Make the agent integrations actually work, and correct the OpenClaw claim

[samrusani]

You asked me to test the OpenClaw skill. Three things came out of it, each verified by running software rather than reading about it.

1. I was wrong about OpenClaw, and the claim was already public

I previously concluded OpenClaw does not support MCP and that a plugin would be required. That came from reading docs.openclaw.ai and never opening their source, which was on the machine the whole time.

$ openclaw mcp probe
- alice: 11 tools
$ openclaw mcp doctor
- alice: ok

OpenClaw 2026.8.1 ships @modelcontextprotocol/sdk 1.30.0 as a production dependency with a stdio transport. One config entry under mcp.servers, no plugin, zero diagnostics. Tools arrive prefixed, so alice_recall is exposed as alice__alice_recall.

The false claim had reached the wiki, the landing page copy, and the announcement drafts, where it sat inside the section arguing for our own candour. All corrected.

2. Neither skill pack was loading, on either host

This is the real defect and it hit Hermes too, which is the platform being promoted on. Three variants of the same pack, side by side in OpenClaw's managed skills directory:

DROPPED  alice-project-memory-skill.md   (what we shipped since July)
DROPPED  SKILL.md, no frontmatter
LOADED   SKILL.md + frontmatter          modelVisible=true

Silent at every log level: a file not named SKILL.md is never a candidate, and a file without frontmatter parses to an empty record with zero issues. Hermes uses the same convention. Both packs were inert from the day they were written.

Both are now <skill-name>/SKILL.md with frontmatter, moved with git mv. Copied straight out of the repo into OpenClaw with no edits, both load with eligible=true.

3. Every documented payload was rejected by the server

Running the skill's own loop against the published v0.15.4 artifact:

MCPToolError: tool 'alice_memory_commit' does not accept additional properties: intent

One cause: the packs name the core eleven-tool surface but carry payloads shaped for the legacy alice_vnext_* tools. intent: "explicit_remember" exists only on alice_vnext_commit_memory, and the core tool rejects unknown properties rather than ignoring them.

defect	files
intent on a core commit	4
sprint output sends content/output_type/propose_memory to a core tool, omits required raw_text	2
proposal examples omit required title	6 blocks
proposal_type is not a shipped field	4 blocks
all 11 context-pack recipes nest scope/options; the core tool takes them flat	11 blocks
lifecycle sent to alice_vnext_confirm_memory and /vnext queues, neither on a default server	3
"create open loops with alice_open_loops", whose actions are list/close/edit/reopen/snooze	1
two /ABSOLUTE/PATH/TO/AliceBot paths the rename missed	2

4. Our release gate was blind to a hole the landing page split would open

Found while answering a separate question about putting a public page on the same host. Every check in validate_caddyfile is a file-wide substring test, so it proved some block demands a client certificate, not that Alice's block does. Reproduced against the shipped validator: strip the whole tls/client_auth stanza from the Alice block, leave those strings in any other block, and all four mTLS checks pass. CI goes green on a configuration serving the memory API to anyone.

Harmless while the example had one site block. The moment a second one exists it is live, and a hostname split for a landing page is a second block.

Now scoped per block. A public block is still allowed, and there is a test for it, as long as it cannot reach Alice.

Guards

Three new files, each proven non-vacuous by reverting the fix:

guard	reverting fails
test_agent_facing_payloads_are_callable.py	33 of 106
test_skill_packs_are_loadable.py	4 of 11
the two Caddy block-scoping mutations	both, against the old validator

The payload guard runs each documented payload through _validate_mcp_arguments_against_advertised_schema, the same function the server calls before any handler runs, against default-surface tools only. It caught two defects I had missed while writing it.

End to end

Payloads parsed live out of the shipped markdown rather than retyped, against the published artifact:

[OK] hermes ambient commit    status=confirmation_required
[OK] hermes explicit commit   status=committed
[OK] hermes low-confidence    status=review_required
[OK] openclaw sprint capture  status=imported
[OK] openclaw project commit  status=committed
--- new session, same data dir ---
[OK] recall count=1, count=1 · context_pack memories=1

Every outcome the docs promise now occurs, including the graded ones.

Full unit suite: 4682 passed, 2 skipped.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.