Tasks: pre-trust the workspace before spawning a harness session

Phase 3 of 3. Derived from requirements.md and design.md. Each task names the requirements it satisfies.

DAG

Tasks

• [x] T1 — trust.py: TrustConfig, TrustResult, safe JSON update. New module with the two dataclasses (from_mapping, merge) and the private _update_json helper: missing→{}, unparseable→error-without-writing, no-change→no-write, temp-file + os.replace, 0600 on creation, per-path threading.Lock. (R3.2, R3.3, R3.4, R1.5)

• [x] T2 — trust.py: ClaudeTrustStore. Path resolution (CLAUDE_CONFIG_DIR, .config.json preference, settings file), project_keys() (exact dir + realpath, never a parent), trust() and accept_bypass_permissions() (settings key + legacy config key). (R1.1, R1.2, R1.3, R2.2, R3.7)

• [x] T3 — adapter hook.HarnessAdapter.__init__(..., trust=None) + prepare_environment(cwd) returning a no-op TrustResult; ClaudeCodeAdapter.prepare_environment delegating to ClaudeTrustStore, with _wants_bypass() honouring auto | always | never and recognising both bypass spellings; build_adapters(harness_args, trust=None). (R2.1, R2.3, R2.4, R3.1, R3.6)

• [x] T4 — dispatcher wiring.RoutingConfig.harness_trust (+ from_mapping), pass it through build_adapters in __init__ and reload, and call the new private _prepare_environment(adapter, work_item, cwd) from _spawn_for (before either runner) and from _respawn_tmux. Never raises, never changes the dispatch outcome. (R1.1, R1.4, R3.1, R3.4)

• [x] T5 — event types.workspace.trusted / workspace.trust_failed in eventlog.EVENT_TYPES. (R3.5)

• [x] T6 — unit tests (cli/tests/test_trust.py). Every bullet in design §7's unit list, including the trust-creep guard and the "no bypass flag ⇒ settings untouched" assertion. Fake HOME via monkeypatch; no test may touch a real ~/.claude.json. (all)

• [x] T7 — integration tests (cli/tests/test_trust_integration.py). The four Gherkin scenarios from design §7, including the ordering assertion (prepared before spawn) and the failure-does-not-fail-dispatch scenario. (R1.1, R1.4, R3.1, R3.4, R3.5)

• [x] T8 — config surface.routing.harnessTrust in .the-loop/cli-config.schema.json and the dogfooded block in both .the-loop/cli-config.yaml and the scaffolding template skills/the-loop/templates/cli-config.yaml, commented with why it exists. (R2.4, R3.1)

• [x] T9 — documentation.cli/README.md, skills/the-loop/reference/observability.md (event types), skills/the-loop/reference/automation.md, and the capability docs (interactive-sessions.md, webhook-triggers.md) with history rows — same-PR rule.

• [x] T10 — docs/decisions/decision-037.md (+ index row): pre-seed the harness's own config before spawning, with the alternatives from design §8 and the residual concurrent-write risk.

• [x] T11 — gates. ruff, pyright, pytest, markdownlint from the repo root; execution log updated.

• [x] T12 — PR + reviewer briefing, explicitly requesting the tier-4 named human security sign-off (security.review.humanSignOffMinTier: 4).

Review follow-ups (added after PR #92 feedback)

• [x] T13 — rebase onto #91 (issue-89). Resolve the _respawn_tmux conflict by placing _prepare_environment above _try_resume, so it precedes every harness start rather than only the fresh-spawn fallback; strengthen the respawn assertion to cover all recorded starts. (R1.4)

• [x] T14 — configurable trust scope, defaulting to the workspace root (owner decision on PR #92). harnessTrust.scope: workspace-root | directory; ClaudeTrustStore.trust(cwd, root) writing the trust key on the root and the onboarding key per directory; is_within / is_too_broad guard rails and Dispatcher._trust_root(); schema + both yamls + README + capability docs + decision-037's "Scope: the owner's call" section; 10 new tests. (R1a.1–R1a.6)