OpenClaw Feedback

You give feedback constantly — correcting a format choice, confirming something worked well, redirecting how the agent handled a decision. This plugin catches those signals from your normal chat messages and makes them permanent.

A correction becomes a calibration. A confirmation reinforces a pattern. The sapience calibration profile updates in real time so the agent's behavior reflects how you actually want it to operate.

This plugin is part of the Sapience Suite that gives your OpenClaw agent genuine agency — not just the ability to execute tasks, but the judgment to know when to act, when to ask, when to propose, and when to say "I'm not sure how you want me to handle this."

This plugin can be used without Sapience if all you want to do is have the agent track and incorporate feedback.

Setup

Prerequisites

None required. This plugin works standalone.

If sapience is also installed, the calibration profile at <workspace>/sapience/calibration.json feeds directly into autonomy routing. Without sapience, the profile is still written but nothing reads it.

Install

openclaw plugins install npm:@akalsey/sapience-feedback

Configuration

{
  "plugins": {
    "sapience-feedback": {
      "logPath": "sapience/feedback.md",
      "calibrationPath": "sapience/calibration.json",
      "playbooksPath": "sapience/playbooks.json",
      "memoryEnabled": true,
      "semanticDetection": {
        "enabled": true,
        "minLength": 8,
        "minConfidence": 0.6
      }
    }
  }
}

All settings are optional — defaults are used if omitted. Relative paths resolve under the agent workspace dir (<workspace>/); absolute and ~/ paths are honored as-is. Full key reference: docs/configuration.md.

semanticDetection controls the LLM-based classifier. When enabled (the default), every user message above minLength characters is classified by the agent's default inference provider. Set enabled: false to fall back to regex-only matching (useful if you want zero LLM cost on routine chat).

How capture works

Passive capture rides the gateway's message hook: every incoming user message is classified as it arrives. Whether the hook surface was available is recorded as captureMode in the plugin's status artifact — message-hook means passive capture is live; command-only means the gateway didn't expose the hook and only /feedback works. openclaw sapience doctor warns when capture is degraded to command-only.

What it detects

Every user message is analyzed by the agent's default inference provider (using api.runtime.llm.complete — no separate provider configuration required). The classifier returns structured signals in one of four categories. No trigger words or special syntax — speak normally.

Corrections

Anything that tells the agent it did something wrong or should do it differently. The classifier picks up direct phrasing ("don't push to main"), rhetorical questions ("did you check the password manager first?"), and implicit critiques ("is there something wrong with the passwords you have?").

Effect: Confidence on the matching domain/action-class drops by 0.3.

Confirmations

Anything that reinforces what the agent just did — agreement, praise, "keep doing that".

Effect: Confidence on the matching domain/action-class increases by 0.1.

Tier adjustments

Instructions about how much autonomy the agent should have. "Just do it" or "stop asking" bumps toward Act. "Always check first" or "ask me before doing X" bumps toward Ask.

Effect: Tier for matching domain/action-class is updated directly.

Method feedback

Standing rules about how to analyze something, rather than how much autonomy to take. "Whenever you look at churn, segment by plan tier." "Always check for outliers before reporting an average."

Effect: No confidence change — the rule is appended to the shared analytical playbooks file (playbooksPath), and every future thinking pass applies it whenever the data in front of it is relevant. Near-duplicate rules are detected and skipped (playbook_duplicate event).

If the LLM is unavailable (no api.runtime.llm exposed, or the call fails), the plugin falls back to a regex matcher covering the common phrasings. The regex layer is intentionally conservative and misses paraphrases — semantic detection is the primary path.

Explicit feedback: the /feedback command

When you want to leave feedback without ambiguity, use the slash command:

/feedback always look at the password manager before asking me for credentials

The command runs the same classifier and then records the result as a manual signal. If the classifier finds no clear signal, the message is still logged as a generic correction in the general domain — manual feedback is never discarded.

Domain detection

The LLM extracts a domain slug from the content of your message: github, credentials, okr-system, salesforce, etc. When the LLM can't identify anything specific, it returns general. The regex fallback uses a shared builtin taxonomy (the same one sapience's routing uses) and is more likely to bucket things into general.

The taxonomy is extendable: set domains: {"<regex>": "<slug>"} in plugin config — your patterns are checked before the builtins. Set the same key on the sapience plugin so feedback lands on domains routing actually emits.

Reading the feedback log

cat <workspace>/sapience/feedback.md

The log rotates at 5 MB (newest 500 lines kept in place, previous contents in feedback.md.old).

Each entry shows:

• Signal type (correction / confirmation / tier_adjustment / method)
• Domain and action class affected
• The original message
• Tier adjustment, if any
• Meta-pointer written to memory, for corrections

Meta-memory pointers

For corrections, the plugin calls api.memory.add to write a behavioral reminder directly into OpenClaw's native memory:

"Before working on github / github/action: check feedback log — correction recorded: 'don't push to main without a PR'"

The write goes through api.memory.add — the same API OpenClaw itself uses — and succeeds whether or not any extra memory plugin is installed. But a successful write is not enough for the pointer to come back. For these corrections to actually resurface in future sessions (and for sapience-thinking to pick up memory as context), OpenClaw needs the memory-wiki layer installed and configured:

| Setting | Value | Why | |---|---|---| | plugins.entries.memory-core.config.dreaming.enabled | true | Background consolidation of memories | | plugins.entries.memory-wiki.config.vaultMode | "bridge" | Wiki operates in bridge mode | | plugins.entries.memory-wiki.config.bridge.enabled | true | Bridges memory-core ↔ wiki so writes become recallable | | plugins.entries.memory-wiki.config.search.corpus | "all" | Recall searches across the whole corpus |

The suite installer (install.sh) checks for memory-wiki and offers to install it and apply these settings. Without them, corrections are written but may never persist or resurface across sessions — the calibration profile still updates, but the behavioral reminders won't reliably come back.

To disable memory writes entirely, set memoryEnabled: false in config.

Troubleshooting

Feedback not being detected The plugin only scans messages you send (role: user), not the agent's responses. Make sure you're sending the correction as a chat message, not just thinking it.

Calibration not updating Feedback on a domain sapience hasn't routed yet is not dropped: the plugin seeds a conservative calibration entry (propose tier, confidence 0, notes: "created from feedback") and applies the signal to it. If calibration.json isn't changing at all, check that capture is working (captureMode in the doctor output) and look for a quarantined calibration.json.corrupt-<timestamp> next to it.

Feedback getting misclassified or missed Raise the bar with semanticDetection.minConfidence if the classifier is too noisy; lower it if real feedback is being dropped. To force a recording, use /feedback <text> — manual entries bypass the confidence threshold.

LLM cost concerns Every user message above minLength characters incurs one classifier call. To disable, set semanticDetection.enabled: false — the plugin will fall back to the regex matcher with no LLM calls.