Agent Architecture Review, Snapshot di validazione

Design for delegation rather than direct manipulation

The workflow is designed around delegation of a cohort validation task rather than manual file-by-file operation: `approve()` records approval intent, then `run_cohort_validate()` creates a `CohortValidationJob`, clones the applicant repo, selects an agentic surface via `select_agentic_surface()`, builds a `ValidationRequest`, and calls `validate_code_against_principles()`. Scope constraints are encoded through selector limits such as `MAX_FILES`, `MAX_BUNDLE_BYTES`, `MAX_FILE_BYTES`, language detection in `_detect_language()`, and repository namespace construction in `_execute_with_job()`.

Ensure that background work remains perceptible

Background work is made perceptible through durable job state in `CohortValidationJob`: `status`, phase timestamps such as `cloning_started_at`, `selecting_started_at`, `bundling_started_at`, `validating_started_at`, `terminal_at`, and typed terminal fields like `failure_kind`, `safe_display_message`, and `retry_eligible`. The runner updates these via `mark_step_started()`, `mark_completed()`, `mark_blocked()`, `mark_failed()`, and mirrors terminal state back to the application through `_mirror_terminal_to_app()`.

Align feedback with the user’s level of attention

Feedback is tiered by attention level: the applicant/application surface receives coarse states such as `validation_queued`, `validation_complete`, and `validation_failed` through `onboarding_state`, while operators can inspect `failure_kind`, `safe_display_message`, `retry_eligible`, and the persisted audit object. Routine progress is represented by status transitions; higher-attention failures are escalated with typed messages such as `invalid_repo_url`, `selector_rejected`, `validate_error`, and `persist_error`.

Apply progressive disclosure to system agency

The code supports progressive disclosure by separating simple operational state from detailed inspection data. Primary state is stored as concise values like `queued`, `cloning`, `validating`, `completed`, `blocked`, `failed`, and `aborted`; deeper evidence is stored in `audit_object`, including `source`, `selection`, `validate`, and `job` sections. `_summarize_validate_log()` deliberately reduces raw validator internals to safe metadata such as top-level keys and entry counts rather than dumping the full log by default.

Replace implied magic with clear mental models

The workflow makes its operating model explicit: `fetch_public_repo()` only accepts public HTTPS GitHub URLs via `_GITHUB_URL_RE`; unsupported repos become `invalid_repo_url`, `clone_failed`, `clone_timeout`, or `repo_too_large`; language support is limited by `_detect_language()` and `LANGUAGE_EXTENSIONS`; and the validation bundle is wrapped with `BOUNDARY_HEADER` plus `ENVELOPE_ADVISORY` stating that repo code is untrusted inert input. These are concrete capability and boundary signals rather than implied magic.

Expose meaningful operational state, not internal complexity

The state model exposes user/operator-meaningful states instead of raw implementation chatter. `CohortValidationJob.status` uses terms like `queued`, `cloning`, `selecting`, `bundling`, `validating`, `completed`, `blocked`, `failed`, and `aborted`; `_mirror_terminal_to_app()` maps those into application-level `validation_complete` or `validation_failed`; and user-safe explanations are kept in `safe_display_message` rather than exposing subprocess output or raw validator logs.

Establish trust through inspectability

Inspectability is backed by concrete trace artifacts. `build_file_envelope()` creates a typed `bridge.files.v2` envelope with per-file content, paths, byte sizes, and SHA-256 hashes, then computes an `envelope_hash`; `_execute_with_job()` records `commit_sha`, selected file hashes, skipped reads, bundle size, `bundle_sha256`, `latency_ms`, validator log signals, `BOUNDARY_CONTRACT_VERSION`, `ENVELOPE_SCHEMA`, and `envelope_hash` in `audit_object`, which is merged into `UserValidationRun.result_json`.

Make hand-offs, approvals, and blockers explicit

Approvals, hand-offs, and blockers are explicit. `approve()` requires an operator confirmation unless `--yes` is supplied, then separately records Firebase creation, sign-in link generation, email delivery, and validation queueing states. Validation blockers are typed through `FAILURE_KINDS` and set with `mark_blocked()` or `mark_failed()`, with distinct outcomes for policy/user constraints such as `invalid_repo_url`, `no_supported_language`, and `selector_rejected` versus tool failures such as `clone_timeout`, `validate_error`, and `persist_error`.

Represent delegated work as a system, not merely as a conversation

Delegated work is represented as a structured system rather than a conversation. The code has persistent system records (`CohortApplication`, `CohortValidationJob`, `UserValidationRun`), explicit phase transitions in `mark_step_started()`, terminal-state functions, retry metadata, abort metadata, timestamps for each stage, and a structured audit object. Execution state is separate from the LLM response and is persisted independently in the job table.