feat(agentic): Phase 3 — local in-process inference (epic #1544) by ooples · Pull Request #1552 · ooples/AiDotNet

ooples · 2026-06-09T17:03:05Z

Draft — do not review yet. Epic #1544 (Phase 3). Stacked on Phase 2 (#1551), which is stacked on Phase 0 (#1545). Rebase down the stack as the lower PRs merge.

Phase 3 goal — the flagship differentiator

Run inference in-process over AiDotNet's own model: no Python, no API, no key. An IChatClient<T> implementation means the exact same agent code (executor / supervisor / swarm / memory) runs against a local model.

Landed (slice 1 — the generation core)

ICausalLanguageModel<T> — minimal next-token-logits seam the real Transformer plugs into (free to keep an internal KV-cache).
IGenerationTokenizer + TokenizerGenerationAdapter over the repo's ITokenizer.
TokenSampler<T> — greedy / temperature / top-k / top-p (nucleus), seedable for reproducibility.
IChatPromptTemplate + ChatMlPromptTemplate.
LocalEngineChatClient<T> — render → encode → autoregressive sample (until EOS / token cap) → decode; non-streaming and streaming (incremental deltas); token-count usage.
9 tests (green net10.0 + net471) on a deterministic fake model + tokenizer, including an AgentExecutor drop-in.

Planned next in this PR

Wire the concrete Transformer/LM behind ICausalLanguageModel<T> (real logits + KV-cache).
Constrained decoding (JSON-schema / grammar at the logits) → native local structured output & tool-calling.
Stop-sequence handling; beam search; quantization hooks.

No null-forgiving operators; enums for closed sets; "For Beginners" docs throughout.

🤖 Generated with Claude Code

vercel · 2026-06-09T17:03:09Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
aidotnet_website	Ready	Preview, Comment	Jun 12, 2026 3:04pm
aidotnet-playground-api	Ready	Preview, Comment	Jun 12, 2026 3:04pm

coderabbitai · 2026-06-09T17:06:17Z

Warning

Review limit reached

@ooples, we couldn't start this review because you've reached your PR review rate limit.

More reviews will be available in 1 hour, 33 minutes, and 1 second. Learn how PR review limits work.

Your organization has run out of usage credits. Purchase more credits in the billing tab to continue.

⌛ How to resolve this issue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

We recommend that you space out your commits to avoid hitting the rate limit.

🚦 How do rate limits work?

CodeRabbit enforces hourly rate limits for each developer per organization.

Our paid plans include higher PR review limits than trial, open-source, and free plans. In all cases, reviews become available again over time. During sustained high-volume PR review activity, CodeRabbit may temporarily slow when the next review becomes available.

Please see our Fair Usage Limits Policy for further information.

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro

Run ID: 8b91a08f-bead-4046-aba2-89f1561cd00c

📥 Commits

Reviewing files that changed from the base of the PR and between cdc00a2 and 69a8271.

📒 Files selected for processing (19)

src/Agentic/Models/Local/AllowedTokenSetConstraint.cs
src/Agentic/Models/Local/ChatMlPromptTemplate.cs
src/Agentic/Models/Local/FiniteStateTokenConstraint.cs
src/Agentic/Models/Local/ICausalLanguageModel.cs
src/Agentic/Models/Local/IChatPromptTemplate.cs
src/Agentic/Models/Local/IGenerationTokenizer.cs
src/Agentic/Models/Local/IIncrementalCausalLanguageModel.cs
src/Agentic/Models/Local/ITokenConstraint.cs
src/Agentic/Models/Local/LocalEngineChatClient.cs
src/Agentic/Models/Local/LocalEngineOptions.cs
src/Agentic/Models/Local/LocalSamplingOptions.cs
src/Agentic/Models/Local/NeuralNetworkCausalLanguageModel.cs
src/Agentic/Models/Local/TokenSampler.cs
src/Agentic/Models/Local/TokenizerGenerationAdapter.cs
tests/AiDotNet.Tests/UnitTests/Agentic/Local/BeamSearchTests.cs
tests/AiDotNet.Tests/UnitTests/Agentic/Local/ConstrainedDecodingTests.cs
tests/AiDotNet.Tests/UnitTests/Agentic/Local/IncrementalDecodingTests.cs
tests/AiDotNet.Tests/UnitTests/Agentic/Local/LocalInferenceTests.cs
tests/AiDotNet.Tests/UnitTests/Agentic/Local/RealModelLocalInferenceTests.cs

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Commit unit tests in branch feature/agentic-phase3-local-inference

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

vercel · 2026-06-09T18:42:24Z

Deployment failed with the following error:

Resource is limited - try again in 24 hours (more than 100, code: "api-deployments-free-per-day").

Learn More: https://vercel.com/franklins-projects-02a0b5a0?upgradeToPro=build-rate-limit

…ion) (#1544) The raw material every self-improvement mechanism learns from: - AgentTrajectory: a captured run (transcript, final answer, iterations, usage, agent name, optional evaluator-assigned Reward, metadata). - ITrajectoryStore (Add/Get/GetAll/Query/Clear) + InMemoryTrajectoryStore. - TracingAgent<T> (IAgent<T>): transparently records each run of any wrapped agent (executor/supervisor/swarm/memory) to the store without changing behavior; supports per-experiment metadata. - 5 tests (green net10.0 + net471): records runs, accumulates + queryable, metadata, get-by-id + reward annotation, clear. net471 nullability fix on GetAsync (Task.FromResult<AgentTrajectory?>). No null-forgiving. Stacked on Phase 3 (#1552). Next: eval harness scoring trajectories with the existing reward models, then learned routing / prompt-opt / online LoRA. Part of epic #1544 (Phase 4). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

ooples · 2026-06-11T13:15:04Z

@coderabbitai full review

coderabbitai · 2026-06-11T13:15:14Z

✅ Action performed

Full review finished.

…ion) (#1544) The raw material every self-improvement mechanism learns from: - AgentTrajectory: a captured run (transcript, final answer, iterations, usage, agent name, optional evaluator-assigned Reward, metadata). - ITrajectoryStore (Add/Get/GetAll/Query/Clear) + InMemoryTrajectoryStore. - TracingAgent<T> (IAgent<T>): transparently records each run of any wrapped agent (executor/supervisor/swarm/memory) to the store without changing behavior; supports per-experiment metadata. - 5 tests (green net10.0 + net471): records runs, accumulates + queryable, metadata, get-by-id + reward annotation, clear. net471 nullability fix on GetAsync (Task.FromResult<AgentTrajectory?>). No null-forgiving. Stacked on Phase 3 (#1552). Next: eval harness scoring trajectories with the existing reward models, then learned routing / prompt-opt / online LoRA. Part of epic #1544 (Phase 4). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

ooples · 2026-06-11T13:56:43Z

@coderabbitai resume

coderabbitai · 2026-06-11T13:56:48Z

✅ Action performed

Reviews resumed.

ooples · 2026-06-12T12:22:20Z

@coderabbitai full review

coderabbitai · 2026-06-12T12:22:26Z

✅ Action performed

Full review finished.

The flagship local-first differentiator: an IChatClient<T> that runs entirely in-process over AiDotNet's own model — no network, no API key. Same agent code drives it as drives OpenAI/Anthropic. - ICausalLanguageModel<T>: minimal next-token-logits seam the real Transformer (or any model) plugs into; implementations may keep an internal KV-cache. - IGenerationTokenizer (+TokenizerGenerationAdapter over the repo's ITokenizer): encode/decode/EOS seam. - TokenSampler<T>: greedy (temp 0 = argmax) + temperature + top-k + top-p (nucleus), seedable for reproducibility; reads logits via Convert.ToDouble so float/double share one path. - IChatPromptTemplate + ChatMlPromptTemplate (role-tagged, opens the assistant turn). - LocalEngineChatClient<T>: renders prompt -> encodes -> autoregressively samples until EOS or token limit -> decodes; non-streaming + streaming (incremental decode deltas); usage = prompt/generated token counts. Drop-in for agents/supervisor/swarm/memory. (Native tool-calling + structured-output constraints via constrained decoding are a follow-up; tools are ignored this slice.) - 9 tests (green net10.0 + net471) on a deterministic ScriptedCausalModel + WordTokenizer: greedy decode + EOS stop, token-limit -> Length, streaming deltas reconstruct full text, AgentExecutor drop-in, sampler argmax/seed-reproducibility/top-k/top-p, ChatML template. No null-forgiving. Stacked on Phase 2 (#1551) — the integration test shows the local engine driving an AgentExecutor. Part of epic #1544 (Phase 3). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…ence (#1544) - NeuralNetworkCausalLanguageModel<T>: adapts a trained NeuralNetworkBase<T> LM (Mamba/GLA/Transformer-LM-head) to ICausalLanguageModel<T>. Encodes context as the one-hot [1,seq,vocab] tensor these models expect, runs a forward pass (ResetState first so recurrent models start fresh), and extracts the final position's logits (handles rank-2 and rank-3 outputs). Full context re-fed each step; KV-cached fast path is a follow-up. - 3 integration tests over a real tiny untrained MambaLanguageModel<double> (greedy, deterministic): adapter logit width + no-NaN, end-to-end generation honoring the token budget, streaming termination. Pinned to CpuEngine + non-parallel collection (documented GPU-autodetect mitigation). - Fixed Engine_IsDropInForAgentExecutor: it relied on the engine's stochastic default sampling while asserting a fixed output; now configures greedy (Temperature 0) so the assertion is deterministic. Root-caused the flake to default-temperature sampling, not a library/GPU issue. - Local suite green + stable (verified repeated runs) on net10.0 + net471. No null-forgiving. Part of epic #1544 (Phase 3). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Native LOCAL guaranteed-structure generation — enforced at the logits, not merely prompted (a capability cloud models can't guarantee): - ITokenConstraint seam: AllowedNextTokens(generated) returns null (free) / a set (restrict) / empty (terminal -> stop). TokenSampler.Sample gains an allowed-token mask honored by both greedy (argmax within allowed) and stochastic (candidates restricted before top-k/top-p). - AllowedTokenSetConstraint (fixed allow-list) + FiniteStateTokenConstraint (finite-state grammar over token ids: start set + per-token transitions; terminal states stop) — the general mechanism a JSON-schema/regular grammar compiles to. - LocalEngineOptions.Constraint wires it into LocalEngineChatClient generation (non-streaming + streaming). - Stop sequences: ChatOptions.StopSequences now halt generation and trim the output at the earliest match (both paths). - 5 tests (green, stable, net10.0 + net471): allowed-set greedy/stochastic masking, FSA forces an exact JSON-shaped sequence despite a model biased to 'garbage', allow-list restricts whole output, stop-sequence halts+trims. No null-forgiving. Part of epic #1544 (Phase 3). JSON-schema->token-grammar compiler (to drive structured output / tool-calling off this framework) is the next follow-up. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

- LocalEngineOptions.BeamWidth (>1) switches non-streaming GetResponseAsync to beam search: explores N hypotheses in parallel, expands each by its top tokens (honoring the ITokenConstraint and stop sequences per beam), prunes to the top-N by length-normalized log-probability, and returns the best completion. Deterministic; streaming stays token-by-token. - LogSoftmax (with allowed-token masking → -inf) + length-normalized scoring + Beam bookkeeping. - 3 tests (green net10.0 + net471): greedy takes the locally-best token ("A"); beam width 2 discovers the globally better "BC" path the same model would miss greedily; beam respects a finite-state constraint ("AB"). No null-forgiving (removed an introduced allowed! via proper narrowing). Part of epic #1544 (Phase 3). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…e fast path (#1544) - IIncrementalCausalLanguageModel<T> : ICausalLanguageModel<T> adds ResetCache / StartSequence(prompt) / AppendToken(id) — the contract a KV-cached model implements to advance one token at a time instead of re-feeding the full O(n^2) context. - LocalEngineChatClient auto-detects the interface and takes the incremental fast path (prime once, then feed single tokens), falling back to full re-feed otherwise. Constraint + stop-sequence + sampler logic shared via a single PickToken helper across both paths. - 2 tests (green net10.0 + net471): the engine drives the incremental path correctly (ResetCache once, StartSequence with the prompt, AppendToken per generated token) and produces output IDENTICAL to the full-refeed path. The remaining half — a real K/V cache inside Mamba/GLA/attention forward — is a model-layer change (the network Predict API has no incremental entry point yet); this lands the engine-side support + verification it plugs into. Part of epic #1544 (Phase 3). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…ompression (#1544) Quantization for the local engine is achieved by quantizing the NeuralNetworkBase<T> with the existing ModelCompression stack before wrapping it — the adapter accepts any such network unchanged. No engine-side code (no duplication of ModelCompression). Documented on NeuralNetworkCausalLanguageModel. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…ion) (#1544) The raw material every self-improvement mechanism learns from: - AgentTrajectory: a captured run (transcript, final answer, iterations, usage, agent name, optional evaluator-assigned Reward, metadata). - ITrajectoryStore (Add/Get/GetAll/Query/Clear) + InMemoryTrajectoryStore. - TracingAgent<T> (IAgent<T>): transparently records each run of any wrapped agent (executor/supervisor/swarm/memory) to the store without changing behavior; supports per-experiment metadata. - 5 tests (green net10.0 + net471): records runs, accumulates + queryable, metadata, get-by-id + reward annotation, clear. net471 nullability fix on GetAsync (Task.FromResult<AgentTrajectory?>). No null-forgiving. Stacked on Phase 3 (#1552). Next: eval harness scoring trajectories with the existing reward models, then learned routing / prompt-opt / online LoRA. Part of epic #1544 (Phase 4). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…1556) * feat(agentic): Phase 4 — trajectory capture (self-improvement foundation) (#1544) The raw material every self-improvement mechanism learns from: - AgentTrajectory: a captured run (transcript, final answer, iterations, usage, agent name, optional evaluator-assigned Reward, metadata). - ITrajectoryStore (Add/Get/GetAll/Query/Clear) + InMemoryTrajectoryStore. - TracingAgent<T> (IAgent<T>): transparently records each run of any wrapped agent (executor/supervisor/swarm/memory) to the store without changing behavior; supports per-experiment metadata. - 5 tests (green net10.0 + net471): records runs, accumulates + queryable, metadata, get-by-id + reward annotation, clear. net471 nullability fix on GetAsync (Task.FromResult<AgentTrajectory?>). No null-forgiving. Stacked on Phase 3 (#1552). Next: eval harness scoring trajectories with the existing reward models, then learned routing / prompt-opt / online LoRA. Part of epic #1544 (Phase 4). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(agentic): Phase 4 — evaluation harness (score trajectories) (#1544) Turns captured runs into a measurable quality signal — the continuous-eval step the learners optimize against: - ITrajectoryEvaluator (trajectory-native scoring seam; higher = better) + DelegateTrajectoryEvaluator (custom scoring functions: exact-match, JSON validity, cost penalties, LLM-as-judge, or an adapter over the reasoning reward models). - EvaluationReport (count + mean/min/max reward + pass-rate at a threshold) — the scoreboard a self-improvement loop watches. - TrajectoryEvaluationRunner: scores trajectories (writing AgentTrajectory.Reward) and aggregates the report; EvaluateAsync(list) + EvaluateStoreAsync(store). - 9 tests total (green net10.0 + net471): exact-match grading + reward annotation, pass-rate threshold, whole-store eval with write-back, empty-set report. No null-forgiving. Reward models (OutcomeRewardModel/etc.) are reasoning-chain-specific + internal; they plug in via an ITrajectoryEvaluator adapter (follow-up) rather than coupling the harness to reasoning types. Part of epic #1544 (Phase 4). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(agentic): Phase 4 — learned routing (contextual reward-weighted bandit) (#1544) Replaces a hand-tuned router with one that learns from graded history which agent performs best — the orchestration improves as rewarded runs accumulate: - LearnedAgentRouter<T> (an IAgent<T>): LearnFrom(trajectories) builds per-agent (optionally per-context) mean-reward stats; routing exploits the best agent, explores unseen candidates first (optimistic), and takes a random agent with probability explorationRate. Optional contextKey learns different best-agents per task type. - Composable: drops in anywhere an agent is expected and, wrapped by TracingAgent, its routed runs feed back into the store — closing the self-improvement loop. - 5 tests (green net10.0 + net471): routes to highest-reward agent after learning, explores unseen first, learns per-context (math vs prose), closed-loop capture under the chosen agent, empty/duplicate guards. No null-forgiving. Part of epic #1544 (Phase 4). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(agentic): Phase 4 — prompt auto-optimization (DSPy-like) (#1544) Evaluation-driven prompt search: pick the system prompt that scores best on a labeled eval set. - PromptOptimizer<T>.OptimizeAsync(candidatePrompts, agentFactory, cases): builds an agent per candidate prompt, runs the eval cases, scores each, returns the best prompt + ranked candidates. Default scorer = case-insensitive substring match vs expected; overridable (e.g., reuse an ITrajectoryEvaluator / reward model / length penalty). - PromptEvalCase (input + expected) + ScoredPrompt + PromptOptimizationResult. - The optimizer is the SELECTION half; candidate prompts (the SEARCH half) can come from a hand-written set, an LLM proposer, or AiDotNet's existing genetic/beam/annealing prompt optimizers feeding their population in. - 4 tests (green net10.0 + net471): selects best prompt (ranked), custom scorer, single candidate, empty-input guards. Tests use a real AgentExecutor whose scripted model keys off the system prompt (no manual result construction, no null-forgiving). Part of epic #1544 (Phase 4). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(agentic): Phase 4 — reward-filtered fine-tuning dataset (online LoRA data prep) (#1544) The data-preparation half of online LoRA self-improvement (reward-filtered behavior cloning): - RewardFilteredDatasetBuilder.Build(trajectories): keeps graded runs at/above a reward threshold and distills each into a (prompt, completion) pair — prompt = conversation up to the final turn, completion = final answer; skips ungraded/degenerate runs. - FineTuningExample (prompt/completion/reward) + FineTuningDataset (examples + mean reward) — ready to hand to the repo's LoRA/FineTuning trainer to produce an adapter for the local model. - 3 tests (green net10.0 + net471): keeps only high-reward runs (prompt excludes completion), skips ungraded/too-short, threshold controls inclusion. No null-forgiving. The actual LoRA training (dataset -> adapter on the local model) is the model-layer step performed by the existing trainer; this is the agentic step that decides WHAT to learn from. Part of epic #1544 (Phase 4). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: franklinic <franklin@ivorycloud.com> Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

vercel Bot deployed to Preview – aidotnet-playground-api June 9, 2026 17:03 View deployment

vercel Bot deployed to Preview – aidotnet_website June 9, 2026 17:03 View deployment

This was referenced Jun 10, 2026

feat(agentic): Phase 4 — self-improving orchestration (epic #1544) #1556

Merged

feat(agentic): Phase 5 — parity polish (epic #1544) #1557

Merged

ooples marked this pull request as ready for review June 11, 2026 00:54

ooples changed the title ~~feat(agentic): Phase 3 — local in-process inference [DRAFT, do not review]~~ feat(agentic): Phase 3 — local in-process inference (epic #1544) Jun 11, 2026

ooples force-pushed the feature/agentic-phase2-multi-agent branch from 072ed9e to 978191c Compare June 11, 2026 12:27

ooples force-pushed the feature/agentic-phase3-local-inference branch from cea00fd to dec8f40 Compare June 11, 2026 12:40

vercel Bot deployed to Preview – aidotnet-playground-api June 11, 2026 12:40 View deployment

vercel Bot deployed to Preview – aidotnet_website June 11, 2026 12:41 View deployment

ooples changed the base branch from feature/agentic-phase2-multi-agent to master June 11, 2026 12:54

ooples force-pushed the feature/agentic-phase3-local-inference branch from dec8f40 to 3cb7e09 Compare June 11, 2026 13:13

vercel Bot deployed to Preview – aidotnet-playground-api June 11, 2026 13:13 View deployment

vercel Bot deployed to Preview – aidotnet_website June 11, 2026 13:14 View deployment

franklinic and others added 3 commits June 12, 2026 10:59

ooples force-pushed the feature/agentic-phase3-local-inference branch from 801f973 to 69a8271 Compare June 12, 2026 15:03

vercel Bot deployed to Preview – aidotnet-playground-api June 12, 2026 15:04 View deployment

vercel Bot deployed to Preview – aidotnet_website June 12, 2026 15:04 View deployment

ooples merged commit 5c37400 into master Jun 12, 2026
11 of 63 checks passed

ooples deleted the feature/agentic-phase3-local-inference branch June 12, 2026 15:53

ooples mentioned this pull request Jun 24, 2026

Epic: Agentic Orchestration — surpass Semantic Kernel & LangGraph #1544

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat(agentic): Phase 3 — local in-process inference (epic #1544)#1552

feat(agentic): Phase 3 — local in-process inference (epic #1544)#1552
ooples merged 6 commits into
masterfrom
feature/agentic-phase3-local-inference

ooples commented Jun 9, 2026

Uh oh!

vercel Bot commented Jun 9, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Jun 9, 2026 •

edited

Loading

Review limit reached

Uh oh!

vercel Bot commented Jun 9, 2026

Uh oh!

ooples commented Jun 11, 2026

Uh oh!

coderabbitai Bot commented Jun 11, 2026 •

edited

Loading

Uh oh!

ooples commented Jun 11, 2026

Uh oh!

coderabbitai Bot commented Jun 11, 2026

Uh oh!

ooples commented Jun 12, 2026

Uh oh!

coderabbitai Bot commented Jun 12, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

ooples commented Jun 9, 2026

Phase 3 goal — the flagship differentiator

Landed (slice 1 — the generation core)

Planned next in this PR

Uh oh!

vercel Bot commented Jun 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

coderabbitai Bot commented Jun 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review limit reached

Uh oh!

vercel Bot commented Jun 9, 2026

Uh oh!

ooples commented Jun 11, 2026

Uh oh!

coderabbitai Bot commented Jun 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ooples commented Jun 11, 2026

Uh oh!

coderabbitai Bot commented Jun 11, 2026

Uh oh!

ooples commented Jun 12, 2026

Uh oh!

coderabbitai Bot commented Jun 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

vercel Bot commented Jun 9, 2026 •

edited

Loading

coderabbitai Bot commented Jun 9, 2026 •

edited

Loading

coderabbitai Bot commented Jun 11, 2026 •

edited

Loading

coderabbitai Bot commented Jun 12, 2026 •

edited

Loading