Agent affordances: the harness gives the model what it needs to act well (six gaps)

[Hmbown]

Ratified 2026-09-23 (CURRENT_DECISIONS §22, private ops). Source: a review of Codewhale from the model's side. Each claim was checked against origin/main 3b19044ea. Every place the harness is ambiguous, silent or lossy turns into a wrong turn, a wasted round trip, or a claim the model can't verify.

The six gaps, one sub-issue each:

1. Decisions: a non-blocking "decisions waiting on you" list, on the same "needs you" surface as approvals (Approvals program: never lost, never silently decided (child approvals, Full Access, auto-review, grants) #6467).
2. Spend and context: the model can see both, and gets a notice before auto-compaction.
3. Deferred tools: the first well-formed call to a deferred tool runs. This is PR feat(tools): run well-formed first calls to deferred tools #6437, already in the 0.10.1 queue.
4. Scratch and leases: a per-session scratch directory named to the model, plus a working-tree lease.
5. Child receipts: every child result carries a machine-built receipt.
6. Loud spawn failures: spawning with an empty or unknown effective toolset fails loudly.

Also in scope: the runtime_handoff.rs:55-58 wording points the model at a "tool details view" it cannot open. Fix it to point at retrieve_tool_result.

Already true and to be kept:

• completion notices wake an idle parent
• child budget disclosure and hand-back
• provenance labels on runtime events
• the read-before-edit guard
• output spill to a file