AgentWorkforce/burn cli-v0.42.0

@relayburn/cli@0.42.0

on GitHub

[0.42.0] - 2026-04-28

Added

• Ghost-surface detector now consumes user-turn text from the content sidecar to recognise slash-command invocations (#172). burn waste --patterns ghost-surface now populates GhostSurfaceInputs.userTurnTextBySession from readContent({ sessionId }) for Claude / Codex sessions before running the detector. As a result, a Claude command invoked via <command-name>/foo</command-name> and a Codex prompt invoked via /openspec-apply no longer surface as ghosts even though they never appear as tool calls. The map is keyed by SourceKind first so a Claude <command-name> marker can't de-ghost an identically-named Codex prompt and vice versa. Sessions whose content sidecar is empty (content.store=off, pruned, or never captured) silently fall back to v1 (tool-call only) behaviour, matching the existing graceful-degradation contract elsewhere in burn waste. The CLI only loads sidecars for sources that have a slash-command notion (Claude / Codex), so OpenCode sessions pay no extra I/O cost.
• burn waste --patterns tool-output-bloat (#168). New cross-harness detector that flags oversized tool_result content riding in cache. Two signals fold into one table: Signal A reads ~/.claude/settings.json and the project .claude/settings.json and flags any env.BASH_MAX_OUTPUT_LENGTH whose token-equivalent (chars / 4) exceeds 15000 tokens (≈ 60000 chars; project overrides user, last-wins); Signal B walks the ledger's tool_result_events stream (#42 substrate) for any session in scope and reports per-(source, toolName) buckets above the threshold (default max(15000 tokens, p95) with a sample-size guard), pricing the next-turn carry cost at the source turn's model input rate. Findings render in the unified --findings table next to retry-loops, failure-runs, etc.; --json carries toolOutputBloats alongside the other detector arrays. Suggested fix surfaces as a WasteAction.paste — a character-unit corrected env line for Signal A (default 60000), a head / tail / grep reminder for Signal B's CLAUDE.md / AGENTS.md.
• burn waste --patterns ghost-surface (#166). New detector kind that flags user-installed surface files (Claude ~/.claude/{agents,skills,commands}/, Codex ~/.codex/{prompts,skills,rules,memories}/, OpenCode opencode.json + project skills folder) whose basenames never appear in the observed-names set mined from the turn stream. Output is a labeled Ghost user-installed surface table with columns source | kind | path | tokens | sessions | cost | note. JSON output (--patterns --json) gains a ghostSurface array; the unified findings payload (--patterns --findings) folds ghost findings in alongside retry-loop / failure-run / etc. and severity-ranks them by usdPerSession (per-session cost, not cumulative). Each finding's suggested fix is a command-style WasteAction for filesystem ghosts (mkdir -p <archive-dir> && mv <path> <archive-dir>/, shell-quoted for safety) and a paste-style instruction for OpenCode opencode.json entries (synthetic JSON-pointer paths can't be mv'd). OpenCode declared-catalog skills are emitted with cost: $0 and a catalog (#54) note in the table to avoid double-counting against the opencode-system-prompt catalog-bloat detector. Slash-command-style invocations are detected via #172 when the content sidecar is available. Bare --patterns (no value) now selects all 10 detectors; the previous count was 8.
• OpenCode ingest persists compaction events (#148). The OpenCode passive ingest path now appends parser-emitted compactions through the existing ledger compaction writer, so burn waste --kind compaction can see OpenCode context compactions with the same event shape Claude and Codex use.