Repository navigation
Update for new version - #148
Merged
Merged
Conversation
Register Opper as an OpenAI-compatible gateway, following the Requesty registry pattern: one ProviderSpec and one ProvidersConfig field. - Default base https://api.opper.ai/v3/compat, detected by provider name or by opper.ai in apiBase. Reads OPPER_API_KEY and PAOS_OPPER_API_KEY. - custom_openai keeps model ids unchanged, so pool names such as claude-sonnet-4-6 and pinned routes such as anthropic/claude-sonnet-4-6 both reach Opper as written. - English and Chinese user manual entries, including the gateway reasoning effort note. - Offline tests for config selection, runtime credentials and request routing with a mocked completion. Disclosure: Opper is my startup. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
feat(providers): add Opper gateway
Downgrade three hash-mismatch hard failures to warnings so local modifications of files inside an installed Forge Skill remain runnable: - AgentTask creation no longer fails when the activated SKILL.md hash differs from the installed Runtime binding's frozen document hash (both hash prefixes are logged); provenance in the binding is kept. - Binding preview no longer fails when the installed skill.yaml version differs from the running Runtime's startup version. - NodeInstaller.load no longer fails when an installed node executable sha256 differs from its receipt; an unusable receipt sha256 still errors. Download/archive integrity, receipt-vs-lock consistency, and the interaction-Skill change rails remain strict. Relates to #134
- Restore the hard failure when the installed skill.yaml version differs from the running Runtime (binding.py): validate_tool() still requires binding.skill_version == runtime.skill_version, so allowing the mismatch at preview only deferred an opaque failure to the first tool call. - Freeze the document the agent actually reads (workspace overlay first) as workflow_document_sha256 on the binding candidate/binding, next to the installed bundle document hash. Task creation warns when the activated document differs from the frozen workflow document or from the installed bundle; it does not block. The field is excluded from the freeze-time equality check so local edits between activation and task creation do not invalidate the binding. - NodeInstaller.load(strict=True) still fails on a modified node executable; forge-node verify now uses it, so the CLI never prints 'SHA-256 verified' for a file that differs from its receipt. The runtime start path keeps the warning-only behavior.
- freeze() now takes activated_document_sha256 from task creation and freezes that hash into the binding, instead of the hash re-read from the file at task creation time. A local edit between activation and task creation no longer re-attributes execution to a version the agent never read; the divergence is logged as a warning. - The workflow document lookup returns None when the resolved document cannot be confirmed (no loader wired, resolution failed, unreadable file) instead of substituting the installed bundle document hash. - satisfies() (install cache acceptance) is strict again: a locally modified executable is not a valid cache hit, so installing the original archive cannot silently return the modified file. Only the Runtime start path keeps the warning-only behavior. - Experience episodes now attribute skill_document_sha256 from workflow_document_sha256 first, falling back to the installed bundle document hash for bindings that predate the field.
…l-file-change fix(skill-runtime):warn instead of blocking when locally modified Skill files hash-mismatch
A structured Gateway rejection raised before anything is dispatched to a provider (concurrency exhausted, admission rejected, endpoint/operation unknown, no active provider) provably leaves no remote state, but start_action recorded every invoke exception as status=unknown. That armed the identical-Action guard, so a 503 rejection blocked the exact resend the retryable codes invite, forcing operators to differentiate arguments on every probe. Classify such rejections as failed via a whitelist of pre-dispatch error codes. FORGE_TOOL_INVOKE_TIMEOUT is deliberately excluded: the Gateway also emits it after dispatch when the synchronous caller response times out, where the provider may still be executing and the outcome is genuinely unknown. HTTP timeouts and transport failures without a structured code, and errors carrying a remote invocation identity, keep the unknown status and the guard.
A rejected-then-retried Action produced records [failed, succeeded], but finalize_task() required every record to succeed, so off and audit modes still reported FAILED. Mark records the gateway provably rejected before dispatch and ignore them when evaluating execution facts and cancel status; a task whose every call was rejected still fails because nothing executed. The gateway emits error.code for Action admission rejections from 1.1.2 onward; older 503 responses remain recorded as unknown.
…atus fix(forge): record pre-dispatch Action rejections as failed, not unknown
…-dev docs: add bilingual TrackB guide and Piper X development demo
docs: add contribution guide focused on TrackB development
docs: streamline developer guides and add simulation quickstart
docs: retain benchmark documentation navigation
docs: add live MuJoCo GIFs and highlight Piper X demo
v1.1.1 - Skill development, execution reliability and the Opper gateway
ci: automate PyPI publishing from stable GitHub releases
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.