CATALOG_FILE points at the catalog JSON (abridged below; manifest is a complete, schema-valid manifest):

{
  "version": 1,
  "plugins": [
    {
      "id": "hello-world",
      "version": "0.1.0",
      "source": "https://plugins.example.com/hello-world-0.1.0.zip",
      "integrity": "sha256:<64 lowercase hex characters>",
      "signature": {
        "publisher": "acme-labs",
        "algorithm": "ed25519",
        "value": "<base64 signature over the manifest>"
      },
      "manifest": { "id": "hello-world", "...": "see plugins/plugin.schema.json" }
    }
  ]
}

The endpoint is owner-only (manage_plugins) and fails closed:

Failure Status
Missing/invalid plugin_id or version 400
No catalog configured, or unknown id/version 404
Invalid manifest or incompatible compatibility range 400
Trust refusal (unsigned under required, forged, unknown publisher) 403
Already installed, or entry point conflicts with a registered plugin 409
Non-member of the workspace 404
Member who is not the owner 403

On success it registers metadata only: the global Plugin row and the workspace PluginInstallation. The response also reports the catalog source and integrity so the resolved pin is visible. It does not download the artifact, execute plugin code, or create capability grants, and trust refusals are appended to the workspace audit trail.

Marketplace trust model (end to end)

The operator-facing summary lives in the security model. In short: the catalog source is allowlisted, every entry is schema/integrity/signature checked, installs are owner-only and dependency-validated, updates are atomic with rollback, and the lockfile pins the resolved set. Catalog data is advertised metadata only — it never executes and never grants a capability.

Concern Implementation Tests
Catalog parsing, id/version/source/integrity validation app/services/plugin_catalog.py tests/test_plugin_catalog_install.py, tests/test_plugin_catalog_remote.py, tests/test_plugin_catalog_deps.py
Read-only discovery + manifest-schema validation app/services/marketplace.py, app/plugins/routes.py tests/test_marketplace_catalog.py
Allowlist / HTTPS / timeout / size / no redirects app/services/plugin_network.py tests/test_plugin_network.py
Ed25519 signatures and trust policy app/services/plugin_signatures.py, app/services/plugin_trust.py tests/test_plugin_signatures.py, tests/test_plugin_trust.py
Transitive dependency resolution and cycles app/services/plugin_deps.py tests/test_plugin_catalog_deps.py, tests/test_plugin_deps.py
Atomic update / rollback app/services/plugin_ops.py, app/services/plugin_updates.py tests/test_plugin_update.py, tests/test_marketplace_e2e.py
Lockfile pinning and serialization app/models/plugin*.py, app/services/plugin_lockfile.py tests/test_plugin_lock_entries.py, tests/test_plugin_lockfile.py

Configuration for each layer is in the configuration reference; remote installation stays disabled until an operator sets a catalog and allowlist.

Marketplace catalog schema (#373)

plugins/marketplace.schema.json is a Draft-07 schema for marketplace entries (id, name, version, description, author, capabilities, permissions, dependencies, compatibility, artifact URL, sha256 and publisher key id). It also accepts the installer catalog shape (id/version/source/integrity/manifest), so existing operator fixtures continue to validate without changing parse_catalog. The reusable helper app.services.marketplace_schema.validate_marketplace_entry loads the schema and returns path-qualified errors via the dependency-free draft-07 subset validator in app/services/schema_validation.py (shared with the manifest schema tests). No new runtime dependency is required.

Plugin lock entries (#394)

PluginLockEntry (plugin_lock_entries) persists the resolved plugin set for reproducible installs: plugin id, pinned version, source, sha256 integrity hash, declared dependencies, the workspace, and its owner. A row is unique per (workspace_id, plugin_id), and lock entries never imply capability grants.

Rows serialize through the existing canonical lockfile builder, so dumps_lockfile([*rows]) produces the same byte-stable document as any other resolved entry and omits volatile audit fields (resolved_at, row/workspace ids):

from app.models import PluginLockEntry
from app.services.plugin_lockfile import dumps_lockfile

rows = PluginLockEntry.query.filter_by(workspace_id=workspace_id).all()
print(dumps_lockfile(rows))

The Alembic migration for the table is tracked by the Phase 9 lockfile migration issue; the model and its serialization are covered by tests/test_plugin_lock_entries.py.

The workspace plugin page renders this lockfile read-only for every workspace member (#364): each pinned plugin row shows its id, version, source, integrity hash, and dependencies. Any member reads it from GET /plugins/api/workspaces/<id>/lockfile; an empty lockfile renders an explicit empty state instead of a blank panel. Configuration values are never part of the lockfile payload, and source URLs have embedded credentials stripped before rendering, so the viewer cannot expose a secret. Mutating plugin routes remain owner-only.

Marketplace browser (#364)

GET /plugins/api/workspaces/<id>/catalog (also available as /marketplace) is readable by any workspace member and returns catalog entries with stable, bounded pagination (page >= 1, limit clamped to 100, deterministic order by plugin id then version). Each entry exposes safe metadata only: name/version/description/author, declared capabilities and permissions, dependency requirements, compatibility, trust state and publisher, artifact source and sha256 integrity, plus per-workspace installed, update_available, rollback_available and actions fields. can_manage and every action are true only for owners; the backend independently enforces manage_plugins on install/update/rollback.

The management page (app/templates/plugins/index.html + app/static/js/plugin_catalog.js) renders this browser with trust and compatibility badges, dependency/artifact details, loading, empty, error and timeout states (shared with plugin_marketplace.js), and owner-only install/update/rollback buttons with surfaced errors. Catalog text is escaped before rendering.

Marketplace discovery API (#360)

The read-only discovery API advertises catalog entries without installing anything. It is separate from the per-workspace installer catalog above:

Route Access Description
GET /plugins/api/marketplace any signed-in workspace member Paginated list with optional q (id/name/description/author), capability, and workspace_id filters.
GET /plugins/api/marketplace/<id> any signed-in workspace member One entry, newest version by default or ?version= pinned.

app/services/marketplace.py loads the catalog from PLUGIN_MARKETPLACE_CATALOG (default plugins/marketplace.json, which ships empty so the offline default is safe). If PLUGIN_MARKETPLACE_URL is set and no local catalog is configured, the document is fetched through app/services/plugin_network.py: HTTPS only, allowlisted host, bounded timeout and body size, and redirects refused.

The discovery API uses the lenient loader introduced in #374/#376: a missing or unreadable local file yields an empty catalog (HTTP 200, never an error), and individual invalid entries are excluded with bounded warnings in the response instead of hiding every valid entry. A malformed JSON document or wrong catalog shape is still a 400 so operators can fix it. The installer catalog used by install-from-catalog remains strict and rejects the whole document. Every advertised manifest is validated against plugins/plugin.schema.json.

Responses also carry a state so callers can distinguish an ordinary empty catalog (empty), an offline source with no previous success (unavailable), and an offline source served from the process-local last known good catalog (stale); stale and unavailable are booleans, and warnings explain the fallback. A non-UTF-8 file is an invalid 400, not a 500. Error payloads use the stable {"error", "code"} shape from marketplace_error_payload, so filesystem paths and parser details are never echoed to the browser (#380).

Each returned entry contains only id, name, version, description, author, capabilities, permissions, dependencies, compatibility, integrity, and source (plus signature when present). Catalog data is untrusted metadata: it is never executed and never grants capabilities. Pagination is bounded (page >= 1, limit clamped to 100) and ordered by id then version. Tests live in tests/test_marketplace_catalog.py, tests/test_marketplace_lenient.py, and tests/test_marketplace_performance.py.

Catalog dependency validation (#391)

Before a catalog install writes anything, CatalogDependencyGraph (app/services/plugin_deps.py) walks the target entry's manifest dependencies recursively through the catalog. It collapses duplicate edges/nodes, returns a deterministic dependency-first order, and reports a full cycle chain such as a -> b -> a. Non-catalog dependencies are not required to be installed before install; the existing enable-time resolver remains authoritative for their installed/disabled/version state. Both the API route and the service/CLI path (install_from_catalog) run this validation.

All marketplace surfaces (plugin list, lockfile, and detail actions) share the empty/error/timeout states in app/static/js/plugin_marketplace.js (#421). Failures show friendly copy and a Retry button, and requests are timeout-bounded so a stalled marketplace never leaves the UI loading forever. The state renderers are covered by tests/js/plugin_marketplace.test.js.

Updating a plugin

POST .../plugins/<plugin_id>/update accepts the same {"manifest": {...}} shape as install, but applies it to an already-registered plugin. The manifest is validated and identity-bound:

  • id must equal the plugin being updated;
  • entry_point must not change (409), so an update can never point an installed plugin at different code;
  • the declared PEP 440 compatibility range is enforced against the running application version before anything is written. An incompatible range is refused with 400 and a message naming both the range and the current version; the registered plugin and its workspace installations are left untouched.

Updates change the plugin's metadata (version, description, capabilities, dependencies, configuration defaults, trust fields) but never capability grants or per-workspace configuration, which stay explicit and workspace-scoped. The plugin must be installed in the authorizing workspace (404 otherwise).

Enable / disable

enable/disable operate on the workspace installation (PluginInstallation.enabled), never on the global Plugin record, and never create or remove capability grants. A disabled installation is denied by the dispatch-time enforcement (see above), so disabled plugins cannot execute. Re-enabling restores delivery while the grant remains valid.

Per-workspace configuration

Each PluginInstallation carries a workspace-scoped config (JSON) that installations read/write through owner-only endpoints:

  • GET .../plugins/<plugin_id>/config — read the workspace configuration.
  • PUT .../plugins/<plugin_id>/config with {"config": {...}} — replace it.

The stored config is seeded from the manifest's configuration at install. Updates are validated (app/services/plugin_config.py) against that declared configuration when one is present: unknown keys and wrong-typed values are rejected with 400, and every stored object is bounded (key count, nesting depth, serialized size). Because config may hold secrets it is owner-only and is deliberately omitted from the plugin list/inspect surfaces.

Structured error reports

Plugin failures are recorded as bounded, safe rows (app/services/plugin_errors.py → plugin_error_reports) instead of only log lines. Each report stores the plugin id, the operation (dispatch:<event_type>, hook:<hook>, …), the exception type, and a length-bounded message. No stack traces, event payloads, or secrets are stored by default.

Recording happens automatically at every plugin invocation boundary inside an app context: loading an entry point (load), instantiating a plugin class, running a lifecycle hook (hook:<hook>), dispatching an event handler (dispatch:<event_type>), and the legacy registry dispatcher (registry-dispatch:<event_type>). It is best-effort and never changes the existing failure-isolation behavior. Reports are read back owner-scoped via GET /plugins/api/workspaces/<ws>/plugin-errors (owner only, workspace-isolated) and operator-wide via flask plugins errors [--json].

Configuration reference

Marketplace and plugin settings are environment variables. The defaults are deliberately fail-closed: catalog installation is disabled, no publisher is trusted, and plugin outbound traffic is denied until an operator opts in.

Variable Safe default Purpose
PLUGIN_CATALOG_FILE empty Path to the validated JSON catalog used by POST .../plugins/install-from-catalog. Empty disables catalog installation.
PLUGIN_MARKETPLACE_ENABLED 1 Enables the read-only discovery API (GET /plugins/api/marketplace). Disabling returns 404.
PLUGIN_MARKETPLACE_CATALOG plugins/marketplace.json Local JSON catalog for the discovery API. Ships empty; set it to advertise validated entries.
PLUGIN_MARKETPLACE_URL empty Optional https catalog URL fetched through the bounded plugin transport when no local marketplace catalog is configured. The host must be on PLUGIN_NETWORK_ALLOWLIST; redirects, non-https URLs, oversized bodies, and non-allowlisted hosts are refused before any request.
PLUGIN_DISCOVERY_DIR plugins Read-only directory scanned for manifest.json files by GET /plugins/discover. Nothing is installed or executed.
PLUGIN_TRUST_POLICY if-present if-present accepts unsigned manifests but always refuses forged or untrusted signatures; required also refuses unsigned manifests. Use required for a hosted marketplace.
PLUGIN_TRUSTED_KEYS {} JSON object of canonical publisher ids to Ed25519 public keys (base64 raw 32 bytes or PEM). Invalid key material aborts startup.
PLUGIN_NETWORK_ALLOWLIST empty Comma-separated outbound hosts. Empty denies every plugin request; * allows any public host and *.example.com allows subdomains.
PLUGIN_NETWORK_HTTPS_ONLY 1 Rejects non-HTTPS plugin requests. Keep enabled outside isolated development.
PLUGIN_NETWORK_ALLOW_PRIVATE 0 Allows private/loopback/link-local targets. Keep disabled in production.
PLUGIN_NETWORK_TIMEOUT 15 Outbound request timeout in seconds.
PLUGIN_NETWORK_MAX_BYTES 2097152 Maximum response body accepted from a plugin request (2 MiB).
PLUGIN_NETWORK_STRICT_DNS 0 When enabled, an unresolvable host is rejected instead of tolerated.
PLUGIN_EXECUTION_TIMEOUT 5 Seconds a plugin event handler or lifecycle hook may run before it is abandoned and recorded as a bounded error. 0 or negative disables the bound.
PLUGIN_CONFIG_MAX_BYTES 65536 Maximum serialized size of a stored per-workspace plugin configuration.
PLUGIN_CONFIG_MAX_KEYS 200 Maximum number of top-level keys in a stored plugin configuration.
PLUGIN_CONFIG_MAX_DEPTH 8 Maximum nesting depth of a stored plugin configuration.
PLUGIN_ERROR_REPORT_LIMIT 100 Maximum error reports returned to a workspace owner (hard ceiling 500).

Catalog cache: there is no active cache TTL setting yet. The catalog file is validated and re-read for each install request, so an operator's catalog edit is visible immediately. A TTL cache and its PLUGIN_MARKETPLACE_CACHE_TTL setting are tracked separately in #377.

Remote catalog (#375): when PLUGIN_CATALOG_FILE is empty and PLUGIN_MARKETPLACE_URL is set, the catalog is fetched through app/services/plugin_network.py. The URL must be https and its host must be on PLUGIN_NETWORK_ALLOWLIST; those checks run before any socket is opened. Redirects are refused, the request uses PLUGIN_NETWORK_TIMEOUT, and the body is capped at the catalog limit (1 MiB). A local PLUGIN_CATALOG_FILE always takes precedence.

Resource limits and failure behavior (#409, #410)

All plugin resource limits are configuration-driven (app/services/plugin_limits.py) and have conservative defaults. Invalid or non-positive values fall back to the default rather than disabling the bound; the execution timeout is the one deliberate exception, where 0 or a negative value disables the bound for local debugging. Defaults, the settings that override them, and the failure behavior are summarized below:

Boundary Default bound Behavior when exceeded
Event handler / lifecycle hook PLUGIN_EXECUTION_TIMEOUT=5 seconds The callback is abandoned (daemon thread), one bounded error report is recorded, and dispatch continues with the remaining handlers.
Stored workspace configuration 64 KiB, 200 keys, depth 8 The write is rejected with 400 and nothing is stored.
Owner error-report read 100 rows The response is truncated to the configured limit (maximum 500).
Outbound plugin HTTP 15 seconds, 2 MiB The request fails closed; see the network settings below.

Failure isolation covers every plugin invocation boundary: loading an entry point, instantiating the plugin class, running a lifecycle hook, and dispatching an event handler or legacy registry handler. A failure at any boundary is contained, leaves the registry consistent, and records one bounded report (load, hook:<name>, dispatch:<event>, or registry-dispatch:<event>). Requests still complete and other handlers still run. Regression tests inject failures at each boundary in tests/test_plugin_failure_isolation.py.

Example production posture:

export PLUGIN_CATALOG_FILE=/etc/ai-code-assistant/catalog.json
export PLUGIN_TRUST_POLICY=required
export PLUGIN_TRUSTED_KEYS='{"acme-labs": "<base64 or PEM Ed25519 public key>"}'
export PLUGIN_NETWORK_ALLOWLIST=plugins.example.com
export PLUGIN_NETWORK_HTTPS_ONLY=1
export PLUGIN_NETWORK_ALLOW_PRIVATE=0
export PLUGIN_EXECUTION_TIMEOUT=5
export PLUGIN_CONFIG_MAX_BYTES=65536
export PLUGIN_CONFIG_MAX_KEYS=200
export PLUGIN_CONFIG_MAX_DEPTH=8
export PLUGIN_ERROR_REPORT_LIMIT=100

Audit trail (capabilities & state)

Security-relevant plugin actions are recorded in the shared, append-only workspace audit log (ActivityEvent, via app/services/plugin_audit.py):

  • plugin.capability.granted / plugin.capability.revoked — explicit grant or revoke through the management API.
  • plugin.enabled / plugin.disabled — workspace installation state changes.
  • plugin.denied — a rejected capability request (e.g. not declared by the manifest) or a workspace-scoped dispatch-time denial (plugin disabled / not installed / missing capability grant).

Each entry carries only safe facts — actor, workspace, plugin_id, capability name, action, outcome, a static reason, and (for dispatch denials) the refused event type. Secrets and event payloads are never stored. Grant/revoke rows are only appended on an actual change (repeat grants are not duplicated), and dispatch-time denial recording is best-effort and never changes the fail-closed delivery decision.

Entries are owner-visible through the workspace audit view (GET /workspaces/api/workspaces/<id>/audit); the member activity feed always excludes this audit subset, and cross-workspace records are never readable outside the workspace they belong to (non-members receive 404).

Writing a plugin

New to plugins? Start with the step-by-step Plugin developer guide. It walks through the layout, manifest, plugin class, install/enable, explicit capability grants, the local verify loop, and common pitfalls, using the runnable examples/plugins/hello_world example.

# plugins/stellar_tools/plugin.py
class StellarPlugin:
    def __init__(self, app=None, manifest=None):
        self.app = app
        self.manifest = manifest

    def on_event(self, event):
        # subscribe via:
        #   from app.services.events import get_dispatcher
        #   get_dispatcher().subscribe("project.created", self.on_event, plugin_id="stellar-tools")
        ...

The entry_point value names module.path:ClassName. The plugin class may be instantiated with app and manifest kwargs (see Plugin.get_instance).

Example plugin

A minimal, runnable example lives at examples/hello-plugin/:

  • manifest.json — a valid manifest (id hello-plugin, capability PROJECT_READ, entry_point plugin:HelloPlugin).
  • plugin.py — HelloPlugin, which subscribes to project.created, logs each event, and records it in received.

Run it from the repository root. Both the repository root (so app resolves) and the example directory (so the entry_point resolves) must be importable:

# Windows (PowerShell)
$env:PYTHONPATH = ".;examples/hello-plugin"
python examples/hello-plugin/plugin.py

# Linux / macOS
PYTHONPATH=.:examples/hello-plugin python examples/hello-plugin/plugin.py

It registers through the real PluginRegistry, subscribes to project.created via the shared EventDispatcher, and prints how many events it recorded. Subscribing never requires a capability grant; grants are enforced only when the host dispatches a live event (see Security). The manifest schema and the load/register flow are covered by tests/test_hello_plugin_example.py.

Stellar notifier example (#415)

examples/plugins/stellar_notifier/ shows an event-driven, read-only integration for Stellar analysis:

  • declares STELLAR_READ and subscribes to stellar.analysis.completed;
  • formats a one-line summary (Project 7 — Stellar detected, confidence 0.98) that is capped at MAX_SUMMARY_CHARS, with each field truncated and the in-memory history bounded to MAX_RECENT_SUMMARIES;
  • copies the payload instead of mutating it and makes no network or database calls, so it runs offline under the mock network;
  • is validated end-to-end by tests/test_stellar_notifier_example.py (schema, load/register, event path, bounding, read-only behaviour, and an offline socket guard).

Security

See security.md for the full security model. In short:

  • Capabilities are explicit and audited (never implicit).
  • Plugin handler failures are isolated from the request lifecycle.
  • Manifests are validated (no arbitrary code identifiers, no unknown capabilities).
  • Publisher signatures are Ed25519 over the canonical manifest body; forged or untrusted signatures are always refused, and required also refuses unsigned entries.
  • Trusted publisher keys are validated at startup and never exposed by the API or CLI (only publisher ids are listed).
  • Catalog installs are owner-only, resolve server-side, register metadata only, and never download code or grant capabilities.
  • Event data carries workspace/user context so plugins can scope work; the Stellar analysis path reuses the Phase 7 content-access gate and fails closed.
  • Event data carries workspace/user context so plugins can scope work; the Stellar analysis path reuses the Phase 7 content-access gate and fails closed.

Contributing

The foundation is intentionally small. Planned, contributor-friendly work is tracked under the Phase 8 - Plugins & Extensions milestone (label phase-8), for example:

  • Dependency resolution and version compatibility checks (implemented — see app/services/plugin_deps.py and app/services/plugin_compat.py).
  • Plugin packaging/distribution and a marketplace (implemented — signed, catalog-pinned installs plus the read-only discovery API; see Marketplace trust model).
  • CLI commands for plugin management.
  • A capability audit trail (implemented).
  • A plugin development guide and an example plugin (implemented — see plugin-developer-guide.md and examples/plugins/hello_world).

Testing

Run the whole suite with pytest (no external services; the event dispatcher is reset between tests):

  • tests/test_plugins_manifest.py, tests/test_capabilities.py — unit tests for manifest validation, capability grants, and role mappings.
  • tests/test_plugins_api.py — the workspace-scoped plugin management API (install/enable/disable/grant) with authorization fail-closed and workspace isolation.
  • tests/test_plugin_trust.py — trusted publisher key loading (base64/PEM), startup rejection of bad configuration, and the secret-free listing surfaces.
  • tests/test_plugin_signatures.py — canonical payloads, detached signatures, forged/unsigned handling, and the embedded-signature regression path.
  • tests/test_plugin_catalog_install.py — catalog parsing/validation and the remote-install endpoint: authorization, trust policy, 4xx failure paths, and no implicit grants or code execution.
  • tests/test_marketplace_e2e.py — offline catalog-to-dispatch coverage: install, explicit enable, dependency ordering, trust refusal, rollback, and no implicit grants.
  • tests/test_plugin_lock_entries.py — the PluginLockEntry model, scoping, unique constraint, and canonical lockfile serialization.
  • tests/test_plugin_audit.py — the capability/state audit trail.
  • tests/test_event_authorization.py — dispatch-time capability enforcement.
  • tests/test_event_wiring.py — real routes emit events to authorized installed+granted plugins.
  • tests/test_plugin_integration.py — the end-to-end flow: a manifest written to a temp directory is parsed/validated, installed through the real API (configuration stored, no implicit grant), the capability is granted explicitly, a handler subscribes to project.created, a real request emits the event, and the authorized handler runs and persists a result. It also proves a failing handler is isolated and that an un-granted (or disabled) plugin never executes.