Constrain controlled CLI sessions and prove native budget overshoot

Assistant: codex
Assistant-Model: gpt-5.6-luna
Assistant-Session: 01a07ff8-19d0-7820-b4d0-1353833cb7fc

docs/native-cli-boundary-proof.md

@@ -0,0 +1,84 @@
+# Native CLI boundary proof — 2026-09-09
+
+The installed Claude Code 2.1.266 binary was exercised against a deterministic
+Anthropic-shaped API in a separate bwrap network namespace. The namespace had
+no provider network route, operator home, real credential or inference service.
+Only the CLI binary, system runtime and a temporary fixture workspace were mounted.
+The executable receipt records the binary SHA-256 and exact adapter-generated argv.
+
+## Findings and resulting change
+
+With `--max-budget-usd 0.01`, the first fixture response reported 60,000 input
+and 10 output tokens. The CLI estimated **USD 0.18015**, then exited with
+`error_max_budget_usd`. Only one messages request reached the fixture. This
+reproduces a stop after the response, not a pre-request maximum-liability check.
+The rein terminal-accounting validator refused success and preserved known cost.
+No amount in this experiment was billed by a provider.
+
+The existing spend ledger remains necessary: it reserves declared per-run
+liability and retains unknown outcomes. The experiment does not establish that
+the declared liability is a valid bound on every possible native request. A
+provider-facing request guard is the next HFACT-WP-0001-T01 implementation return;
+raising the native threshold or accepting a newer binary cannot substitute for it.
+
+The adapter also previously used `--allowedTools` with `acceptEdits`. Those flags
+alone did not constrain the available tool inventory or repository-provided
+customizations. Runs with native USD or turn controls now require CLI >=2.1.266
+and use `dontAsk`, explicit builtin tools derived from registered rules, `--bare`,
+empty settings sources, empty strict MCP configuration, an MCP deny rule, and no
+session persistence. Legacy unconfigured sessions retain their existing behavior.
+The worker prompt explicitly requires reading applicable AGENTS.md/CLAUDE.md;
+these files are instructions, not permission to load executable configuration or
+expand grants. The factory's protected runtime must be rebuilt and re-proven for
+the new minimum; this source change installs or admits no runtime.
+
+The real binary's bare mode exposed Bash, Read and Edit, a subset of the explicit
+inventory. A synthetic tool sequence successfully ran permitted `git status` and
+created a new file through Edit. An unapproved shell write was denied. A fixture
+project containing a broad Bash permission override, a SessionStart hook and an
+MCP command could activate none of them; its ambient CLAUDE.md was not auto-loaded.
+The second messages request contained the expected allowed/denied tool results.
+
+These tests prove the specified direct-tool and configuration boundaries. They
+are not an arbitrary-code sandbox proof: permitted Git operations can invoke
+repository-controlled helpers, and file edits can influence later commands.
+Filesystem, credential and network confinement remain sandbox-owner controls.
+To enforce spend against such indirect execution, the workload must not possess
+a provider credential plus an unmetered route to the provider. A run-scoped,
+owner-controlled request path must cover every forwarded request and retry.
+
+## Reproduction and acceptance
+
+`REIN_REAL_CLAUDE=1` enables `tests/test_native_cli_boundary.py`. The default binary
+comes from `claude` on PATH; `REIN_CLAUDE_PROOF_BINARY` can select an exact installed
+file. The fixture deliberately requires version 2.1.266; revalidate and record a
+new receipt before claiming the same proof for another version. The adapter's
+minimum-version gate does not itself prove arbitrary future binaries.
+
+```text
+REIN_REAL_CLAUDE=1 REIN_REAL_BWRAP=1 python -m pytest tests/ -q
+```
+
+The reproducible fixture is `tests/native_cli_fixture.py`; optional
+`REIN_CLAUDE_PROOF_OUTPUT` writes bounded evidence without prompts, headers,
+credentials or response bodies. The actual CLI checks are opt-in because ordinary
+CI does not provision that binary. Deterministic native launch/accounting tests
+are now included in both mandatory runtime and recovery gates. Paired receipts:
+`docs/evidence/2026-09-09-native-cli-overrun.json` and
+`docs/evidence/2026-09-09-native-cli-tools.json`.
+
+No paid factory request, profile promotion, credential reader or production
+rollout was authorized or performed. REINAH-WP-0003-T05/T06 and HFACT-WP-0001-T01/
+T03/T04/T05 retain the request-admission, exact identity, protected placement and
+natural queue/model execution returns.
+
+## Upstream contract reviewed
+
+The [CLI reference](https://code.claude.com/docs/en/cli-reference) distinguishes
+available tools from auto-permission rules and describes the native stop flag.
+The [headless guide](https://code.claude.com/docs/en/headless) documents bare-mode
+configuration isolation and the API-key authentication path. The
+[cost guide](https://code.claude.com/docs/en/costs) identifies CLI totals as local
+estimates, not authoritative billing, and notes the residency-cost correction in
+2.1.239. Reviewed 2026-09-09; the local executable findings above are narrower
+than a claim about all provider billing or future CLI releases.