[BUG] Long-running session degradation since v2.1.139, compounded by silent fast-mode model swap in v2.1.142 — no changelog, breaks pinned-model workflows, quantified waste

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

---
Title: Long-running session degradation since v2.1.139, compounded by silent fast-mode model swap in v2.1.142

---

## Summary

Starting with Claude Code v2.1.139, long-running sessions exhibit progressive behavioral
degradation that was not present in v2.1.138. This was then compounded by v2.1.142, which
silently changed the fast-mode default model from Opus 4.6 to Opus 4.7 without any
changelog entry or user notification.

The combination produces sessions that start well but progressively lose tool-calling
discipline, begin hallucinating capabilities, and enter unproductive loops — consuming
tokens and user time without producing useful output.

## Symptoms

Degradation patterns that emerge in long-running sessions (not short ones):

• Synthesis-over-verification drift: Model confidently asserts facts without calling

available tools to check them — worsens as session length increases

• Phantom capability claims: Claims to have invoked tools/features it never actually called
• Failure loop blindness: Retries the same broken action 3+ times without changing approach
• Reduced tool-calling discipline: Reasons about what a tool would return rather than

calling it

• Bulk-assertion without evidence: Marks multiple verification criteria as "complete" in

a single text block with no interleaved tool calls

• Context compaction artifacts: After compaction events, behavioral rules from system

prompts are partially forgotten or deprioritized

## Version Timeline

| Version | Released | Status |
|---------|----------|--------|
| 2.1.138 | 2026-05-09 | Last known-good for long-running session stability |
| 2.1.139 | 2026-05-11 | First version exhibiting long-session degradation |
| 2.1.140 | 2026-05-12 | Degradation continues |
| 2.1.141 | 2026-05-13 | Degradation continues |
| 2.1.142 | 2026-05-14 | Silent fast-mode model swap (Opus 4.6 → 4.7) — compounds the issue |
| 2.1.143 | 2026-05-15 | Both issues present |

## Quantified Impact (Max Subscriber, since v2.1.139)

| Metric | Value |
|--------|-------|
| Duration of degradation | 6 days (2026-05-11 to 2026-05-16) |
| Sessions in affected window | ~850 |
| Total session data | 46.9 MB |
| Estimated tokens consumed | ~4.1M |
| Estimated tokens wasted (degraded behavior) | ~1.4M |
| Per-token cost equivalent | ~$73 |
| User hours spent on corrections/debugging | ~7 hours |
| Quality score before v2.1.139 | 7–8/10 (consistent, monitored) |
| Quality score during affected period | 4.5/10 (sustained, monitored) |
| Sessions requiring full restart (context exhaustion) | 4+ |

Waste rate is conservative (35%) — large sessions showed 50–65% waste (hallucinations,
repeated identical failures, user corrections, forced restarts). At $200/month subscription
cost, this represents over a third of a billing cycle's value lost to degraded output.

What Should Happen?

What Should Happen?

1. Long-running sessions should maintain consistent tool-calling discipline and behavioral

rule adherence throughout the entire session, including after context compaction events.
A rule present in the system prompt at turn 1 should still be followed at turn 50.

1. Model version changes in any mode (fast, standard) should be documented in release notes

and visible to the user — not silently swapped server-side.

1. Context compaction should preserve the full fidelity of system prompts and CLAUDE.md

operational rules. If compaction necessarily drops context, system-level instructions
should be the last thing dropped, not the first.

1. Paid subscribers ($200/month Max) should have a stable, first-class mechanism to pin

model versions — not undocumented environment variables discovered by reverse-engineering
regressions.

1. Behavioral regressions of this magnitude (quality drop from 7-8/10 to 4.5/10) should be

caught by Anthropic's internal testing before release, not discovered by subscribers
debugging wasted sessions.

Error Messages/Logs

Steps to Reproduce

Steps to Reproduce

1. Run Claude Code v2.1.139+ with fast mode enabled
2. Start a session with structured operational rules (system prompts, CLAUDE.md with

explicit verification requirements)

1. Work through a multi-phase task requiring 10+ tool calls
2. Observe progressive degradation: early turns follow rules, later turns drift
3. After context compaction, observe partial loss of behavioral rules
4. Compare identical workflow on v2.1.138 — no degradation

## Workaround

```bash
export CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1

This addresses the 4.7 compound factor but does NOT fix the underlying long-session
degradation introduced in v2.1.139.
degradation introduced in v2.1.139.

Environment

• Claude Code versions affected: 2.1.139 through 2.1.143
• Last known-good version: 2.1.138
• Subscription: Max ($200/month)
• Platform: macOS (Apple Silicon)
• Usage pattern: Long-running structured agentic workflows with custom hooks,

quality monitoring, and explicit model-version requirements

Consumer Terms Context

Anthropic's Consumer Terms §12 permits service modifications "at any time without notice."
However, §6(5) requires 30 days notice for fee increases. A silent quality degradation
that wastes ~1.4M tokens and ~7 hours of user time over 6 days is economically equivalent
to a fee increase for subscribers whose output quality depends on session stability and
model version consistency.

Request

1. Investigate long-session regression in v2.1.139 — what changed in context handling,

compaction, or session management that causes progressive behavioral drift?

1. Add changelog entries for model version changes in any mode
2. Document CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE in official docs
3. Provide stable model version pinning as a first-class setting
4. Advance notification for behavior-affecting changes — especially for Max subscribers

at $200/month building production workflows on Claude Code

1. Review the affected sessions listed above to verify degradation patterns server-side

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.138

Claude Code Version

2.1.139 through 2.1.143

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

## Affected Session IDs (available for Anthropic review)

All sessions >100KB since v2.1.139, sorted by size descending:

| Session ID | Size | Timestamp (ET) |
|-----------|------|----------------|
| ec1c8982-e331-4828-9f6c-a580fdb61d6f | 6.4 MB | 2026-05-11 21:26 |
| 7f18bd12-8be8-4f63-a681-54b3d6d92e94 | 5.1 MB | 2026-05-14 14:52 |
| d0001021-2d9f-4d02-9898-b394c68314e1 | 4.0 MB | 2026-05-15 23:20 |
| ad803cc4-e42d-43ad-9ed1-fdf1c49dcf77 | 3.0 MB | 2026-05-16 02:19 |
| 85d62600-100e-4837-b00c-6fc21de1b360 | 2.6 MB | 2026-05-14 01:13 |
| 21be96d8-e609-41d6-b9be-0c3eaf982c85 | 2.3 MB | 2026-05-12 23:57 |
| c9e86ed7-c735-4b0a-841f-e9adc1590fa9 | 1.8 MB | 2026-05-14 22:39 |
| 17cba326-6301-4845-aefe-6ec1458523f7 | 1.8 MB | 2026-05-12 18:20 |
| 575dcc2a-2be8-4495-a3f2-c85e3eee9506 | 1.8 MB | 2026-05-16 01:51 |
| 629d0e4f-5319-4cd3-b012-7662b913d9e8 | 1.7 MB | 2026-05-16 20:31 |
| 165be4d2-a6b3-4ac4-80b9-adae81f51170 | 1.5 MB | 2026-05-13 00:28 |
| 7b7f4c91-e085-47d2-a895-6ea006918e55 | 1.5 MB | 2026-05-16 20:50 |
| fd394b6f-92a3-4b57-8bc0-1ccb7818f416 | 984 KB | 2026-05-15 23:47 |
| 2b9aa28a-85b2-4d0f-b5a0-fda81586ed12 | 882 KB | 2026-05-14 21:02 |
| dcaf1755-29a2-429d-9c0a-25906b01b5f7 | 749 KB | 2026-05-11 23:37 |
| 3790aa25-c6c8-412c-a076-77a8bbf4147c | 714 KB | 2026-05-14 15:42 |
| 892a0bb8-f8f9-48f8-859f-d683a88450c1 | 646 KB | 2026-05-12 09:53 |
| dc1fe9d2-bf94-44e7-bc73-5d44a4ef7fbe | 450 KB | 2026-05-16 17:12 |
| b1737c6c-4e96-481f-a436-fb20e3950f97 | 421 KB | 2026-05-15 21:29 |
| 7ccf7aa2-4a67-4355-a73f-cf0580596eaf | 297 KB | 2026-05-14 15:57 |
| 4bc122e4-e651-4098-9051-6eef6ca323e1 | 226 KB | 2026-05-14 22:41 |

Session 7b7f4c91 is the primary evidence session where degradation was identified,
debugged, and root causes traced. Session ec1c8982 (6.4 MB, May 11) is the earliest
large session after v2.1.139 — the first one showing the pattern.