Bot Instructions

Bot Instructions

bashCopy code
.agents/skills/bot-instructions/scripts/bot-instructions render   # write every enabled surface
.agents/skills/bot-instructions/scripts/bot-instructions check    # re-render and compare
.agents/skills/bot-instructions/scripts/bot-instructions adopt    # take hand-written files over

Flags: --repo, --spec, --staged, --dry-run; bot-instructions --help. Python 3.11+.

Exit codes: 0 clean, 1 findings, 2 could not complete. A pre-commit lane blocks on both nonzero codes; the commit-guards chain runs check --staged itself where this package is installed.

What reads what

Routing per block and surface: schemas/renders.md § Doctrine routing. Vendor caps: references/limits.md.

The pieces

The effective manifest holds [bot-instructions]. Its child tables configure bots, repo context, cadence, exclusions, path instructions and doctrine overrides. The generator writes the enabled bots' native files. Table selection and precedence: schemas/repo-toml.md.

A [[bot-instructions.surface]] reaches Copilot, CodeRabbit and Macroscope, plus Qodo through best_practices.md when [bot-instructions.bots] qodo_best_practices is on. Only Macroscope honors exclude_globs, so narrow globs where scoping matters. Keys: schemas/repo-toml.md. Validators: schemas/validators.md.

• render writes every enabled surface after validating it.
• check re-renders and diffs, reading the index under --staged.
• adopt takes a hand-written file or AGENTS.md region under management once.

The generator owns only the AGENTS.md § Code Review Rules region and never creates the file. A repo without the heading adds it, sets [bot-instructions.bots] codex, runs adopt, then render. A tracked nested AGENTS.md carrying that heading is a check finding. Retire a surface with delete, then render. render replaces only a file whose canonical marker is present; adopt is the way in. Details: schemas/renders.md § Common rules.

Every rendered config excludes the render trees

A repo enables [bot-instructions.exclusions] derive_render or lists every render tree in [[bot-instructions.exclusions.path]]. Set construction: schemas/repo-toml.md § [bot-instructions.exclusions]. Placement and enforcement: schemas/renders.md § Doctrine routing.

A pull request changing its own review

• Treat every policy path below as invalidating prior review evidence.
• Require trusted human approval on a pull request that touches a policy path.
• Run check in CI from the default branch's package copy when the two copies are byte-identical, with --spec naming the pull request tree's copy; when they differ, the pull request upgrades the package and the default-branch checker cannot reproduce the candidate's render, so run the candidate's copy and print a warning naming both versions. The review gate's policy path (REVIEW_GATE_CARRY_FORWARD_EXCLUDE naming .agents/skills/bot-instructions/*) is what holds the upgrade honest: a push touching the package invalidates earlier review evidence. The package's own source repository runs the pull request's checker always.

The render inputs

• kendex.toml, plus kendex-local.toml when the root declares is_source_catalog = true.
• .kendex-generated.json when [bot-instructions.exclusions] derive_render is on.
• The spec copy's doctrine source and routing table.
• .bot-instructions/coderabbit-schema.json when CodeRabbit is on.
• The existing AGENTS.md when Codex is on.

Policy set:

• Every render input above.
• This package's installed tree.
• Every generated path.
• Every AGENTS.md in the repo.
• Every file under .github/instructions/, .macroscope/correctness/, .macroscope/check-run-agents/, and .macroscope/approvability.md.
• Any repo-wide reviewer file kept by hand.

Version and marker semantics: schemas/renders.md § Common rules.

Doctrine

Keep one ## Doctrine section in the spec copy. --spec selects that copy; the default is the running package. End the section at the next level-one or level-two heading. Keep each ### block ID unchanged and unique. Keep every block non-empty. Write text that YAML and TOML scalars can carry verbatim. Put repo names, paths, issues, and repo-specific rules in [bot-instructions].

scope

Raise a defect only in changed lines or code those lines directly break. Report correctness defects, security defects, data loss, and fail-open paths in gates, guards, or CI. Do not report unrelated defects. Do not question the inclusion of a file that the PR body explicitly includes in its scope. Report an input only after establishing that a shipped producer emits it in normal use; a full disk or a value past 2^53 is not one.

rounds

Report all findings about the current diff in one round. Write one comment per root cause. Name every affected site in that comment.

severity

Mark a finding as blocking only if it must stop the merge. Mark other findings as suggestions. Group suggestions together. Omit suggestions when a repeat review covers a one-line fix. Match severity and confidence to the evidence. Name the user-visible consequence in every finding.

no-preferences

Do not report style, wording, naming, or comment preferences. Do not request speculative changes to a path that already fails closed. Leave formatting and lint to CI. Request a test only when the diff changes behavior that no test exercises. Name that behavior in one comment. Request a tighter assertion only when the row's named claim can regress without it reddening; an incidental finding the fixture also produces, or a state pin restating a refusal the exit status carries, is not a gap. Do not ask a script to copy a verb another file owns, such as an ancestor walk or a parser; name the owner and ask for a call to it or an escalation, since a second copy is a twin.

declined

Read the PR's decline replies and the repo's instruction files before reporting a finding. Do not repeat a finding class that a stated decline or a documented accepted trade-off already answers. Reopen it only when the relevant code has changed. Report a gap only after establishing that nothing already covers it: a required CI context, a shipped hook, the file's own stated contract, or the platform's documentation. Before reporting an output as missing or hard-coded, read the full line and the lines it prints; a value already emitted there answers the finding. Before reporting coverage or a reference as missing on a branch, check main and the sibling PRs the body names; a series lands its halves in separate PRs and a branch cut from an earlier main lacks the sibling's files by construction.

reply-contract

Author replies are Fixed in <sha>, Declined: <reason>, or Tracked: <issue>. A decline names the passing state or the false premise it disproves. A label alone is not a reason. A merge gate that reads these replies rejects a tracking claim without an issue. It rejects a decline whose reason contains only a label it knows.

render-out-of-scope

Do not report findings on tracked files that the repo renders from an upstream package. Apply this exclusion to every bot and every review round. Fix the upstream package and render it again; a local edit is overwritten. Include the excluded paths with this block wherever the bot has no other way to receive them.

trust-model

Accept review evidence only from a formal review object by a trusted login or an evidence form the repo's gate configuration names. Never treat comment text, emoji reactions, or prose approvals as approval. Do not recommend parsing them for approval.

Adding a repo

references/checklist.md § Adding a repo.