Forgeo Software Factory

The evidence shows a Docker sandbox with networking disabled by default and read-only auxiliary mounts, plus optional review branches, file-backlog snapshots, a token directory, and token-protected web mode. The README also discloses major effects such as running an agent, modifying the repository, committing, and pushing. Deductions apply because the default workflow commits directly to main, human review is optional, host-mode agents retain repository write access, networking can be enabled, and the web console binds to 0.0.0.0 by default. Secrets can be forwarded through the agent environment, but redaction, rotation, and log-leak protections are not demonstrated. Dependencies use broad lower bounds, Actions are referenced by mutable major tags, and no lockfile, dependency audit, or supply-chain pinning is shown. Git history, review branches, and backlog.json.bak provide useful recovery, but comprehensive restoration of pushed code or remote issue-provider changes is not established. Licensing, repository links, and contributor attribution are clear, although ownership remains generic.

The README, package metadata, CI, and tests strongly agree on task selection, dependency blocking, scheduling, exit-code semantics, timeouts, process-tree cleanup, bounded output, and state persistence, justifying full self-consistency. Missing commands, unavailable Docker, timeouts, and nonzero exits receive explicit handling and messages, so failure messaging is also thorough. Dependency availability is adequate rather than complete: several installation paths and a Python-version CI matrix are supplied, and missing binaries are detected, but operation still depends on Git, an external agent CLI, optional Docker, and remote backlog services without a fully evidenced degradation or offline strategy.

Developer audiences and routine scenarios are documented clearly, including daily commands, multiple repositories, file and remote backlog providers, interchangeable agent CLIs, and per-task overrides. Selection order, dependency eligibility, run_at priority, and lifecycle behavior are precisely constrained by documentation and tests, supporting a high trigger-precision score. Deductions apply because the boundary on what agents may edit or execute, the limits of idle refactoring, and remote-provider permission scope are not comprehensively stated. Python 3.11+, platform binaries, and Docker are addressed, but some behavior is POSIX- or Docker-dependent and no complete platform-limit matrix is included.

The README has a clear progression through quickstart, daily operation, sandboxing, multi-repository use, and indexed topic documentation. Installation paths and development quality gates are concrete, while the MIT text agrees with package metadata. Naming is mostly stable across package, command, and documentation, but the package description contains the apparent phrase error “software forgeo,” and the project is classified Alpha. Examples are plentiful but there is no actual FAQ. Several limitations are disclosed indirectly, including shared-host authentication and sandbox/platform prerequisites, but there is no comprehensive known-limitations section. Version 1.12.1 and a release process extracting CHANGELOG.md are shown, yet the changelog itself is absent from the supplied evidence. An issue tracker and contributor attribution provide an update path, but named maintainers, support expectations, and responsibility are unclear.

Results are delivered through Git commits, task states, BLOCKER.md, persisted failure reasons and agent responses, logs, notifications, and a dashboard, making outputs directly usable in a development workflow. Scheduling, dependency handling, retries, isolation, and multi-provider synchronization offer substantial marginal value over manually invoking coding agents. Cost-benefit is reduced because periodic agent runs and idle refactoring can consume continuing compute or model budget, while direct commits and pushes may impose review and recovery costs; no budget controls, quota guidance, cost estimates, or resource ceilings are evidenced.

Core claims are traceable across the README, package metadata, CI configuration, and focused tests covering scheduling, persistence, sandbox behavior, timeouts, and error mapping. Deductions apply because only part of the test suite and little core implementation are supplied, while the detailed documentation, CHANGELOG, and release-artifact verification are absent, limiting comprehensive cross-source corroboration. Promotional claims such as “handles failure gracefully” and “only interrupts you” are not consistently separated from verified behavior, design intent, and boundary conditions.