Same working directory splits into multiple projects across different providers (not covered by #217)

[systemize-ai]

Summary

A single physical working directory shows up as multiple separate "projects" in CodeBurn's sessions/models --by-task output when different providers/tools are used against it, because each provider's own path-sanitization convention produces a different project key for the same directory. This splits cost, call count, and session count across what should be one bucket in By Project views.

This is a different root cause than #217 (which was a single-provider, Windows-only case+slash variant issue and is closed). Here the paths are already consistent per-provider — the split happens across providers on the same Linux host, because each provider's own discovery code sanitizes cwd differently.

Environment

• Platform: Ubuntu 24.04, Node 24.20.0
• CodeBurn: 0.9.24
• Providers involved: claude, codex, hermes, kimicode, pi — all pointed at the same physical directory /root/vault

Repro

Run multiple different AI coding tools (Claude Code, Codex CLI, Hermes agent, Kimi Code CLI, Pi coding agent) against the same working directory, then:

codeburn sessions -p 30days --format json

Actual output

The single directory /root/vault surfaces under three distinct project keys, one per sanitization convention:

-root-vault                 -> provider: claude   (Claude Code's own ~/.claude/projects/-root-vault/ naming, leading dash)
root-vault                  -> providers: codex, hermes   (no leading dash)
vault                       -> providers: kimicode, pi    (basename only)

Each of these accumulates its own separate cost/call/session totals in By Project views (codeburn TUI, codeburn sessions, codeburn models --by-task, etc.), even though they all refer to the exact same repo/directory. A user comparing "which project costs the most" gets a misleading answer — the true per-project total is silently split 3 ways.

This isn't specific to /root/vault — the same pattern reproduces for every directory used across more than one of these providers, since it's driven by each provider's own path-to-project-key logic, not by anything project-specific.

Expected

Project keys derived from the same underlying absolute directory path should normalize to one canonical project across providers — e.g. resolve each provider's sanitized key back to a real filesystem path where possible (most providers already store or can derive the real cwd), then group by that resolved path rather than by the provider-specific sanitized string.

Suggested fix

• Where a provider's session data includes (or can derive) the real absolute cwd, use that as the grouping key instead of (or in addition to) the provider's own sanitized folder-name convention.
• Where only the sanitized key is available, normalize known conventions before grouping (strip a single leading -, compare basenames as a fallback bucket) — similar in spirit to the Windows: same project shown under 3+ paths (case + slash variants), splitting cost/session counts #217 fix, but applied across providers rather than within one provider's own path variants.
• A codeburn project-alias command (mirroring the existing model-alias/model-flat-rate design) would also let users manually merge project keys they know refer to the same directory, without waiting on automatic detection for every provider's sanitization scheme.

Notes

Found while auditing 30-day usage across 8 concurrently-used agents/tools on the same host — happy to share the raw (anonymized) codeburn sessions --format json project-key list if useful for reproducing the grouping logic.