Keep autonomous prompts separate from persisted user messages while retaining
conversation history and queued user attachment routing. Carry verified
checklist evidence across segments and evaluate only changed criteria.
Cancel model streams without interrupting checkpoint database writes. Fence
late worker admission during shutdown, persist accepted queued messages and
attachments, and leave interrupted execution leases recoverable.
Add regressions for prompt selection, cumulative evidence, cancellation I/O,
queued input durability and shutdown recovery. Verify 250 focused tests,
200 real General Assistant conversation rounds, 12 checkpoints across 13
autonomous segments, and pause/resume/disconnect/restart boundaries.
- Add a database-backed supervisor with durable scheduling, fenced leases,
cooldowns, bounded worker concurrency and expired-lease restart recovery.
- Default new goals to persistent execution with zero meaning unlimited
cumulative budget; preserve legacy goals and explicit positive limits.
- Yield bounded graph segments to the supervisor instead of ending unfinished
goals at graph-local continuation limits. Require persisted checklist
evidence before accepting completion, including concurrent criterion edits.
- Share conversation admission across interactive and background execution;
preserve partial replies, usage and queued user input during interruption.
- Persist Stop and missing-input pauses, respect approval boundaries, and
commit resume and approval-denial transitions with correct transactions.
- Retry identifiable transient failures with backoff; retain visible pauses
for budget limits and errors that require review instead of replaying tools.
- Expose owner-authorized execution status and reconnectable scheduling events;
add H2, MySQL and Kingbase migrations, API types and bilingual documentation.
Validation: 298 focused backend tests passed, including persistence, restart
scheduling, approval races, cancellation, admission and existing runtime tests.
Frontend type checking, bundled-doc parity and ID precision checks passed.
V188 is registered and all three dialects have unique migration versions;
the migration-map audit still reports 95 pre-existing missing registrations.
Scope: single-backend native runtime. Recovery checks existing state before
repeating effects; this does not promise exactly-once external tool execution.
Stopping a conversation previously disposed the outer reactive stream without reliably cancelling synchronous tool callbacks running on worker threads. Shell commands, skill scripts, and Playwright sessions could therefore outlive the visible chat turn.
Propagate cancellation through run-scoped hooks, interrupt active tool execution and parallel batches, terminate subprocess trees, close per-conversation browser sessions, and wait briefly for final stream persistence before acknowledging Stop.
Expose an explicit interrupting state in the chat input to block duplicate stop clicks and show progress, with regression coverage for cancelling an in-flight synchronous tool callback.
Normalize stream idle timeout overrides, make WebChat orphan cleanup use exact run handles so stale callbacks cannot mutate replacement runs, and preserve approval replay usage metadata with regression coverage.
Apply a Reactor inter-frame idle timeout at the streaming chat chokepoint so half-open provider body streams surface through the normal retry/failover path. Also keep the HTTP timeout documentation accurate and cover the behavior with focused tests.
load_skill returns SKILL.md in full by design — it is the model's behavioral
contract, and pagination by default would let the model silently miss later
mandatory sections. read_file / readSkillFile / load_skill are therefore on
the spill-exclusion list so their output is never replaced by a disk pointer.
The exclusion only covered half the path. In spillRawOrTruncate, an excluded
tool's result came back from persistIfOversized unchanged (no spill), failed
the SPILL_MARKER_PREFIX check, and fell through to truncateToolResult(8000) —
so an 8261-char SKILL.md was hard-cut through the middle and stamped with a
'[TRUNCATED: ... middle omitted]' marker. Weaker models ignore the attached
fidelity note and fabricate the removed span, inventing tool calls against
endpoints the skill never described.
- spillRawOrTruncate now returns retrieval-excluded results raw; the per-turn
aggregate budget stays the backstop.
- Outsized SKILL.md degrades to resumable pagination instead of an unbounded
inline dump. Never a lossy middle-cut.
