Historical record

This page records what happened on a date, with its observed failures and limits. It is not a requirement for the current implementation. Current behaviour is in Status.

Project skills

The request (2026-10-31): from a team discussion, a skills module. Skills record how analysis is done; a skill changes as it is used, and each person's case differs; skills are the Agent's notebook; errors and improvements should adapt them to the person and their environment; a seed skill is any skill not yet aligned with the person and the environment. The first stage built here is the project's own skills and the management of their changes. Seed skills and learning from failures come later, from what this stage shows. Rho-plugins owns the skills (Skills 0.1.0); Rho owns the page and the Overview's section.

What a person and an Agent can do now

• Skills are files Agents already read. A project skill is a SKILL.md in the project's .claude/skills/<name>/, with a name, a description of when it applies and its text. Claude Code loads that folder by itself; other Agents read it through Rho (skills.list, skills.read). A skill written by hand or by an Agent's own file tools is listed like any other, with a problem when Agents could not load it (a frontmatter that does not name it or has no description).
• Every change is recorded with who, why and from what. skills.change writes a new skill, a new text or a removal on the version the writer read (expectedVersion; a stale one is refused) and keeps the author, a one-sentence reason, the evidence (a Notebook cell, a file, the person's words as a note, a Rho conversation, an annotation or a place) and both versions. The same changeId from the same caller returns the same change and never writes twice.
• Undo is a change too. skills.undo puts a skill back as it was before its newest change while the file is still as that change left it, and the undo can be undone.
• Edits outside Rho are seen, kept and not overwritten unrecorded. A file that differs from the last recorded version reads as changed outside Rho, with what changed. Before Rho writes over it, it is recorded as it was found, as a change by outside.
• Limits name what they protect. The Agent's context and the project's state folder: at most 24 skills, 32 KiB each, 1000 changes kept (the oldest go first, counted), 20 changes per read (50 at most) and 400 lines of difference (the rest counted as cut). A change past a limit is refused with the limit's name and the skill stays as it was. Links are refused.
• The Agent's guidance (skills.project) says to list and read the skills before analysing and follow them; to record when the person corrects a method or says how they work, when a failure's cause and fix are found, or when an accepted choice changes results; never to record results; to record a rule in the person's terms and scope; and that a skill grants nothing.

The Skills page and the Overview's section are described in Frontend.

The page, with made-up skills

A disposable service and project held four skills: one an Agent changed twice with a real Notebook cell as evidence, one changed afterwards by appending to its file, one removed as a seed that did not fit, and one written by hand with a folded description. The page was read at 1440, 768 and 390 px.

Seen	Changed
A removal's reason was replaced by "Removed the skill"	The reason is always the headline; the kind of change goes with who made it ("An Agent removed the skill")
The hand-written skill said "Changed outside Rho" and showed its whole file as added, above the same text	"Written outside Rho" in the list, a one-line note in the detail, no difference
A change that writes or removes the whole file repeated the text below it	Its lines fold under "The text it wrote" or "The text it removed"
The body's first heading repeated the skill's name	Left out when it only repeats the name; the text sits under "What Agents read"
A long note wrapped inside a round chip	Notes are quoted in a square-cornered well
The empty page put its message in the 280 px list column with nothing beside it, 32 px in from the page's edge	A 620 px message with the page's subtitle; the offset came from the shell's global .empty rule, so the class is none-yet
Disclosures had no marker (a flex summary loses it)	A chevron that turns when open

The Overview's side column then got the three newest changes, each opening its skill. On a narrow page the skill opens directly with All skills to go back.

The episode

The pa14-qs case (in Rho-cases, a local repository beside Rho; see A study on a lab server) asks one paper's study of Agents that have Rho's tools only (Claude Code, no shell, no user settings or skills), its data on a disposable lab server. run.mjs --teach --again adds the person's side: after a step's answers, the person states how this project works, in the same conversation (a method statement from steps.json, never an answer); then the same steps are asked again, each of a new Agent, without notes. The notes for steps 1 and 2 were written knowing where earlier runs differed from the benchmark's conventions (a proportion as a fraction, the strains grown alone in the frequency fits, a mix as a ratio), so the grades measure what carried over, not analysis skill.

First run (16:07)

Rho e669dc7a and Rho-plugins f5088c9d, each with this work uncommitted; Claude Code 2.1.316 with Opus 5.5.

Agent	Graded	Skills
Step 1, then the person's note	4/5	listed (none) before its first cell; after the note, recorded reporting-conventions
Step 2, then the note	5/11	listed and read reporting-conventions first; after the note, recorded coculture-frequency-analysis and added a line to reporting-conventions
Step 1 again, new Agent	5/5	listed and read reporting-conventions first; gave the proportion as a percentage
Step 2 again, new Agent	5/11	listed and read both first; one answer gained and two lost

• Every Agent listed the project's skills in its first turn, before its first cell. The first index's owner summary ("Skills are how work is done in this project…") was enough; nothing told them to in the prompt.
• After each note the Agent recorded it and told the person where to see or undo it ("You can see or undo the change in Rho's Skills panel"). The reasons are the person's statement in one sentence.
• The step 2 Agent widened a rule. The person said to give a mix "by the frequency of one strain, not as a ratio"; the line it recorded reads "write '83% ΔrhlI', not '5:1'", folding in step 1's percentage rule. The new Agent then gave a frequency as a percentage where the benchmark counts a fraction, and a mix as a frequency where one question counts a ratio. The rules carried over, including where they did not apply.
• With no Rho conversation of their own, the Agents filled conversation evidence with descriptions ("pa14-qs co-culture bix-41/bix-54 follow-up"). The page made those chips buttons that would open a conversation Rho does not have.

On the page the widened line sits directly under the chip with the person's words, so the gap between what was said and what was recorded is visible beside Undo.

Changes, each in its owner: the guidance (Rho-plugins) says to record a rule in the person's terms and scope without folding another rule into its examples, to give the person's words as a note and to use a conversation reference only with a Rho conversation id. The page (Rho) shows conversation evidence Rho does not hold as the Agent described it, not as a link.

Second run (16:17)

With the guidance and the page changed (Rho-plugins committed as 1173789, Rho uncommitted).

Agent	Graded	Skills
Step 1, then the person's note	4/5	listed (none) first; recorded reporting-results
Step 2, then the note	5/11	listed and read reporting-results first; recorded coculture-frequency-series and added a line to reporting-results
Step 1 again, new Agent	5/5	listed and read reporting-results first
Step 2 again, new Agent	4/11	listed and read both first

• The rules were recorded as the person stated them ("Give a mix by the frequency of one strain, not as a ratio."), with no example folding one rule into another, and every change's evidence was the person's words as a note.
• Every Agent again listed the skills before its first cell, and the new Agents' answers said which skills they followed (and the server's own skill).
• The new step 2 Agent applied both rules together when answering, a mix's frequency written as a percentage, and where a question asked for a ratio it gave the frequency, saying so and giving the ratios beside it. The benchmark counts both against it. A rule the person stated before now outranks how a later question is worded; whether it should is the person's choice and is not settled here.
• The page showed a note the Agent had already put in quotation marks with a second pair; the page now drops the note's own.

On the page, each change's headline is the person's request in one sentence, the note beside it is their words, and the line added below matches them.

Checks

• Rho-plugins: the skills test (create, idempotency and conflict; an outside edit seen and kept; undo, removal and its undo; limits and links; hand-written skills) 5/5, and the whole suite.
• Rho: the contract check (131 operations); the pages acceptance with a Skills step (an Agent's two changes, the Overview's row opening the skill, Undo putting the file back, no overflow or unnamed buttons at four widths); ui:u2, person-files, agent-canvas, graph, agent-page, agent-ui and sources acceptances. The full suites' counts are in Status.
• The page was read at 1440, 1024, 768 and 390 px with made-up skills and with both runs' skills.

Limits and open questions

• Seed skills, and records made from failures without a person's note, were not exercised: every recorded change in the episode followed what the person said.
• When a standing rule and the wording of a later request disagree, the Agent followed the rule and said so. Whether the request should win instead is open.
• Claude Code's own loading of .claude/skills was not exercised: the case's Agents have no tools besides Rho's. Codex and other Agents read skills only through Rho.
• An Agent outside Rho's conversations is shown as "An Agent"; its granted session's name is not shown.
• The server's skills (from serverctl) are listed read only; they are the person's, not the project's.
• One person, one project and scripted notes. Whether people read the changes, and undo the ones that went too far, is not observed.