Skip to content

ObneyAI/orc / orc-skills/orc-tools

orc-tools

Verified

Build 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.

references/bounded-workspace-acceptance.md

View on GitHub
# Bounded workspace acceptance and read authority

Apply these distinctions when the agent can edit the implementation exercised by its tools:

- Keep acceptance assertions in the trusted owner, outside the editable code's interpreter. Distinguish failed assertions from premature child exit or invalid transport.
- Audit parent artifact collection independently of child confinement: rooted no-follow traversal, explicit rejection of links/special files and bounded reads also apply to manifests and hashing.
- Separate complete source-read authority from edit concurrency. An unchanged file's complete observed bytes remain usable after an unrelated file changes. The proposed edit still needs the current workspace revision and target hash; a changed target needs a matching full read. Validate the entire batch before mutation.
- Verify the actual final protected acceptance and exact owned snapshot/native revision. Schema-valid output and successful transport do not establish business acceptance.

## Evidence and limits

The retained-read root verification (historical local experiment; private archive not bundled) reproduced the old false conflict and independently exercised seven actual host calls: retained A after unrelated B changed was accepted; changed A without a matching read, stale global edits and mixed-coverage batches were refused. This was focused host/capability evidence, not a new full ORC end-to-end run.

The host root verification (historical local experiment; private archive not bundled) preserves the parent scratch-manifest canary failure and corrected confinement proof. The hosted preparation review (historical local experiment; private archive not bundled) records source-freeze and zero-call preparation separately from live completion.

The closed workspace baseline in workspace outcome report (historical local experiment; private archive not bundled) used 38 physical Luna calls and 226 generated fields. Actual protected acceptance passed 11/11; only three permitted implementation files changed. Root matched the owned final snapshot to the native final revision and qualified owned runtime/host shutdown. No custom model parser was added. The model corrected its initially mis-sequenced probe; the original observations remain evidence.

The independent capture audit (historical local experiment; private archive not bundled) found request/wire/canonical-write correspondence for all 38 calls. Registered and wire schema descriptions were not literally identical; assertions matched. A separate native `llm/predict` return artifact was not captured: canonical writes are the published parsed boundary. This audit alone does not establish semantic acceptance.

These are one bounded fixture/model baseline and targeted controls. They do not establish arbitrary-code security, large-codebase performance, all-model reliability, crash-atomic filesystem edits or app crash recovery. An explicit continuation is a separate experiment; no continuation outcome is inferred here.

## External tool receipts during fresh transfer

Validate real nonempty tool receipts as well as generated arguments. If an external JSON contract declares a typed dictionary with `additionalProperties`, preserving it as a described Malli `:map-of` is different from forcing a closed record with zero named fields. A consumer converter that does the latter can reject a successful confined execution after it has already produced effects. Do not treat that rejection as malformed model output or repeat the effect blindly.

The fresh-author transfer result (historical local experiment; private archive not bundled) preserves an actual failed workspace attempt caused by this conversion. Its minimal generic dictionary correction then passed real confined scratch-output/canonical-readback verification without a paid rerun. That closes the boundary reproduction, not the unfinished workspace repair. The broader experiment's earlier11/11 repair results remain separate evidence. No per-model input/output helper was added.
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: 115kUpdated Claude CodeCodexVerifiedInstall
add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation.Stars: 115kUpdated Claude CodeCodexVerifiedInstall
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: 115kUpdated Claude CodeCodexVerifiedInstall
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: 115kUpdated Claude CodeCodexVerifiedInstall