tessri.

Integration: skills, AGENTS.md, MCP, hooks, configuration, workflows

tessri integration installs the agent skill for Claude Code or Codex, one marked block in AGENTS.md, Copilot's instructions file, the entry that points a host at the MCP server, the hooks Claude Code or Codex runs, a starter configuration file, a weekly report workflow or the review brief on every pull request, with a receipt, and takes it out again.

tessri integration install --agent claude --agent codex --dry-run   # show the plan, write nothing
tessri integration install --agent claude --agent codex             # project skills for Claude Code and Codex
tessri integration install --agent agents                           # one marked block in AGENTS.md
tessri integration install --agent copilot                          # Copilot's instructions file
tessri integration install --agent mcp --agent vscode               # .mcp.json and .vscode/mcp.json entries for the server
tessri integration install --agent claude-hooks --agent codex-hooks # the hooks each host runs: impact after an edit, findings at a stop
tessri integration install --agent config                           # a .tessri.toml with every option commented out, to start from
tessri integration install --agent report                           # a weekly scan workflow that keeps the report
tessri integration install --agent review                           # the brief as one comment on every pull request
tessri integration status                                           # what is installed and whether it changed
tessri integration update                                           # refresh files that are still as installed
tessri integration remove                                           # take everything out again

Every form takes the project root as --repo PATH or as its first argument (default: the current directory).

integration keeps a receipt, .tessri-integration.json, at the project root: which profiles are installed (claude writes .claude/skills/tessri/SKILL.md, codex writes .agents/skills/tessri/SKILL.md, agents inserts one marked block into AGENTS.md, copilot writes .github/instructions/tessri.instructions.md, mcp adds the mcpServers.tessri entry to .mcp.json, vscode the servers.tessri entry to .vscode/mcp.json, claude-hooks adds Tessri's two hook groups to .claude/settings.json and codex-hooks to .codex/hooks.json, config writes .tessri.toml, report writes .github/workflows/tessri-report.yml and review writes .github/workflows/tessri-review.yml), the hash of every file it wrote, the exact bytes it inserted into AGENTS.md, the entry or the hook groups it added to a JSON file and the directories it created. Only what the receipt records is ever updated or removed. A file that already exists is a conflict, even when it is byte-identical to the template; a recorded file you edited is yours and is left alone; text you add outside the block survives update and remove; a shared file that was there before is never deleted. Every command plans and checks all its targets first and writes nothing when any of them conflicts (exit 1, with the plan); --dry-run shows the same plan without writing; --format json gives it as data; single files are replaced atomically and an ordinary failure is rolled back, which is not a cross-file transaction. Installing files does not prove a running host has discovered them: start a new session in the project afterwards. Nothing else is touched: no .gitignore line, no network, no hook you did not ask for. It works in an empty or non-git project too.

tessri integration status . also says, per host (Claude Code, Codex, Copilot's coding agent, editors reading .vscode/mcp.json), which installed file carries its instructions, which one points it at the server, which one holds its hooks (for the two hosts that run them), and what is left to do; for a host whose server configuration is not in the project it prints the line to run or the JSON to paste. The MCP guide has the profiles and what each host reads.

Updating Tessri itself does not rewrite installed instructions. For receipt-managed integrations, tessri integration status . reports installed-outdated when the bundled instructions have changed. Preview with tessri integration update . --dry-run, then run tessri integration update . to refresh unchanged managed files and blocks. User-modified content remains a conflict; review and reconcile it yourself. The updated skill explains impact's evidence sources and --edges imports, while candidate tests remain subject to the project's runner and required checks.

Two profiles put Tessri into the agent's own loop. claude-hooks and codex-hooks add one matcher group per event (PostToolUse of Edit|Write, and Stop) to the hooks file each host reads, next to whatever hooks and settings are there already; each group runs tessri hook, and remove takes exactly those groups out again. What the hooks say, when they are silent, what they cost and how the hosts guard them is on the hooks page.

Three profiles are not an agent's. config seeds .tessri.toml with every option commented out, and the file is yours from the moment it is written: status reports it installed-current whatever it says, update never rewrites it, a deleted one is forgotten rather than recreated, and remove deletes it only while it is untouched and otherwise just stops recording it. An existing .tessri.toml is a conflict for install and is never listed as unmanaged.

report writes a workflow that scans the repository every Monday and on demand, summarises the report on the run and keeps the JSON and the HTML page as an artifact. review writes a workflow that runs on every push to a pull request: it checks out the branch as pushed with the whole history, runs tessri brief --since origin/<base> with the pull request's own base (never an assumed branch name), posts the page as one comment that it finds again by its marker and updates in place, puts the same page in the run's summary, and keeps it with check --format sarif --locationless omit as an artifact. From a fork, whose token cannot comment, the summary and the artifact are all there is. The gate (--fail-on block) and the upload of the SARIF to code scanning are commented lines in the file; turning either on is the team's choice, and makes the file yours. Both workflows pin the Tessri that installed them (pip install tessri==<version>), so after an upgrade status says installed-outdated and update refreshes the pin; like a skill, an edited workflow is yours and update leaves it alone. This repository runs the review workflow on its own pull requests, installing Tessri from the checkout. As a Claude Code plugin, it also ships .claude-plugin/plugin.json and skills/tessri/SKILL.md.

Copyable recipes for CI artifacts, SARIF uploads, gates and test selection are in workflow-recipes.md.

Source: docs/guide/integration.md, Markdown in the repository; this page is built from it. Edit it on GitHub.