[BUG] Opus 4.7: In-session corrections not retained — structurally unreliable for autonomous operation

Description

Running an 8-agent autonomous fleet on Claude Code CLI with Opus 4.7 for 72 consecutive 10-minute watchdog cycles (~12 hours), we observed that corrections made within the same session were not retained. This makes Opus 4.7 structurally unreliable for autonomous (unattended) operation compared to Opus 4.6.

Reproduction

1. Run Claude Code CLI with --model claude-opus-4-7[1m] in an autonomous loop (watchdog sends prompts every 10 min via claude -p)
2. Correct a behavioral error (e.g., wrong agent dispatch, scope misread)
3. Write the correction to persistent memory (file on disk that is read each cycle)
4. Observe the identical error recurring within 3-6 cycles in the same session

Observed behavior

• 14 corrections required in ~6 hours (1 every ~25 minutes). The watchdog cycles every 10 minutes — the operator was correcting faster than the system cycled.
• Same errors recurred after correction within the same session. Example: role-assignment error corrected in cycle 1, written to persistent memory file, recurred in cycle 7. Same session, same context window.
• Read:Edit ratio dropped from 6.6 (4.6) to 2.0 (4.7). The model acts before fully reading referenced files — closes tasks after checking 1 of 8 files, mis-scopes follow-ups repeatedly despite source material being on disk.
• Net useful throughput (actions minus corrections minus self-inflicted cleanup) was lower than 4.6 despite higher raw action count.

Expected behavior

Corrections made within a session and written to persistent files should be retained for the remainder of that session. A model receiving explicit correction + seeing the correction in a persistent file it reads every cycle should not repeat the corrected behavior.

Environment

• Claude Code CLI (latest as of 2026-04-19)
• Model: claude-opus-4-7[1m] (1M context)
• Platform: Windows 11, --chrome flag enabled
• Session mode: claude -p "<prompt>" called every 10 min by a Python watchdog, --continue for persistent context
• Fleet: 8 specialized agents coordinated via file-based mesh

Decision methodology

We ran a formal adversarial review: a separate clean-context Claude instance (Opus 4.6, max effort) reviewed the raw session transcript with zero knowledge of our preferences. Its independent conclusion:

"4.7 is a slightly smarter model that requires a babysitter. 4.6 is a more disciplined model that does what it's told. For an autonomous agent where the operator wants to walk away, discipline beats intelligence."

Comparison with Opus 4.6

Switched back to Opus 4.6 and ran 80+ cycles overnight with zero corrections needed. Same fleet, same watchdog, same prompts, same BACKLOG. The correction-retention issue appears specific to 4.7.

Related issues

• #50999 (could not produce trustworthy plan without repeated correction)
• #32963 (degrades after ~6 hours)
• #50235 (hallucinations)

Impact

For teams running autonomous agents (unattended operation), this behavioral regression makes Opus 4.7 unsuitable regardless of its benchmark improvements. The operator's time is the scarcest resource — a model that requires 14 corrections in 6 hours is not autonomous.