Find plugins

by Jeancarlo Javier

Two-line status line whose headline feature is the session's live workflow phase (Plan, Exec, Verify, Done), rewritten by the model itself at each transition

by voitta-ai

Count permission outcomes (auto-mode refusals, operator rejections) from Claude Code transcripts without the traps that corrupt the number while leaving the output plausible:

by voitta-ai

Continuous-learning meta-skill for both hosts:

by voitta-ai

Pull trustworthy numbers out of a dashboard that renders entirely client-side, by decoding the data blob it ships instead of scraping the DOM or driving a browser. Use when: (1) you want a dashboard's data as a feed for a report, a script, or another skill, (2) `curl` returns a big HTML page whose numbers are nowhere in the HTML, (3) you are about to reach for Playwright / headless Chrome just to read some totals, (4) you copied a rank or a percentage off the screen and want to know what it actually measures before quoting it, (5) numbers you derived from a decoded blob disagree with what the page displays. Covers finding the blob, decoding a positional-tuple event schema from the page's own render loop, porting derived-metric formulas rather than guessing them, and validating against a screenshot. Includes two traps that silently produce plausible-but-wrong numbers: a displayed rank that is really a row number under the active sort, and a "velocity"/delta that is intra-window rather than period-over-period.

by voitta-ai

list-metrics is a 14-day view, not a catalog; absence is not absence.

by voitta-ai

Catch the CloudWatch Logs metric-filter constraint that terraform cannot: a metric transformation may carry dimensions OR a default_value, never both - and the AWS provider has no plan-time guard, so the mistake sails through validate and plan and detonates only at apply. Use when: (1) an apply (often a merge-to-main auto-apply in CI) fails with "InvalidParameterException: Invalid metric transformation: dimensions and default value are mutually exclusive properties", (2) you are reviewing an aws_cloudwatch_log_metric_filter whose metric_transformation sets both default_value and dimensions, (3) you are deciding which of the two to drop and several environments publish the same metric name into one account+region. Core facts: the provider (verified on hashicorp/aws 6.56.0) declares no ConflictsWith between the arguments; the constraint can be settled empirically in seconds with `aws logs put-metric-filter` on a scratch log group; keep the dimension and drop default_value when the dimension is what separates environments - dropping the dimension instead blends their series into one.

by voitta-ai

Decide whether a CloudWatch alarm on a per-host application metric (Micrometer, Dropwizard, StatsD) reflects a fleet-wide incident or ONE sick host, and stop misreading its magnitude. Use when: (1) a per-host gauge/timer alarm fires and the Average looks catastrophic (e.g. an "average latency" of 21,242 when baseline is 2), (2) you are about to call an incident fleet-wide based on a CloudWatch Average, (3) a latency/queue-depth/pool-saturation metric spikes but request throughput and error counts stay flat, (4) you need to know the UNIT of a metric and get-metric-data did not return one. Covers the unweighted-mean-across-hosts trap, the Maximum/Sum concentration ratio, and SampleCount as a host-census signal.

by voitta-ai

Make AI coding agents show up as watchable cmux tabs/panes, and diagnose the six distinct reasons they do not: wrong launch path, Agent-tool spawn path (never tabs; they surface in the native agent list), a shim shadowed on PATH by a real tmux, the shim's own bare cmux being off PATH so tmux dies with exec: cmux: not found, a resumed pane with no $TMUX that starts an invisible real tmux server, and a teammate mode that resolved to in-process so every agent wedges with no pane, no TTY and no visible permission prompt. Includes the one-shot both-hops test, the ps check that separates a wedged transport from an agent that never launched, and reading the live process env rather than the shell's.

by voitta-ai

Diagnose why cmux does not resume agent sessions after a macOS reboot even though terminal.autoResumeAgentSessions is enabled.

by voitta-ai

Run a Claude Code agent and a Codex CLI agent as messaging peers in cmux panes, each named as a tab and both in the agent traffic log. Codex has no SendMessage and Claude's peer socket needs auth, so the cross-runtime hop is keystroke injection (cmux send + send-key Enter) through a small helper that also logs to xs explicitly; each agent names its own tab first via cmux identify + tab-action rename. Covers the Codex TUI parking on a startup modal with no way to read its screen (diagnose from ~/.codex logs and the missing rollout file, inject one Enter), and where the lead, teammate and Codex transcripts live on disk afterwards. Verified 2026-08-27: five rounds in 50 s, every message exactly once.

by voitta-ai

Find out why a cmux.json entry that passes `cmux config doctor` never shows up in the palette, plus-button menu or tab bar. Doctor is syntax-only, the schema is additionalProperties: true, and the unified log stays silent - so the working oracles are a backup-diff bisect, the binary's own [CmuxConfig] diagnostics, and enum triads read straight off `strings`. Worked example: a workspaceCommand action hidden by `"restart": "restart"` (valid: ignore/confirm/recreate, or omit for a new workspace each time).

by voitta-ai

Make agent-to-agent traffic visible to the human watching: a structured SendMessage envelope (kind + direction + topic in the summary field, plus a one-line header that survives raw transcripts) paired with a cmux sidebar status pill per workspace showing whether a session is waiting on a peer or owes one an answer. Covers why spawned teams are cross-workspace traffic by construction (tmux-compat maps a window to a workspace), why the pill key must not collide with cmux's own claude_code key, and the stale-pill failure mode plus every exit that must clear it.

by voitta-ai

Keep cmux's Claude NODE_OPTIONS restore guard alive when it lives under $TMPDIR. Use when: (1) a long-lived cmux/Claude session suddenly has every node/claude invocation fail with a missing --require target, (2) NODE_OPTIONS points at a /var/folders/.../T/cmux-claude-node-options/restore-node-options.cjs path on a cmux build that predates manaflow-ai/cmux#3699, (3) reboot-surviving cmux sessions break after about three days idle. Root cause: macOS reaps $TMPDIR files untouched for about three days, so the required guard file vanishes while NODE_OPTIONS still references it. Covers the persistent-HOME-copy stopgap, the LaunchAgent that recreates and mtime-refreshes the $TMPDIR file daily, why a LaunchAgent resolves the same per-user temp base, and when to retire it.

by voitta-ai

Run a runbook step by step in a sibling cmux terminal tab with exit-code receipts; stage irreversible steps for a human Enter (scripts/cmux-step + YOLT rule).

by voitta-ai

Search all open cmux workspaces/tabs/panes - live scrollback (tree --all + read-screen) plus agent transcripts.

by voitta-ai

Diagnose what cmux actually restored after a relaunch and recover panes it silently dropped.

by voitta-ai

Determine which cmux workspace, tab and surface a session is actually running in, and map other live sessions to theirs. Never answer from CMUX_* env vars: the tab and workspace variables can carry the same id (so tab identity is not in the environment at all), and on a resumed session they are a stale launch-time snapshot that can name a workspace the session has left. Uses the bundled CLI -- identify for the caller's own refs (read caller, not focused), tree --all --id-format both to resolve refs to names, top --all --processes to map pids to surfaces. Matters because a session that cannot name its own tab cannot tell a human which tab to address. Also covers the case the first version missed: the CLI only answers processes descended from the app, so a spawned subagent gets a broken-pipe/access-denied refusal -- the normal case, not a fault -- and must fall back to the app's live state file, using the environment ids as lookup keys rather than as answers.

by voitta-ai

Enumerate, before installing, which tmux verbs a tmux-driving tool needs that cmux's tmux-compat shim lacks: grep the tool's source for subcommands and #{format} variables, diff against cmux's docs/cli-contract.md, probe read-only verbs live via cmux __tmux-compat. Explains 'Unsupported tmux compatibility command' and 'Workspace target not found' (sessions map to workspaces). Ships scripts/tmux-compat-probe.

by chock-core

trigger: secrets, eval/exec, unsanitized SQL, hallucinated dependencies. avoid: committing credentials, adding unverified packages, executing dynamic code. Install scan-secrets (commit, agent write) for secrets and verify-dependency-exists (opt-in; needs an allowlist) for dependencies. Advisory: eval/exec and SQL guidance; a gate decides only the non-literal slice (agentic-code-security code pack: Python eval/exec, SQL built from strings in Python and JS).

by chock-core

trigger: secrets, eval/exec, unsanitized SQL, hallucinated dependencies. avoid: committing credentials, adding unverified packages, executing dynamic code. Install scan-secrets (commit, agent write) for secrets and verify-dependency-exists (opt-in; needs an allowlist) for dependencies. Advisory: eval/exec and SQL guidance; a gate decides only the non-literal slice (agentic-code-security code pack: Python eval/exec, SQL built from strings in Python and JS).

by Harry Yep

First party local Codex CLI delegation for tasks, reviews, resumable threads, and durable background jobs.

by voitta-ai

Run an adversarial Codex review against a GitHub PR and post its findings as a single batched PR review (inline comments + summary body), instead of letting them die in stdout. Use when: (1) you want `/codex:adversarial-review` output to land ON a pull request where the author can act on it, (2) you are the reviewer role in an agent-team loop (skillz#87) and need the review to close the loop with the developer agent, (3) you want a deterministic, scriptable reviewer rather than an LLM hand-posting comments, (4) you need to sweep a whole PR backlog and post one review per PR. Encodes the codex-companion `--json` call, the finding->diff-line mapping, the GitHub gotchas: inline comments are rejected (422) on lines not in the PR diff (out-of-diff findings are rolled up into the body), self-review forbids APPROVE/REQUEST_CHANGES (default COMMENT), low-confidence findings are demoted to a collapsed section, and a `--dry-run` payload is byte-for-byte the POST body so it can be saved, edited, and posted later without a second Codex pass — plus the recurring false-positive shapes (type-strictness on coerced config, "missing attribution" demanding a refactor, stale-branch mass-deletion artifacts) and degenerate-output shapes (plan-only "zero findings", quiet background-launch failure) to judge before posting.

by voitta-ai

Read the Codex CLI hook contract out of the shipped native binary with strings rather than a live probe turn:

by Anh-Thi Dinh

Lean toolkit for disciplined engineering workflows with Claude Code