Pre-release. Not a stable release and not evidence of effect in real use.

What is new

• otw forget --all --yes deletes every stored correction, the per-session state and the diagnostic log (config.json stays). otw forget --project [DIR] --yes deletes only one working directory's corrections and session state. Without --yes it only shows counts. It never follows links or touches anything outside the Off the Wall folder, and a capture authorized before the deletion cannot append afterwards (tests). Provider-side copies and Claude Code/Codex histories are not affected.
• Documentation brought in line with what was actually checked: one Codex smoke run and one headless Claude Code run; interactive Claude Code use is still unverified. Experiment records 5b and 7 added.

No change to recall behavior. An opt-in lexical gate against recall noise was tried and removed again (it did not generalize to a larger store); see docs/data/exp7/plan-and-log.md.

Plain-language overview: https://toranalytics.github.io/en/otw-what-it-does/
Full changes: CHANGELOG.md. Install: /plugin install off-the-wall --marketplace toranalytics/off-the-wall

Behaviour change

Wide recall is now on by default: for prompts of 20+ trimmed characters, up to 6 eligible same-project quotes are shown ranked by similarity (zero-overlap records excluded). Prompts under 20 characters keep the 0.18 floor. Turn it off with "recallWide": false in ~/.otw/config.json. Cost: up to 6 quotes (about 300 tokens) per prompt.

Other changes since 0.2.0

• Fail-closed config, revocable capture, hook visibility and Codex trust count in otw status (0.2.1), no routing re-injection on session resume (0.2.2).
• Windows-only test flake fixed (test change only).

Evidence, with its limits

Synthetic only (Claude Code with haiku, headless, store and tasks written by other models, plans fixed before the runs). 11 personal-preference tasks, 3 runs each: followed the preference 1/33 with no recall, 9/33 with the former narrow default, 17/33 with wide recall (replicated on a second store: 16/33 vs 8/33). A harm check on 14 tasks with no applicable preference: 40/42 without recall, 41/42 with wide recall. Not tested: real histories, other models, interactive use, the short-prompt guard. Records, scripts and the weaknesses found by independent review: docs/.

Marked pre-release on purpose: no real-user evidence yet. No npm package is published; install from the marketplace: /plugin install off-the-wall --marketplace toranalytics/off-the-wall.

• Renamed to Off the Wall: plugin off-the-wall, CLI bin/otw.mjs, data in ~/.otw (OTW_HOME). 0.1.0 kept its
  data in ~/.ownwords; to keep your corrections, move it once: mv ~/.ownwords ~/.otw.
• Plugin manifests carry homepage, repository, license and keywords; the Codex manifest fills its publishable fields.
• otw report ends with a link to a prefilled Show and tell discussion carrying your top three kinds and their
  counts (no quoted words), so results can be compared across users.

Stop telling your coding agent the same thing twice. First release, for Claude Code and Codex.

What it does

• Quotes your past corrections (up to 6: most similar to the request, then the latest), word for word, before each answer, and asks the agent to predict your objection and avoid it.
• otw report shows what you corrected most in the history already on your machine, with your own words. --seed starts recall from it.
• Subagent routing: the orchestrator names a model for every subagent (Claude: haiku / sonnet / opus; Codex: detected from your account's model list). An unnamed call is sent back once per turn.

Measured (one person's history; a first result)

• Quotes vs nothing attached: 25 wins, 11 ties, 3 losses on 39 replayed requests. One-line distilled rules: 14 / 17 / 8, no better than random rules.
• Routing on one delegation task: $0.171 and $0.173 against $0.300 and $0.439, same result.

Install

/plugin install off-the-wall --marketplace toranalytics/off-the-wall

codex plugin marketplace add toranalytics/off-the-wall && codex plugin add off-the-wall@off-the-wall

Then trust the hooks once with /hooks in Codex.

Try the report without your own history:

git clone https://github.com/toranalytics/off-the-wall && cd off-the-wall
node bin/Off the Wall.mjs report --host claude --claude-dir examples/claude-history --days 0