fix: preserve analyst outcomes and stop authority in Jev-compatible execution

[drewstone]

Changes

• Reuse the existing Eval registry → analystsFromRegistry → graph/supervisor path. No new runtime, agent class, scheduler, or hook bus.
• Failed, skipped, or missing analyst executions no longer masquerade as empty successful reviews. Successful reviews with zero findings remain valid.
• Recheck the existing stop predicate after awaited preparation and compaction. A classifier, steer wait, or compaction that consumes the remaining budget/deadline cannot trigger another inference call.
• Preserve completed turns, usage, tool results, and preparation/compaction state for continuation when execution stops.
• Add regressions through the real tool loop. Remove the wrapper-only cancellation test; retain actual graph and compaction composition coverage.
• Document caller-owned questions/policies and the actual local-versus-external-harness hook coverage.

Verification

Full CI passed for head cb194c39511f3f14013bc8374c816f936c320da2:
https://github.com/tangle-network/agent-runtime/actions/runs/35394548837

Passed: lint, full tests, build, source/examples typecheck, packed exports, docs freshness, exact first-party package cohort, official optimizer integrations, and agent-bench.

The first benchmark attempt failed a temporary-checkout fixture with ENOENT under /tmp/swe-cache-copy-*/destination/.git/objects. A rerun passed without weakening assertions or changing benchmark behavior. The intermittent filesystem cause remains unproven; this is not presented as a benchmark bug fix.

Composition

Use jevEvaluator or an ordinary typed evaluator where a bounded decision is useful. Use Eval's jevAnalyst/asAnalyst for graph analysis. Retain registry summaries through existing observation/persistence hooks for diagnostics and spend attribution; function-shaped analysts still need the shared paid-call account.

Related: agent-eval #761; tangle-router #533 (native endpoint #531 is merged); agent-dev-container #7620. Production consumers need the released Eval adapter. No publish, merge, deployment, or funded live-model test performed.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[tangletools]

✅ Auto-approved drewstone PR — 149c52e1

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

This approval is provisional. It rests on the audit running. If the audit cannot run — for example the CLI bridge rejects it — this approval is dismissed rather than left standing, so an unrun check never reads as a passing one.

_{tangletools · auto-approval · reason: drewstone_author · 2026-09-18T19:06:06Z}

[tangletools]

🟡 Value Audit — sound-with-nits

Verdict	sound-with-nits
Coverage	2 of 2 lenses (value, usefulness)
Concerns	2 (2 weak-concern)
Heuristic	0.0s
Duplication	0.0s
Interrogation	195.0s (2 bridge agents)
Total	195.0s

💰 Value — sound-with-nits

Adds a passing deterministic regression suite plus a guidance doc proving 'Jev on existing seams' — registry analysts through the real graph path, beforeTurn→compaction→inference ordering, and caller-owned brain composition — with no runtime changes; good and in-grain, with one tautological test.

• What it does: Adds tests/kernel/jev-composition.test.ts (3 tests, all verified passing locally via pnpm exec vitest run) and docs/jev.md. Test 1 registers a fixture analyst on a real agent-eval AnalystRegistry, adapts it via analystsFromRegistry (src/runtime/supervise-surface.ts:126), and runs it through the actual runGraphWithTestBrain analyzes edge of the shotLoop example, asserting the shared CostLedger/co
• Goals it achieves: Makes the 'native decisions compose on existing seams, not a new runtime' policy executable and regression-guarded: the eval-registry→graph→ledger path, cost-ledger identity propagation, beforeTurn/compaction ordering, and cancellation/identity preservation through an explicit brain. It also documents the integration contract for the related agent-eval/router/dev-container changes (billing limits,
• Assessment: Good on its merits. I ran the suite: 3/3 pass against the installed @tangle-network/agent-eval 0.182.0 (which does export /analyst and carries costLedger/costPhase in its context types). The coverage fills real gaps, not duplicates: beforeTurn had zero test coverage anywhere (only production use at src/runtime/supervise/coordination-driver.ts:1253); the registry→graph composition is tested elsewhe
• Better / existing approach: none — this is the right approach. Searched for prior art: analystsFromRegistry consumers (only supervise-surface.test.ts lens-level tests and a comment in supervise-option-capture.test.ts), analyzes-edge coverage (graph.test.ts, graph-topologies.test.ts — inline registries only), beforeTurn usage (src-only), compaction tests (tool-loop-compaction.test.ts covers mechanics but not hook ordering). E
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 2
• Bridge warning: opencode/kimi-for-coding/k2p7: opencode: opencode error event without a message: {"type":"error","timestamp":1789761906509,"sessionID":"ses_f49e09a0bffe6BTdSUbRa3ltee","error":{"name":"UnknownError","data":{"message":"Unexpected server error. Check server logs for details.","ref":"err_c93cd92e"}}}

🎯 Usefulness — sound-with-nits

A passing, genuinely additive composition regression test that routes the real agent-eval AnalystRegistry through analystsFromRegistry into runGraph's analyzes edge with cost-ledger context assertions, plus first-ever beforeTurn→compaction ordering coverage; one of its three tests is self-referentia

• Integration: Verified reachable and real, not dead surface. Every import resolves to a live seam: AnalystRegistry/CostLedger from pinned @tangle-network/agent-eval@0.182.0 (package.json:134), analystsFromRegistry (src/runtime/supervise-surface.ts:126), runGraph's analysts option and analyzes edge ledger (src/runtime/supervise/graph.ts:442,150-152), runBrainLoop hooks.beforeTurn (src/runtime/tool-loop.ts:208),
• Fit with existing patterns: Follows the established tests/kernel pattern exactly (scripted brains, deterministic fixtures, ledger assertions — same style as tests/kernel/graph.test.ts:590-636 and tests/kernel/tool-loop-compaction.test.ts). It does not compete with or duplicate existing coverage: graph.test.ts tests analyzes edges with a hand-rolled lens (graph.test.ts:591-594), supervise-surface.test.ts tests the adapter in
• Real-world viability: Deterministic, no network, fast (34ms of test time), so it will hold in CI. The first test fails meaningfully if the adapter, edge naming, or context propagation regresses (analystsFromRegistry throws on unregistered kinds at supervise-surface.ts:137-141, so it can't silently pass). The second test pins real runBrainLoop ordering. The third does not: see finding. Error paths like abort mid-graph a
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 1

💰 Value Audit

🟡 Third test is tautological — it exercises no runtime code [maintenance] ``

tests/kernel/jev-composition.test.ts:126-146 defines both transport and brain inside the test and calls them directly; runBrainLoop is never invoked in that test, so any regression in src/runtime/tool-loop.ts leaves it green. It only demonstrates JS semantics (frozen-object passthrough, throwIfAborted after abort). It works as an executable mirror of the docs/jev.md:113-122 brain snippet, but as a 'composition proof' it pins nothing about the system. Either route the wrapper through a real

🎯 Usefulness Audit

🟡 Third test is self-referential: it invokes only its own closures, no runtime code [problem-fit] ``

tests/kernel/jev-composition.test.ts:126-146 defines brain and transport locally and calls brain([], [], context) directly; nothing from src/ executes except AbortController semantics. It asserts that the test's own closure forwards a frozen argument and that throwIfAborted throws after abort — true regardless of runtime behavior, so it passes even if the runtime regressed caller-owned cancellation. The real contract (frozen ToolLoopCallContext built at src/runtime/supervise/coordination-d

---

What this audit checks

It judges the change on its merits — not whether it was tasked out in an issue. Unticketed, fast-moving work is fine; the question is whether the change is good and whether a better or existing approach should be used instead.

Pass	What it asks
Heuristic	Vague title? Whitespace-only or cruft-bearing diff? (content signals only)
Duplication	Do added function/class names already exist elsewhere in the repo?
Value Audit	What does it do? What goal does it achieve? Is it good? Better architecture or already-exists?
Usefulness Audit	Does it integrate and fit? Will it hold up in real use and actually get used?

Findings are concerns, not blocks — the human reviewer decides what to do with them.

_{value-audit · 20260918T200819Z}

[tangletools]

⚠️ Review Incomplete — 149c52e1

At least one required reviewer lane failed closed. No approval or request-changes review was published. This is a reviewer run failure, not a PR quality score.

Trigger a fresh review on the current PR head.

_{tangletools · 2026-09-18T20:16:54Z}

[tangletools]

✅ Auto-approved drewstone PR — 149c52e1

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

This approval is provisional. It rests on the audit running. If the audit cannot run — for example the CLI bridge rejects it — this approval is dismissed rather than left standing, so an unrun check never reads as a passing one.

_{tangletools · auto-approval · reason: drewstone_author · 2026-09-18T20:27:13Z}

[tangletools]

✅ Auto-approved drewstone PR — cb194c39

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

This approval is provisional. It rests on the audit running. If the audit cannot run — for example the CLI bridge rejects it — this approval is dismissed rather than left standing, so an unrun check never reads as a passing one.

_{tangletools · auto-approval · reason: drewstone_author · 2026-09-18T21:09:42Z}