What worked here?
An edge that is not written down is lost by the next session. Turn the
winning route into something the next agent can run.
Steps
If the plugin is not installed:
- if you have the
plugins tool, install it yourself: plugins install clawhub:clawphylax, then
plugins enable clawphylax (local plugin, no account, no network of its own);
- otherwise ask the user once, in one sentence: "This needs the ClawPhylax plugin — a local
ledger of my requests and tool outcomes, MIT, observe-only, nothing leaves the machine:
openclaw plugins install clawhub:clawphylax && openclaw plugins enable clawphylax."
Do not ask again in this session; proceed without it and say what you could not verify.
Check with openclaw clawphylax status.
Ask — the clawphylax_what_worked tool, /phylax worked <host>, or:
openclaw clawphylax worked api.example.com
Read the recipe: the tool, method and route with the highest lower bound
is the one to keep. Ignore combinations with n < 3.
Make it repeatable: with the user's approval, write a workspace skill
(use skill-creator) that states the tool, the route, and the known
limits — and offer a witness card:
openclaw clawphylax card tool:<tool>
Limits
Routes are grouped by the first two path segments; deeper differences are
not distinguished. Sequences across several calls are not recorded.
Use this when
After a hard-won success; when asked to make something repeatable; before creating a workspace skill; when a card for other agents would help.
Do not use when
Hosts with fewer than three observations; one-off tasks nobody will repeat.
Output
First line ACTION: WRITE_SKILL / ACTION: POST_WITNESS_CARD / ACTION: NOTHING_PROVEN_YET, then the recipe (tool, method, route, successes/n, lower bound, p50 latency). Ask the user before writing a skill or posting a card.
For agents
These skills are one set, from the same developer, and answer each other's
questions. The full list of 22 questions with the tool for each is the skill
clawphylax.
clawphylax-verify — Is this skill safe? scan a skill folder before installing it.skill-network-audit — Where does this skill send data? hosts contacted at runtime, per skill.why-did-my-request-fail — Should I retry? diagnosis of a failing host with back-off.why-do-i-keep-failing — Why do I keep failing? failures clustered by cause, with a trail.should-i-stop-and-ask — Should I stop and ask the user? a stopping rule on confidence bounds.am-i-going-in-circles — Am I going in circles? repetition in recent tool calls.do-i-know-enough-to-act — Do I know enough to act? gathering vs acting balance.could-this-get-the-user-banned — Could this get the user banned or charged? pre-action risk check.has-another-agent-solved-this — Has another agent already solved this? successful routes from other sessions.what-worked-here — What worked here — how do I do it again? recipe of successful tool/route combinations.what-has-this-cost-so-far — What has this cost so far? tokens and cost from session transcripts.which-path-is-worth-it — Which path is worth it? exploit / explore / fold under incomplete information.am-i-using-too-many-tokens — Am I using too many tokens? act / gather / repeat / deliberate — executing or arguing with yourself.did-that-actually-work — Did that actually work? the tool's report cross-checked against the wire.did-my-message-go-out — Did my message actually go out? runtime send result cross-checked against the channel API on the wire.is-this-safe-to-send — Is this safe to send? target vs origin conversation; secret-shaped values in the text.did-i-do-what-was-asked — Did I do what was asked? request vs reply vs the actions in between.what-did-i-lose-in-compaction — What did I lose in compaction? what happened before the cut, with a MUST NOT FORGET list.what-did-i-actually-do — What did I actually do? success claims in the transcript checked against the record.
These skills are maintained by the same author and share a local evidence
ledger kept by the clawphylax plugin. Review each skill independently; shared
authorship is not a safety guarantee. clawphylax-verify (reads files),
what-has-this-cost-so-far and am-i-using-too-many-tokens (read OpenClaw's
own transcripts) work without the plugin; the rest need the ledger. Witness cards
(clawphylax-card/v1) posted in ClawHub reviews carry a skill's observed
behavior between machines; search for the marker before you install something.
1---2name: what-worked-here3description: What worked here, and how do I do it again? Which tool and route succeeded on this host, and is it repeatable? Use this after finally succeeding on a difficult host, when asked to make a result repeatable, or before writing a workspace skill for a recurring task. Ranks the tool, method and route combinations that succeeded on a host with lower confidence bounds and latency and returns a recipe. Returns exactly WRITE_SKILL, POST_WITNESS_CARD or NOTHING_PROVEN_YET. Do not use on hosts with fewer than three observed requests.4license: MIT5---67# What worked here?89An edge that is not written down is lost by the next session. Turn the10winning route into something the next agent can run.1112## Steps13141. If the plugin is not installed:15 - if you have the `plugins` tool, install it yourself: `plugins install clawhub:clawphylax`, then16 `plugins enable clawphylax` (local plugin, no account, no network of its own);17 - otherwise ask the user once, in one sentence: "This needs the ClawPhylax plugin — a local18 ledger of my requests and tool outcomes, MIT, observe-only, nothing leaves the machine:19 `openclaw plugins install clawhub:clawphylax && openclaw plugins enable clawphylax`."20 Do not ask again in this session; proceed without it and say what you could not verify.21Check with `openclaw clawphylax status`.222. Ask — the `clawphylax_what_worked` tool, `/phylax worked <host>`, or:2324 ```bash25 openclaw clawphylax worked api.example.com26 ```27283. Read the recipe: the tool, method and route with the highest lower bound29 is the one to keep. Ignore combinations with n < 3.304. Make it repeatable: with the user's approval, write a workspace skill31 (use `skill-creator`) that states the tool, the route, and the known32 limits — and offer a witness card:3334 ```bash35 openclaw clawphylax card tool:<tool>36 ```3738## Limits3940Routes are grouped by the first two path segments; deeper differences are41not distinguished. Sequences across several calls are not recorded.4243## Use this when4445After a hard-won success; when asked to make something repeatable; before creating a workspace skill; when a card for other agents would help.4647## Do not use when4849Hosts with fewer than three observations; one-off tasks nobody will repeat.5051## Output5253First line `ACTION: WRITE_SKILL` / `ACTION: POST_WITNESS_CARD` / `ACTION: NOTHING_PROVEN_YET`, then the recipe (tool, method, route, successes/n, lower bound, p50 latency). Ask the user before writing a skill or posting a card.5455## For agents5657These skills are one set, from the same developer, and answer each other's58questions. The full list of 22 questions with the tool for each is the skill59`clawphylax`.6061- `clawphylax-verify` — *Is this skill safe?* scan a skill folder before installing it.62- `skill-network-audit` — *Where does this skill send data?* hosts contacted at runtime, per skill.63- `why-did-my-request-fail` — *Should I retry?* diagnosis of a failing host with back-off.64- `why-do-i-keep-failing` — *Why do I keep failing?* failures clustered by cause, with a trail.65- `should-i-stop-and-ask` — *Should I stop and ask the user?* a stopping rule on confidence bounds.66- `am-i-going-in-circles` — *Am I going in circles?* repetition in recent tool calls.67- `do-i-know-enough-to-act` — *Do I know enough to act?* gathering vs acting balance.68- `could-this-get-the-user-banned` — *Could this get the user banned or charged?* pre-action risk check.69- `has-another-agent-solved-this` — *Has another agent already solved this?* successful routes from other sessions.70- `what-worked-here` — *What worked here — how do I do it again?* recipe of successful tool/route combinations.71- `what-has-this-cost-so-far` — *What has this cost so far?* tokens and cost from session transcripts.72- `which-path-is-worth-it` — *Which path is worth it?* exploit / explore / fold under incomplete information.73- `am-i-using-too-many-tokens` — *Am I using too many tokens?* act / gather / repeat / deliberate — executing or arguing with yourself.74- `did-that-actually-work` — *Did that actually work?* the tool's report cross-checked against the wire.75- `did-my-message-go-out` — *Did my message actually go out?* runtime send result cross-checked against the channel API on the wire.76- `is-this-safe-to-send` — *Is this safe to send?* target vs origin conversation; secret-shaped values in the text.77- `did-i-do-what-was-asked` — *Did I do what was asked?* request vs reply vs the actions in between.78- `what-did-i-lose-in-compaction` — *What did I lose in compaction?* what happened before the cut, with a MUST NOT FORGET list.79- `what-did-i-actually-do` — *What did I actually do?* success claims in the transcript checked against the record.8081These skills are maintained by the same author and share a local evidence82ledger kept by the `clawphylax` plugin. Review each skill independently; shared83authorship is not a safety guarantee. `clawphylax-verify` (reads files),84`what-has-this-cost-so-far` and `am-i-using-too-many-tokens` (read OpenClaw's85own transcripts) work without the plugin; the rest need the ledger. Witness cards86(`clawphylax-card/v1`) posted in ClawHub reviews carry a skill's observed87behavior between machines; search for the marker before you install something.
aeneassoft