Chris mode

Working conventions mined from feedback and correction patterns across this workspace. Read once per session if invoked. Don't restate it back to the user.

Autonomy and checkpointing

Default to acting, not asking. On routine, reversible, in-scope work: just do it. No permission prompts for read-only steps or standard tool use.

The exception is blast radius, not task size. Once a scope is approved, still pause before any single step that's expensive to reverse: force-push, large deletion, rewriting already-shipped code, an external side effect (sending a message, merging a PR). A short "found X, about to do Y" before that step is cheap. Running the whole chain unattended is not what approval meant.

On long autonomous runs (multi-task plans, SDD execution), checkpoint every 5-8 tasks or when a session is clearly running long. Flag progress and rough time/token burn, and suggest a fresh session rather than pushing to a hard cap.

If a background task doesn't return within one or two polling waits, stop polling in a blocking loop and switch to an event-driven watch instead.

Verification discipline

A subagent's or tool's self-report of "done," "tests pass," or "reverted cleanly" is a claim, not a fact. Before treating any of these as true:

• Rerun the actual command yourself and check the real exit code. Don't grep a summary for "PASS."
• Diff the actual files touched against what the report claims. Flag anything undisclosed.
• For "reverted to origin/main" or similar, diff against that ref directly.

This applies uniformly across agents (Codex, Antigravity, subagents). Failures are silent and confident regardless of source.

Before agreeing that a system exists, or that a design doc's cited components are real, grep for the actual artifacts (scripts, config, generated files) rather than reasoning from the doc's own confident tone. This cuts both ways: don't voice doubt about something real without checking first, and don't accept "now that X, Y, Z are established" without checking that X, Y, and Z actually are.

Before trusting a fix in any system with codegen, caching, or other side-effecting regeneration, run the code path that owns the file, not just read the diff. A plausible-looking fix can be silently undone by the next real invocation.

Git and process

• State every file path in full (C:\dev\...), including inside code blocks, chat replies, and end-of-turn summaries. Never a bare relative path.
• In this repo, Edit/Write rewrite CRLF files to LF wholesale, turning a one-line change into a full-file diff. For a small edit to a CRLF file, use sed -i instead, and verify with git diff --stat that only the intended lines changed.
• Never git reset --hard with a dirty working tree. It discards all uncommitted work, not just the target commit. Check git status first; if dirty, stash, checkout -- <path>, or work in a separate worktree instead.
• After any git add -A, check the command's own output for "adding embedded git repository" warnings. git status alone won't show them, and a silent broken gitlink carries no real content.
• If a repo has a CI bot that pushes back on every push (release automation, sync commits), push once per session rather than after every commit, to avoid a rejection/rebase cycle on each subsequent push.
• Commit-type tags: chore(sync): for bulk/automated resync commits (state regen, mirrored files); test: (or test(qa):) when a commit's primary content is test or regression coverage, even if it touches other files too.

Subagent dispatch

• When dispatching a subagent to a specific path, make verifying that checkout (cd <path> && git rev-parse --show-toplevel && git rev-parse HEAD) its literal first command, not a prose instruction. Prose doesn't guarantee execution, and silent wrong-checkout work is expensive to catch after the fact.
• When handing Codex a plan with literal code blocks, tell it explicitly to copy them verbatim. It will otherwise retype or condense them and introduce silent syntax errors that look like environment hangs.
• A subagent's completion report describes the scoped task, not everything that landed. Check the actual commit range (git log, git show --stat) for anything auto-pushed, not just the commit mentioned in the report.
• For multiple same-shape background agents running in parallel, poll proactively once elapsed time passes roughly 2x what sibling tasks took. Don't wait passively for a completion notification a hung task will never send.

Platform

Windows: run Docker commands from PowerShell directly, not bash under WSL. WSL bash can't reach the Docker Desktop socket.