Fix CF-7 Qwen3.5 NoKvSlot prompt overflow

[hardcoreerik]

Summary

• enforce exact loaded-model token accounting when CF-7 packs Reviewer evidence
• reserve the full answer allowance and MultiHop chase headroom against the 8,192-token context
• use exact counts for the general CF invocation gate whenever the model is loaded
• queue the same token array for native inference and cap completion to remaining context cells
• correct the §7e history and compatibility docs without changing the historical invalid results

Root cause

The B3 Reviewer request exhausted its own attention KV context. The fully rendered Qwen3.5 prompt plus requested completion exceeded n_ctx=8192; the existing chars/4 estimate plus safety margin under-counted the JSON-heavy prompt. The matching llama.cpp signature is failed to prepare attention ubatches followed by failed to find a memory slot: ggml-org/llama.cpp#20049 (comment)

This is distinct from the recurrent SeqMax defect fixed in PR #56. SeqMax=40, recycle threshold 24, hard limit 40, thinking suppression, disposal behavior, and LLamaSharp 0.27.0 are unchanged.

Validation

• targeted Context Fabric/native runtime regression tests — 3 passed
• dotnet test OrchestratorIDE.UnitTests\OrchestratorIDE.UnitTests.csproj --no-restore --verbosity minimal — 564 passed, 4 skipped
• dotnet build OrchestratorIDE.slnx --no-restore --verbosity minimal — succeeded; existing warnings only
• CI Build (Windows) — passed

Live gate status

Pending. The dedicated %APPDATA%\OrchestratorIDE\Models-CF7 directory and both historical GGUFs are currently absent. The docs deliberately remain at “root-caused; fix implemented; full Q8_0/Q4_K_M validation pending.” Once the exact model files are restored, the planned validation is one Q8_0 replay followed by sequential full 120-question Q8_0 and Q4_K_M gates with THEORC_KVCACHE_DIAGNOSTICS=1, requiring zero NoKvSlot/attention-slot failures.

Review status

• Grok 4.5 — two MINORs addressed (MultiHop exact-token chase reserve and exact general invocation gating); two consecutive final raw verdicts were CLEAN. The wrapper exited nonzero only because Grok prefixed each CLEAN with a sentence.
• CodeRabbit — completed review of 0b97c758…33d6eab2; no actionable comments, zero review threads, status passed

Summary by CodeRabbit

• Bug Fixes

  • Improved prompt token accounting using exact runtime-provided counts when available.
  • Prevented completions from exceeding the model’s remaining context capacity.
  • Improved evidence admission and budget checks to reduce context-overflow failures.
  • Preserved safe fallback behavior for runtimes without exact token counting.

• Documentation

  • Updated Context Fabric bug history, test results, model compatibility notes, and packaged-app model path guidance.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 6380d37b-6487-4a92-9c39-ff1598157f40

📥 Commits

Reviewing files that changed from the base of the PR and between 33d6eab and 9f975e3.

📒 Files selected for processing (3)

• docs/CF_TEST_RESULTS.md
• docs/CONTEXT_FABRIC_BUG_HISTORY.md
• docs/CONTEXT_FABRIC_INFRASTRUCTURE_NOTES.md

---

📝 Walkthrough

Walkthrough

Context Fabric now uses runtime-provided exact prompt token counts for evidence admission, budget checks, and metrics. Native completion generation is capped by remaining model context, with fallback estimates for runtimes without loaded-model tokenizers. Tests and capacity documentation were updated accordingly.

Changes

Exact prompt token accounting

Layer / File(s)	Summary
Native token contract and completion bounds OrchestratorIDE/Core/Runtime/IRoleRuntime.cs, OrchestratorIDE/Core/Runtime/LLamaSharpRuntime.cs	Role runtimes tokenize rendered prompts with the loaded model, report exact counts, and cap completion generation to remaining context capacity.
Context Fabric exact-count integration OrchestratorIDE/Services/ContextFabric/*	Evidence packing and invocation gates use exact counts when available, while metrics prefer exact counts over estimates.
Runtime and budget validation OrchestratorIDE.UnitTests/*	Tests cover exact prompt metrics, evidence-pack inclusion, MultiHop reservation, scripted token counting, and completion-limit boundaries.
Capacity defect documentation updates docs/CF_TEST_RESULTS.md, docs/CONTEXT_FABRIC_BUG_HISTORY.md, docs/CONTEXT_FABRIC_INFRASTRUCTURE_NOTES.md	Results and compatibility notes describe the exact-token fix, revised capacity incidents, validation outcomes, and MSIX path virtualization.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant ContextFabric
  participant Runtime
  participant LoadedModel
  participant Metrics
  ContextFabric->>Runtime: CountPromptTokens(rendered history)
  Runtime->>LoadedModel: Tokenize rendered prompt
  LoadedModel-->>Runtime: Exact prompt token count
  Runtime-->>ContextFabric: Prompt token count
  ContextFabric->>ContextFabric: Apply evidence and context budget gates
  ContextFabric->>Runtime: Generate within remaining context
  Runtime-->>Metrics: Report prompt token usage

Loading

Possibly related PRs

• hardcoreerik/TheOrc#34: Modifies BuildEvidencePack evidence-selection logic.
• hardcoreerik/TheOrc#40: Modifies Context Fabric token-budget gating for prompt-capacity handling.
• hardcoreerik/TheOrc#42: Modifies evidence budgeting and MultiHop-related inclusion behavior.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch codex/fix-cf7-nokvslot-storm-7e

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

[hardcoreerik]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.

[hardcoreerik]

@coderabbitai review

[coderabbitai]

✅ Action performed

Review finished.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.