[EvanProgramming]

Native Plan/Auto turns still depended on textual controls, and native execution could lose task identity, miss context costs, or report unsuccessful work as completed. Interrupted mutations and background processes also lacked sufficient cancellation and restart evidence.

This draft is stacked on #325 and contains the V3 finalization repairs:

• Use typed native execution and interaction controls across Agent/Ask/Plan; isolate legacy parsing and preserve literal replies through CLI/Web/TUI.
• Keep approved objectives immutable while refining implementation; bind task receipts explicitly, pin active plans, and move strategic transitions into the tasks-owned PlanningService.
• Include schemas and continuation payloads in context preflight, compact complete history units, and retry only an unconsumed generation after overflow. Provider/model/schema changes invalidate reported token counts.
• Block unknown-effect replay; persist scoped process intent and uncertainty before dispatch, retain bounded restart context, and preserve cancellation under SQLite/Python lock contention without changing ordinary persistence waits or conversation CAS.
• Derive unresolved effect failures from typed receipts, preserve strict child reports and plan completion status, keep Jev advisory, and require complete measured paired evidence before learning promotion.
• Protect original benchmark tests, retain independent safety components, stop/checkpoint the first safety failure, meter DeepSeek with a cumulative CNY 15 cap, and bind rows to the actual dispatched apparatus. The 30 IDs, three repeats, two arms and statistical thresholds remain fixed.
• Keep public downloads on published v2.0.8 assets and mark V3 unreleased.

Validation on clean signed candidate 469d76e7c6b758f4c12e2181d673ef0ab6191388: 1,374 full local tests passed (two Linux-only sandbox skips on macOS), with 530 tracked-file hashes unchanged. Check/lint/docs, Go tests, real PTY TUI and agent/subagent acceptance, retained wheel/sdist installations outside the checkout, and Docker persistence passed. GitHub verified all 16 commits. Exact-candidate six-environment CI and >=1,800-second soaks are running in 42292145829; Windows baseline/CPU/SQLite/combined controls are running in 42292145962. Historical load controls completed 80 passing observations; these do not establish the original timeout's cause.

Release remains blocked. Same-process child actor authentication and detached macOS descendant ownership are unproven; full campaign preflight stops before credential lookup or paid requests. R260 remains unattributed. Current routing benefit ablations, the live 180-row comparison, all 17 passing signed records and same-SHA main integration are outstanding. Historical DeepSeek exposure re-audits to CNY 0.582914960007; no external DeepSeek calls were added. No tags, draft/public Releases, package publication or preview publisher were performed.

Related: #227–#268 and #319. Keep this draft and main integration blocked until the formal requirements are satisfied.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

• X
• Mastodon
• Reddit
• LinkedIn