5. Preference memory with provenance#
Corrections ("don't add comments", "use bun") are detected on prompt submit and offered as one-keystroke memory candidates, scoped to the project or globally.
Each memory records why and when. When one is applied, the agent says so:
"applying your rule: no semicolons". Memory you can't inspect is memory you
can't trust; /memory lists, edits and deletes.
7. Honest completion#
A built-in Stop self-check: before finishing, compare the original request
against what actually changed, and state anything skipped, unverified or failing.
No silent narrowing of scope. If four of five items are done, that is the report — not "done".
8. Verification on by default#
After edits, run the project's test or lint command — declared in AGENTS.md, or
detected from package.json scripts or a Makefile — and report the output
verbatim. Not a summary of the output. The output.
10. Cost and context transparency#
The status line always shows context percentage and session cost. --max-cost
(or maxCostUsd in settings) sets a session ceiling, checked before each model
call rather than after — stopping once the money is gone is not a budget, it is
a receipt. Reaching it prompts to stop, raise the ceiling, or remove it; a
headless run stops and exits 3, because it has no terminal to ask. Cost is
computed from catalog pricing including cache reads and writes.
Context shapers#
Run in order before each call, cheapest first: cap individual tool results, prune old tool results to one-line stubs, then auto-compact at ~80% of the model's window — a model-written summary plus the last K turns verbatim, open todos, and files touched. Nothing rewrites history: compaction records a cut and a preamble, and the request is rebuilt from those.
Inspect and control#
/memory lists preferences and provenance; /memory forget <id> deletes one. Capture a correction with Ctrl-R for the project or Ctrl-G globally. Nothing is stored without that action. /context reports context shaping, /compact appends a summary, and /cost [usd] inspects or changes the session ceiling. 0 removes a ceiling. A model call may take cost above a ceiling because the check happens before the call, not during token generation.