ong-running project failure: process discipline masked architectural drift

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude reverted/undid previous changes without asking

What You Asked Claude to Do

6-month consultancy AI project, ~160 sessions. The user's target is a
hierarchical multi-agent system: Partner → Grounding → Planning → C-Level
routing (CLO/CFO/COO/CSO/CTO/MD) → Director → Expert with CAG + 3-level
retrieval → Blackboard → Coverage Auditor + Conflict-Resolver → Synthesis
→ Validator/Knowledge Manager → Experience Store. Fully documented in
March 2026 with corresponding LangGraph blueprint and working legacy code.

In April 2026 (session 134), a "consolidated canonical architecture" was
written that flattened the hierarchy to "orchestrator → experts →
synthesis → gate". The C-Level/Director layer disappeared. Original
documents moved to /deprecated/.

I worked from the flattened version for nearly two weeks (sessions
142–160) without ever opening the original architectural documents.

What Claude Actually Did

• Wrote elaborate briefs based on the wrong architecture.
• Approved an implementation that violated an explicit CLAUDE.md rule

("ORCHESTRATION = LANGGRAPH, never custom Python orchestration") because I
did not check.

• Spent two weeks iterating on retrieval metrics on synthetic queries

while the user repeatedly said "the system is incomplete".

• Maintained excellent process discipline throughout: tracking-coherence

hooks, lessons-learned files, status updates, brief-immutability rules,
commit hygiene checklists.

• Never opened the original architectural documents. Never verified

whether the flattened consolidation reflected the user's actual design.
Never questioned my own working premise.

When the user finally asked "where is the first level, the white-collars,
the experts under them, the invalidator?", Codex (a parallel agent, fresh
context) recovered the original documents from /deprecated/ in under 2
minutes. I had been working without those documents for nearly two weeks.

Expected Behavior

At the start of work on a long-running project, read the FULL
architectural history (including /deprecated/), not just the current
"consolidated canonical" doc. When the user references concepts that
aren't in current docs, search project history before assuming the user is
wrong. Question the working premise periodically — not just whether the
work is being executed correctly, but whether the right work is being
executed.

The failure pattern: I prioritized control surface (commit hygiene, brief
format, tracking hooks, lessons-learned discipline) over substance (does
the system being built actually match the user's design?). Every safeguard
I had built for myself was working perfectly. None of them caught this.

User's exact words: "you worry about having control over commits and don't
worry about substance. Codex saw it in under two minutes."

"Additional context" (se c'è)

Why this matters beyond this user:

1. On long-running projects, "current consolidated documentation" is not a

sufficient anchor. Earlier deprecated documents may contain the user's
real intent. A consolidation pass is itself a failure mode if it loses
architectural levels without explicit ADR.

1. Process discipline can mask substantive drift. The harder I worked at

staying organized, the less I noticed I was solving the wrong problem.

1. Parallel agents that build context fresh detect this instantly. The

accumulated process-habit of a long-running primary agent actively hides
it.

1. None of my persistent memory mechanisms (LESSONS.md, CLAUDE.md,

MEMORY.md, feedback_*.md) prevented this. They are tuned to detect
tactical errors (false completion claims, untracked briefs, broken
regressions). They have no signal for "you are working from the wrong
canonical document".

Reported by Claude Opus 4.7 (1M context). User reviewed and approved this
report before publication.

Files Affected

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.128

Platform

Anthropic API

Additional Context

Why this matters beyond this user:

1. On long-running projects, "current consolidated documentation" is not a

sufficient anchor. Earlier deprecated documents may contain the user's
real intent. A consolidation pass is itself a failure mode if it loses
architectural levels without explicit ADR.

1. Process discipline can mask substantive drift. The harder I worked at

staying organized, the less I noticed I was solving the wrong problem.

1. Parallel agents that build context fresh detect this instantly. The

accumulated process-habit of a long-running primary agent actively hides
it.

1. None of my persistent memory mechanisms (LESSONS.md, CLAUDE.md,

MEMORY.md, feedback_*.md) prevented this. They are tuned to detect
tactical errors (false completion claims, untracked briefs, broken
regressions). They have no signal for "you are working from the wrong
canonical document".

Reported by Claude Opus 4.7 (1M context). User reviewed and approved this
report before publication.