fix: improve compatibility with current ChatGPT MCP behavior

[uniplanck]

Summary

This draft proposes a compatibility and runtime-reliability update for the practical MCP execution regression observed after the recent ChatGPT model rollout.

The issue was not limited to GPT-5.6. In the same environment, GPT-5.5 also became materially slower or stalled during multi-step DevSpace operations, despite working normally before the rollout. This PR does not make claims about undocumented OpenAI internals; it reduces eager payload size, repeated context expansion, unnecessary tool round trips, and environment-dependent command failures.

Related to #68.

Main changes

• compact open_workspace payloads with bounded instruction excerpts and explicit full reads
• exact-path reads for advertised instruction files without widening filesystem access
• optional compound inspection tools and workspace-scoped approved-shell aliases
• safe standard PATH augmentation without sourcing .zshrc, .zprofile, or another login-shell file
• execution-cost diagnostics for duration, Tool calls, errors, retries, character volume, and approximate text tokens
• bounded runtime diagnosis and compatibility smoke checks integrated into the existing Bash Tool
• model-facing Tool catalog kept compact; Finder is registered as an app-only Tool
• workspace-scoped Open in Finder actions for MCP App result cards
• optional built-in agent profiles, skill matching, design-audit tools, and generic MCP App integration improvements
• configuration documentation and automated tests

Runtime commands

The diagnostic features do not add model-facing MCP Tools. They are intercepted through the existing Bash Tool:

• devspace-runtime diagnose [--github] [command ...]
• devspace-runtime smoke
• devspace-runtime costs
• devspace-runtime finder <path>

Finder paths are validated against the active workspace root. GitHub authentication checks report only an authenticated/unauthenticated classification and do not return credentials.

Compatibility kit

The branch includes compatibility-kit/openai-model-compatibility-2026-07 with six ordered patches, an apply preflight, a verification script, architecture notes, and security/privacy documentation. Bundle version is 2026.07.2.

1. compact workspace and usage foundation
2. safe tools and compound inspection
3. agents, skills, and MCP App integration
4. runtime diagnostics and execution-cost aggregation
5. integration through the existing Tool surface
6. Finder result-card action

The updater performs no commit, push, publish, deploy, or user-configuration mutation.

Upstream-first basis

• base repository: Waishnav/devspace
• base branch: current upstream main
• upstream main SHA used during refresh: 6ccefbf6213c56056a98ff52d7bdb27c081d13b9
• the fork branch is only the delivery branch for this upstream PR

Privacy and branding

The public branch contains no personal paths, private repositories, credentials, private endpoints, personal logs, or custom private branding. Generic test paths are used. The upstream package name and version remain unchanged.

Validation

• npm ci: passed, 0 vulnerabilities
• compatibility-kit privacy verification: passed
• npm run typecheck: passed
• npm test: passed
• npm run build: passed
• runtime smoke test: 7 passed, 0 failed, 0 skipped
• runtime diagnosis found Git, Node, npm, shell, and GitHub CLI through the safe PATH logic
• execution-cost snapshot verified on the active server process
• Finder opened workspace directories and revealed files; workspace escapes were rejected
• patches 0004-0006 passed git apply --check --whitespace=error-all against the exact post-0003 source state
• source and documentation changes pass git diff --check

The generated patch-container files encode added blank lines in diff form, so running git diff --check on the patch artifacts themselves reports the same container-level whitespace warnings already present in the earlier patch artifacts. The applied source passes whitespace checks, and patch preflight uses --whitespace=error-all successfully.

Dogfooding note

The investigation, sanitization, implementation, runtime testing, compatibility-kit update, validation, branch delivery, and PR maintenance were executed from ChatGPT instructions using DevSpace itself. This describes the workflow and is not an endorsement or security guarantee.

This remains a draft because the scope is broad. The runtime additions are included in this PR because they depend directly on the compatibility foundation already proposed here; creating a separate PR before this branch merges would duplicate most of the same source changes.

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: fa1e1da3-7076-482c-86e8-c877e345ae42

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[jamesx0416]

Does this fix Error in message stream? I started getting this error about an hour ago

[jamesx0416]

Works now for some reason

[uniplanck]

Thanks for the update. Since it started working again without applying this PR, the message-stream error was likely a temporary ChatGPT-side issue rather than the same DevSpace compatibility regression.

This PR mainly addresses tool payload size, runtime reliability, and multi-step MCP execution behavior. It may help with some tool-call failures, but I cannot confirm that it fixes that specific message-stream error.