V7.17.1 Current System Architecture and Security Boundaries¶
Status:
active. This page describes the current V7.14.6 package architecture. Earlier design and release evidence is retained only for historical traceability.
1. Layers¶
Base Plugin: 10 Skills -> Main Agent -> Engineering task
Optional managed enhancement
+ Global AGENTS / 7 Reviewers
+ Account Hooks: 8 event types / 9 registrations
+ Project Context Runtime / index / budget
+ TaskOutcomeEvent V3 -> Assessment / Proposal
+ Human Decision -> Independent Implementation Task
The package version is V7.14.6. Names such as TaskOutcomeEvent V3, Operation v2, and Evolution Policy identify component contracts or data formats; they do not mean an older package is installed.
The base Plugin has no Hooks and does not start runtime state. The following lifecycle, isolation, budget, and governance sections apply only after installing the optional enhancement. Its verified host profile registers eight event types and nine account entries because PreToolUse has separate dispatch and write matchers.
2. Skill routing¶
Choose one primary domain Skill per phase:
backend-engineering: server applications, APIs, business logic, transactions, concurrency, and workers;frontend-engineering: browsers, WebViews, renderers, state, and interaction;ai-engineering: model calls, RAG, agents, evaluation, inference, and multimodal generation;data-middleware-infrastructure: databases, caches, messaging, search, storage, GPU resources, containers, and networks.
Logging, quality delivery, independent review, technical documentation, long-running memory, and controlled evolution are supporting capabilities loaded by phase. See the V7.9 domain Skill architecture and V7.9 Skill trigger matrix for detailed boundaries.
3. Project and data isolation¶
Every TaskOutcomeEvent V3 binds at least:
project_id;repo_fingerprint;session_id / turn_id / task_id;event_id.
Observation first verifies the hash chain or HMAC, checks project_id + repo_fingerprint, deduplicates by event_id, and then aggregates by Task. A project or repository identity mismatch fails closed instead of mixing cross-project records into one conclusion.
4. Hook and model boundaries¶
PreToolUse checks the automatic sub-agent model ceiling before dispatch. SubagentStart and SubagentStop record minimal runtime facts, while the other observation Hooks form lifecycle events; Interrupt remains host-controlled and does not write legacy gate-cancellation state. The Hook guard is a workflow protection, not an unbypassable platform security boundary.
Model evidence keeps two meanings separate:
dispatch_policy_status = whether the pre-dispatch request obeyed the approved ceiling
host_model_identity = not collected, inferred, persisted, or exported
Only the approved dispatch profile, permit, and reserved cost participate in governance. Host runtime model identity and reasoning effort are outside the data contract.
5. Reviewer isolation¶
read-only in a Reviewer TOML file expresses configuration intent only. If the parent session is writable and no valid system-denial evidence exists, the review is only logical-readonly. It can be system-readonly only when the whole parent session is read-only or a controlled probe is actually denied by the system. Self-review is not independent review.
6. Privacy and integrity¶
Lifecycle data retains only minimal structured metadata required for governance. Raw prompts, complete responses, source bodies, patches, diffs, tokens, secrets, authorization data, cookies, API keys, and private keys must not be persisted by default.
The event chain uses forward SHA-256 integrity and can add HMAC and detached seals when configured. The SessionEnd Hook only constructs a capped, body-free sanitized event and dispatches without waiting; it performs no event-chain I/O or synchronous pipe write. A detached worker validates stable lifecycle identity at the shared queue boundary, semantically deduplicates and persists the event, then seals the chain through a v2 signed job bound to event_id. Evolution rejects an unsealed seal_required tail. Active events, read-only segments, and archive manifests preserve project identity and chain-head continuity; corruption, cross-project leakage, or inconsistent references fail closed.
7. Proposal authorization model¶
A Proposal is governance advice only:
PENDING_REVIEW
├─ REJECTED
├─ DEFERRED
└─ ACCEPTED
↓
IMPLEMENTATION_LINKED
↓
VALIDATION_RECORDED
↓
CLOSED
A proposal superseded by newer evidence may become SUPERSEDED. No state changes execution_authorization from NONE. Human acceptance still requires a new implementation task and fresh authorization for modifications, commits, pushes, or publication.
8. Current and historical documentation¶
- Enter current guidance through the documentation hub, where it is marked V7.9.
- Upgrade-source versions, migration mappings, and component-contract versions may appear in current guidance only when their purpose is explicit.
- Earlier release notes, validation reports, and design documents remain available for traceability but do not establish current installation, runtime, or acceptance state.
- Historical detail pages are excluded from default site search so outdated commands cannot be confused with current operating instructions.
Optional project gate and effective loading¶
The registration contains eight entry points, while project gates default to disabled. In V7.14.6, UserPromptSubmit is asynchronous observation, Stop is neutral observation, and Interrupt remains host-controlled. Canonical apply_patch advances repository-external Operation v2 through PreToolUse/PostToolUse only for an explicitly enabled policy. A creates an origin and is denied, B atomically binds READY, and completion accepts only B's receipt. Unconfigured or disabled policies remain neutral.
Unconfigured or disabled ───────────────→ native writes preserve host behavior
Legacy enabled + native write tool ────→ deny LEGACY_WRITE_ORIGIN_UNAVAILABLE
UserPromptSubmit / Stop / Interrupt ───→ do not consume legacy GateTask control state
Configured enabled=true, a loaded Plugin, and semantic suitability are separate conclusions. Legacy prepare/finish/check and GateTask v1 receipts cannot authorize Operation v2 writes. An open session can retain an older Plugin snapshot, so verify actual behavior in a fresh task after upgrade. Hook permission and tool side effects are not atomic, and shell/MCP entry points do not have guaranteed pre-write interception.
All 11 Codex CLI versions in the frozen window have per-version official source evidence for UserPromptSubmit async behavior; Desktop and other real hosts still require separate readback. See the capability index and workflow entry for legacy gate enable/disable entry points, but this patch does not restore the legacy task lifecycle as write authorization.