Repository navigation
feat: add fail-closed cross-client capability negotiation - #7
Conversation
Verification evidenceLatest head:
The result is static capability evidence only; it deliberately does not claim live host enforcement. |
Competitive-review validationLatest head The design adapts the fail-closed capability-matrix pattern observed across AgentSync, wshobson/agents, and anywhere-agents, while preserving ORRERY's stricter exact-model/effort/isolation semantics. No third-party code was copied. Static capability evidence remains explicitly distinct from live-host proof. Per the requested workflow, this PR remains open and must not be merged. |
76f60a9 to
d1e86c9
Compare
|
Merge-order note — conflicts with #8. Both this PR and #8 rewrite the same The dangerous resolution is the quiet one: keep one test file, drop the other, and CI stays green because the missing tests simply never run. Whichever merges second needs its Ping me and I'll rebase the second one. |
d1e86c9 to
27d1fca
Compare
|
Conflict resolved — rebased onto The conflict was the one flagged above: both this PR and #8 rewrote the same "plan": "bun tools/capability-planner.ts",
"tools:review": "bun tools/tool-surface-review.ts --baseline tools/schema/tools.policy.json",
"test": "bun test ... tools/capability-planner.test.ts tools/tool-surface-review.test.ts"Verified the resolution actually holds rather than just compiling —
|
Problem Orrery gained three read-only diagnostics in three separate pull requests — the capability planner (#7), the tool-surface review (#8), and the semantic compatibility doctor (#9). Each arrived with a sentence appended wherever its own change happened to touch the README, so two of them ended up as orphan paragraphs below the development gate and the third existed only as a comment inside a code block. Read individually they look like three unrelated utilities. They are not: they are the same capability at three different moments — before you choose a client, after you install, and whenever the tool surface moves. That framing was nowhere on the page, so the most checkable thing about this project was also the least visible. Approach Add "Interrogate it yourself" between "What it will not do" and "When not to use this", so the honesty block reads in one run: what it refuses to do, how you can check that for yourself, and when you should not use it at all. The section is organised by moment rather than by tool, and each entry leads with what the command *refuses* to do, because that is the part with value. A path-only portability matrix would report six clients as working; the planner names the ones that would quietly downgrade advisor isolation and refuses them. The doctor separates what it observed on disk from what it cannot know about runtime behaviour. The review turns "the digest moved" into whether the thing that moved can take a new argument, hold state, or has dropped a `readOnlyHint`. It closes by inverting the usual precedence: if a diagnostic disagrees with the README, the README is wrong. Verification Every documented invocation was run, not copied from a pull request description. `plan` with the documented flags returns `bestFit: codex` and refuses cursor and the remaining clients. `tools:review` returns `changed: false`, `permissionExpansion: false`. `doctor` reports schema 2 with `readOnlyMechanism` and per-role adapter state. `bun run ci` is green. Impact Documentation only.
Problem Four pull requests merged into main without a changelog entry between them: the semantic compatibility doctor (#9), the explainable tool-surface review (#8), the capability planner (#7), and the toolchain bump (#15). The Unreleased section described only the Astra lane, so the three most user-visible additions since 0.6.0 existed in the code and in the README but nowhere in the release record. A changelog that silently omits four merges is worse than no changelog, because it is read as complete. Approach Document all four, then cut Unreleased as 0.7.0 and bring the three declared versions into agreement. Minor rather than patch: the release adds a second native implementation lane and three read-only diagnostics. Nothing is removed and no interface breaks, so it is not a major. Each entry records what the thing refuses to do, since that is the part with value: the planner refuses hosts that would silently weaken requested effort or advisor isolation; the review turns "the digest moved" into whether the change is permission-bearing; the doctor no longer reports a marker-only or weakened adapter as healthy. The toolchain entry records why the bump was applied by hand, so the next person does not reopen the Dependabot question. Verification `bun run ci` is green: verify.sh three-role checks pass, 82 tests across 6 files, validate, sbom (now emitting orrery-0.7.0.cdx.json, 18 packaged files, 0 runtime dependencies), and release-check. `bun run tag:check -- v0.7.0` passes, so the tag and all three manifests agree. Impact The release record now matches what shipped. Ready to tag v0.7.0 once merged.
1 participant
Summary
Adds a read-only, versioned capability planner that evaluates a requested portability contract before configuration or adapter writes:
This PR is independently based on
e8a9604c8ac2b15de76d48080604c9151e76ec1d.Competitive rationale and provenance
Tools such as AgentSync, MCPM, and Vercel Skills make cross-client installation convenient. Large agent catalogs such as wshobson/agents use generated adapters and graceful degradation. Orrery's differentiated gap is deciding before installation whether a target can represent a requested security/effort contract without silent loss.
The implementation is original and derived only from Orrery's existing
CLIENT_PROFILESregistry. No competitor code or data was copied. Host semantics remain linked to primary documentation such as Claude Code subagents and VS Code custom agents.Verification
bun run testEvidence limit
The planner reads the versioned static registry. It does not probe installed host versions and explicitly refuses to present a compatible declaration as proof of the model, effort, or sandbox used at runtime.