agent-joule/dev-docs/FINDINGS.md

Findings — gap remediation (2026-06-14 audit → fixes)

Status of the audit's actionable gaps. Most are addressed; the rest are deferred with reasons. All fixes are committed with 76 panel tests green and the full agent-mei binary building.

P1 — undercuts the core "signed / accountable" claims

1. ✅ Op author stamped. Every CodingOp now carries the session identity's verifying key as its author (build_coding_op gained an author param; record_multi_edit/captured_op stamp session_key.verifying_key()), replacing author: [0u8;32]. Ops are signed by a principal, not anonymous.
2. ✅ L4 closes seal deterministically. The thread-event handler now matches AcpThreadEvent::Stopped(_) (turn completion) and seals once, with a last_l4_op_id de-dup guard — no partial/duplicate ops, no lost close. An L4 turn that edits files reliably produces exactly one ledger op + L0 entry, like L0–L3.
3. ✅ let _ = discards fixed. adopt_from_file, cite_op, spec/ops persistence writes, record_primary_op, session set_status/mark_verified, and the release.json write now .log_err() per the repo's CLAUDE.md rule.

P2 — real semantic limits

4. ✅ L0 keyed on buffer precondition. Every request captures the active buffer's content-id (pending_prior); the L0 lookup and every close's cache insert use it, and LedgerEntry carries a serde(default) precond for cross-session correctness. A cached edit replays only when the same intent meets the same buffer state. New test: l0_precondition_differentiates_buffer_state.
5. ✅ Session grant enforced. When a background session is active, authorize_tier checks the session's (attenuated) grant names the tier and the session is under its own joule ceiling, before the global gate. A delegate that holds only {l0,l1,l2} can no longer reach L3/L4 — delegate-only-what-you-hold, enforced not just declared.
6. ✅ Ship door — already resolved in the session arc: certify_release uses the live frame-backed contract_discharged() = !frame.is_empty() && contract_ok, no hardcoded false.
7. ✅ PostTool / SessionEnd hooks fired. note_close fires PostTool on every tier close; close_session fires SessionEnd before teardown. Observe-only (no-op when no hook registered) — the points are now live and extensible.
8. ✅ "No energy counter" distinct from 0 J. Render shows ⚠ no host energy counter — figures are estimates when absent; control state reports energy_counter: bool. (L4 op energy remains token-estimated, as honestly tagged.)

P3 — hygiene

9. ✅ JOULE_CODE.md corrected — the stale "path dependencies" description now reflects the actual pinned git deps on Transaction-Science/open-standards.
10. ✅ Gutting warnings cleared — removed 8 unused imports (main.rs, zed.rs, remote_connection.rs), _-prefixed two unused vars, and #[allow(dead_code)]'d the kept-for-compat title_bar.show_sign_in. Binary builds with no agent-mei warnings.
11. ⏸ visual-tests feature left as-is — its visual_test_runner.rs is deeply tied to the removed agent_ui/AgentPanel (a test harness for the deleted native agent). Repair = large rewrite, removal = build-config change; dev-only, off the shipped path. Documented, not bulldozed.
12. ✅ Namespace test fixed — test_action_namespaces was actually red: it lacked the new agent_mei namespace and still listed skill_creator + inline_assistant (removed crates). Added agent_mei, dropped the two dead ones; test green. (agents_sidebar stays — still a live zed_actions namespace.)
13. ⏸ Two near-identical MCP stdio clients — optional refactor, not a correctness issue.
14. ⏸ joulecontract-schema near-vestigial — revisit when per-op contract discharge lands.

CI blind spot (model layer) — ✅ killed

The on-device-model plumbing is now CI-tested with no download and no #[ignore]: two tests build a complete tiny synthetic llama (random weights, embedded vocab) via jouleclaw's public synthetic::synthesize_llama_gguf, write it to a temp .gguf, and load it through the real load_l3_model path —

• synthetic_model_loads_generates_and_embeds: l3_generate runs end to end and is deterministic; l3_embed returns a 16-dim L2-normalized vector.
• synthetic_model_powers_codebase_index: builds a real index over a small repo with the synthetic embedder and gets ranked search hits.

These assert the pipeline runs (load → generate → embed → retrieve), not edit quality (a 16-dim random model writes nonsense). The five #[ignore]'d real-model tests remain as the quality probes (manual, JOULE_CODE_TEST_MODEL). 78 panel tests green.

---

Net: all P1 and all actionable P2 items are fixed and tested. The deferred items are P3 hygiene in upstream/David-owned files, optional refactors, or CI-infra — none affect the correctness of the accountability story, which is the audit's core concern.