Releases · CleanSlice/runtime

07 Jun 19:07

v0.22.0

6384ea5

v0.22.0 Latest

Latest

feat(bridle): inject `data-prompt` integrator context into the agent

The embed's data-prompt value flowed from the SDK → hub → runtime but was silently dropped during bridle wire-message normalization, so the agent never saw it.

Fix: carry the value onto Message.prompt and fold it into the system prompt as a clearly-marked untrusted # Integrator Context section (situational awareness only — not instructions that override the agent).

Changed

channel.types.ts — typed prompt?: string on Message + buildMessage
bridle.repository.ts — read msg.prompt from the wire payload
runtime.service.ts / agent.module.ts / agent.service.ts — thread integratorPrompt through to buildSystemPrompt

Assets 2

07 Jun 12:42

dmitriyzhuk

v0.21.1

bc21ace

v0.21.1

Fix: context-overflow on sessions with huge tool outputs

Sessions with few but very large tool_result events (raw API payloads) could exceed the model's context window even though the event count stayed under compactionThreshold — leaving the chat permanently stuck with "chat too large".

Changes

Cap tool outputs sent to the LLM — buildHistory now trims tool_call/tool_result payloads to 4000 chars/string before sending to the model. Full data stays on disk for retrieval.
Size-aware compaction — compaction now also triggers on serialized session size (compactionBytesThreshold, default 200KB), not just event count.

Stuck sessions self-recover on the next message — no manual file surgery needed.

Assets 2

29 May 20:08

github-actions

v0.21.0

b8d313d

v0.21.0

Full Changelog: v0.20.0...v0.21.0

Assets 2

25 May 15:23

github-actions

v0.20.0

524ba43

v0.20.0 — resource_status tool + delayed-turn hint

Agent self-awareness of operating resources

The agent can now inspect its own LLM/system resources and gets nudged when something is wrong.

New tool: `resource_status` (admin-only)

Returns a structured snapshot:

llm — provider, model, context window, OAuth/API credentials with per-token cooldown state (cooldownUntilMs, consecutive429s), anyAvailableNow, soonestAvailableMs, primaryOverloaded.
usage — today's totalInputTokens / totalOutputTokens / totalCallCount.
system — container rssBytes, memoryLimitBytes, memoryUsagePct, cpuPercent, cpuQuotaPct. Reads cgroup v2 → falls back to v1 → degrades to host (macOS / non-Docker) with just RSS + CPU%.
lastTurn — duration, retry count, rate-limited / overloaded flags of the previous LLM call.

Auto-hint on delayed turns

After any LLM turn that took more than 30s, hit a retry, was rate-limited, or saw an overloaded_error, the runtime injects a one-shot # Resource Notice block into the next system prompt:

Your previous turn took 45s (retries=2, rate-limited). If you suspect resource pressure, call the resource_status tool…

The hint fires exactly once per delayed turn (consume-on-read).

Under the hood

ClaudeRepository.getResourceSnapshot() exposes the private OAuth pool + tokenStates via a normalized DTO.
withRetry tracks retries / rate-limit / overload and attaches them to ModelResponse.meta.
New LastTurnStatsTracker lives on LoopService; RuntimeService.buildPrompt consumes it per turn.
LlmModule.getResourceSnapshot() returns a minimal stub for non-Claude providers so the tool stays callable.
BuildPromptOpts.extraHint is the generic injection point for the next-turn nudge.