jeopi-cli 16.2.30 → 16.3.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +16 -0
- package/dist/cli.js +2341 -2277
- package/dist/types/config/settings-schema.d.ts +12 -2
- package/dist/types/goals/runtime.d.ts +16 -0
- package/dist/types/goals/verifier.d.ts +9 -0
- package/dist/types/session/agent-session.d.ts +22 -7
- package/package.json +12 -12
- package/src/autoresearch/index.ts +60 -6
- package/src/config/settings-schema.ts +14 -2
- package/src/goals/runtime.ts +51 -0
- package/src/goals/verifier.ts +100 -0
- package/src/prompts/agents/goal-verifier.md +51 -0
- package/src/sdk.ts +2 -0
- package/src/session/agent-session.ts +51 -8
- package/src/task/agents.ts +2 -0
package/CHANGELOG.md
CHANGED
|
@@ -2,6 +2,22 @@
|
|
|
2
2
|
|
|
3
3
|
## [Unreleased]
|
|
4
4
|
|
|
5
|
+
## [16.3.0] - 2026-07-12
|
|
6
|
+
|
|
7
|
+
### Added
|
|
8
|
+
|
|
9
|
+
- Goal-mode completion now spawns an independent `goal-verifier` sub-agent (fresh context, no exposure to the maker's transcript/reasoning) to grade a `goal({op:"complete"})` claim against the objective and the actual repo diff, after the existing deterministic evidence gate passes. New `goal.verifyCompletion` setting (default `true`). A non-`okay` verdict (`iterate`/`reject`) bounces the completion once with the verifier's justification and required fixes; a second `complete` call with no new mutation/verification evidence in between passes (the same bounce-once escape hatch pattern as the existing evidence gate, tracked independently so a wrong verifier verdict can't permanently block completion).
|
|
10
|
+
- Autoresearch mode now auto-resumes a pending (run-but-never-logged) experiment on the next session load — `session_start` and `/resume`-reason `session_switch` re-arm the same auto-continuation the live in-process loop already used, closing the gap where a process restart required manually re-invoking `/autoresearch` to pick a run back up. Bounded to pending runs completed within the last 24 hours, and skipped entirely when messages are already queued.
|
|
11
|
+
|
|
12
|
+
### Changed
|
|
13
|
+
|
|
14
|
+
- `advisor.includeThinking` now defaults to `false`: the advisor no longer sees the primary agent's raw thinking blocks by default, only messages/tool calls/results/diffs — reduces anchoring on the maker's own already-committed reasoning trail instead of bringing an independent angle. Still toggleable via Settings → Advisor.
|
|
15
|
+
- Raised `SHUTDOWN_CONSOLIDATE_BUDGET_MS` (the `/quit`/`/exit` shutdown budget for the mnemopi consolidate pass and the hindsight retain-queue flush) from 1500ms to 8000ms — the prior budget was tighter than a single LLM fact-extraction round-trip (1-3s), so it lost the race on effectively every session with any memory-worthy content.
|
|
16
|
+
|
|
17
|
+
### Fixed
|
|
18
|
+
|
|
19
|
+
- The Hindsight retain-queue flush on session dispose (`AgentSession.dispose` → `HindsightSessionState.flushRetainQueue`) made unbounded `fetch()` calls with no timeout, so an unreachable or hung Hindsight server could block `/quit`/`/exit` indefinitely. Now bounded by the same shutdown budget used for mnemopi consolidation when one is supplied; remaining items are dropped with a logged warning on timeout instead of hanging the process.
|
|
20
|
+
|
|
5
21
|
## [16.2.30] - 2026-07-10
|
|
6
22
|
|
|
7
23
|
### Added
|