[BUG] Agent ignored spec across 6 sessions (9 days), repeatedly broke codebase, produced zero usable work

Summary

Over 6 sessions spanning 9 days (March 20–28, 2026), Claude Code agent was tasked with integrating a new UI repo (Finaldash) into an existing Site repo for an admin dashboard + client portal system. The agent repeatedly ignored the provided specification, made destructive mistakes requiring cleanup sessions, ran out of context 3+ times, and ultimately produced zero usable work toward the stated objective.

Session Timeline (90MB+ of transcripts)

Session 1 — March 22 (16K transcript)

• Vercel commit identity confusion
• Agent committing under wrong identity

Session 2 — March 20–25 (23MB transcript, context exhausted)

• User requested client portal setup and admin dashboard streamlining
• Agent built portal pages but introduced fabricated/made-up data and incorrect UI
• User: "a lot of this shit is made up. I dont like that"
• Session ran out of context — required continuation

Session 3 — March 25–27 (53MB transcript, context exhausted 3x)

• Continued from Session 2 via context summary
• Agent produced TWO different admin views causing confusion
• User: "WHY WERE WE IN TWO DIFFERENT ADMIN VIEWS"
• User: "because this is confusing me and its my damn company"
• Portals had incorrect data, wrong tone, fake content
• Session ran out of context MULTIPLE times — required 3+ continuations

Session 4 — March 27 (384K transcript)

• Runway2Roy client website / digital wedding invitation work
• Prior session's agent "made a mess" of the site requiring cleanup

Session 5 — March 27 (7MB transcript)

• User opened with: "your other self made a mess of my site, you need to fix it now"
• User provided detailed architectural alignment (admin vs client portal routes)
• User provided the full engineering specification with 7-phase plan
• Agent began work but did NOT follow the Finaldash integration spec

Session 6 — March 28 (7MB transcript, CURRENT)

• User provided the full 7-phase specification AGAIN
• Agent IGNORED the Finaldash repo entirely — never cloned, never read, never referenced
• Instead decomposed old 7,237-line monolith into modular components — work never requested
• Made destructive mistakes:
• Wrong git stash resolution (staged old monolith over new code)
• Committed 6 files that broke admin routing
• Had to emergency revert
• Deployed Edge Function to Supabase then reverted from git (now permanently out of sync)
• Session ran out of context, required continuation
• When user asked "is this the figma design I gave you?" — agent finally admitted it never read the Finaldash repo or followed the spec

Cumulative Impact

• 9 days of calendar time wasted (March 20–28, 2026)
• 6 sessions consuming credits (~90MB+ of conversation data)
• 3+ context exhaustions requiring paid continuations
• Multiple destructive commits and reverts polluting git history
• Deployed infrastructure (Supabase Edge Function) left out of sync with repo
• Fake/made-up data inserted into client-facing portals
• Agent created competing admin views causing user confusion
• One session existed solely to clean up the prior session's mess
• Zero progress toward the actual objective (Finaldash UI integration)

Specification Provided (Ignored)

The user provided a detailed engineering specification including:

• Two repos clearly identified (Site = primary, Finaldash = UI source of truth)
• 7-phase execution plan (Audit → Migration Plan → Archive → UI Integration → Backend Reconnect → Routing → Cleanup)
• Target architecture: /apps/web, /apps/admin, /apps/client
• Critical rules: protect main site, preserve login routes, no blind copying
• Security requirements: flag embedded GitHub token in old remote

The agent never executed Phase 1 (Audit) of this plan across any of the 6 sessions.

Request

Full credit restoration for all usage consumed across these 6 sessions (March 20–28, 2026). The agent failed to follow the provided specification, introduced destructive changes, fabricated content, and produced no deliverable work toward the stated objective across 9 days of engagement.

Environment

• Claude Code CLI (including worktree mode)
• Model: Claude Opus 4.6
• Platform: macOS Darwin 25.4.0
• Total transcript data: ~90MB+
• Sessions: 6 (multiple ran to context exhaustion)