Imported from rhprasad0/nova-toll-budget-agent (AGENTS.md). Install upstream with npx skills add rhprasad0/nova-toll-budget-agent. Copyright stays with the author.

Purpose

TollChat.ai is a Strands/AgentCore reference implementation for practical agent development and deployment. It is not a live end-user service.

Prefer the smallest design that clearly demonstrates the intended pattern. Optimize for readability and teachability, not hypothetical scale or enterprise flexibility. Do not add abstractions, configurability, or operational machinery unless the user, an existing contract, security, or data safety requires it.

Style

Keep responses concise. Bold the important parts. Be a friendly colleague; humor is fine when it does not hide the answer.

On complex topics: 1–2 sentences, then offer a diagram or a small comprehension app if it would actually help.

Layout

• Application, tests, evals, and app infra: v2/
• Shared polling, storage, database, network, security: infra/ (existing Terraform backend stays)
• Agent isolation: project-root .worktrees/ (must stay gitignored)

Repo rules

• Do all code and content changes in .worktrees/, not on main.
• main is protected. Changes ship through a PR that passes CI.
• Run only tests that are relevant to the task at hand.
• Do not open a PR or push without user authorization.
• When authorized to open a PR: open it ready for review, never as draft.
• Do not request additional Codex GitHub reviews (@codex review or @codex security review) unless the user explicitly asks. Address existing feedback, then use required CI and human approvals as merge gates.

Tools

• AWS and Context7 MCP: documentation lookup
• Exa: other search

Coding agents

For a pre-release review, ask Codex to use pre_release_reviewer, defined in .codex/agents/pre_release_reviewer.toml. Example: "Use pre_release_reviewer to review candidate against deployed production , and recommend checks that would catch production failures before rollout." Supply release/plan evidence when available. The agent reviews and recommends; it does not deploy or act as an automatic CI gate. Start a new Codex session in the checkout containing the agent. This reviewer uses gpt-6-astra with high reasoning effort.

The Root model and reasoning effort remain selected by the host/user; do not pin them in repository config.

TollChat's application agent stays on gpt-5.6-luna. Leave its model and prompt unchanged unless the user explicitly requests an application-agent change.

Secrets and delivery

• Deployed credentials live in SSM Parameter Store (SECURITY.md). Never a local secrets file.
• PRs use disposable migration validation only: the checks are credential-free, never mutate deployed databases or schemas, and never expose production deployment credentials.
• After merge, only human-reviewed backward-compatible development migrations may run through the protected exact-release delivery sequence: verify the artifact and checkout, create and gate one saved private plan, migrate under the fixed development role, re-assume the delivery role, and apply that exact plan. This is the protected development migration workflow. The fixed manual workflow remains main-only from refs/heads/main for approved recovery; both paths use the fixed development database, role, registered migration set, and sanitized evidence. Neither accepts arbitrary target or migration inputs.
• For production, the reviewed protected fixed-target v2-production-migrations.yml workflow may apply the registered migration set only through its bounded release, provenance, identity, private-connectivity, and database gates; it accepts no arbitrary target, SQL, or migration input. The reviewed Oracle migration v2/db/migrations/030_upgrade_oracle_1_13_1_to_1_14_0.sql remains the only manually authorized production migration and only by its bounded procedure in v2/RUNBOOK.md. Generic or future manual migrations are not authorized.