fix(optimization): support published GEPA package

[drewstone]

What changed

• run the standard GEPA engine through the published gepa[full]==0.1.4 API
• retain the pinned official source revision for composed and source-only engines
• normalize both upstream result shapes without implementing search locally
• test published and source installs independently, including real two-run resume
• release as npm and Python package version 0.126.5

Proof

• Python source suite: passed
• Python published-GEPA suite: 25 passed, 1 intentional source-only skip
• TypeScript suite: 3,380 passed, 3 skipped
• clean built agent-eval-rpc==0.126.5 wheel + published GEPA + TypeScript bridge: passed
• lint, typecheck, examples typecheck, build, package export check, and actionlint: passed

[tangletools]

✅ Auto-approved drewstone PR — 951c7761

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.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-24T19:16:52Z}

[tangletools]

⚠️ Review Interrupted — 951c7761

The review runner stopped before publishing a final verdict: webhook_restarted.

State	Detail
Interrupted	webhook restarted

No review verdict was produced for this run. Trigger a fresh review on the current PR head if the PR is still open.

_{tangletools · #431 · model: kimi-for-coding · updated 2026-07-24T19:20:18Z}

[drewstone]

@tangletools review now

[tangletools]

✅ Auto-approved drewstone PR — 951c7761

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.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-24T19:21:56Z}

[tangletools]

🟢 Value Audit — sound

Verdict	sound
Concerns	0 (none)
Heuristic	0.0s
Duplication	0.0s
Interrogation	93.8s (2 bridge agents)
Total	93.8s

💰 Value — sound

Extends the GEPA bridge to run against the published gepa[full]==0.1.4 PyPI wheel (not just the pinned git revision) via a clean API-shape resolver; coherent, in-grain, no existing equivalent to reuse.

• What it does: Adds a gepa_api.py resolver (GepaApi dataclass + load_gepa_api()) that sniffs whether the installed GEPA exposes OptimizeAnythingConfig ('engine' shape, source revision) or GEPAConfig ('launcher' shape, published wheel), then branches _engine_config() to map the agent-eval-normalized inputs (maxEvaluations, maxProposerCostUsd, maxConcurrency, stopAtScore) into the correct kwargs for each shape. Re
• Goals it achieves: (1) Remove the install-friction of requiring a git URL for the common case: users can now pip install gepa[full]==0.1.4 for the standard gepa engine. PyPI metadata cannot express a Git dependency, so the previous source-only requirement was a real blocker for the published agent-eval-rpc wheel. (2) Preserve the pinned source revision for composed recipes (sequential/adaptive/best-of/vote/omni) a
• Assessment: Coherent and well-scoped. The GepaApi abstraction centralizes the 'which shape is installed?' decision in one load_gepa_api() and the bridge consumes it uniformly via api.optimize_anything / api.composition(name) — replacing 5 scattered try/except ImportError blocks. The two-branch _engine_config is unavoidable: the two GEPA APIs genuinely differ in kwargs (max_evals vs max_metric_calls, stop_at_s
• Better / existing approach: Searched for an existing API-shape resolver to reuse: none. optimizer_bridge_common.inspect_optimizer_runtime (optimizer_bridge_common.py:143) resolves package provenance, not API shape. skillopt_bridge.py:44-51 does a simpler importlib.import_module with no shape-sniffing because SkillOpt only has one shape. The two-branch _engine_config cannot be collapsed into a unified mapping without an adapt
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 2
• Bridge warning: opencode/kimi-for-coding/k2p7: bridge stream ended without value-audit content

🎯 Usefulness — sound

Adds published gepa[full]==0.1.4 wheel support to the GEPA bridge by normalizing both upstream API shapes behind one resolver, verified end-to-end through the real TypeScript caller in CI.

• Integration: Fully wired and reachable. The bridge's single caller, gepaOptimizationMethod (src/campaign/gepa-optimization-method.ts:350), spawns agent_eval_rpc.gepa_bridge, which now calls load_gepa_api() once at startup (gepa_bridge.py:82). The resolver picks launcher shape for the published GEPAConfig or engine shape for the source OptimizeAnythingConfig (gepa_api.py:38-52); both branches are
• Fit with existing patterns: Fits the established compat-layer grain. The new gepa_api.py mirrors how gepa_compat_0_1_4.py and skillopt_compat_v020.py already pin behavior to a specific upstream version/shape. It normalizes the two divergent upstream config+result shapes behind one GepaApi dataclass rather than forking the bridge or duplicating search logic (the bridge still delegates all search to GEPA). No competing
• Real-world viability: Holds up beyond the happy path. Result-shape divergence is handled for both installs: _selected_score falls through best_score -> val_aggregate_scores[best_idx] (gepa_bridge.py:419-432) and _result_evaluations falls through total_evals -> total_metric_calls (gepa_bridge.py:518-521). The launcher path maps stopAtScore to ScoreThresholdStopper and rejects non-gepa engines with a cl
• Model: opencode/zai-coding-plan/glm-5.2
• Bridge attempts: 1

No concerns — sound change, no better or existing approach found. ✅

---

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 · 20260724T192351Z}