Official technical research: two agents, one owned project
Retrieved: 8 November 2026. Researcher: Codex source-review subagent. Status: documentation research, not a live installation or workflow rehearsal. No Claude session, paid API job, credential exchange, deployment or account change was run.
What the evidence changes
The durable core is a project owner giving bounded work to two agents, storing decisions and files in a repository, checking changes and recovering safely. It is not an automatic conversation between Claude Code and Codex merely because both can access GitHub.
Current official documentation differs materially from older tutorials:
- The current Codex command reference says
codex mcp-server and its standalone binary have been removed and points to App Server. A historical MCP-server bridge must not be taught as the current default (T08).
- Current Claude Code can read
AGENTS.md, subject to version, configuration and the presence of CLAUDE.md. It is inaccurate to teach that Claude always ignores it. An explicit @AGENTS.md import remains a documented compatibility route (T04).
- Claude and Codex permission systems differ. Instructions describe expected behavior; runtime permissions control access. Shell restrictions do not automatically describe every connector or browser boundary (T02, T07).
- Codex app-managed worktrees may start detached, whereas the cited Claude worktree recipe creates a checkout on its own branch. Do not teach one identical branch workflow for every surface (T03, T09).
- A session resume restores a conversation; it is not proof that every process, remote action or tool job resumed successfully. Reconcile actual files and status before continuing (T05, T08; course inference).
Evidence method and limits
I read the installed OpenAI Docs skill, searched official documentation, opened the selected primary pages, and used targeted section retrieval for the claims below. I did not claim to read every line of the long documentation sites. Search results from mirrors, forums and third-party tutorials were excluded as technical authority. Where an older official URL redirects, the current destination is recorded.
The register supports product mechanics, not market demand, learning effectiveness, CPD accreditation, or a guarantee that any learner's account exposes the same features. Prices, plan eligibility, CLI versions, configuration and UI labels must be checked immediately before recording and before each course release. No account-specific access or quota was tested here.
Reuse policy for this research: original explanations and short paraphrases with links. Public documentation is not permission to copy complete pages, screenshots, video or branding into a paid course. Review the relevant owner's terms before reproducing assets. Rights-cleared recording and asset review remain separate. No downloaded source corpus is included.
Source and claim register
Every row was retrieved on 2026-11-08. “Verified” below means the named official section supports the bounded statement, not that a command was executed.
T01 — Claude Code terminal setup
- Issuer / source: Anthropic, Quickstart, “Before you begin”, “Install Claude Code”, “Start your first session”.
- Verified: terminal setup is distinct from web, desktop and editor surfaces. Installation method varies by OS; the page gives
claude --version as a successful-install check and starts the tool from a project directory. Native and package-manager update behavior differs.
- Course use: setup lab and troubleshooting card. Have learner record OS, installation route, version, project path and signed-in identity.
- Limit: no command or login tested. Avoid a permanent model/plan list in narration. The page's current subscription list is not evidence of the learner's entitlement.
T02 — Claude permissions
- Issuer / source: Anthropic, Configure permissions, “Permission modes”, “Manage permissions”.
- Verified: manual/default, plan, acceptEdits and auto have different behaviors; bypass skips broad permission checks. Permission rules are enforced by the client, while instructions influence behavior.
- Course use: classify a read, local edit, external send and deployment before acting; show the actual permission screen in a test folder.
- Limit: auto availability varies. Do not instruct learners to grant blanket access to make the lesson proceed. A written prohibition is not an enforced technical boundary.
T03 — Claude isolated work and resume recipes
- Issuer / source: Anthropic, Common workflows, “Resume previous conversations” and “Run parallel sessions with worktrees”.
- Verified: the documented worktree recipe uses a separate checkout/branch and needs an existing commit. Separate named worktrees prevent concurrent edits in the same working files. The continuation command selects the most recent conversation in the current directory.
- Course use: demonstrate sequential work first; later split two non-overlapping tasks into isolated worktrees.
- Limit: isolation does not merge results or prevent logical incompatibility. Dependencies and runtime ports can still need separate setup; do not promise conflict-free parallelism.
T04 — Shared instructions without duplicated project law
- Issuer / source: Anthropic, How Claude remembers your project, “AGENTS.md” and “Share one file with other coding tools”.
- Verified: current support is conditional; default behavior prefers applicable CLAUDE files. The documented
@AGENTS.md import offers a shared-file route. Ordinary prose saying “read AGENTS.md” is not the same loading mechanism. Earlier versions differ.
- Course use: keep stable rules in AGENTS.md and a small Claude compatibility file; keep changing status in a separate handoff record. Verify actual loaded context.
- Limit: not a universal configuration prescription. Existing organization rules take precedence; inspect before adding files.
T05 — Claude session recovery
- Issuer / source: Anthropic, Manage sessions, opening explanation and resume command table.
- Verified: CLI sessions are saved conversations associated with project directories. Continue selects the latest in the current directory; resume supports selecting a saved session. Surfaces keep different session lists.
- Course use: record project path, session identifier/name, latest commit, unsaved changes, completed checks, active jobs and next action before closing.
- Limit: the handoff fields are course recommendations. A saved transcript is not a Git backup or a guarantee of restored remote jobs.
T06 — Claude checkpoints are not universal rollback
- Issuer / source: Anthropic, Checkpointing, “Automatic tracking”, “Limitations”.
- Verified: checkpoint tracking centers on file-editing tools. A file changed only by shell commands is not tracked; external/concurrent changes are normally outside coverage unless they affect tracked files.
- Course use: recover a disposable file, compare with a Git checkpoint, explain why sending or deleting something remotely is a different operation.
- Limit: do not teach rewind as an undo-everything button. Preserve current work before choosing a recovery path.
T07 — Codex permissions and separate boundaries
- Issuer / source: OpenAI, Permissions, “Filesystem permissions” and local-command/connector boundary discussion.
- Verified: read, write and deny grants differ. Local execution permission profiles operate alongside approval policies. MCP, connectors, browser/computer use and cloud environments have their own controls.
- Course use: learner makes an access inventory for each surface and confirms what the current task can read/write/send.
- Limit: the page describes evolving permission profiles and legacy compatibility. Avoid copying a large config into learners' accounts. Teach checking the active setup.
T08 — Codex commands and obsolete bridge
- Issuer / source: OpenAI, Developer commands, CLI, “codex resume”, “codex exec”, “codex mcp-server”. Older
/codex/cli/reference/ redirected here.
- Verified: interactive and noninteractive resume routes differ. The current reference explicitly removes
codex mcp-server and directs users to App Server. Managed-worktree flags have command-specific restrictions.
- Course use: supported session picker/resume demonstration; optional advanced orchestration appendix, not a core bridge dependency.
- Limit: CLI documentation is not proof of installed-version support. Check
--help in the recorded version. No controller or App Server integration was implemented in this research.
T09 — Codex worktrees
- Issuer / source: OpenAI, Worktrees, creation and management sections.
- Verified: app-managed worktrees use the selected branch's HEAD and normally start detached; current app behavior can carry uncommitted changes. Worktrees occupy separate directories and have lifecycle/cleanup controls.
- Course use: learner confirms exact checkout and starting revision before assigning each agent a task; keeps results before cleanup.
- Limit: this describes managed app worktrees, not every manual Git or remote workflow. Never assume a branch name or assume uncommitted edits were excluded.
T10 — Codex installation and first task
- Issuer / source: OpenAI, Codex CLI, “Get started with Codex CLI”.
- Verified: current documentation provides platform installation choices, project-directory launch and sign-in selection; it recommends Git checkpoints around tasks. The CLI exposes status, permissions, model and review controls.
- Course use: a separate Codex setup card, then a read-only explanation of the learner's project.
- Limit: installation and actual account access not tested. The current general quickstart redirects toward desktop guidance, so do not conflate desktop installation with CLI installation.
T11 — Codex AGENTS.md loading
- Issuer / source: OpenAI, Custom instructions with AGENTS.md, “How Codex discovers guidance”.
- Verified: instruction discovery builds a chain from global and project guidance; empty files are skipped and there is a configured combined-size cap (documented default 32 KiB).
- Course use: short stable project rules, a named current handoff, and an instruction-loading verification exercise.
- Limit: “the file exists” does not prove complete loading or compliant behavior. Do not put long transcripts or every research source in the instruction file.
T12 — Codex review scope
- Issuer / source: OpenAI, Code review, “Review local changes”, “Choose a review scope”, “Share feedback and follow up”.
- Verified: local review can inspect selected differences without editing the working tree. The review pane may include edits by others. A chat review is distinct from posting, submitting a review or merging.
- Course use: specify exact base/commit/files, expected behavior and evidence needed; inspect each finding and rerun affected checks after repairs.
- Limit: findings are not proof of correctness. Separate reviewers can share blind spots; founder acceptance and realistic user testing remain necessary.
T13 — Claude MCP configuration and trust
- Issuer / source: Anthropic, Connect Claude Code to tools via MCP, installation, server status and workspace trust sections.
- Verified: server scope, transport and authentication matter. A cloned project cannot automatically approve its own configured servers while untrusted; status inspection distinguishes connection and policy failures.
- Course use: one tool selected for a named task, official configuration, account/scope check, then one harmless read-only call with recorded result.
- Limit: configured, authenticated, connected and successful are different states. Do not treat installation as usage evidence or a plugin listing as permission to spend/send.
T14 — Codex MCP support
- Issuer / source: OpenAI, Model Context Protocol, configuration and active-server sections.
- Verified: configuration supports local stdio and HTTP connections; authentication and environment selection affect availability.
/mcp exposes active servers in the terminal UI. Tool startup/execution timeouts exist.
- Course use: record provider, purpose, scope, price/allowance, input sensitivity and a successful result. Show a disconnected-tool fallback.
- Limit: support in one client does not establish equivalent configuration, credentials or tools in another. No external server was installed or tested for this research.
T15 — MCP is a connection protocol
- Issuer / source: MCP project, Architecture overview, “Participants”, “Primitives”, discovery and execution examples.
- Verified: a host connects clients to servers; tools/resources/prompts are exposed capabilities. Discovery and execution are distinct, and capabilities can change.
- Course use: plain analogy: agent is the worker, server offers a tool, connection exposes it; the owner still specifies the job and checks the result.
- Limit: the protocol does not establish source truth, agent independence, business authority, subscription entitlement or a useful course outcome. Avoid teaching protocol internals to beginners.
T16 — MCP authorization boundaries
- Issuer / source: MCP project, Security Best Practices, confused-deputy and token-passthrough sections.
- Verified: tokens must be intended for the receiving service; forwarding inappropriate tokens can bypass controls and undermine accountability.
- Course use: use official sign-in, inspect scope, keep credentials out of prompts and repository files, revoke a connection when no longer needed.
- Limit: these are protocol security concerns, not certification of a named server. Learners are not expected to implement OAuth or audit server code.
T17 — GitHub as the shared record
- Issuer / source: GitHub, GitHub flow, full short article.
- Verified: branch-based collaboration applies beyond software, including documents; branches, changes, review and merge are distinct steps. Web, CLI and Desktop routes exist.
- Course use: one repository, one visible current branch/revision, small commits and a reviewable change proposal.
- Limit: a remote repository is not automatic agent messaging. Committing locally, pushing remotely and merging the default branch are separate results.
T18 — GitHub authentication
- Issuer / source: GitHub, About authentication to GitHub, command-line authentication.
- Verified: HTTPS and SSH use different routes. GitHub CLI can authenticate through a browser; password-based Git authentication is not the supported route.
- Course use: browser authentication for the learner's account, then verify repository access without displaying tokens.
- Limit: a browser login alone does not prove CLI access, organization SSO, write permission or approval authority.
T19 — Pull-request review is a decision
- Issuer / source: GitHub, Pull request reviews, review decision types and required reviews.
- Verified: comment, approve and request-changes are distinct review outcomes; reviews precede merge and may be governed by repository requirements.
- Course use: learner reads the changed deliverable, tests it, resolves findings and makes an explicit accept/request-changes decision.
- Limit: automated or AI approval is not course-founder approval. Do not assume branch protection is configured merely because the workflow says review is required.
T20 — Conflicts need reconciliation
- Issuer / source: GitHub, Merge conflicts, full short article.
- Verified: conflicting edits can arise from same-line changes or edit/delete combinations; conflicts must be resolved before merge.
- Course use: controlled text-file conflict in disposable practice, preserve both intended requirements, compare result and retest.
- Limit: conflict-free merge does not prove compatible business logic. Do not use “accept ours/theirs” as an unexplained repair instruction.
T21 — Usage is not context or elapsed time
- Issuer / source: Anthropic, Manage costs effectively,
/usage, stale usage display, spinner token count and agent-team costs.
- Verified: subscription usage and locally estimated API-token cost differ. Usage data can be stale. The spinner's token figure describes turn output, not full context occupancy. Multiple agents consume usage.
- Course use: record observed usage and elapsed time; plan a small pilot; checkpoint when a limit interrupts. Avoid guaranteed completion times.
- Limit: no exact learner cost, quota or model entitlement established. Do not use historical session speed to promise account-independent overnight execution.
Course-owned procedures derived from these sources
These are instructional recommendations, not vendor guarantees:
- Prefer a sequential builder/reviewer handoff for the first project. Supply the same objective, inputs, constraints, acceptance tests and exact revision to both. Evidence, not model agreement, decides whether a finding is real.
- Keep HANDOFF.md for current state, DECISIONS.md for consequential choices and REVIEW.md for findings/checks. These filenames are a course convention; vendors do not mandate them.
- Before a restart, record files, uncommitted changes, active jobs and next action. After restart, inspect status before resuming; avoid duplicate operations.
- Parallelize only independent outputs in isolated checkouts. Reconcile deliberately and test the combined result. For beginners, two concurrent workers are enough.
- First MCP lab: trusted read-only source, synthetic data, one bounded call, inspect output and provenance. If unavailable, complete the project from a supplied dated snapshot and label the live capability unverified.
- A local calculator is complete only after numeric edge cases, accessibility and expected decision outputs have been checked. It is not a payment, booking or public production system.
Recording and release checks still required
- Rehearse both tools on a clean test project; record actual version, OS, model, account route and limitations without secrets.
- Test setup cards on intended Mac and Windows paths; do not imply a filmed Mac command was tested on Windows.
- Rehearse exact AGENTS.md/CLAUDE.md loading, permissions, review scope, worktree starting state and resume behavior.
- Choose the actual MCP lab provider; verify account access, free/paid allowance, data handling and one live call. Preserve a no-connection learning alternative.
- Rehearse local calculator tests and a controlled conflict/recovery. Keep production data, publishing and paid automation out of this course's default exercise.
- Recheck feature documentation immediately before recording. Specifically revalidate App Server versus removed MCP-server commands and changing desktop product labels.
- Expert technical review and learner usability testing remain release checks. This documentation review does not replace either.
Addendum: selected documentation MCP lab
T22 — Official public OpenAI Docs MCP
- Issuer / source: OpenAI, Docs MCP, “What it provides” and client setup. Retrieved 8 November 2026.
- Verified from documentation:
https://developers.openai.com/mcp provides read-only documentation search and page content. It does not invoke the OpenAI API on the user's behalf. The guide gives Codex and Claude Code configuration and verification routes.
- Instructor-test evidence reported by course supervisor: direct HTTP discovery, search and fetch succeeded, with local receipts
reference/MCP-TEST.json, MCP-SEARCH-TEST.json and MCP-FETCH-TEST.json. This source-review subagent did not perform those calls or independently inspect those receipts. The supervisor checked CLI add-command syntax, without changing existing host configuration.
- Course use: L09 fetches AGENTS.md instruction-discovery guidance and records actual tool output. This replaces the earlier generic-provider suggestion in the design advice; GitHub remains the shared project home.
- Limit: protocol success does not certify live Claude/Codex host integration. Rehearse integration before filming; do not claim permanent pricing or universal account availability.
Source-review authoring handoff
This researcher subsequently authored L03, L04, L09, L10, L11 and L12 and their V03/V04/V09/V10/V11/V12 narration scripts. Those are original course materials, not translations or copies of vendor pages. Local word/section checks were performed; no independent content reviewer has passed them in this task. No media was generated. The six scripts require actual recording, environment rehearsal and final timing. Generic tool-failure scenarios are explicitly simulated; no invented exchange is labelled a real agent test.