release notes · v0.5.16
skylakegrep 0.5.16 — bounded agent latency and stronger scoped evidence
0.5.16 is a reliability and agent-performance release. It keeps the 0.5.15 agent presets, but makes their runtime contract stricter: machine-readable agent calls now use bounded local-model timeouts, skip heavyweight foreground refresh lanes, return a wider compact evidence set by default, and filter low-evidence JSON candidates before they reach the next LLM step.
The goal is simple: when an agent calls skygrep, it should get useful
structured context quickly, and if the semantic index is still warming up,
the command should return bounded evidence or an honest empty result instead
of making the user wait indefinitely.
What changed
- Agent and JSON calls now use explicit bounded budgets:
SKYGREP_AGENT_ROUTER_TIMEOUT_Sdefaults to1.5,SKYGREP_AGENT_MODEL_TIMEOUT_Sdefaults to3.0, andSKYGREP_AGENT_CASCADE_TIMEOUT_Sdefaults to8.0. - Human interactive searches still allow deeper foreground work by default:
SKYGREP_FOREGROUND_MODEL_TIMEOUT_Sdefaults to12.0, andSKYGREP_CASCADE_TIMEOUT_Sdefaults to30.0. - Local Ollama generation and router calls now pass request timeouts all the way down, so a slow model call degrades to fallback behavior instead of blocking the agent loop.
- Machine-readable agent paths no longer run expensive foreground refresh, symbol, or graph-prior maintenance before returning JSON context. Background indexing still continues.
--agent-contextnow returns top-8 compact evidence by default, matching the documented long-form contract:--json --content --detail standard --top 8 --no-rerank.- JSON agent results are filtered through a configurable evidence floor
(
SKYGREP_AGENT_MIN_EVIDENCE_SCORE, default0.50) to reduce misleading low-confidence context. - Cold lazy indexing walks large directories incrementally under the foreground budget instead of blocking on full-tree discovery.
- Fast intent routing keeps pathlike phrases intact and adds more generic implementation-location prototypes, improving code-location queries without adding topic-specific trigger lists.
Agent command contract
Use the same presets introduced in 0.5.15. The difference in 0.5.16 is that the presets are now bounded in the runtime, not only documented:
skygrep --agent-fast "where is token refresh implemented?"
skygrep --agent-context --include "src/**" "what does token refresh do?"
skygrep serve --port 7878
skygrep --agent-daemon --agent-context "what does token refresh do?"
Escalate only when the next step needs it:
skygrep --content --detail full --include "docs/migration-plan.md" "show the deployment steps"
skygrep --answer --content "summarize the retry policy"
Benchmark receipts
Release validation used only generic repository-maintenance queries and placeholder examples. The private local prompts and paths used during interactive debugging were not written into public release artifacts.
Targeted 0.5.16 smoke matrix on the current repository:
| Case | Shape | Result |
|---|---|---|
| T1 | scoped path-only | 1.945 s, 100 % expected-path coverage |
| T2 | scoped evidence | 5.117 s, 100 % coverage |
| T3 | router-timeout evidence | 4.840 s, 100 % coverage |
| T4 | cascade-timeout evidence | 4.805 s, 100 % coverage |
| T5 | setup-instruction discovery | 3.059 s, 100 % coverage |
| T6 | deep known-file read | 4.406 s, 100 % coverage |
| T7 | broad implementation lookup | 4.920 s, 100 % coverage |
| T8 | absent-concept negative control | 4.554 s, correctly returned no evidence |
The top-8 --agent-context retest for a multi-file implementation query
returned the expected auxiliary file that top-5 could miss before this release.
Compatibility
- No index rebuild is required.
- Existing
skygrep setupmanaged blocks should be refreshed so agents learn the current top-8--agent-contextcontract and daemon-first guidance.skygrep setup --checkaudits that without modifying files. - Users who intentionally want longer agent waits can raise the
SKYGREP_AGENT_*_TIMEOUT_Svariables, but the default is optimized for agent tool loops where stale or ambiguous evidence should fall back quickly.
Verification
- Full test suite:
358 passed, 2 warnings, 20 subtests passed. git diff --check: clean.skygrep setup --check: all managed snippets current.- Source privacy scan and distribution privacy scan are required before publishing.
Privacy
All public examples in this release use fictional placeholders such as
token refresh, migration-plan.md, and retry policy. No real user
prompt, local absolute path, private filename, screenshot, document category,
name, or email is included in the release notes, README, docs site, wheel, or
sdist.