Summary

• Tighten the shared agent grounding contract so numeric evidence must be copied exactly from observed sources.
• Cover counts, sizes, dates, timestamps, durations, currencies, percentages, quotas, and ids.
• Add prompt regression coverage for the main dynamic prompt path and the narrow sub-agent renderer.

Problem

• The agent can paraphrase, round, or recompute numerical claims after reading tool output.
• That can produce wrong storage amounts, counts, durations, and dates even when the source data was available.
• Issue Agent hallucinates numbers — storage, counts, durations #3603 explicitly allows a scoped system-prompt rule as a first fix for this failure mode.

Solution

• Extend the existing central GROUNDING_BODY prompt contract instead of adding a separate post-processor.
• Require exact preservation of numeric evidence from tool results, user messages, or cited memory.
• Instruct agents not to round, convert units, rewrite relative times, or recalculate unless the user asks and the calculation is shown from observed values.
• Assert the new contract is present in dynamic prompts and inherited by sub-agent prompts.

Submission Checklist

• Tests added or updated (happy path + at least one failure / edge case) per Testing Strategy — added prompt regression coverage for exact numeric evidence and sub-agent inheritance.
• Diff coverage ≥ 80% — changed lines are covered by focused Rust prompt tests; full merged diff-cover runs in CI.
• Coverage matrix updated — N/A: shared prompt-contract change, no feature row added/removed/renamed.
• All affected feature IDs from the matrix are listed in the PR description under ## Related — N/A: no matrix feature IDs affected.
• No new external network dependencies introduced (mock backend used per Testing Strategy).
• Manual smoke checklist updated if this touches release-cut surfaces (docs/RELEASE-MANUAL-SMOKE.md) — N/A: no release-cut manual smoke surface changed.
• Linked issue closed via Closes #NNN in the ## Related section.

Impact

• Runtime/platform impact: core agent prompt text only; no frontend, persistence, RPC, or migration changes.
• Performance impact: negligible, two additional static grounding bullets in the system prompt.
• Compatibility impact: sub-agents inherit the same numeric-grounding rule through the existing shared prompt constant.

Related

• Closes: Closes Agent hallucinates numbers — storage, counts, durations #3603
• Follow-up PR(s)/TODOs: N/A

AI Authored PR Metadata (required for Codex/Linear PRs)

Linear Issue

• Key: N/A, GitHub issue contribution.
• URL: N/A, GitHub issue contribution.

Commit & Branch

• Branch: fix/numeric-tool-output-guardrails
• Commit SHA: 4512d962

Validation Run

• pnpm --filter openhuman-app format:check
• pnpm typecheck
• Focused tests: cargo test --manifest-path Cargo.toml openhuman::agent::prompts::tests -- --nocapture
• Rust fmt/check (if changed): cargo fmt --manifest-path Cargo.toml --all --check; cargo check --manifest-path Cargo.toml
• Tauri fmt/check (if changed): pre-push hook ran cargo fmt --manifest-path app/src-tauri/Cargo.toml --all --check and cargo check --manifest-path app/src-tauri/Cargo.toml

Validation Blocked

• command: N/A after setup.
• error: N/A after setup. Local pre-push initially failed because app dependencies and vendored Tauri submodules were missing; resolved with pnpm install and git submodule update --init --recursive.
• impact: Final push hooks passed. Local environment still warns that Node is v22.22.3 while the repo requests >=24.0.0.

Behavior Changes

• Intended behavior change: agents are instructed to preserve numeric claims exactly from observed evidence instead of paraphrasing, rounding, unit-converting, or recomputing them by default.
• User-visible effect: answers grounded in tool output should be less likely to misstate counts, sizes, timestamps, dates, durations, and quotas.

Parity Contract

• Legacy behavior preserved: existing tool-use, anti-fabrication, action, and output-style grounding rules remain in the same shared prompt contract.
• Guard/fallback/dispatch parity checks: tests assert the rule is appended through SystemPromptBuilder::build and the narrow sub-agent renderer.

Duplicate / Superseded PR Handling

• Duplicate PR(s): None found for #3603 before publishing.
• Canonical PR: This draft PR.
• Resolution (closed/superseded/updated): N/A.

Summary by CodeRabbit

• Bug Fixes

  • Enhanced numeric evidence handling to preserve exact values without rounding or unit conversion unless explicitly requested; conflicting numeric sources are now identified rather than resolved with approximations.

• Tests

  • Added verification tests for numeric evidence preservation in system prompts.