v0.9.20

Added

• Completion contracts and verification targets: history.rs now infers task kind, observation requirements, and verification targets (URLs, paths, project scopes) for each turn, allowing the loop to track when work still needs a final read-only confirmation step.
• Structured tool-call semantics: New ToolCallSemantics, ToolCallOutcome, target hints, and verification modes let tools report whether a call observed state, mutated state, or both. Added shell-command semantics classification for run_command and terminal.
• Project-scope inheritance and enforcement: Turn context now extracts and normalizes project scopes from user text/history, injects _project_scope into delegated agents, and blocks out-of-scope tool calls unless multi-project work was explicitly requested.
• Goal failure summaries: fail_goal now persists a failure_summary, and the agent can build failure notifications from stored summaries or failed task details.
• Deferred-no-tool recovery metrics: New policy counters track forced tool-call retries, deferred-action detections, fallback model switches, and terminal deferred-no-tool error markers.

Changed

• Completion gating tightened: Consultant completion now blocks success responses until required verification happens, or returns an explicit "verification pending" reply when tools are unavailable in the current phase.
• Goal completion guardrails: manage_goal_tasks(action="complete_goal") now refuses summaries that admit partial progress or pending verification, while still requiring every concrete task to be resolved first.
• Tool execution plumbing upgraded: Tool execution now propagates structured semantics/metadata, merges fallback semantics for legacy plain-text tools, and uses those semantics to update completion progress.
• Project path normalization: Project-scope resolution now promotes nested directories to project roots and resolves configured project aliases consistently across follow-ups and delegated runs.

Fixed

• False-success replies after writes: Requests that mutate files, URLs, or project state no longer claim completion before a matching verification step.
• Delegation scope spoofing: Executor-spawned _project_scope arguments are now overwritten with the trusted parent scope instead of honoring model-supplied values.
• Pre-tool deferral loops: Repeated "I'll do it" text-only replies before the first tool call now trigger a hard tool-call directive, a fallback-model retry window, and clearer blocker text when no tools are available.

Full Changelog: v0.9.18...v0.9.20