ObneyAI/orc / orc-skills/orc-tools
orc-tools
VerifiedBuild and use ORC tools through authored tool leaves, guarded code callers, described catalogs, typed dynamic or finite arguments, recursive specialist chains and parallel actions. Use when selecting, invoking, composing or debugging tools, including researcher-generated tool nodes.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 15.2 KB |
| agents/openai.yaml | config | 249 B |
| references/bounded-workspace-acceptance.md | markdown | 4.0 KB |
| references/catalogs-and-arguments.md | markdown | 10.1 KB |
| references/current-invocation.md | markdown | 6.3 KB |
| references/current-proofs-and-limits.md | markdown | 9.6 KB |
| references/researcher-tools.md | markdown | 14.1 KB |
| references/tools-as-behaviors.md | markdown | 8.8 KB |
references/current-proofs-and-limits.md
View on GitHub# Current tool proofs and remaining limits
## Evidence boundary
At ORC `1851212396cced2fc4533fda30b824bd4b691390` and Grain `dbf5b5229aad09785e40f71e96adf65c90f971d5`, independent public workflow experiments used actual Grain commands/processors, tenant-scoped SQLite, projection caches, caller contexts and readback. Provider fixtures prove deterministic mechanics; live calls and real MCP prove the specifically observed integrations. No engine changes were made.
## Working nominal paths
- A semantic route led to generated query arguments, discovery, a delegated note specialist, its reader grandchild, generated grounded note arguments, an authorized command and exact projected record. It used three real conversational calls.
- A native finite path chose among actual discovered source IDs, invoked an authored reader, selected a bounded audience and recorded the exact source. It used two real Jev calls, without free-text generation by the decision model.
- Two independent delegated reader tools overlapped through a two-party barrier, returned distinct observations and supplied one real conversational synthesis before an authorized command. This path has two direct children; it does not alone prove a nested delegate inside a parallel branch.
- Fresh JVMs with new caches rebuilt all three exact domain records and left their stores unchanged: 259, 178 and 157 events respectively. This is normal-shutdown projection reconstruction, not a crash or external exactly-once proof.
Declared arguments, raw results, flat tool-leaf mapping, nested code arguments, contract rejection, missing callers and useful effects were independently exercised. The source regression selection passed 131 tests / 453 assertions. Campaign ownership/capture tests passed 14 / 36. These counts do not replace consumer acceptance obligations.
## Shared budget admission works; cancellation has an open defect
The current full consumer tool suite has **15 tests / 80 assertions, zero failures, one error**. Its earlier 15 / 84 green result predates strict family trace readiness and is not the current acceptance result. **15 obligations, 15 covered, 14 satisfied, 1 unmet.**
The bounded failing family probe used actual workflows/SQLite with injected prediction and prohibited physical HTTP requests. Two provider invocations completed. A grandchild read the exact observed source; the next attempted model call was refused by the root family budget. No domain mutation occurred, and public family usage counted the two fixture calls once.
Two remaining lifecycle gaps were exposed:
1. A cancelled active child has durable cancellation facts but no canonical trace. Trace assembly subscribes to normal tree completion, not cancellation. Raw events are evidence, not a satisfied trace-publication obligation.
2. Child cancellation interrupts the same leaf worker while it dispatches root cancellation. The root cancellation can already be committed in SQLite when the interrupted caller fails before event publication/result delivery. The root caller then waits until timeout or is labeled wedged rather than receiving prompt budget cancellation.
The bounded public reproducer has **one test / seven assertions, four passes and three failures**. Keep it failing for a separate engine repair. Do not relax readiness, count timeout/wedged failure as prompt cancellation, or fabricate normal completion facts for cancelled work. Verify every started terminal family execution's trace and original identity, cancellation return/delivery, exactly one terminal fact and no late effects. Reliable durable terminal reconstruction must include cancellation when opportunistic notification was lost.
Nominal recursion and parallel work still succeeded. This defect does not mean budget admission fails or tools are generally broken. Execution leases protect recovery ownership but do not themselves supply coherent family cancellation publication or durable judge jobs.
## Real MCP and semantic completion
The unmodified filesystem-MCP researcher test used real native/conversational providers: **one test, 13 passing assertions and one failing assertion**. Tool execution, invalid-argument rejection before dispatch, missing-file error handling, checkpoint caller identities and researcher calls worked. Generated code chose the first filename matching a disjunction that included any text-file extension, read unrelated evidence, and returned incomplete prose. The model accurately reported that its supplied file lacked the requested information. A string completion schema did not establish task completion.
A separately copied consumer with stronger instructions and actual evidence/retry requirements passed **one test / 14 assertions**. This is a changed consumer policy, not proof the original passed or that a prompt universally prevents poor selection.
Generalize the lesson: choose relevance semantically from descriptions and actual observations, use exact IDs only after that assessment, continue or recover when required evidence is absent, and verify the final task postcondition. Do not use filename/extension/phrase heuristics as semantic capability or evidence selection. A safe tool invocation plus schema-valid final prose can still be insufficient work.
The earlier researcher discovery/read/save/generated-transform and same-root fresh-process recovery paragraphs in [researcher tools](researcher-tools.md) remain historical at ORC `490816267b76b0428c53437052980a25f5905209`; they are separate from this new MCP proof. Neither proves reconciliation after an external effect is accepted before its receipt is recorded.
## Fresh-author provenance counterexample
Both new nested tool consumers completed their observed live tasks. A separate wrong-source-ID probe exposed that one consumer accepted a schema-valid result for a different discovered locator, then generated and persisted a note about that result. The contract validator correctly enforced the declared shape; the missing relational ID check belonged to the consumer. The first evaluation remains failed on this obligation. A separate repaired copy rejects mismatched identities on both first and retry reads before generation or mutation. Generalize this into the explicit retrieval identity rule above; do not ask a semantic checker to replace deterministic provenance validation.
## Parallel nested consumer counterexamples
Two fresh consumers combined independent parallel specialists with a nested tool delegate and actual native/chat nodes. Public deterministic suites passed older5/48 and revised6/61. Both later produced useful live model-authored results with four tick projections and exact Grain records; the older original main first selected clarify, and its separately explicit read-only planning input succeeded. Those observed outcomes remain distinct.
Adversarial host-boundary checks exposed different consumer mistakes. The revised tree recorded inherited context at its gate but forwarded nil to the actual host; its unchanged default caller ignored context and hid the problem. The older tree accepted literal MCP isError=true alongside schema-valid fields. Its error-only fixture merely failed the data schema. Each frozen grade remains6/7. Separate public RED→GREEN repairs passed revised7/69 and older6/53 without changing original outputs/grades. Read [current invocation](current-invocation.md) for an explicit context-closing builder and MCP flag handling; assert the actual host arguments, not only configuration or gate logs.
Family tokens reconciled across registered calls. The pinned conversational OpenRouter adapter drops provider-reported cost before ORC receives it, so family cost can be incomplete. Preserve and label safe provider-level cost evidence separately; do not infer zero-cost generation from missing metadata. This is an adapter repair rather than a tool-budget or family-token arithmetic defect.
## Versioned workspace evidence and diagnostic access
A later controlled consumer exercised actual root/delegate/code tool callers, concrete described contracts, scoped Grain projections, complete-read edit authority, whole-batch stale refusal, a one-shot independently announced feature, exact protected acceptance receipts and currentness invalidation. Reviewer-supplied replacements were explicit test controls; no model authored that repair. A subsequent cold reconstruction and one new acceptance verified chronological latest-result selection after a consumer reducer bug. These native mechanics do not establish fresh-author transfer or all-model semantic quality.
When a diagnostic is previewed, an artifact path alone does not make the omitted evidence available to the agent. Provide complete evidence within admitted context or a described receipt-scoped retrieval capability with exact stream identity, raw-byte hash, decoding and progressing continuation offsets. Keep original bytes and historical receipt/revision identity; retrieving a log does not make an old acceptance current or grant source-edit authority. A final later chunk need not mean one chunk contained the entire stream. Index historical receipts rather than repeatedly embedding full code/log bodies; retain the responsibility's relevant original artifacts separately.
Artifact capture must not bypass tool-process cleanup. One actual confined-worker fault probe reproduced a save error leaving its timed-out child alive; independent cleanup in a nested finally killed/joined the child while preserving the capture failure. The native root proof then closed all actual Futures/bodies/processes before its store/cache. This supports those bounded mechanisms, not arbitrary-code security or filesystem crash-atomic effects. No per-model parser or semantic repair helper was added.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |