Guildhall 0.7.0

Guildhall 0.7.0 is the release where the product starts behaving less like a task runner and more like a legible project room. It adds project planning, scoped lever settings, codebase orientation through the Corpus Map, workspace-aware execution, richer context packets, and a calmer project shell for decisions, recovery, and imported work.

What changed

• Project planning: Guildhall can describe work as a build map with survey, blueprint, framing, trade work, inspection, change orders, and punch list. The UI avoids literal map-style zoom and instead gives perspective shifts: project frame, active room/slice, task detail, and the current next action.
• Proportionate detail: advanced planning detail is available when a project needs it, but small projects should not be forced to manage phases, settings, or coordinator structure before the system can help.
• Scoped lever settings: project and global behavior settings now have a clearer inheritance model. Project settings can stay Same as global setting until a local override is justified, and repeated project-level preferences can later be promoted to global defaults.
• Settings UI cleanup: advanced settings moved from raw snake-case dumps into friendlier groups and controls. Enum-like choices use selectable controls instead of disabled-looking pseudo-buttons.
• Corpus Map integration: Guildhall builds a compact project map for agent context instead of dumping the whole repo. It captures file fingerprints, areas, imports, symbols, likely abstractions, read-next guidance, design-system signals, and verification hints.
• Context indexer role: model assignments now include contextIndexer. Deterministic Corpus Map refreshes stay cheap; semantic enrichment can use a smarter, schema-checked model when the project needs architecture summaries and worker guidance.
• Semantic map repair: malformed or schema-invalid semantic output gets a deterministic cleanup attempt and one model repair pass before Guildhall marks the map stale.
• Worker context packets: agents receive focused packets with task state, accepted plans, worktree/checkpoint evidence, likely files, Corpus Map guidance, design-system summaries, memory, decisions, and review rubrics where relevant.
• Design-system reuse guardrails: UI work is expected to inventory existing primitives, tokens, and patterns before editing. Repeated one-off ideas can trigger just-in-time abstraction consideration instead of bespoke controls.
• Open-model recommendations: the docs now capture the current development findings for open and open-weight model lanes. GLM is the recommended contextIndexer semantic lane; DeepSeek remains the cheap general challenger; Qwen remains the stronger default worker lane from the current replay set.
• Workspace projects: ./guildhall.yaml can describe a workspace that coordinates multiple buildable child projects. Task worktrees, bootstrap, and gates bind to the child project instead of blindly inheriting parent commands.
• Workspace import cleanup: imported notes are treated as source material and draft candidates, not automatically useful runnable tasks. The review UI better explains what will be imported, why source material matters, and when Guildhall is shaping an imported draft.
• Thread and Work state semantics: the shell now separates Needs you, Agent-active, Shaping, Recovery, Queued, and Drafts instead of using overloaded active/working/queued counts.
• Partial-progress recovery: a worker that touched durable files and then failed is shown as Needs recovery, not as both failed and queued. The drawer shows the recovery state and the activity evidence.
• Run-control acknowledgement: stopping a run immediately moves the top-bar control into a disabled Stopping... state so the click visibly landed.
• Mobile project navigation: the project rail is no longer reserved on narrow screens. Mobile uses a full-screen navigation overlay triggered by the hamburger menu.
• Button and shell cleanup: status buttons, neutral buttons, badges, rail controls, and project-shell spacing were tightened around shared components instead of accumulating one-off styling.
• Rich artifact protocol seed: the first schema and validation tests for guildhall-html-v1 are in place so future visual planning/review artifacts have a constrained, auditable lane instead of leaking arbitrary HTML through Markdown.
• Docs versioning workflow: the docs build now prepares a /next/ snapshot from current docs while the public root continues to default to the last published release.
• Serve shutdown fix: guildhall serve now handles interrupt shutdown cleanly instead of hanging after printing that Guildhall is shutting down.

Why it matters

The main product improvement is context discipline. Guildhall is now better at giving agents enough project knowledge to reuse the right abstraction without turning every prompt into a full repository dump.

That matters for both quality and cost. Workers get clearer pointers to existing helpers, packages, UI components, design tokens, tests, and project decisions. Reviewers can reject parallel implementations when the right primitive already existed. The coordinator can ask a small settings or split question only when the project complexity justifies it.

Proof

This release was shaped against real local projects, especially Narrative Harness and Looma + Knit. The proof covered:

• importing a documentation-heavy product project and turning notes into useful starter work;
• detecting when imported notes are still draft material rather than ready worker tasks;
• running a Corpus Map and context-indexer comparison across documentation, small-code, design-system, and hard-architecture tracks;
• verifying that GLM produced the best combined semantic context-indexer result in the current development ladder;
• confirming that Thread, Work, and the drawer agree on recovery/shaping state;
• checking that the project shell stays usable with pinned rail and mobile navigation;
• validating docs version behavior so Next can move ahead of the latest published npm release;
• keeping future experiments out of the release claim.

Validation

• pnpm test
• pnpm typecheck
• pnpm build
• pnpm docs:build
• focused Svelte tests for Thread, Work, Settings, workspace import, project shell, task drawer, mobile rail behavior, and Corpus Map integration
• browser walkthroughs against real local project routes

Still not claimed

0.7.0 does not claim future runtime or artifact experiments. This release establishes guardrails and context foundations so larger features have a safer place to land later.