file_read: offset past EOF now returns a clear error instead of silently
returning empty content, which caused models to think files were empty
and retry with different offsets.

list_dir: add "total" field to every response so models know the real
directory size when truncated (count capped at 500 but total reflects
the actual entry count).

grep: switch map → filter_map so unparseable rg output lines (e.g.
binary-file notices) are dropped rather than emitted as malformed
{content: "..."} objects missing the file and line fields. Also moves
.take(100) after filter_map to ensure up to 100 *parseable* matches.
Adds the first test module for grep.rs (6 new tests total across the
three files).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…y, session

commands/mod.rs had grown past 3,000 lines. Move the command handlers into
four focused submodules — config_cmds (/config, /backend, /model, /verbose…),
context_cmds (/context, /compact, /reset, /checkpoints), memory (/index,
/map, /memory, /remember, /forget), and session (/new, /undo, /session,
/resume, /export, /path) — leaving dispatch and the command list in mod.rs.
No behavior change.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Add turn_trace: every turn appends structured events (tool calls with
redacted args, approvals, compaction, warmup, timing) to a sidecar at
.sessions/<session-id>.events.jsonl, enabled by default via
display.eventLog.enabled. API keys and sensitive object keys are redacted
before anything is written.

/trace on|off surfaces nested subagent/critic tool calls as indented lines
in the TUI — previously their activity was invisible (events swallowed) —
without flooding the parent context. Tool calls now carry a depth field, and
the subagent/evaluator tools forward their inner events when tracing is on.

The end-of-turn status line gains a timing breakdown (TTFT, model, tools,
approval, total), the loader shows which tool is running, compaction of
oversized tool output is now reported to the user with the original size,
and /export <session> events copies the event log. Also prints a context
pressure notice as the prompt budget nears the model's effective limit.

/export current events copies the sidecar; /new and /resume reset it to the
active session.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

…CI job

small-harness --eval <fixture> [--model M] [--json] runs a bundled agent
eval fixture from the shell and exits 0 on pass / 1 on fail, so evals can
gate CI. A new optional macos CI job runs two fixtures against Ollama
nightly or when a commit message contains [eval]; it is continue-on-error
so a flaky local model never blocks merges.

Add agent_integration_test: drives the real agent loop against a mock
OpenAI-compatible SSE server (no live LLM) covering a tool-call round trip
plus eval checks, and the hit_step_limit cutoff flag.

Two fixes surfaced while wiring this up: the rubric heading parser now
matches "(weight:" case-insensitively on raw bytes instead of byte offsets
from a lowercased copy (which can diverge for some Unicode chars), and the
HTTP client gets a 10s connect timeout so a dead backend fails fast instead
of hanging — without capping long streaming completions.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>