[nvdorman]

Summary

• resolve compaction limits from active-model metadata and include tool schemas in request-size estimates
• normalise per-call context size by provider instead of double-counting cached prompt tokens
• preserve cache-read/write counters separately for billing and analytics
• prevent persistent terminal sessions from inheriting non-POSIX shells such as fish
• reset dashboard activity timing per tool/turn state and make Stop interrupt the backend
• inherit the parent project for background workers and honour isolated worktrees
• add regression coverage for context/cache semantics, compaction estimation, shell completion, and sub-agent workspace inheritance/isolation

Reproductions

Cached context gauge

In session ses_52aed4a75b7997bc6d6d, an OpenAI-compatible usage record reported input_tokens=145878 and cached_tokens=145408. OpenAI defines cached tokens as a subset of prompt tokens, but the gauge added both and showed roughly 279k/256k (100%). The real context was about 145.9k/256k (~57%). Provider adapters now expose a normalised context size: OpenAI/Gemini use total prompt tokens, while Anthropic includes its separately reported cache read/write tokens.

Terminal hang

In session ses_206f9da8, $SHELL=/bin/fish rejected the POSIX $? completion sentinel, so completed commands waited until the 300-second timeout.

Background sub-agent workspace

In parent session ses_52aed4a75b7997bc6d6d, task task_e5b480030e8ab5c79713 created child ses_73594af85e6d0c9e0274 in the global Antares workspace instead of /home/nvdorman/captcha-solver, making the first file tools fail as outside the workspace. Background delegation now shares the synchronous workspace/isolation path.

Verification

• go test ./internal/llm ./internal/agent
• go test ./internal/agent ./internal/worktree ./internal/tools ./internal/server
• make check
• bun run build

[enowdev]

Merged — thank you, @nvdorman. This is a well-reasoned fix with real reproductions and solid test coverage, and I reviewed it carefully before integrating.

The per-provider context normalization is exactly right: OpenAI/Gemini/Codex report cached tokens as a subset of the prompt (so they must not be added), while Anthropic reports cache read/creation separately (so they must be summed) — verified across all four adapters. The ContextSize() helper with its conservative InputTokens fallback is the clean way to express that. The estimateRequestTokens change to count tool schemas is a genuinely good catch — large tool packs really can delay compaction until the provider rejects the request. The sub-agent workspace refactor also tightens write confinement on the isolation-failure path, and the fish-shell fix is spot on.

Integration note: this landed alongside a parallel in-progress fix for the same runaway-gauge symptom, so I merged by hand rather than a straight fast-forward. Your ContextSize() value won on the usage emit (it's the more provider-aware one), the tools parameter was kept on maybeCompact, and it was reconciled with a per-model context-window lookup (so e.g. glm-5.2 reports its real 1M window). Everything builds and the full llm/agent/tools/worktree/server suites pass. 🙏