fix: Division by zero protection is ineffective#2
Conversation
…udget totals are small but non-zero. Both base_total and adjusted_total are updated to use this reasonable minimum. The change keeps the adaptive scaling bounded as intended.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5b8c405a53
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
|
||
| fn apply(&self, budget: &[ContextSourceConfig]) -> Vec<ContextSourceConfig> { | ||
| let base_total = budget.iter().map(|entry| entry.max_pct).sum::<f32>().max(f32::EPSILON); | ||
| let base_total = budget.iter().map(|entry| entry.max_pct).sum::<f32>().max(1.0); |
There was a problem hiding this comment.
Clamping base_total to 1.0 changes semantics for valid configs where budget percentages sum to less than 1.0 (allowed by context validation), so scaling is no longer anchored to the real pre-adaptation total. In those cases, adaptation can inflate per-source caps and increase total allocated share beyond what the config specified, which breaks the invariant this function is supposed to preserve and can materially change context allocation behavior.
Useful? React with 👍 / 👎.
- Plugin should_block checked in execute_tool_request before SecurityPolicy - PluginRegistry loaded at session start in PraxisRuntime (cli/core, daemon, tests) - Curator::run_cycle() called in execute_reflect after learning, gated by LiteCapability::Curator - LiteMode gains disable_curator field + LiteCapability::Curator variant - PRAXIS_GAPS.md updated: #2 code exec, #3 plugin wiring, #6 curator wiring, #9 voice/TTS, #12 tool policy, #29 model switching — all marked with real impl details - cargo check clean (0 errors)
New modules implementing all gaps from GAP_ANALYSIS_HERMES_OPENCLAW.md: - #2 Sessions spawn (src/session/) — programmatic session creation for kanban workers - #3 Curator (already existed, verified wiring) - #4 429 fallback (src/backend/retry.rs) — exponential backoff with jitter - #5 Cron extensions (src/tools/cron_ext.rs) — no_agent script mode, wake_gate - #6 Live Canvas (src/canvas/) — streaming HTML workspace surface - #7 LanceDB memory (src/memory/lance.rs) — vector-backed semantic recall - #8 Auto-reply (src/messaging/auto_reply.rs) — rate-limited proactive responses - #9 i18n (src/i18n/) — 9 languages, builtin + custom translations - #10 Plugin marketplace (src/plugins/marketplace.rs) — remote discovery + install Also added: - Discord/Slack inbound polling (src/messaging/inbound.rs) - Plugin registry stub (src/plugins/registry.rs) Fixed all compiler warnings: - Unused imports removed (session, kanban, auto_reply, lance, marketplace) - Dead code annotated with #[allow(dead_code)] for API surface items - Deprecated tungstenite::read_message -> read - Unnecessary mut bindings cleaned - cargo fmt applied
1 participant
Description
Replacing f32::EPSILON with 1.0 prevents extreme scale factors when budget totals are small but non-zero. Both base_total and adjusted_total are updated to use this reasonable minimum. The change keeps the adaptive scaling bounded as intended.
Finding: CRITICAL — Division by zero protection is ineffective
File:
src/context/adaptive.rsRisk Assessment
Validation
cargo testGenerated by Archiview · Autonomous AI Audit