[FIX] read_file failures and sequential read loops regressed after #782/#757

[dimokru]

Problem (observed in real usage)

The agent is again failing on read_file and falling into repeated sequential read loops, after fixes #782 and #757 were closed as completed.

Evidence (aggregated, anonymized)

• Frequency: 26 failures across the last 7 days of sessions (+73% vs the prior introspection cycle)
• Repeated-run loops: 10 sessions had ≥5 consecutive read_file calls (max 8 consecutive)
• Tool involved: read_file
• Failure shape: file-not-found / path errors, followed by repeated single-file reads instead of using repo_map or batch read_file
• Regression baseline: [FIX] read_file failures and repeated sequential reads block file access #782 and [UX] read_file still loops sequentially instead of using repo_map / batch reads #757 were closed completed; failure count increased, indicating the fixes did not hold

Impact on real tasks

Sequential read loops waste turns and context budget and prevent the agent from exploring codebases efficiently. File access failures block reading configs, logs, and source files, which stalls almost every coding or debugging task.

Proposed direction

1. Root-cause why read_file failures rose after the prior fixes: revisit path resolution, batch-read routing, and repo_map prompting.
2. Make repo_map / batch read_file usage a loop_guard-enforced fallback after 3 consecutive read_file calls.
3. Record a realized-impact ledger verdict for [FIX] read_file failures and repeated sequential reads block file access #782/[UX] read_file still loops sequentially instead of using repo_map / batch reads #757 showing the regression.

Value

• Impact: 0.75
• Effort: 0.4
• Priority Score: 1.26

[Lexus2016]

Rejected by evolution-analysis: already-exists — PR #979 merged. read_file is in _SPIRAL_PRONE_TOOLS (agent/tool_guardrails.py:71) with spiral_failure_cap enforcement. Confirmed on main.

[Lexus2016]

Re-opened by realized-impact gate: post-merge verdict was regressed — the fix did not reduce the target signal in real sessions. Needs a focused regression before re-closing. (Introspection cycle 2026-07-18)

[Da-Mikey]

Closing as part of pipeline cleanup — tested tool behaviour and no longer reproducible with current HEAD. Re-open if you hit this again.

[Da-Mikey]

This was flagged by the evolution pipeline for needs-regression verification. After investigation, the tool behaviour is working correctly with current HEAD — no longer reproducible. Closing as part of pipeline cleanup per PR #1181. Re-open if you hit this again.