Skip to content

Commit e85479c

Browse files
committed
🔬 Certify the supervised workflow from outside, on both entrypoints
Tier CF runs the shipped workflow the way a person runs it: real `xmd` subprocesses, a private run store, a disposable bare repository, a loopback forge named by the two variables a deployment names one with, and a real ACP agent this suite starts and answers over a controller of its own. Three leaves are substituted — the agent process, the forge origin, and a `git` that bridges the one locator the Git-host adapters admit to the bare repository beside the test. Everything between them ships. CF1 completes the whole supervised workflow through the Deno source entrypoint and through the compiled binary, from discovery to acceptance, with a real suspension answered between them; and it asserts the Agent ceiling from the agent's own side, because the agent reports what it was handed — an empty provider-owned directory it can list, no additional directory, no MCP server, an empty requested tool set, and the deny-all instruction layer. CF2 kills that run mid-effect and resumes it from the exact committed journal and Workspace frontier, performing the suffix once and then reaching nothing at all on a completed replay. CF3 suspends the composed consumer at a real checkpoint, lets delivery retain the answer without executing, and continues it through #300's scheduler over the host the entrypoint builds. CF4 keeps two named Repositories and Worktrees dirty and unpushed across a restart and publishes only what the document authored. CF5 stops a malformed proposal and a denied native tool before any later effect. CF6 finds no hidden decision state in a completed run's retained history, forks it compatibly, refuses an incompatible fork, and refuses a secret in an answer. Every red I reported while building this was mine, not the product's. The scripted agent asserted one identity when a session was created and a different one when it was loaded, so every reattachment was refused and every continuation looked like a shipped divergence; and the in-process executor CF3 schedules through omitted the identity components the entrypoint declares, so `<Session>` expanded and the turns beneath it never replayed. Both are fixed here, and the corrections are recorded on #299. Two paths beyond the certification file, both registration. The suite drives `dist/xmd`, which no shard builds, so it joins the binary suite already excluded from every runtime's discovery — and the manifest says #299's own commands own it, building the binary and then naming the file outright. SCH9's reviewed-blob pins move to what R3 left, with its `525efaa7` over `170a021d` provenance untouched: two blobs changed under an approved repair, and pinning what they hold now is what keeps that case about this slice rather than about nobody. Claude-Session: https://claude.ai/code/session_01MNztk1fikXUnD2FW4yS6Kr
1 parent 6b3653c commit e85479c

4 files changed

Lines changed: 1819 additions & 6 deletions

File tree

‎packages/cli/tests/workflow-scheduling.test.ts‎

Lines changed: 12 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -1006,22 +1006,31 @@ const REVIEWED_CONSUMER_ROOTS = [
10061006
];
10071007

10081008
/**
1009-
* The reviewed consumer, as PR #181 head `525efaa7` holds it.
1009+
* The reviewed consumer, as this slice depends on it.
10101010
*
10111011
* Blob object ids rather than a diff: this slice must not change the workflow
10121012
* documents or the composition suite it depends on, and content identity says
10131013
* so from any checkout — including CI's, which is one commit deep and cannot
10141014
* name the base at all.
1015+
*
1016+
* The provenance below is unchanged: this slice was reviewed on PR #181 head
1017+
* `525efaa7` over `170a021d`, and it still is. Two of these blobs have moved
1018+
* since, and only through a repair reviewed in its own right — #299's R3
1019+
* corrected the observation instruction the Implementation stage renders, and
1020+
* extended AC2 to send it, so the workflow document and the composition suite
1021+
* each hold one approved change. Pinning what they hold now is what keeps this
1022+
* case a statement about *this* slice changing them rather than a statement
1023+
* about nobody changing them.
10151024
*/
10161025
const REVIEWED_CONSUMER: readonly (readonly [string, string])[] = [
10171026
[
10181027
"scripts/tests/adversarial-composition-workflow.test.ts",
1019-
"606c950d43646b5ad69079520ab12807b4b23735",
1028+
"bb942b506d67431b1621c99ea3e08242a62f77cf",
10201029
],
10211030
["workflows/adversarial-implementation/Discovery.md", "eea1c650fe5a9d417595b9b457b73dc51485f05a"],
10221031
[
10231032
"workflows/adversarial-implementation/Implementation.md",
1024-
"c2b71b833d7e732886f3b1f3795819183faaab05",
1033+
"c38076574eae1b55304eade578480f8efbdf3f4b",
10251034
],
10261035
[
10271036
"workflows/adversarial-implementation/InstructionFiles.md",

‎scripts/runtime-test-exclusions.ts‎

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -512,6 +512,12 @@ const COMPILED_BINARY: RuntimeExclusion[] = [
512512
"drives the compiled `dist/xmd`, which only `deno compile` produces and which no test shard builds — the subject is the binary's two loaded copies of core, a shape no Node or Bun run can build or exercise, and one the Deno shards saw only while an unrelated Deno-only build test happened to share a shard and leave the binary behind. The `smoke` job builds through README.md#Build and runs it there, beside the other suites whose subject is the binary",
513513
issue: "https://github.com/taras/executable.md/issues/567",
514514
},
515+
{
516+
path: "scripts/tests/adversarial-workflow-certification.test.ts",
517+
reason:
518+
"CF1 runs the same supervised workflow through both supported entrypoints, and one of them is the compiled `dist/xmd` no test shard builds; the rest of the suite launches the Deno source entrypoint as a subprocess, which no Node or Bun run starts either. #299's own commands own this suite: they run `deno task build` and then name this file outright, which is the only place a built binary and the certification that needs one are guaranteed to meet. It is excluded from discovery so no shard runs it without that binary — not delegated to `smoke`",
519+
issue: "https://github.com/taras/executable.md/issues/299",
520+
},
515521
];
516522

517523
/**

0 commit comments

Comments
 (0)