feat: add Grok Build runtime (plain-mode adapter fields + registry entry)#27
feat: add Grok Build runtime (plain-mode adapter fields + registry entry)#27tynamite wants to merge 9 commits into
Conversation
…untimes) Extends the manifest spec with the fields a finite one-shot headless runtime needs — prompt_flag / interactive_prompt_flag (flag-bound prompts), continuity_args (cold-start session init/resume), and event_protocol (grok-headless-v1) — and registers Grok Build 1.0.0. Spec changes are additive with snake_case + camelCase aliases; legacy manifests deserialize unchanged. New validation mirrors coven's loader: event_protocol and capabilities.stream are mutually exclusive, continuity_args must declare a usable init or resume launch, its session_id_flag requires preassigned_session_id, and the preassigned-session rule now accepts a continuity session flag. JSON Schema updated in lockstep (schema_examples drift guard extended with a grok-shaped adapter and examples/grok.json). The manifest matches the trusted recipe reviewed in tynamite/coven#1 byte-for-token: --single prompts, streaming-json output, --session-id / --resume continuity, and argv-list sandbox mapping driving both --permission-mode and --sandbox. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ct8hD3ixehKLeaSu7jGwnA
… install hint Normalize the preassigned-session ValidationError field to the canonical snake_case key (capabilities.preassigned_session_id) so error fields map onto manifest JSON consistently, and reword the Grok Build install_hint to lead with npm and the official install guide instead of embedding a pipe-to-shell command, matching the other registry entries. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ct8hD3ixehKLeaSu7jGwnA
Per Codex review: a streaming adapter receives the pre-assigned id through stream_args, so a continuity-only session_id_flag validated but could never reach the streaming process. Streaming adapters now require stream_args.session_id_flag; the continuity flag satisfies the rule only for non-stream adapters (the Grok Build shape). Regression test added; docs updated to match. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ct8hD3ixehKLeaSu7jGwnA
…conjure A registry index written by a newer spec version used to fail parsing entirely on older consumers: deny_unknown_fields on the manifest types meant one new field poisoned every runtime in the document, and the notify-downstream sync makes that scenario real, not hypothetical. Parsing is now tolerant — unknown fields are ignored and an unrecognized event_protocol value degrades to a per-adapter Unknown marker instead of failing the index — while authoring strictness moves to the new pure unknown_manifest_fields checker, which conjure's shared manifest loader enforces (with a friendlier hint when the input is actually a registry index). validate_manifest rejects Unknown protocols so authored manifests must name a protocol their target spec knows; the sandbox object keeps strict field matching because its untagged forms are disambiguated by field sets. Docs cover the two-layer rule and the v0.1.3 live-index caveat; drift-guard and forward-compat tests added. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ct8hD3ixehKLeaSu7jGwnA
Per Copilot review: stream_args and continuity_args now validate against their own allowlists so a key nested in the wrong block is reported by the authoring typo guard, and the adoption doc anchors the v0.1.4 behavior split to the release that first contains tolerant parsing instead of implying the unbumped in-repo version. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ct8hD3ixehKLeaSu7jGwnA
Per Codex review, tolerant parsing had two conjure-side holes: validate --registry silently blessed typo'd entries (the strict check only guarded load_manifest), and load-rewrite flows like registry yank would destructively strip newer-spec fields and collapse future protocol values to 'unknown' when an older conjure touched a newer index. The spec's unknown-field checker is refactored into a shared unknown_adapter_fields (which also flags event_protocol values that parse to Unknown), the registry crate gains unknown_index_fields covering index/entry/adapter levels, and load_registry now refuses any unrecognized content — typos fail validation, and mutation of a newer-spec index fails with an upgrade hint instead of losing data. Tolerant parsing remains the read-only consumer path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ct8hD3ixehKLeaSu7jGwnA
…tocol schema Per Copilot review: unknown_index_fields iterates borrowed objects instead of cloning the runtimes map and entry arrays, the CLI loaders deserialize from the already-parsed Value instead of parsing twice (load_registry had the same pattern), and the schema's eventProtocol definition accepts explicit null like every other optional scalar field. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ct8hD3ixehKLeaSu7jGwnA
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 3ee734af0d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| let unknown = coven_runtime_registry::unknown_index_fields(&value); | ||
| if !unknown.is_empty() { |
There was a problem hiding this comment.
When an older conjure runs registry list against an index containing a newly added adapter field or protocol, run_list calls this shared loader and the unknown-field check rejects the entire index. Listing is read-only, so this defeats the tolerant parsing introduced for read-only consumers and prevents even known runtimes from being displayed; use tolerant RegistryIndex parsing for registry list while retaining this strict check for validation and mutation flows.
Useful? React with 👍 / 👎.
| #[serde(default, alias = "promptFlag", skip_serializing_if = "Option::is_none")] | ||
| pub prompt_flag: Option<String>, |
There was a problem hiding this comment.
If an authored manifest sets prompt_flag (or interactive_prompt_flag) to "" or whitespace, the schema and validate_adapter both accept it, but consumers distinguish Some(flag) from the positional-prompt fallback and will construct a malformed flag-bound prompt. This makes conjure validate bless an adapter whose launches fail; validate these new optional flags as non-blank when present, as is already done for session and sandbox flags.
Useful? React with 👍 / 👎.
Per upstream Codex review on OpenCoven#27: registry list is read-only, so it now loads the index tolerantly — one entry written by a newer spec must not make every runtime unlistable. Mutating and authoring flows (validate --registry, registry build, registry yank) keep the strict loader. prompt_flag and interactive_prompt_flag reject blank values in both the Rust rules and the JSON Schema (minLength) — a present-but-blank flag would make consumers emit a malformed '=<prompt>' argument instead of falling back to a positional prompt. Regression tests for both. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ct8hD3ixehKLeaSu7jGwnA
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 85e9b45649
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| /// with no positional prompt slot (e.g. Copilot's `--prompt`, Grok Build's | ||
| /// `--single`). `None` means the prompt is the final positional argument. | ||
| #[serde(default, alias = "promptFlag", skip_serializing_if = "Option::is_none")] | ||
| pub prompt_flag: Option<String>, |
There was a problem hiding this comment.
Adding this field and the other new RuntimeAdapter fields breaks every downstream struct literal and exhaustive destructuring pattern, even though the fields are optional for serde. The compatibility documentation explicitly identifies the next release as v0.1.4, but a patch update within the 0.1 series is expected to remain source-compatible; consumers can therefore receive compilation failures from an ostensibly compatible upgrade. Release this API expansion as 0.2.0, or provide a construction strategy that does not require callers to enumerate fields.
Useful? React with 👍 / 👎.
| if let Some(continuity) = &adapter.continuity_args { | ||
| if !continuity.has_init_launch() && !continuity.has_resume_launch() { |
There was a problem hiding this comment.
When continuity_args contains usable init data, a present but blank resume_flag passes this aggregate check because has_init_launch() is already true and there is no per-field validation. The serialized adapter still contains Some(" "), so consumers that distinguish presence from absence can construct a malformed flag-bound resume invocation instead of using the positional/absent fallback. Reject resume_flag when present but blank, as is now done for the prompt flags.
Useful? React with 👍 / 👎.
The companion coven PR landed as a pure manifest recipe running Grok Build's plain output mode — no event-protocol bridge in core. Align this registry PR with that outcome: - Drop the EventProtocol enum, its two validation rules, its schema definition, and the unknown-protocol-value scan; the generic event_protocol mechanism moves to a maintainer design issue and can return with it. - Re-point the grok manifest, example, and canonical index entry at the shipped plain-mode contract (`--output-format plain` prefix args, official-installer install hint, protocol-free description). - Keep everything the recipe actually exercises: ContinuityArgs, prompt flags, blank-flag validation, unknown-field strict/tolerant layers. Forward-compat tests that used an unrecognized protocol value as their fixture now exercise the same tolerance through unknown fields. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ct8hD3ixehKLeaSu7jGwnA
|
Heads-up on the rescope just pushed (cd92006): the companion coven PR
Diff is −204/+57 versus the previous head. fmt / clippy -D warnings / |
tynamite
changed the title
|
@BunsDev Update now that the companion PR has landed: OpenCoven/coven#410 merged This PR is already rescoped to match that outcome (cd92006), removing
|
2 participants
Adds Grok Build to the runtime registry as a plain-mode one-shot
adapter, plus the spec fields it needs — matching the recipe-only
adapter in OpenCoven/coven#410 exactly.
Spec additions (coven-runtime-spec)
ContinuityArgs— one-shot session continuity (init/resume prefixargs,
session_id_flag,resume_flag), mirroring coven's type.prompt_flag/interactive_prompt_flag— flag-bound prompts(
--single=<prompt>) for runtimes with no positional prompt slot.launch; blank prompt flags are rejected; preassigned-session-id may be
satisfied by the continuity path.
spec versions; authoring (
conjure) stays strict viaunknown_manifest_fields/unknown_index_fields. Mutating registryflows load strictly; read-only
registry listloads tolerantly so onenewer entry can't make every runtime unlistable.
Registry
registry/runtimes/grok/1.0.0.json+examples/grok.json: the GrokBuild adapter in plain output mode (
--output-format plain, Grok'sown headless default — final response text on stdout, errors on
stderr, no translation layer), with sandbox argv mapping and
continuity args identical to the coven trusted recipe.
conjure registry build.Deliberately not included
No
event_protocolfield or enum: the coven-side event bridge waswithdrawn in favor of the recipe-only adapter per maintainer review on
OpenCoven/coven#410. The generic mechanism belongs to a follow-up design
issue and can re-enter the spec with that discussion.
Verification
cargo fmt --check,cargo clippy --workspace --all-targets -- -D warnings,cargo test --workspace --locked(92 tests) all pass.Schema and examples validated in lockstep (
schema_examplessuite).