v0.2.0 Manual Review Implementation¶
This work implements the operator's notes in the current alpha-development worktree. It preserves operator state and earlier working changes; no production database or live formalization run was migrated. The package still declares alpha status. Passing these checks does not declare a beta or stable release.
Review Notes And Result¶
| Note | Implemented result and evidence |
|---|---|
| AGENTS.md / CLAUDE.md | Kept and explained in Contributing. They guide development of Horizon; dispatched projects use pinned API catalogs. CLAUDE imports the shared source. |
| .gitignore | Grouped protections, removed redundant frontend build and blanket lockfile rules, protected private environment variants and retained reusable examples and tracked dependency locks. |
| Migrations | Kept required Alembic history. Startup checks the revision; explicit migration supports fresh PostgreSQL installations and upgrades. Added revision 0029 rather than rewriting applied history. |
| Horizon skill introduction | Identifies Horizon, links its repository and explains consulting a matching local checkout. PATH locates tools, not source. |
| Documentation site | Added MkDocs, strict build, navigation, root-document inclusion, source links and an artifact-building Pages workflow. Publication is a separate manual workflow action after source review. |
| Agent-assisted configuration | README recommends working with a coding agent and reviewing the generated installation plan. |
| Dashboard references | Added project catalog/search/pagination, selected detail, source links, BibTeX display/download, permission-aware registration/editing, conflict handling and durable uncertain-write retries. See References. |
| README organization | Installation, security/sandbox, resource parallelism and state paths now have explicit sections. |
| Dashboard demo | Added a generated synthetic fixture and separate build of the actual dashboard, with read-only transport and repository-subpath browser checks. See demo. |
| Additional tools | Added an AST review inventory with documentation, branches/nesting, optional coverage and explicit review-ledger inputs. Statistics select review work; they do not establish correctness. |
| Version tags | Created v0.2.0-alpha.1 for the already published alpha commit 0df5571. It excludes later development work. Release guidance explains pinning and future tags. |
| Context briefing | Added Agent context, distinguishing bounded read-only database views, initial prompts and native history. |
| Independent code review | Separate readers reviewed complete source bodies and recorded exact catalogs, findings and tests. The implementation specification was derived from executable code without design prose as evidence. See coverage below. |
| Hardcoded prompts | Moved behavioral instructions into Markdown assets. New immutable bundles pin their text/digests; the instruction browser displays them. IDs, serialization, authorization and status/error text remain code. |
| ORCHESTRATOR | Fresh orchestrated launches and new orchestrator assignments are rejected. Saved legacy profiles remain isolated in execution/legacy_automation.py; planners and maintainers own diagnosis in new runs. |
| Milestone-specific execution | New projects default to agent-led graph planning. Milestones are labeled ordinary nodes. Maintainer phase decisions need source evidence, not a mandatory frozen/compiler-certified milestone baseline. Strict existing projects retain explicit compatibility policy. |
Agent-Led Planning And Lean Verification¶
Roadmap PRs remain the accepted source for objectives, graph routes, milestones and their amendments. Ordinary source paths, dependency DAG validity, revision conflicts, permissions and review provenance remain checked. An idle existing strict project can be switched explicitly to graph workflow by a human; active run changes are rejected. Source edits and migration do not silently reinterpret accepted legacy contracts.
Objective formalization can start directly from its objective. Phase acceptance records the maintainer's evidence and chosen next inputs; revision 0029 allows a transition only when it matches that recorded acceptance. This also repairs an older immutable-field trigger that blocked objective phase advancement.
Generic library audits now use worker/lean_verify.py, projects/lean_checks.py
and projects/verification_jobs.py, with /api/v3/lean/verifications and
/api/v3/lean/checks. They require an exact-source ready library checkout and
host lease, independent of milestone manifests. Historical storage names, routes
and flags retain old receipt/lease identities. Audits reject admissions in
changed declarations and their transitive dependencies; semantic correctness
and merge authority remain separate review decisions.
Prompt And Context Provenance¶
The Markdown renderer substitutes allowlisted tokens once. JSON braces, shell variables and inserted user content remain literal. New sessions receive all templates and hashes in their immutable catalog. Retained sessions use those pinned templates; missing or corrupted new-bundle entries fail explicitly. Historical catalogs retain their original pinned core/function fields, with an explicit compatibility fallback for fragments that older releases never pinned. Prepared reviewer manifests keep their rendered text and source identities.
Generated mission and recovery instructions are rendered at record creation and stored with their owner; editing installed assets does not rewrite those persisted instructions. Objective continuations avoid repeating full role contracts, and ledger excerpts enforce the aggregate character budget including markup and identifiers. Context views continue to expose links for omitted detail.
Independent Review Evidence¶
| Scope | Evidence |
|---|---|
| Core entrypoints, API, scheduler, missions and persistence | Core review, catalog |
| Project/domain services, operations, dashboard projections and migrations | Domain review, catalog |
| Instructions, reviews, integrations, providers and search | Backend review, catalog, prompt review |
| Workers, configuration and agent client | Worker review, catalog |
| React sources and Vite configuration | Frontend review, catalog, References |
| Installer and commit hook | Installation catalog; hook failures preserve the original message and scratch is cleaned |
| Developer scripts, documentation and demo | Tooling/demo review, catalog |
The coverage reconciliation matches all 219 source files against their current hashes: 166 Python files with 1,210 named functions, 51 TypeScript/configuration files with 1,130 function bodies, and two shell entrypoints. Tests, generated builds and upstream dependencies are outside this source-body review. The independent core reader produced the specification before reading design prose; other readers reviewed comments as part of normal code assessment.
Catalog entries record actual body inspection at a source hash. They are not proofs or replacements for runtime tests. Changed source requires a fresh review of the changed body. Large lifecycle modules were inspected individually; they were not mechanically split merely to reduce branch counts or file sizes.
Deployment Status And Limits¶
The site and demo build locally and as reviewable CI artifacts. The repository's existing public Pages site is an older dashboard snapshot; this task does not replace it with uncommitted source. Publish the reviewed Documentation workflow on main to deploy the new site. No production service, database, provider configuration or live workspace was changed.
The catalogs document remaining limits, including abrupt verification-daemon death before durable helper ownership, connector schema drift, lexical Lean presentation and semantic mission paraphrases. Live-provider sessions, large-project resource behavior and rootless deployments require disposable operator trials before release. See the validation record for the completed integration checks and exact skipped scope.