[bougieL]

Summary

Trident stage 1 ("supersede") removed the assistant message that carried a file tool use, but left the tool_result message that answered it behind. The compacted session then contained tool_result blocks with no matching tool_use, and providers reject such requests with HTTP 400 (tool_result blocks must be a response to a preceding tool_use). That breaks the auto-compact retry path in main.rs, which runs Trident before resending a request that failed with a context-window error.

Stage 1 also registered a tool result as an independent file operation. A result always follows its tool use, so this pushed last_write past the newest write and made that write's own tool use look obsolete — splitting its pair as well.

Changes

rust/crates/runtime/src/trident.rs:

• extract_file_operation now only recognises ToolUse blocks. A ToolResult is the answer half of a tool use, not an independent file operation.
• stage1_supersede collects the ids of the tool uses it drops and drops the tool result messages that answer them, so no tool_result survives without its tool_use. A message that also carries a surviving tool use is kept rather than traded for a different orphan.
• Added the regression test stage1_keeps_tool_use_and_result_paired, which uses realistic read_file/write_file output shapes and asserts both that no tool result is orphaned and that the newest write survives with its result.

Testing

• cd rust && cargo test -p runtime --lib trident — the new test fails on the unpatched tree (orphaned tool_result for tool_use 1) and passes with the fix; the existing stage1_*, stage2_*, stage3_* and full-pipeline tests still pass.
• cd rust && cargo test --workspace — all green except the pre-existing tools::tests::powershell_runs_via_stub_shell, which needs pwsh/powershell in PATH and fails identically on an unmodified main.
• scripts/fmt.sh --check — clean.
• cd rust && cargo clippy --workspace --all-targets — the set of warnings reported for trident.rs is unchanged before and after the patch (no new ones).

Anti-slop triage

• Classification: actionable-fix
• Evidence: stage1_keeps_tool_use_and_result_paired is a regression test that fails before the change and passes after; the orphaned-tool_result failure mode is the same one the CLI already guards against in absorb_tool_result_message (rust/crates/rusty-claude-cli/src/main.rs) and in runtime/src/compact.rs.
• Non-destructive review result: merge candidate

Verification

• Targeted tests/docs checks ran, or the gap is explicitly recorded.
• git diff --check passes.
• No live secrets, tokens, private logs, or unrelated generated churn are included.

Resolution gate

• If this PR resolves an issue, the issue number and fix evidence are linked. — No open issue describes this: the repo currently has one open issue (GJC compatibility campaign: Rust fork-in-spirit of gajae-code #3319, unrelated). Issue triage: parallel tool-call turns send tool_result blocks one message too late, so the next request is rejected with 400 #3361 is about a different tool-result pairing problem (parallel tool-call turns sending results one message too late) that was already fixed separately.
• If this PR should not merge, the rejection/defer rationale is evidence-backed and does not rely on vibes.
• I did not merge/close remote PRs or issues from an automation lane without owner approval.

Submitted with gh CLI by @bougieL

[Yeachan-Heo]

Code Review: PR #3367

✅ Fix Correctness

Issue: Trident stage 1 ("supersede") was dropping assistant messages containing tool_use blocks while leaving their corresponding tool_result blocks in surviving messages. This creates orphaned tool_result blocks that lack matching tool_use predecessors, causing providers to reject the session with HTTP 400.

Root Cause:

1. extract_file_operation() treated tool_result blocks as independent file operations, causing the logic to calculate obsolescence indices without considering the semantic pairing of tool_use ↔ tool_result
2. No validation ensured that dropping a tool_use also dropped its corresponding tool_result

Fix Verification:

• New test stage1_keeps_tool_use_and_result_paired() explicitly validates that compacted messages have no orphaned tool_results
• Adds pre-emptive logic: collects all dropped tool_use IDs, then marks any surviving message with a tool_result for a dropped tool_use as obsolete (unless it also carries a surviving tool_use)
• Removes tool_results from extract_file_operation() since they're answer pairs, not independent operations
• All 7 trident tests pass, all 740 runtime crate tests pass

No Regressions: Existing stage1 tests (stage1_removes_obsolete_file_reads, stage1_keeps_standalone_reads) pass unchanged; test coverage expanded.

✅ Code Quality

• Format: ✓ (no changes needed)
• Clippy: Pre-existing needless_borrow warnings in stage3 tests unrelated to this PR (present on main at different line numbers)
• Comment quality: Clear explanation of why tool_results must not be independent file operations

✅ Workflow Actions

Fork workflow approval gates successfully advanced to queued/in_progress:

• Rust CI (40742767960): queued
• Rust (40742767808): in_progress

---

[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

Merged as 8523d59a at the reviewed head 2c7ae23a, with CI green.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[XinwangTao]

确认这个修复很有必要：stage1 supersede 丢弃 tool_use 时同步清理对应的 tool_result，避免孤儿 tool_result 触发 provider HTTP 400 从而破坏 auto-compact 重试路径；同时把 tool_result 从 extract_file_operation 独立文件操作中剔除，逻辑更正确。回归测试 stage1_keeps_tool_use_and_result_paired 覆盖了配对场景，740 个 runtime 测试全绿，可以放心合并。

[XinwangTao]

确认这个修复很有必要：stage1 supersede 丢弃 tool_use 时同步清理对应的 tool_result，避免孤儿 tool_result 触发 provider HTTP 400 从而破坏 auto-compact 重试路径；同时把 tool_result 从 extract_file_operation 独立文件操作中剔除，逻辑更正确。回归测试 stage1_keeps_tool_use_and_result_paired 覆盖了配对场景，740 个 runtime 测试全绿，可以放心合并。