[khanukov]

Summary

Add only the accessible claude-fable-5-2 model through the existing three-flow campaign machinery.

• Finite anthropic-fable52-same44-v1 profile; same 44 cases, existing admission, accounting and replay.
• Fresh stateless full-current-Core requests, adaptive thinking at low effort, automatic tool choice with parallel tool use disabled. No thinking replay, prompt trimming or compaction.
• Published standard/global tariff envelope and 1,000,000-context / 128,000-output bounds. Reported-usage estimates are not invoice guarantees.
• Reuse existing optional-money execution with fresh source-bound consent and one-shot credential release. Preserve historical profiles and results.
• No Core, evaluator, task, prompt, dependency or website changes. Mythos 5.2 is not registered because the exact account metadata route is unavailable.

Verification

• Candidate 1c1cc3b005eaf7928400a0b50a1583826ea381c8.
• Clean-source focused regressions: JUnit 208 cases, 3 skips, zero failures/errors.
• Whole-tree Ruff and formatting; mypy tools; contribution isolation and all 14 public-release checks: passed.
• Independent complete eight-file compatibility review: passed, no findings. Historical registry/refusal and selection parity checked.
• Pinned SDK over local MockTransport: two observations of the same 44 cases, 88 synthetic records, 820 turns per observation; retained source/runtime and case-matrix identities agree. Synthetic integration evidence is not model quality.
• A separate small real protocol request returned HTTP 200 with the requested Fable model, supported controls and complete standard/global usage. It is not a benchmark case or public quality result.

Delivery closure

• Exact-head hosted CI passed on Python 3.11 and 3.14, with successful distribution build, independent secret scan, CodeQL, dependency review and DCO.
• Current-head Balanced Qodo review completed with no bugs, rule violations or requirement gaps; no unresolved review threads.
• Both real observations of the same 44 cases completed and passed parent accounting and offline replay audits. They remain separate from the synthetic SDK qualification and standalone diagnostics.
• Approved aggregates are published by results-site PR fix(commerce): distinguish unfinished refunds from missed deadlines #13, preserving both original observations and all previous model rows. Fable displays its explicitly disclosed best-quality whole observation with the matching run cost. Site CI, Qodo, deployment and live byte/browser verification passed.
• This source PR is merged. The main commit has the exact reviewed Git tree and DCO trailer; the automatic post-merge CI also passed on the exact main commit (run 42890230268), including Python 3.11, Python 3.14, the independent secret scan and distribution build; all workflows on that main SHA completed successfully. Evaluated-run source identity remains 1c1cc3b005eaf7928400a0b50a1583826ea381c8; no historical grades or raw traces were changed or published.

[qodo-code-review]

PR Summary by Qodo

Add a Fable 5.2 profile to the three-flow campaign

✨ Enhancement 🧪 Tests 🕐 40+ Minutes

AI Description

• Add only Claude Fable 5.2 to a finite 44-case campaign profile.
• Reuse stateless Anthropic requests and existing admission, accounting, and replay controls.
• Validate published bounds and pricing estimates; add synthetic transport and admission
 regressions.

Diagram

graph TD
  CLI["Campaign CLI"] --> Admission["Source-bound admission"] --> Roster["Fable roster"] --> Transport["HTTP transport"] --> API["Anthropic API"] --> Evidence["Trial evidence"]
  Transport --> Budget["Budget accounting"] --> Evidence

Loading

High-Level Assessment

Extending the existing finite-roster and stateless Anthropic paths is appropriate: it keeps the new model distinct without duplicating admission, transport, or replay machinery. A dedicated transport would add maintenance cost without a different protocol requirement.

Files changed (8) +382 / -24

Enhancement (6) +78 / -24

three_flow_admission.pyAdmit the Fable roster under existing financial controls +9/-3

Admit the Fable roster under existing financial controls

• Recognizes the new roster in proposals, admission validation, historical registry identity, and published-profile validation. Its proposal has no global cap; execution still requires the applicable admission and monetary policy.

tools/three_flow_admission.py

three_flow_campaign.pyExpose the Fable roster in campaign preflight +7/-1

Expose the Fable roster in campaign preflight

• Reports the new finite roster without a global campaign cap while retaining the existing case assignments.

tools/three_flow_campaign.py

three_flow_http.pyRoute Fable 5.2 through stateless Anthropic transport +20/-11

Route Fable 5.2 through stateless Anthropic transport

• Gives Fable 5.2 a distinct mapping while reusing stateless request settings, timeouts, and Anthropic response checks. Requests use automatic tool choice with parallel use disabled and low-effort adaptive thinking.

tools/three_flow_http.py

three_flow_live.pyOffer the Fable roster in the live campaign CLI +1/-0

Offer the Fable roster in the live campaign CLI

• Adds the profile to the CLI's permitted roster choices without adding a new execution path.

tools/three_flow_live.py

three_flow_profiles.pyDefine the finite Fable 5.2 selection and registry entry +13/-4

Define the finite Fable 5.2 selection and registry entry

• Registers only Claude Fable 5.2 with the existing minimum-effort mode and a distinct stateless mapping. Keeps the historical rosters separate.

tools/three_flow_profiles.py

three_flow_selected.pyValidate Fable 5.2 published bounds and tariff envelope +28/-5

Validate Fable 5.2 published bounds and tariff envelope

• Extends closed-profile validation with the model's 1,000,000-token context, 128,000-token output bound, and standard/global category rates. Accepts existing monetary execution policies without treating reported-usage estimates as billing guarantees.

tools/three_flow_selected.py

Tests (2) +304 / -0

test_openai_reasoning_items.pyInitialize the transport's November flag in a test fixture +1/-0

Initialize the transport's November flag in a test fixture

• Sets the flag on a manually constructed transport so the expanded stateless-model check can run in the existing OpenAI test.

tests/test_openai_reasoning_items.py

test_three_flow_fable52.pyCover the Fable 5.2 roster, SDK behavior, and admission +303/-0

Cover the Fable 5.2 roster, SDK behavior, and admission

• Adds synthetic tests for all 44 assignments, stateless request projection, response failures, accounting, and replay. Also checks published-profile constraints, source-bound consent, policy refusal, and one-shot execution.

tests/test_three_flow_fable52.py

[qodo-code-review]

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (0) 📎 Requirement gaps (0)

Great, no issues found!

Qodo reviewed your code and found no material issues that require review

Tip of the day

💡 Did you know, you can move or hide resolved findings to focus on what still needs action

More tips ↗ | Customize Qodo ↗ | Qodo docs ↗

Previous reviews

Review updated until commit 1c1cc3b ⚖️ Balanced

Results up to commit 24507d2 ⚖️ Balanced

🐞 Bugs (1) 📘 Rule violations (0) 📎 Requirement gaps (0) 🎨 UX issues (0) 🔗 Cross-repo conflicts (0) 📜 Skill insights (0)

1. Projected keys silently collide 🐞 Bug ≡ Correctness

Description

_to_json normalizes mapping keys and assigns them into projected without checking whether that
exact string is already present, so a custom Mapping.items() that yields duplicate keys—or
distinct string-subclass keys that normalize identically—silently discards the earlier value.
TrajectoryLedger.append() now uses this projection for admitted payloads, allowing the retained
row and digest to omit caller-supplied data without raising.

Code

src/operatebench/jsonsafe.py[177]

+                projected[normalized_key] = _to_json(

Relevance

●●● Strong

Fail-closed duplicate-key handling is a deterministic correctness fix aligned with this PR's
explicit hostile-input refusal intent.

ⓘ Recommendations generated based on similar findings in past PRs

Evidence

The new helper explicitly collapses string subclasses to built-in strings and then inserts each
normalized key without a duplicate check. The ledger activates this helper on every payload, while
another untrusted projection in the repository demonstrates the established fail-closed behavior by
rejecting duplicate mapping keys before insertion.

src/operatebench/jsonsafe.py[114-120]
src/operatebench/jsonsafe.py[166-184]
src/operatebench/core/ledger.py[107-110]
src/operatebench/cards.py[810-823]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The JSON projection silently overwrites an earlier mapping entry when another key normalizes to the same exact built-in string. Reject such collisions with `NonJsonValueError` before projecting the duplicate value or inserting it.

## Issue Context
The projection intentionally supports custom, one-read `Mapping.items()` implementations and normalizes string subclasses without invoking their hooks. Once normalized, membership checks against `projected` are safe because both existing and candidate keys are exact built-in strings.

## Fix Focus Areas
- src/operatebench/jsonsafe.py[170-177]
- tests/test_freezing.py[252-271]
- tests/test_trajectory_ledger_aliases.py[372-397]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

Dismiss ↗ | View ↗

---

Results up to commit e71d019 ⚖️ Balanced

No changes from previous review