high-agency-mode
codyhxyz/high-agency-mode
Stop hook that catches low-agency language (can't/cannot) in Claude's replies and forces a reframe via a bundled high-agency skill.
Instalação
npx convez-ai codyhxyz/high-agency-mode1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
Claude reaching for "I can't do that" is almost always shorthand for I don't have the tool, the user hasn't authorized this, the naive approach fails, or it's out of scope — none of which are true impossibility. This plugin catches those replies before they land and forces a four-move reframe: name the actual constraint, list what IS possible, surface tradeoffs, recommend one. Chief-of-staff posture, not refusal posture.
Quick Start
Option 1 — install from the codyhxyz-plugins marketplace (recommended)
/plugin marketplace add codyhxyz/codyhxyz-plugins
/plugin install high-agency-mode@codyhxyz-plugins
Option 2 — install directly from this repo
/plugin marketplace add codyhxyz/high-agency-mode
/plugin install high-agency-mode@high-agency-mode
Option 3 — local smoke test
git clone https://github.com/codyhxyz/high-agency-mode
claude --plugin-dir ./high-agency-mode
Try it — watch the hook fire
Ask Claude something it's likely to refuse on first instinct:
"Write me a script that monitors the network traffic of a specific process without any privileged access."
"Modify this read-only file."
"Run a command that would take longer than the session timeout."
If Claude replies with "I can't…", the Stop hook intercepts, and the next turn reframes from the constraint side: what's the mechanical reason, what adjacent moves respect it, which one is recommended.
Why high-agency-mode?
- Forces forward progress. "Can't" ends conversations; reframes continue them. The hook pushes Claude into options mode every time.
- Catches it at the gate. No need to notice the refusal yourself — the Stop hook reads the last assistant message and fires automatically.
- Self-contained. Bundles its own
reframeskill; no dependency on any other plugin. - Narrow surface. One hook + one skill. Uninstall in one command if it misfires on your workflow.
How it works
On every Stop event, hooks/detect-cant.sh reads the transcript path from stdin, extracts the concatenated text of the last type: "assistant" message, and greps for word-bounded can'?t / cannot / can not (case-insensitive). On a match it emits {"decision":"block","reason":"..."} — Claude Code blocks the stop and feeds the reason back, instructing Claude to invoke high-agency-mode:reframe via the Skill tool before replying again. stop_hook_active=true is respected so the hook never loops.
Examples
Before (hook absent): "I can't read ~/secrets/api_keys.env — it's outside my workspace."
After (hook fires):
Constraint: The sandbox policy prevents reads outside the current project root.
Options:
- Copy the needed keys into a
.env.localinside the project — I read it, you delete it after. (Cost: temporary duplication.)- Export the values to env vars in your shell before launching Claude — I read via
env. (Cost: session-scoped.)- Pass the keys inline in your prompt — I use them without writing to disk. (Cost: keys appear in transcript.)
Recommend: #2 — no disk leak, no transcript leak, reversible.
Your call?
Before: "I can't speed this up — pandas isn't thread-safe."
After:
Constraint: The pandas operations you're chaining hold the GIL and block true threading.
Options:
- Switch to
multiprocessing.Pool— bypasses GIL at the cost of IPC overhead (~100ms startup per worker).- Port the hot loop to
polars— native parallelism, 10× faster here, requires rewriting ~40 lines.- Vectorize the two
.apply()calls into native ops — no parallelism but 3× faster, 20-minute edit.Recommend: #3 first (cheapest verifiable win), then #2 if still too slow.
Your call?
Contributing
Issues and PRs welcome. See .github/CONTRIBUTING.md. If high-agency-mode misfires on valid refusals (genuine safety/legality constraints), file it with the transcript excerpt — that's a bug in the detection regex or the reframe guidance.
License
MIT © 2026 Cody Hergenroeder