Task evidence and recovery
Keep an objective, acceptance criteria, source-bound validation, and phase history together across Claude Code and Codex. This minor release adds the task workflow and includes the workshop and runtime improvements delivered since 4.3.72.
New capabilities
/taskin Claude Code and$taskin Codex connect current proof to acceptance criteria, detect stale source or artifacts, and preserve phase checkpoints and explicit interrupted-action reconciliation.- The workshop adds a read-only task view with evidence, status and next action. Background verification keeps other dashboard requests responsive.
- Seeded comparison schedules distinguish actual, fixture and unknown outcomes across no-agent, single-agent and ForgeFlow trials.
- Memory dependencies and task-linked feedback withhold stale or contradicted guidance.
- Fleet contracts check ownership, pinned baselines and distinct resource declarations. Local CI maintenance deduplicates failures, applies bounded repairs, validates results and prepares draft handoffs.
Reliability fixes
- Oversized task updates preserve the last readable record instead of exceeding its 16 MiB read limit.
- Committed, staged, unstaged and untracked paths are inspected separately so later edits cannot conceal ownership violations.
- Task scans share source and artifact checks, bound concurrent workers and deliver timeout errors promptly while cleanup finishes.
Also included since 4.3.72
The refreshed workshop and Ember companion, session dashboard opening, visual HTML/PDF guide, rewritten wiki onboarding, Codex model inheritance, research divergence tools, memory and command-learning controls, runtime hardening, and deterministic CI.
Update and boundaries
Use /update-forgeflow in Claude Code or $update-forgeflow in Codex, then restart the host and any running dashboard to load the updated runtime. A Git worktree with an initial commit is required for task evidence. Existing workflows without task records remain supported.
Task recovery restores saved phase context, not hidden model state. Fleet validates declarations without reserving infrastructure. Maintenance prepares local drafts without publishing a PR. Comparison fixtures are not claims of model superiority or user adoption. Initialized Git submodules remain unsupported by whole-repository task snapshots.
Validation
The task implementation and corrective review passed 195 test commands, including TypeScript and both installed host packages, and 25 browser tests. Regression cases cover record growth, Git-layer cancellation, real archive responsiveness and native-worker deadline cleanup.