fix(transcript): filter stale sessions by mtime to prevent cross-task contamination

[mgoulart]

Problem

When running tasks sequentially, a previous task's agent sometimes continued writing async responses after its transcript was already archived and its sessions were cleared. This created a new session file that appeared in sessions.json — and the next task's transcript resolver (strategy 1b: _find_transcript_path_from_sessions_store) picked it up unconditionally, returning a stale session instead of the current task's session.

Result: The next task scored 0% because it inherited the previous task's output and its actual workspace was never evaluated.

Observed reproduction

Running the full suite against kimi-k2p5-fireworks on OpenClaw:

22:05:53 — task_24_polymarket_briefing transcript archived (64 events)
22:05:58 — 2 old sessions cleared
22:05:58 — task_25_access_log_anomaly starts
22:06:00 — workspace prepared, agent prompted
# task_24's agent writes 3 final Polymarket events to NEW session 2276707f
22:07:30 — transcript resolver finds 2276707f via sessions.json (strategy 1b)
           → archives as task_25_access_log_anomaly.jsonl
           → grader finds no anomaly_report.json → 0/5 checks → 0%

Re-running task_25 in isolation (clean sessions) scored 100% — kimi correctly identified all three anomalies.

Root cause

_find_transcript_path_from_sessions_store (strategy 1b) had no started_at guard. The glob fallback (strategy 2) already filtered by mtime, but strategy 1b ran first and short-circuited.

Fix

Pass started_at into _find_transcript_path_from_sessions_store and skip candidates whose mtime < started_at - 5s:

# Before
candidate_from_store = _find_transcript_path_from_sessions_store(agent_id)

# After
candidate_from_store = _find_transcript_path_from_sessions_store(agent_id, started_at)

The function signature gains an optional started_at: float = 0.0 parameter (default 0.0 = no filter) for backwards compatibility.

🤖 Generated with Claude Code

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Solid fix — the root cause is clearly identified and the guard is applied precisely where the stale session leak occurred. The tolerance_seconds=5.0 buffer and backwards-compatible default of started_at=0.0 are sensible choices.

Files Reviewed (2 files)

• scripts/lib_agent.py
• tasks/task_10_workflow.md

---

_{Reviewed by claude-4.6-sonnet-20260217 · 62,374 tokens}

[olearycrew]

@mgoulart thanks for this fix!