You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Added OpenCode-compatible arbitrary and multiple context-price tiers, selected independently for every completed provider step before the legacy 200K fallback.
Added reconciliation between message-derived step-finish telemetry and OpenCode's session aggregate, with visible warnings when token buckets or recorded cost disagree.
Added explicit snapshot-boundary, retained-content, active-revert, missing-metadata, and approximate-tokenizer warnings so incomplete data cannot look exact.
Added adaptive precision for small USD values and per-million rates, preserving values such as $0.00001234 and $0.00875/M in displayed formulas.
Added dedicated regression coverage for session reconciliation, current OpenCode skill formatting, tokenizer fallbacks, multi-tier pricing, report isolation, and cyclic child-session graphs.
Changed
Reframed reports into three distinct layers: locally tokenized retained content, recoverable OpenCode-recorded usage, and visibly labeled explanatory estimates.
Renamed API-call totals to completed provider steps and documented that the step invoking TokenScope, later report reads, and the final response are outside the snapshot.
Updated pricing estimates to use live OpenCode provider metadata first, with the bundled models.dev-derived catalog retained only as a warned fallback.
Updated skill accounting for OpenCode v1.17.18: the verbose permission-filtered catalog belongs to system context, while the static skill-tool description is counted separately from its schema.
Changed repeated skill accounting to tokenize and sum each persisted result instead of multiplying the first result size by the call count.
Changed zero-cost and mixed-cost reporting to distinguish OpenCode-recorded cost from public API-rate estimates without assuming a subscription or provider invoice.
Changed report output to unique, atomically written files in private per-invocation directories under OpenCode's temporary path, leaving analyzed worktrees untouched.
Changed updater behavior to reinstall declared dependencies before rebuilding, preventing stale plugin SDK and tokenizer packages from surviving an update.
Documented the v1.17.18 accuracy target and the remaining limits around generated system prompts, provider-transformed tools, compaction, reverts, and bundled-pricing freshness.
Fixed
Fixed sub-200K and multi-tier cost estimates, strict threshold selection, tier-before-legacy-fallback precedence, per-step tiering, and zero-valued omitted tier cache rates to match OpenCode.
Fixed cost formulas and labels so reasoning uses the output rate, fresh/cache-read/cache-write buckets remain non-overlapping, and displayed rates agree with the calculated totals.
Removed invalid system-prompt remainder inference that could double-count assistant history and removed the nonexistent fixed project-tree estimate.
Bounded heuristic context components so they cannot exceed the observed first cache-write bucket, which is now labeled as an allocation anchor rather than a context-window measurement.
Fixed OpenCode-hosted Claude and other non-OpenAI models being tokenized as OpenAI, mapped newer GPT/o-series models to the GPT-4o tokenizer family, and preserved real model names when tokenization is approximate.
Fixed silent tokenizer fallbacks by reporting unknown encodings, load failures, and encode failures before using an approximate count.
Fixed retained-content handling for ignored:true text, completed compacted tools, and replayable errored or interrupted tool results.
Fixed tool-schema reporting that implied current metadata was the exact historical enabled toolset; active-agent permissions, MCP additions, and provider transforms are now called out explicitly.
Fixed skill-catalog reconstruction that double-attributed an obsolete catalog-bearing tool description and made unavailable skill/agent metadata visible.
Fixed recursive subagent attribution by preferring the session's agent field and stopping duplicate or cyclic traversal.
Fixed unsafe pricing-prefix matches that could assign a related model's price without a valid model-version boundary.
Fixed concurrent report overwrites, repository pollution, shell-path handling, and stale fixed-filename instructions.
Security
Restricted report directories to owner-only permissions and report files to mode 0600 on Unix.
Added ownership, symlink, and writable-parent validation before writing potentially sensitive session reports under OpenCode's allowed temporary path.