cllama: tolerate model-emitted managed tool aliases without hash suffix

[mostlydev]

Problem

In Tiverton on May 12, 2026, Boulton had claw-wall.search_channel_context and claw-wall.get_channel_messages correctly compiled into .claw-runtime/context/boulton/tools.json; cllama also logged tool_manifest_loaded with tools_count=25 for Boulton.

The provider-facing tool names are safe aliases with hash suffixes, for example:

• claw-wall_search_channel_context_2919442f
• claw-wall_get_channel_messages_c83b6078

When the operator asked Boulton to try those retrieval tools, the model attempted calls to the unsuffixed sanitized names:

• claw-wall_search_channel_context
• claw-wall_get_channel_messages

cllama did not recognize those as managed tools, classified them as runner-native, and returned the tool calls to Hermes. Hermes then printed Unknown tool 'claw-wall_search_channel_context' and sent a local tool-error message back to the model.

This makes correctly compiled managed tools appear "declared but not wired" whenever a model drops the safety hash suffix from a sanitized managed-tool alias.

Evidence

• Remote deployment: ssh tiverton, repo ~/tiverton-house.
• .claw-runtime/context/boulton/tools.json contains both claw-wall managed tools with HTTP execution metadata and bearer auth.
• .claw-runtime/context/boulton/channels-allowlist.json contains channel 1464509330731696213.
• cllama logs around 2026-05-13T01:50Z show tool_manifest_loaded for Boulton with tools_count=25, then no tool_call mediation event.
• Boulton Hermes session session_20260512_214931_2dd825d2.json shows assistant tool calls named claw-wall_search_channel_context and claw-wall_get_channel_messages, followed by Hermes-local errors: Tool ... does not exist. Available tools: clarify, delegate_task, ... send_message, ....

Expected

If a model emits a unique unsuffixed sanitized alias for a managed tool, cllama should resolve it to the corresponding managed tool instead of handing it to the runner. Exact canonical names and exact hashed aliases should keep working. Ambiguous unsuffixed aliases should stay unresolved rather than guessing.

Scope

• Update cllama managed-tool resolution to accept unique sanitized aliases without the hash suffix.
• Add tests proving hashless aliases are mediated and ambiguous hashless aliases are not.
• Keep provider-facing schema names unchanged; this is a tolerant inbound resolver, not a schema rename.
• Do not change runner-native tool semantics or clawdapus release pins in the feature PR.

Related

• claw-wall: busy-channel restart backfill does not guarantee 24h awareness #238 fixed the separate claw-wall 24h backfill gap.
• cllama: guarantee bounded 24h channel awareness independent of cursors and restarts #232 introduced the channel-awareness and retrieval tools.

[mostlydev]

Implementation split after the Tiverton/Boulton repro:

• proxy: accept unique hashless managed tool aliases cllama#15 fixes the runtime resolver: unique hashless aliases like claw-wall_search_channel_context now resolve to the managed tool, ambiguous hashless aliases do not, and fallback use logs managed_tool_hashless_alias:<canonical>.
• claw: show callable managed tool aliases #241 fixes the upstream prompt cause Claude pointed out: generated CLAWDAPUS.md listed only canonical dotted names while the provider schema exposed hashed provider-safe names. The generated guidance now lists the callable provider-safe name first and retains the canonical name as metadata.

This is not a Tiverton pod config issue and not missing cllama registration. Boulton's tools.json, channel allowlist, and cllama tool_manifest_loaded log all showed registration was present. The failure was name alignment/resolution: the model emitted hashless sanitized names, cllama did not classify them as managed tools, and Hermes then rejected them as nonexistent runner-native tools.

Release/deploy implication: #241 can merge independently for better generated guidance. The full tolerant runtime fix requires merging/releasing mostlydev/cllama#15, then updating Clawdapus to consume the released cllama tag/pin in a release-boundary change.

[mostlydev]

Shipped in v0.17.1.

• cllama v0.6.6 — runtime resolver accepts unique hashless managed-tool aliases, logs fallback use as managed_tool_hashless_alias:<canonical>, [channel-tool] result header reinforces the provider-safe schema name on every mediated turn.
• clawdapus v0.17.1 — generated CLAWDAPUS.md lists the provider-safe callable managed-tool name first with the canonical dotted name retained as parenthetical metadata; pins bump to cllama v0.6.6 + infra v0.17.1 so the runtime fix lands end-to-end.

Downstream deployments pick this up on next claw pull + claw up -d.