AGF Local Coding Session Finder

The product is primarily local and read-oriented, with no account, cloud synchronization, or extra resident process. JSON, MCP, and resume-plan are expressly described as read-only or review interfaces; summary search is opt-in, and the documentation warns about private or untrusted text, CSV formula prefixes, and the fact that project scope is not an OS sandbox. Resume, deletion, configuration, and shell-profile effects are documented; bulk deletion has a distinct mode and confirmation screen, while deletion is disabled for providers where it could corrupt related state. Locked dependencies, RustSec auditing, constrained CI permissions, and release verification provide strong dependency-security evidence. Deductions apply because no recovery or trash strategy is documented for completed deletions and no automatic rollback is described for shell-profile edits. Attribution reaches the repository, MIT copyright holder, and upstream projects, but publisher identity is unverified and no clearer maintainer or security contact is supplied.

The README, Cargo metadata, CI, release workflow, and PTY tests consistently describe the version, supported platforms, capabilities, and requirements. Rust 1.88, the C compiler requirement, optional MCP feature, prebuilt packages, and required provider executables are documented. User-visible scan and settings-save failures are described, while the supplied tests explicitly cover scan-failure status and terminal cleanup. No deduction was made for absent runtime results because execution-only properties are outside this static assessment.

Interactive users, scripts, agent tools, and MCP clients receive distinct interfaces through the TUI, versioned JSON, resume-plan, and stdio MCP. The source documents 16 providers, cross-platform paths, environment overrides, no-color and limited-color terminals, shell differences, and optional features. Boundaries between reading, launching, and deletion are explicit, as are unsupported providers and unemulated configuration layers. Precise commands, agent filters, an explicit Cursor executable override, and refusal to treat a generic agent executable as Cursor make triggering well scoped. The supplied files support full marks across these criteria.

The README has strong organization covering installation, upgrades, commands, controls, configuration, storage, limitations, contribution, and licensing. Command and provider naming are stable, including a documented muse-code alias. Numerous examples are present, but there is no dedicated FAQ or systematic troubleshooting section, so examples_and_faq loses one point. The complete MIT text matches package metadata. Version metadata, semantic tag checks, release gates, and a CHANGELOG reference exist, but the CHANGELOG contents were not supplied, preventing full credit for versioning_changelog. Issues, PRs, and a contribution path are identified, but no explicit maintainer, security-reporting contact, or maintenance commitment is shown, so maintenance_responsibility also loses one point.

The tool consolidates discovery of local sessions from many agents and constructs native resume actions, directly addressing the stated cross-tool retrieval problem. Its TUI, quick resume, versioned JSON, MCP interface, project filtering, and reviewable resume plans produce outputs suitable for both people and automation. Reusing existing local stores, requiring no account or cloud service, offering prebuilt binaries, and allowing MCP to be omitted support strong marginal value and favorable cost-benefit. Unexecuted performance and correctness claims were not used as positive evidence.

Core claims can be traced across the README, Cargo.toml, CI, release workflow, and test fixtures. Platform coverage, version requirements, locked dependencies, provider count, read-only behavior, failure presentation, and terminal compatibility receive cross-file corroboration. The documentation clearly labels unverified upstream overrides, unemulated configuration layers, bounded previews, non-sandbox scope, and unsupported providers, separating supported facts from limitations and inference. Because this was a static review without independent execution, confidence remains low.