Claw tool-name aliases

Claw accepts the exact spellings Read, Bash, and Grep as read_file, bash, and grep when the bundle declares requires.iterion at or above bundle.ToolAliasesSince — 3.179.0, the release that first ships the resolver. Canonical names continue to work in a plain .bot or on older engines. This is engine-version-dependent behavior in both DSL profiles; it is not a profile-1 lowering that an older reader can reproduce.

The pin cannot rot again (it named 3.144.0, then 3.146.0, and both rotted — 3.144.0, 3.144.1, 3.145.0 and 3.146.x shipped while the resolver sat on this branch): the floor is now an entry of the ONE syntax-floor table (pkg/bundle/profile.go, #1276), and TestSyntaxFloorsNameReleasesThatExist holds it like every parser floor — while uncut, exactly the next minor above the changelog's newest release; once cut, the release's notes must carry the word "alias", or the test reddens with the re-pin instruction. Re-deriving the floor at merge is therefore not a memory task; run devbox run -- go test ./pkg/bundle/ -run TestSyntaxFloorsNameReleasesThatExist and do what it says. The release cut then writes the number itself: release-it runs cmd/release-floors --apply before it stages the release commit (internal/floorsalign), which rewrites a pin still at the next minor to the version being released and refuses the release when the notes it just rendered do not carry the word — the same rule as the test, run once more at the one moment the number is known.

The manifest opt-in

# manifest.yaml
requires:
  iterion: ">= 3.179.0"

Declaring the floor is what turns the resolver on; the same declaration is what the floor predicate at push admission, validate's C252, dsl migrate and the scaffold ask for when a bundle's sources spell an alias in tools:, tool_policy:, a tool node's command: or recovery: agent_tools: — so an author is told the floor is missing before the bundle is written to a runner that cannot serve it.

agent inspect:
  backend: claw
  model: "openai/gpt-5.6-sol"
  tools: [Read, Bash, Grep]
  tool_policy: [Read, Bash, Grep]

The entire resolution order matters:

1. A registered exact name wins.
2. Qualified MCP names (mcp.server.Read, mcp__server__Read) retain their existing meaning. They are never builtin aliases.
3. A bare name matching one MCP tool resolves to that tool. Matching two MCP tools is an ambiguity error, even if a builtin alias has the same spelling.
4. Only a remaining unresolved exact alias uses its canonical builtin.

The compiler leaves the source names intact. It cannot see the project .mcp.json and plugin servers that are connected after compilation. C135 therefore warns with the canonical spelling and manifest remedy; the runtime resolves against the complete connected catalog. Without the declared floor it keeps the old resolver and refuses an alias that would otherwise select a builtin before calling a model. A unique MCP Read still works without an alias opt-in, just as before.

This applies to agent/judge tools:, the same list when a fallback route selects Claw, and Verified Action recovery.agent_tools. Workflow/node tool_policy patterns use the same registry precedence when aliases are enabled: a policy Read selecting an MCP grants that MCP, never read_file. An ambiguous alias in a matched policy denies the call, including beside *. The model receives one canonical tool schema even if both alias and canonical names were listed.

capabilities: contains host rights such as board.read and runs.read, not tool names. capabilities: [Read] remains malformed (C081); aliases do not grant host rights. CLI backends retain their existing tool and permission semantics.

Each engine derives this opt-in from its own bundle on Run, Resume and branch dispatch. A child without a sufficient floor does not inherit an opted-in parent's setting. The existing bundle admission gates reject an older engine; the queue schema and AST retain their original shape. Rollback therefore requires reverting the alias-using bundle as well, or the older image will correctly refuse it.

Verification

The local regression test TestToolAliasesThroughRealEngineAndClawTools compiles a workflow, runs the real engine and Claw executor, and calls real file, search and shell tools against a temporary proof file. Only the provider's choice of calls is scripted. Missing/insufficient floors produce zero provider calls. Separate tests cover zero/one/two MCP suffixes, exact builtin priority, qualified names, policy isolation and duplicate schemas. The same real tools are exercised through a Claw fallback, a rung-4 recovery agent satisfying a filesystem postcondition, and a fresh engine resuming a failed run.

On 2026-09-14 the compatibility probe passed with a real new SubmitLaunch producer and the unchanged 3.143.0 runner at 882c76afbb: the old reader accepted queue schema 17, the serialized AST and the immutable bundle snapshot emitted by the producer (no fallback bundle store), then executeRun refused the bundle as failed/BOT_REQUIRES_NEWER_ENGINE before executing tools or models. This is evidence of refusal under version skew, not an assertion that release 3.144.0 already contains the feature.