Compatibility
Two agents have been frozen mid-call, killed, and resumed on the bench. That is evidence of generality, not proof of universality, and this table only grows when a gate runs.
The tables below are programs run in plain sessions. Agent sessions run a certified runner; which ones are certified is on Runners.
Every row is proven on the evidence bench; hosted parity publishes per row as the fleet's own gates run, and no row inherits a green it did not earn there.
95 green tags on this release, 1 not green · full list in the archive
the historical record on Proof
soak-3-red: red. 71 of 72 checkpoints: tick 47 failed when the firecracker snapshot exceeded its fixed client deadline (BACKLOG-39); no public row moves and there is no soak-3-green tag. SUPERSEDED by soak-3b (2026-09-19): the deadline defect it found is BACKLOG-39, fixed and deployed on 2026-09-08, and the re-run on the fixed engine held 72 of 72 checkpoints with a max of 24.3 s against a deadline that now scales. This entry keeps its red because it is a completed measurement of a defect that existed; the record is not rewritten.
1b-1-green · 1b2d-green · ac-1-green · ad-1-green · ad-2-green · ad-3-green · ad-4-green · af-1-green · af-2-green · agents-prod-green · az-0-green · az-1-green · az-2-green · az-3-green · az-3b-green · az-4-green · az-5-green · b37-green · b38-green · b38-prod-green · b39-green · b44-green · b44-prod-green · b48-green · b48-prod-green · billing-ledger-prod-green · browser-console-prod-green · byok-prod-green · ckpt-resume-prod-green · cli-1-green · cp-1-green · cp-2-green · cp-3-green · cp-c1-green · cp-c2-green · cp-c3-green · cp-c4-green · cp-c5-green · cp-c6-green · cr-1-green · cr-2-green · cr-3-green · cr-4-green · cr-5-green · cr-6-green · cr-7-green · cr-8-prod-green · cr-9-green · deploy-pv-green · em-1-green · email-prod-green · er-1-green · er-1-prod-green · er-1c-green · er-1c-prod-green · er-1d-green · er-1d-prod-green · fo-1-green · fo-1b-green · fo-1b-prod-green · fork-prod-green · hosted-attach-prod-green · hosted-prod-green · hr-1-codex-green · hr-1-openrouter-generic-green · kill-guard-green · l-1-green · launch-ready-1 · product-prelaunch-audit-baseline-2026-09-08 · pv-1-green · pv-2-green · pv-3-green · pv-3-prod-green · pv-3b-green · rs-1-green · s1-green · s11-green · s2-green · s3-green · s4-green · s5-green · s6-green · s7-green · self-serve-prod-green · soak-2-green · t0-c-p-green · t0-p-green · v0.1.0-mvp · v0.1.1 · v0.1.2 · v0.1.3 · v0.1.4 · v0.1.5 · v0.1.6 · v0.1.7
Complete compatibility matrix
| agent | absorption claim | evidence |
|---|---|---|
| Claude Code | claimed | record Historical client-specific result. A successful provider replay does not guarantee that every client retries. |
| aider 0.86.2 | not claimed door held 4 of 4, walked through 3 of 4 | record Historical client-specific result. A successful provider replay does not guarantee that every client retries. |
Rows are never edited by hand. Every tag links to the archive on Proof.
Coding harnesses
One badge cannot mean that running, recovering, replaying and switching all work. Each operation below is its own column and carries its own answer. A cell says supported only when it can name the gate that proved it for that harness; otherwise it says not claimed, which is different from a no. This matrix is supported by a dated public test summary.
| harness | run | checkpoint and resume | interactive attach | fork | stored-response replay | model or provider change | access status |
|---|---|---|---|---|---|---|---|
| codexcodex exec --skip-git-repo-check | supported record | supported record | not claimed | not claimed | supported record | not claimed | available hosted, with your own provider key |
| openrouter-genericaider --yes --no-auto-commits | supported record | supported record | not claimed | not claimed | supported record | not claimed | available hosted, with your own provider key |
What each column means, and what earns it
- run
- the harness starts and executes work inside a session. derived: the level gate ran this harness in a microVM to prove the rest, so a green level gate proves it ran
- checkpoint and resume
- the session freezes, is killed, and comes back with state intact. the durable level, which is exactly this claim
- interactive attach
- a terminal into the running session. no harness gate exercises attach. It is a session verb and is not proven per harness, so no harness claims it here.
- fork
- branching one checkpoint into diverging sessions. fork is proven on the engine, not per harness. Nothing proves a given harness survives being forked, so no harness claims it here.
- stored-response replay
- a stored provider response is replayed rather than re-purchased. the absorbed level, which is exactly this claim
- model or provider change
- pointing the harness at a different model or provider. the metered level proves traffic is ROUTED and billed through the gateway for the configured provider. It does not prove a change of provider or model, which nothing measures, so it is not claimed.
- access status
- who can use it today. stated from the recipe surface, not from a gate: every harness in this matrix runs on the hosted service with the account own provider key.
Two operations, interactive attach and fork, are not claimed for any harness: they are proven on the engine but no harness gate exercises them, and this table will not borrow one proof for another.
Your row is earned by a gate run, not written by marketing.
Your agent not here? The whole adaptation surface for agent #2 was one line inside the guest image. Become a design partner and your row appears when its gate goes green.
Comparisons.
Five decision matrices: what each tool is for, when to choose it over us, and where KeepState's cells cite their live claim status.