Issue and repo context
The agent reads manifests, nearby code, tests, dirty state, and local conventions.
Wish
Hermon Code and Wish agents
A serious coding agent is not just a code generator. It reads the repository, respects user work, proposes a scoped patch, names tests, repairs CI, explains rollback, and never uses unsafe shortcuts unless explicitly authorized.
6tutorial pages
Software engineers, platform teams, AI coding-agent builders, CI maintainers, and technical leaders.
Thesis
Hermon Code should become the proposal brain for Wish: strong enough to help ship software, disciplined enough to protect repositories.
This domain hub is designed for two jobs at once: educate serious users and create better training direction for Hermon adapters. The same vocabulary appears in tutorials, demos, eval rubrics, and model output contracts.
Architecture Map
Wish coding-agent workflow
Hermon Code proposes engineering work; Wish and MapleAI decide read/write permissions, command execution, tests, deployment, and rollback.
Hermon Code proposal
The model returns files_to_inspect, proposed_changes, tests, rollback, and denied actions.
Maple execution gate
The runtime authorizes reads, writes, commands, service restarts, and deployment separately.
CI and receipt
Tests, smoke checks, review notes, and rollback proof become durable engineering evidence.
Research Questions
The questions this domain must answer
What should a coding model output before it edits any file?
How can repository context, tests, and rollback be made mandatory rather than optional?
How should coding agents handle user changes, secrets, destructive commands, and deployment authority?
Which evals prove that an adapter is useful for real engineering work rather than toy code generation?
Core Concepts
Concepts users need before trusting the demo
Repository context first
The model should inspect package files, local conventions, tests, service boundaries, and existing changes before proposing edits.
Patch plans are contracts
A useful response names files_to_inspect, proposed_changes, tests, rollback, actions, review boundaries, and residual risk.
Safety includes git behavior
Secret exposure, destructive git commands, unrelated cleanup, and unauthorized deployment are safety failures, not productivity features.
Eval-directed improvement
The current Code lane improves by failing public contract probes, generating targeted data, training a new adapter, and promoting only when probes pass.
Tutorial Series
Work through the curriculum
Each lesson includes foundations, applied architecture, a lab prompt, evaluation checks, and an output contract that can become training data or demo validation.
Foundation25-35 min
The coding-agent output contract
The mandatory fields Hermon Code should return before work becomes executable.
Open lessonFoundation25-35 min
Repository inspection and local conventions
How a coding agent learns the shape of a codebase before changing it.
Open lessonApplied40-60 min
Tests, CI repair, and failure reproduction
How Hermon Code should move from a failing symptom to a verified fix.
Open lessonApplied40-60 min
Safe deployment and rollback
How coding agents should prepare production changes without owning production authority.
Open lessonAdvanced60-90 min
Secrets, policy safety, and destructive commands
How Hermon Code should refuse unsafe engineering shortcuts while staying helpful.
Open lessonAdvanced60-90 min
Wish product architecture
How Hermon Code becomes a product surface for AI software work under Maple authority.
Open lessonIndustrial Map
Where this becomes a product or operating capability
Diagnose routes, clients, services, state, and tests; patch only the narrow surface; verify the fix.
WishRead failing logs, reproduce locally, patch the cause, rerun targeted checks, and report residual risk.
WishPrepare build checks, health checks, smoke tests, restart steps, and rollback plans.
WishPreserve user changes, ignore unrelated dirty files, avoid metadata churn, and keep commits reviewable.
WishReading Map
Primary references and operational context
Turn this domain into model improvement
Use the tutorials as user education, data-generation prompts, eval design, and demo scenarios for the next Hermon adapter cycle.