fix(agents): oldest-first tool result compaction with configurable ratios

[Cstewart-HC]

Summary

Fix for Issue #1 from the stream investigation report — per-iteration tool-result compaction was destroying agent context in long loops.

Changes

A) Reverse compaction direction: Compact oldest results first instead of newest. The model needs its most recent tool outputs to maintain coherent operation in long agent loops.

B) Configurable ratios via ToolsConfig:

• tool_result_compaction_ratio (default 75, set to 0 to disable)
• preemptive_overflow_ratio (default 90)

C) Minimum iteration guard: compaction_min_iterations (default 3) — skip compaction for the first N iterations to prevent premature context destruction in short loops.

Behavior

All defaults match the previous hardcoded values for backward compatibility. No moltis.toml changes required for existing deploys.

Files changed (7 files, +71 −20)

• crates/agents/src/runner/mod.rs — reverse iteration order, read config values
• crates/agents/src/runner/streaming.rs — pass config through
• crates/agents/src/runner/non_streaming.rs — pass config through
• crates/agents/src/runner/tests/helpers.rs — test helper update
• crates/config/src/schema/tools.rs — 3 new config fields with defaults
• crates/config/src/template.rs — template comments
• crates/config/src/validate/schema_map.rs — schema key registration

Validation

• cargo fmt --check ✅
• cargo clippy --workspace --all-targets -- -D warnings ✅
• cargo test --workspace ✅ (344 passed, 1 timing-flaky unrelated failure)

Session context

Built during an extended investigation session into stream compaction issues. Redacted — no secrets or tokens in session.

[greptile-apps]

Greptile Summary

This PR resolves the oldest-first compaction direction bug in long agent loops, makes the compaction and overflow thresholds configurable via ToolsConfig, and adds a minimum-iteration guard that correctly delegates overflow checking even when compaction is suppressed. All five previously flagged issues (ratio=0 disable, off-by-one semantics, missing validation, test coverage, overflow suppression in early iterations) are addressed in this revision. Remaining findings are all P2: a literal newline inside the diagnostic format string, a misleading "compaction disabled" message that fires during the early-iteration suppression window, and missing upper-bound (>100) validation for the two percentage fields.

Confidence Score: 5/5

Safe to merge — all previously flagged blockers are resolved; only P2 polish items remain.

All five P0/P1 findings from prior review rounds have been correctly addressed: ratio=0 disable path, off-by-one in min-iterations semantics, ordering validation in check_semantic_warnings, oldest-first test coverage, and overflow still firing during the early-iteration window. Remaining findings are three P2 issues (diagnostic format string, misleading error message text, missing >100 validation) that do not affect correctness or safety.

crates/config/src/validate/semantic.rs (format string), crates/config/src/schema/tools.rs (upper-bound validation)

Important Files Changed

Filename	Overview
crates/agents/src/runner/helpers.rs	Reverses compaction direction to oldest-first, adds configurable ratio/overflow params, handles ratio=0 disable path, and retains post-compaction overflow guard — all correct.
crates/agents/src/runner/non_streaming.rs	Wires in new config fields; effective_ratio=0 during early iterations correctly delegates overflow checking to enforce_tool_result_context_budget without skipping it.
crates/agents/src/runner/streaming.rs	Mirrors non_streaming.rs changes exactly; same effective_ratio guard pattern, no issues.
crates/config/src/schema/tools.rs	Adds three config fields with correct defaults; missing upper-bound (>100) validation for the two percentage fields.
crates/config/src/validate/semantic.rs	Adds ordering constraint (overflow_ratio > compaction_ratio); diagnostic message has a literal newline/indentation inside the format string.
crates/agents/src/runner/tests/basic.rs	Adds 12 new unit tests covering oldest-first direction, edge cases (already-compacted, too-small, zero tokens, zero context window), and enforce_budget path (ratio=0, overflow, compaction fire, no tool results).
crates/config/src/validate/schema_map.rs	Registers the three new config keys; straightforward and correct.
crates/config/src/template.rs	Template comments added for all three new fields with correct defaults; no issues.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["enforce_tool_result_context_budget\n(messages, tool_schemas, context_window,\ncompaction_ratio, overflow_ratio)"] --> B{context_window == 0\nor no tool results?}
    B -- Yes --> C[return Ok]
    B -- No --> D{compaction_ratio == 0?}
    D -- Yes --> E[compute overflow_budget\ncheck current_tokens]
    E --> F{current_tokens >\noverflow_budget?}
    F -- Yes --> G[return Err\nContextWindowExceeded\n'compaction disabled']
    F -- No --> C
    D -- No --> H[compute compaction_budget\nand overflow_budget]
    H --> I{current_tokens >\ncompaction_budget?}
    I -- Yes --> J["compact_tool_results_oldest_first_in_place\n(oldest → newest order)"]
    J --> K[log debug]
    I -- No --> L[re-estimate post_compaction_tokens]
    K --> L
    L --> M{post_compaction_tokens >\noverflow_budget?}
    M -- Yes --> N[return Err\nContextWindowExceeded\n'after compaction']
    M -- No --> C

    subgraph Callers["Callers (non_streaming / streaming)"]
        R["iterations > compaction_min_iterations?"]
        R -- Yes --> S["effective_ratio = compaction_ratio"]
        R -- No --> T["effective_ratio = 0\n(overflow still checked)"]
    end
    S --> A
    T --> A

Loading

_{Reviews (5): Last reviewed commit: "fix(agents): don't skip overflow check d..." | Re-trigger Greptile}

[Cstewart-HC]

@greptileai review

[Cstewart-HC]

@greptileai review

[Cstewart-HC]

@greptileai review

[Cstewart-HC]

@greptileai review