Read the evidence grade
Supported means a named implementation and deterministic evidence exist for the stated scope. Limited means partial fidelity or conditional support. Unavailable means no verified adapter for that operation. Unknown means missing evidence, never zero activity or implicit permission.
This matrix describes reviewed pre-release source, not installed-version parity, provider policy or live account readiness. Local means the Commandem local adapter, not an arbitrary provider CLI running on your machine.
Core conversation capabilities
| Operation | Claude | Codex | Gemini | Antigravity | Local |
|---|---|---|---|---|---|
| List/view owned history | Supported | Supported | Supported | Unavailable; native archives are not imported | Supported, default namespace |
| Named-account archives | Supported | Supported | Supported | Unavailable; named Profiles are disabled | Unavailable |
| Native owned-session resume | Supported | Supported | Supported | Supported with a known conversation ID; no archive picker | Supported, default namespace |
| Trace / Replay fidelity | Limited; richer tool events | Limited; bounded tool inputs and output bodies, execution status unknown | Limited; bounded saved tool inputs and exact response-body observations, execution status unknown | Unavailable for native archives | Limited |
| Verified multi-turn Incognito | Supported; artifact purge scope applies | Unavailable in the reviewed adapter | Unavailable in the reviewed adapter | Unavailable in the reviewed adapter | Unavailable |
| Automatic continuity | Supported within owned paths | Limited; identity checked, classification gaps | Limited; identity checked, classification gaps | Limited; exact active conversation ID | Limited; default namespace |
Codex archive evidence: Trace and Replay retain saved function/custom tool inputs and output bodies from the selected Profile, project and session. A result links to a call only when its nonempty ID identifies a unique preceding call in admitted evidence. Missing, duplicated, malformed or unsupported records remain explicit gaps; output text does not prove execution success, failure or duration. Reads cover the first 8 MiB and at most 2,000 events, with overlong lines and text explicitly limited. This is deterministic archive/CLI evidence, not complete native-tool or live-provider certification.
Gemini archive evidence: Trace and Replay retain saved tool-call inputs and matching saved response bodies from the selected Profile, project and session. A response links only when its nonblank call ID and exact tool name identify one unique preceding call in complete admitted evidence. Saved status, response prose and completion timestamps do not prove execution success, failure or duration. Message-ID replacements, whole-message snapshots and target-inclusive rewind select current saved evidence; missing, duplicated, malformed, unsupported or bounded evidence keeps explicit gaps and cannot establish a unique result attribution. The selected read covers at most 8 MiB, 1 MiB per JSONL line and 2,000 folded messages; projection retains at most 2,000 events and 8,000 runes per text field. These are selected-read and projection limits, not an aggregate discovery or whole-archive resource guarantee. Deterministic local archive/CLI evidence does not certify live Gemini execution, complete native-tool coverage, unseen writers, History/Matrix behavior or privacy parity.
Usage is a separate evidence surface
- Claude: account-scoped cached rate-limit events can provide reported windows. Absent or old observations are not real-time headroom.
- Codex: a read-only account quota RPC can provide account-specific windows; failed or unsigned-in profiles remain unknown.
- Gemini: activity and exhaustion observations are available, not a complete quota-percentage meter.
- Antigravity: native usage is conversation-cumulative. The adapter leaves turn-level usage unavailable; it does not present a quota-percentage meter.
- Local/API: server telemetry and receipt accounting vary. Estimated transcript value and subscription quota are not interchangeable.
The generic parity baseline and newer fleet-usage implementation cover different scopes; this page deliberately does not repeat the older baseline's blanket quota limitations as current fact.
Handoff and privacy are not parity shortcuts
Cross-provider transfer is reviewed context delivery, not foreign native resume. Inspect provenance and privacy before a handoff. Unknown training/retention controls remain unknown. Incognito describes local/provider session persistence, not a promise about training.
Antigravity chat and its boundaries
The separate Antigravity adapter uses the native agy CLI. It streams replies and tool results, binds a fresh conversation to the selected project, and continues the exact conversation ID returned by that provider. A known ID can be passed to commandem chat <project> --agent antigravity --resume <conversation-id>; native archive listing and import are unavailable.
Watched, non-Plan turns use a workspace PreToolUse hook and an owned per-turn socket for shared approval cards. The gate checks shared command deny rules and can reuse grants for mapped actions. This path uses native permission bypass so the hook can answer; Default asks for uncovered matched calls, Accept edits auto-allows concrete mapped edits and asks once when a matched file target is uncertain; explicit Full access retains automatic approval of remaining non-floor calls, including outbound commands. Plan, unwatched and unattended turns retain native permissions. The hook matches six named shell/file tools and does not provide complete tool or network confinement. Named Profiles remain disabled because no isolated account root is verified.
Gate fixes and remaining boundaries: a matched hard-floor command denies before shared-grant lookup. The missing-target change makes Accept edits request an explicit one-time decision for uncertain mapped file targets, with no shared-grant reuse or creation for that call; it does not change concrete-edit automatic approval or Full access. The original six failures/three healthy controls and floor-only five failures/four passes remain recorded. The unchanged original nine controls now pass. Separate matching human Allow/Deny and no-saved-grant controls qualify the explicit one-time answer path; the original five missing-target cases alone demonstrate timeout denial. These are local gate controls, not proof of native malformed payload emission or live tool execution.
Approval-surface limits: an empty reusable pattern suppresses standing-scope choices in chat, Approvals inbox and Notifications. Home retains a tool-name fallback and can still offer Always/write policy independently, so this Go change does not establish one-time-only UI everywhere. Shared file-grant derivation covers Write/Edit, not MultiEdit/NotebookEdit; no complete file-policy parity is claimed.
Only run_command's argument shape has recorded live canary evidence. Five file-tool mappings and the integrated Commandem path remain unverified against a live agy turn. Model catalog access does not establish chat, tool or approval readiness.
Boundary and recovery: hook installation failure stops an approval-active turn; inspect the reported error and native sign-in or policy. Shared grants apply only to supported mapped actions and do not change native policy. Archive-based Trace, Replay and Matrix history do not gain Antigravity coverage from live chat; Incognito and turn-level usage remain unavailable.
Other adapters and integrations
Additional providers, API endpoints, tools and collaboration adapters have separate contracts. Their presence in a selector or catalog is not certification of history, quota, Incognito or two-way sync. Inspect the in-app capability details and unavailable states before dispatch. A deterministic fixture is not a live private-service smoke test.