A sandbox-placed root cannot use its coordination tools, so no director has ever run on the fleet: 13 runs, 0 children, 0 winners

[drewstone]

The measurement

Every retained pursuit record on one operator host, 385 runs with a readable spawn journal, deduplicated across seven discovery-lab checkouts:

root placement	runs	longest	runs ≥ 3 h	children ever spawned	winners
cli (local cli-bridge)	372	397 min	16	up to 132 in one run	many
tangle-sandbox	13	69 min	0	0	0

Eleven of the thirteen sandbox-rooted runs have a terminal record. All eleven settled no-winner with spentTotal.iterations: 0, and not one spawned a child. One burned 7,525,451 input tokens (7,394,880 of them cache reads) in a single 30-minute iteration before settling with nothing.

Long runs exist only in the cli column. What has never worked on the fleet is not a long run; it is a director — a root that recurses.

Two failures, both on the coordination path

Quoted from the roots' own retained artifacts.

1. MCP configured, tools not exposed to the harness (mech-interp-foundations-opencode-fleetroot-20260915a, opencode on the sandbox):

The MCP server is configured but not exposed as functions in my harness — I'll need to reach it via HTTP.

It reached the server over hand-built JSON-RPC and confirmed All coordination verbs are available via JSON-RPC — so the server was healthy and the credential worked. The tools simply were not callable. The rest of the turn went to plumbing:

Parameter is question. Fixing and retrying … Quoting bug — extra brace in my inline JSON … my helper has a payload-building bug. Rewriting it safely with Python constructing the JSON

Each coordination verb cost three to five shell round trips whose output is unstructured text, which is where the 7.5 M-token single turn comes from.

2. Coordination rejects the credential with 401 (mech-interp-foundations-pi-20260914e, -20260915k, pi on the sandbox):

MCP coordination server is down (401) — will retry later.

The coordination MCP endpoint is rejecting all credentials (401), so I'll set up a poller (already running) and meanwhile do the actual owed research work locally so real evidence exists.

Both directors fell back to working alone. Whether these two symptoms share a cause — a coordination bearer materialized into the sandbox harness config but not honoured by the coordinator — or are independent, was not tested; this report is observational over retained records.

Why it stays invisible

The run does not crash. It settles no-winner at zero iterations after spending real money, so a status reader sees "the director settled without a result" and reaches for prompt or budget explanations. Widening the budget cannot help a run that never spent an iteration.

There is also a harness-capabilities skill in three Lab checkouts whose purpose is teaching a director to discover that its tools are missing and hand-roll around them. That is a workaround standing in for a fix, and it is why this has gone eleven runs without a bug report.

Ask

1. A root placed on tangle-sandbox should get the coordination verbs as native harness tools, the same ones a cli-placed root gets. If the materialized MCP config cannot be surfaced as callable functions in that harness, the placement should refuse at admission rather than produce a director that cannot direct.
2. Resolve the 401 on the coordination endpoint from inside a sandbox environment, or report it as an admission failure instead of leaving the root to discover it mid-turn.
3. A root that completes zero metered iterations and spawns zero children is worth a distinct settle reason. no-winner reads as a research outcome; this is an infrastructure refusal.

🤖 Generated with Claude Code