956732d4a8
Eval probes wrote fixtures, receipts and evidence to hard-coded /tmp paths and two live A/B tasks literally instructed the model to work in /tmp. They now derive locations from tempfile.gettempdir() / os.tmpdir() (env overrides kept), usage examples use relative output names, and the desktop e2e screenshot dirs, perf scripts and the short-session repro fixture stop naming /tmp. Also adds the explicit encoding= the windows-footgun check wants in the touched files.
Delegation grouping schema receipts
probe.py TREE OUTPUT [--independent] assembles real eager CLI tool definitions
in a fresh, credential-free temporary home with networking forbidden. Run it
in separate interpreters against base and fix with the same arguments.
Requires tiktoken (o200k_base); counts use compact OpenAI function JSON.
This measures schema footprint, not model quality or billed savings.
Base: b2aa855b626ff8688eb34b95c60ee8b6a4af3679.
| Policy | Base delegate tokens | Fixed delegate tokens | Change |
|---|---|---|---|
| Default off | 927 | 826 | -101 |
| Explicit on | 927 | 913 | -14 |
Only delegate_task changed in either same-config comparison. Total eager
CLI schemas were 7586 → 7485 (off), 7586 → 7572 (on). The field remains
available when enabled; its parameter schema is unchanged. Repeated assembly
is byte-stable. The two invariants also check static/previous schema immutability,
legacy task normalization, and grouped/ungrouped delivery partitioning.
The default-off exposure assertion failed on base before implementation.
Real-model N=1 child smoke
Pinned to configured Nous google/gemini-3.7-flash; no provider fallback.
A separate temporary home held only copied Nous authentication and minimal
explicit configuration (no personal skills/memory). One forced parent tool call
using the real default-off registry schema produced a one-task array without
group. The actual delegation handler then built and ran one real AIAgent child,
which completed in one API call with GROUP_GATE_OK. Parent tool bytes were
unchanged across child execution. This is a schema-acceptance/execution smoke,
not a comparative quality evaluation or proof of asynchronous delivery timing.
Before inference, the live catalog was read and an approximately $0.016 budget estimated for two calls (20k input + 1024 output allowance).
| Inference | Input | Output | Reasoning (included in output) | API-reported usage.cost |
|---|---|---|---|---|
| Parent schema call | 796 | 104 | 75 | $0.000987 |
| Child | 708 | 42 | 37 | $0.0006885 |
| Total | 1504 | 146 | 112 | $0.0016755 |
Parent request ID: gen-1788889571-fBLF8eStu3UAevfGtvgD.
The raw provider usage exposes the same values as upstream_inference_cost;
these are API-reported costs, not an independently reconciled Portal invoice.
The child result's local estimator instead reported $0.000551 with
cost_status: estimated; do not substitute that estimate for the raw receipt.
Local full receipt: /tmp/delegation-group-live-receipt.json.