# Changelog ## [Unreleased] ## [0.80.23] - 2026-07-19 ### Added - Added `historicalReplay.mode: "recommended"` to the frozen evidence strategy: the researcher now recommends non-blocking replay evidence with a reason instead of freezing it as a release blocker. Scheduled runs execute recommendations automatically; request-triggered runs evaluate first, then park at `awaiting-evidence` with a `pending-verification.json` decision record so the user chooses 执行/跳过/拒绝 (Inspector decision card with `e`/`s`/`r` hotkeys, auto-prompted in the TUI, or `/evo verify [run-id] execute|skip|reject`). - Added `verification.approval` config (default `ask`): set `auto` to execute recommended verifications inline on request runs too. - Added a detached `__worker-resume` background worker: an approved execute decision runs the missing replay, re-evaluates, and applies the release policy without blocking the TUI (wires up the previously dormant `resumeEvolutionEvidence`). - Added `retry [--from research|building|validating]`: resume a terminal run with fresh research, from the frozen plan at the builder (`resumeFrom: "building"` skips research and reuses the frozen corpus), or from the built component at validation, with background building resumption wired into the TUI and CLI. - Added `/evo model [role]`: interactively pick the model and thinking level for an evolution phase, persisted through the fail-closed control-config writer. - Added a 120s default execution budget for read-family tool calls (read/grep/find/ls) in headless model runs: an over-budget call is aborted and returned to the model as a tool error instead of hanging the phase. ### Fixed - Fixed required replay execution to be contract-driven: a frozen plan that requires `paired-replay` now executes it inline for any proposal tier with replay scenarios, instead of only T2 — previously a T0/T1 candidate whose plan required replay was evaluated against evidence the harness never produced and was rejected as invalid. - Fixed durable preference ids to be enforced at the schema level: a session-sourced preference id must be the deterministic instruction hash (`pref-`), so a builder can no longer stage ids derived from unrelated identifiers such as inbox UUIDs. - Evaluator passes now receive deferred recommended profiles explicitly and must not count their absence as a candidate defect. - Fixed code candidates to always patch Evo-Pi's own source repository: worktree staging and unknown-ABI building previously anchored to the triggering session's launch directory, so a background reflection started from a non-repository directory failed with "Resolve repository root failed" (and one started inside another project would have patched that project). `repositoryCwd` remains as an explicit override. - Added `grants`: lists every component's persisted capability grants with per-capability usage against its budgets (calls, cost, tokens) plus in-flight operations and reservations. - Added `history []`: read-only view of the bundle transition audit log (initialize/keep/rollback/pause and proposal decisions) with digests, proposal ids, and reasons. - Added `triage [now]`: shows the streaming-triage cursor and backlog, and `triage now` forces an immediate scan of the new session digests regardless of the cadence. - Added `go --scheduled` as the single entry point for the guarded cadence-gated evolution attempt; `scheduled-improve` remains a deprecated alias. - Added command-surface coverage for the workflow ecosystem: `packs [init