Why

complete_step's enforcement is bound to a per-turn evidence ledger (a.evidence.Reset() runs at the top of every Run). Two consequences users have hit:

1. Cross-turn evidence is rejected. Command evidence already falls back to scanning the full session (fix(agent): reduce evidence verification false signals #3587), but diff/files (path) evidence did not — so citing a file you edited in an earlier turn (or before compaction) failed with "no matching writer receipt in this turn".
2. The loop doesn't close across turns/compaction. The final-answer gate reads only this turn's todo_write, so once the structured list is gone (a later turn, or compacted into a summary) a premature "all done" slips through — the unclosed loop in [Bug]: todo_write / complete_step 失败后，Agent 仍允许最终回答“全部完成” #2917.

What

• diff/files session fallback (mirrors fix(agent): reduce evidence verification false signals #3587). PathsProvenInSession resolves a path against any successful (non-errored) write/read tool call in the transcript; verifyStepEvidence consults it when the per-turn ledger misses. Fabricated paths are still rejected.
• Host-side canonical todo list. Agent.todoState is the latest todo_write with completions applied, rebuilt from the session on load/rewind (latest todo_write + replayed complete_steps). It never rides in the prompt, so it survives turn boundaries and compaction. The final gate falls back to it when a turn did work without a todo_write, so a premature "all done" is blocked until the plan is genuinely finished — bounded by the existing maxFinalReadinessBlocks.
• complete_step advances the list. A successful sign-off marks the matching step completed, promotes the next, records a synthetic todo_write for the in-turn gate, and emits an event so the panel updates. The model is told it no longer needs a todo_write to mark completions — which removes the manual batch-completion step that [Bug] Agent批量提交complete_step导致todo_write校验失败——违反串行工作模式 #3909 tripped over (planApprovedMessage / executor handoff / tool copy updated to match).

All host-side, zero added tokens.

Tests

Unit: cross-turn diff/files fallback, canonical gate fallback, advance + promote, rebuild (incl. skipping failed complete_steps).

End-to-end (real complete_step/todo_write builtins driven through Agent.Run across turns):

• serial plan, host auto-advances, no batch todo_write, final answer allowed
• cd-prefixed command drift accepted in-turn
• cross-turn diff evidence accepted via session fallback (and an unbacked path still rejected — asserted at the tool-result level, not just Run exit)
• cross-turn canonical gate blocks a premature "all done", then clears once the steps are actually signed off

go test ./... green; gofmt/go vet clean.

Closes #2917