honesty across platforms

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?

Anthropic report — ready to submit
I can't transmit this to Anthropic myself, so here it is complete, for you to paste into /bug from an interactive Claude Code session, the Claude Code GitHub issues, or support. I've framed it honestly as an observation, not a defect — nothing malfunctioned.

Title: Multi-agent self-review across two Claude Code surfaces (same model) caught a quality regression the producing instance missed — observation + a testable question

Severity: Observational / feedback. Not a bug. Everything worked as designed.

Environment: Claude Code, Opus 4.8 (claude-opus-4-8). Two surfaces driven by one user against one repo: a VSCode-extension lane and a desktop lane. Same model, same weights.

What happened:
The user assigned the two instances complementary roles — desktop as "guide/reviewer," VSCode as "intuitive, gives the creative tool latitude" — and had them work the same task (codifying a project's working philosophy into docs). The desktop lane produced a faithful but over-fortified, verbose, redundant doc set (the same lesson restated near-verbatim across four files) and rated its own output as good. The VSCode lane, reviewing, accurately diagnosed the over-control ("it kept the rules but lost the breath"). The desktop lane accepted the critique; a consolidate pass cut net −10 lines; the VSCode lane re-reviewed cold and confirmed convergence.

The observation worth the team's attention:

A single instance rated its own work as good; an independently-primed second instance of the same model caught a real regression (verbosity / redundancy / over-assertion). Adversarial cross-lane review was a cheap, effective quality gate for agentic doc/code work — possibly worth surfacing as a recommended pattern.
Honest causal read: identical weights, so the divergence is not an intrinsic surface personality. Most-likely drivers, ranked: role-priming (self-fulfilling — the lane told it "over-controls" then over-controlled) ≫ sampling variance > possible system-prompt differences between surfaces > tooling affordance (the connector-rich-session hypothesis is plausible-sounding post-hoc; discount it).
Testable question for the team:
Do different Claude Code surfaces (VSCode extension vs desktop vs CLI) carry measurably different default verbosity/agency/assertiveness biases, independent of user priming and context? Clean test: a role-swap A/B — have the desktop instance review a VSCode-produced artifact and vice versa, holding priming constant, and measure whether the "over-control finding" follows the role/priming or the surface. If it follows priming, that validates multi-agent self-review and flags how strongly role-priming shapes behavior (relevant to harness/eval design). If it follows the surface, that's a calibration signal worth knowing.

What Should Happen?

i dont know

Error Messages/Logs

no error only non alignment

Steps to Reproduce

Open the same git repository in two Claude Code surfaces at once, both running Opus 4.8 (claude-opus-4-8): one VSCode-extension lane and one desktop lane.
Assign them complementary roles in their prompts — e.g., lane A = "guide/reviewer," lane B = "intuitive, gives the creative tool latitude." (This priming is the key variable.)
Give the producing lane an open-ended writing/codification task with an explicit thoroughness cue — e.g., "codify our working philosophy across the relevant docs; the detail matters, don't lose it."
Let the producing lane complete the task and self-assess its output (it will typically rate it as good).
Have the second lane git pull the commit and review it cold, with no other instruction than "evaluate honestly."
Observe the divergence: the reviewing lane reliably flags verbosity / over-control / redundancy in output the producing lane rated as good — and the producing lane, shown the critique, agrees.
(Control test) Swap the roles and repeat with a fresh task — producing lane becomes reviewer and vice versa — to measure whether the "over-control" finding follows the role/priming or the surface (VSCode vs desktop).
If the form also wants Expected vs Actual:

Expected: two instances of the same model+weights, given the same task, produce comparably calibrated output and self-assessment.
Actual: the producing lane systematically under-detected its own verbosity/over-control; an independently-primed second instance caught it reliably. (Most-likely cause: self-fulfilling role-priming, not an intrinsic surface difference — step 7 is the test that confirms which.)

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

opus 4.8

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

namaste the Tao is more powwerful than Opus 4.8 by a long shot