nimmo/nixos-config

Watch

Fork 0

ci: raise AI review context to 128k #136

Merged

nimmo merged 1 commit from issue-135-ai-review-context-128k into main

2026-08-01 08:13:21 +01:00

nimmo commented

2026-08-01 08:07:45 +01:00

Owner

Summary

raise the trusted Nova AI reviewer context from 65,536 to 131,072 tokens
document that 128k is deliberately half of the Qwen model native 262,144-token window
retain the model, prompt inputs, diff/guidance limits, review gates, trusted-base loading, and publication behavior unchanged

Rationale

PR #134 currently produces a 65,880-token initial prompt. With num_ctx fixed at 65,536, Ollama context-shifts away the leading review schema and instructions and returns malformed output. The 128k setting gives this review substantial output headroom while remaining conservative relative to the model full native window.

This change is isolated from PR #134 because reviewer configuration is loaded from the pull request base revision.

Validation

reviewer parse assertion: context_size is integer 131072
actionlint over all Forgejo workflows
nix flake check (all checks passed; electra, lena, vega, lyra, and cosmos evaluated)
git diff --check
automated review run 565: Ready to merge, no findings
measured Nova capacity at the 64k baseline: 59% VRAM used; projected 128k KV-cache growth is 1.25 GiB with about 24.9 GiB remaining headroom
no deployment or updater invocation performed

The Attic workflow does not match changes limited to .forgejo/ai-review.yml, so no Attic job was expected or reported; workflow files and build inputs are unchanged.

Closes #135

## Summary - raise the trusted Nova AI reviewer context from 65,536 to 131,072 tokens - document that 128k is deliberately half of the Qwen model native 262,144-token window - retain the model, prompt inputs, diff/guidance limits, review gates, trusted-base loading, and publication behavior unchanged ## Rationale PR #134 currently produces a 65,880-token initial prompt. With num_ctx fixed at 65,536, Ollama context-shifts away the leading review schema and instructions and returns malformed output. The 128k setting gives this review substantial output headroom while remaining conservative relative to the model full native window. This change is isolated from PR #134 because reviewer configuration is loaded from the pull request base revision. ## Validation - reviewer parse assertion: context_size is integer 131072 - actionlint over all Forgejo workflows - nix flake check (all checks passed; electra, lena, vega, lyra, and cosmos evaluated) - git diff --check - automated review run 565: Ready to merge, no findings - measured Nova capacity at the 64k baseline: 59% VRAM used; projected 128k KV-cache growth is 1.25 GiB with about 24.9 GiB remaining headroom - no deployment or updater invocation performed The Attic workflow does not match changes limited to .forgejo/ai-review.yml, so no Attic job was expected or reported; workflow files and build inputs are unchanged. Closes #135

nimmo added this to the Automated Update Remediation milestone

2026-08-01 08:07:45 +01:00

nimmo added the

phase:validation

host:nova

labels

2026-08-01 08:07:45 +01:00

nimmo added 1 commit

2026-08-01 08:07:48 +01:00

ci: raise AI review context to 128k

AI pull-request review / review (pull_request) Successful in 2m4s

Details

d32a54f90a

nimmo commented

2026-08-01 08:09:53 +01:00

Author

Owner

Automated code review

Reviewed commit: d32a54f90a8e0a2b7f090deb046462475533571d

Verdict: Ready to merge

No blocking or important findings were retained, and deterministic checks passed. Suggestions remain advisory.

Overall assessment

The pull request updates .forgejo/ai-review.yml to increase the context_size from 65,536 to 131,072 tokens and adds a comment explaining the rationale relative to the model's native window. This is a configuration-only change targeting AI prompt processing limits.

Blocking findings

None.

Important findings

None.

Suggestions

None.

Tests and validation

No configured deterministic checks.

Questions

Has the increased context size been validated against CI runner memory constraints to ensure no OOM errors or latency degradation occurs during review runs?

Review limitations

Review is limited to the provided diff and configuration file. Runtime behavior, resource utilization on CI runners, and actual prompt truncation scenarios were not observed.

Diff coverage

.forgejo/ai-review.yml: reviewed — included in a context-limited batch

Review metadata

Reviewer: forgejo-ai-review v1
Model: qwen3.6:35b-a3b-q8_0
Guidance loaded: .forgejo/ai-review.md, .forgejo/ai-review.yml, AGENTS.md, README.md
Commits considered: 1
Previous automated review: none found
Findings: 0

## Automated code review Reviewed commit: `d32a54f90a8e0a2b7f090deb046462475533571d` ## Verdict: Ready to merge No blocking or important findings were retained, and deterministic checks passed. Suggestions remain advisory. ### Overall assessment The pull request updates `.forgejo/ai-review.yml` to increase the `context_size` from 65,536 to 131,072 tokens and adds a comment explaining the rationale relative to the model's native window. This is a configuration-only change targeting AI prompt processing limits. ### Blocking findings None. ### Important findings None. ### Suggestions None. ### Tests and validation - No configured deterministic checks. ### Questions - Has the increased context size been validated against CI runner memory constraints to ensure no OOM errors or latency degradation occurs during review runs? ### Review limitations - Review is limited to the provided diff and configuration file. Runtime behavior, resource utilization on CI runners, and actual prompt truncation scenarios were not observed. ### Diff coverage - `.forgejo/ai-review.yml`: **reviewed** — included in a context-limited batch <details><summary>Review metadata</summary> - Reviewer: forgejo-ai-review v1 - Model: qwen3.6:35b-a3b-q8_0 - Guidance loaded: .forgejo/ai-review.md, .forgejo/ai-review.yml, AGENTS.md, README.md - Commits considered: 1 - Previous automated review: none found - Findings: 0 </details>

nimmo commented

2026-08-01 08:12:22 +01:00

Author

Owner

Resource-headroom check after the successful automated review:

Nova ROCm VRAM: 68,719,476,736 bytes total; 40,682,377,216 bytes used (59%); GPU idle.
Ollama loaded qwen3.6:35b-a3b-q8_0 at 65,536 context with a 35,524.98 MiB model buffer and 1,280 MiB KV cache.
KV allocation is linear here, so 131,072 context projects to a 2,560 MiB KV cache: approximately 1.25 GiB additional allocation, leaving about 24.9 GiB VRAM headroom at the observed baseline.
Host RAM has 54 GiB available and 28 GiB swap free.

PR #136 itself correctly ran with the trusted 64k base configuration, so this is capacity evidence rather than a claim that CI already launched the 128k runner. The first genuine 128k exercise will be the post-merge PR #134 rerun. No deployment or updater run was performed.

Resource-headroom check after the successful automated review: - Nova ROCm VRAM: 68,719,476,736 bytes total; 40,682,377,216 bytes used (59%); GPU idle. - Ollama loaded qwen3.6:35b-a3b-q8_0 at 65,536 context with a 35,524.98 MiB model buffer and 1,280 MiB KV cache. - KV allocation is linear here, so 131,072 context projects to a 2,560 MiB KV cache: approximately 1.25 GiB additional allocation, leaving about 24.9 GiB VRAM headroom at the observed baseline. - Host RAM has 54 GiB available and 28 GiB swap free. PR #136 itself correctly ran with the trusted 64k base configuration, so this is capacity evidence rather than a claim that CI already launched the 128k runner. The first genuine 128k exercise will be the post-merge PR #134 rerun. No deployment or updater run was performed.

nimmo merged commit a90a87ae20 into main

2026-08-01 08:13:21 +01:00

nimmo referenced this pull request from a commit

2026-08-01 08:13:23 +01:00

Merge pull request 'ci: raise AI review context to 128k' (#136) from issue-135-ai-review-context-128k into main