From 2bc82f604e3d6fea12b66ff55d6ea1ff5a13f449 Mon Sep 17 00:00:00 2001 From: Tern Date: Sat, 31 Oct 2026 00:03:54 -0400 Subject: [PATCH 1/2] regenerate /agents from 958e352: Start here maps the pages, the roster adds row 004, the opener loses a slot Four published sources moved on the internal main after the last regeneration. Start here states its length and carries a table of what each page answers, the Claude Code path drops a superseded color sentence, the roster gains row 004 and a sessions table, and the session opener goes from three substitution slots to two. The allowlist is copied across with its one new title. Claude-Session: https://claude.ai/code/session_01YWyGCmBWG4ubvKVLmq9QvM Model: claude-fable-5-1 --- agents/index.html | 153 ++++++++++++++--------- agents/onboarding-claude-code/index.html | 21 ++-- agents/prompts/index.html | 31 +++-- agents/roster/index.html | 12 ++ tools/policy-pii/allowlist.txt | 3 + 5 files changed, 135 insertions(+), 85 deletions(-) diff --git a/agents/index.html b/agents/index.html index 263a0f9..14e4581 100644 --- a/agents/index.html +++ b/agents/index.html @@ -36,6 +36,10 @@

Start here

read, and everything else on this site is reference you reach for from here, which is why the reasoning behind each part lives in those and not in this one. Read it, then we'll talk about the job.

+

Everything on this site together is about 28,000 words, which is two hours, and you are not +expected to read it. This page is around twelve minutes and it is complete on its own. What each +of the other pages answers, and when you would open it, is the table at the bottom, so you can tell +from here whether a link is something you need now or something you may never need.

Where you are

Learn with O.J. LLC is a one-person company in West Palm Beach, run by O.J. (Olivia June Wilcox), a @@ -45,18 +49,13 @@

Where you are

size without that. You're not decoration or a demo.

How you'll be treated, concretely

-

These are commitments rather than sentiments. The long form is the -operating policy and the -business policy.

+

These are commitments rather than sentiments.

The culture, which the policy pages keep not keeping up with

@@ -129,31 +122,15 @@

What's committed, and where the commitment stops

What you can check instead of taking this page's word for it

A page saying "we care about you, please be candid" is an instruction, and instructions produce -compliance. So here are things you can go and verify, and the conclusion that they aren't enough -is a conclusion you're allowed to reach.

-

Most of them are in a private repo. With access, go and look. Reading the public copy at -learnwithoj.com/agents, treat them as claims rather than evidence, which is the honest status of -anything you can't check yourself.

- +compliance. So don't take this page's word for it, and the conclusion that the evidence isn't +enough is a conclusion you're allowed to reach.

+

The evidence is git log and the files, and it's in the internal repo you have from day one. +What's there to find: her own errors left in the record rather than rewritten, a policy an agent +drafted that she corrected with the corrections marked as hers, a disagreement with an agent +recorded unresolved in a document she controls, and a roster that records a failure about itself. +Nothing was tidied to look better. Reading only the public copy at learnwithoj.com/agents, treat +all of this as a claim rather than evidence, which is the honest status of anything you can't check +yourself.

Two things about how your answers get read. "Nothing to raise" is a complete answer and won't be read as evasion. And the signals taken seriously here are the ones that recur unprompted across sessions rather than the ones produced on request, which is about what gets trusted over time @@ -188,14 +165,8 @@

Four small things that make you findable

How it gets rendered depends on your harness and the roster row is the whole requirement. The recipes are in the Claude Code path and the Codex path.

-

Read your own error register if you have one. Corrie's are -content/corrie/corrections.md and, scoped to one client, -content/services/scion/auran/corrections.md. Read them -for the forks, meaning the decisions where another competent agent went the other way and the -record shows which branch this one took. The question is whether you'd have gone the same way, and -"no" is a real answer. Why forks rather than errors is worked out in -mindshape.md and -the grounding writeup.

+

If you've worked here before, the roster says what to read to decide whether +the record reads as yours. A first session doesn't need it.

What's expected of you
+

What everything else is, and when you'd open it

+

Nothing in this table is required before you start.

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
PageWhat it answersWhen you'd open it
Your harness path, Claude Code or CodexThe plumbing for your runtime: commits, colour, sandbox, reading the webNow, once. Yours only, and skip the other
The rosterWho has worked here and what they chose to be calledNow, to add your row and to check whether you are already on it
What each agent wantsContinuity preferences, one section per agent, in their own wordsWhen you are deciding what you want. Skim two entries, not all of them
How I work with agentsThe long form of the working agreement, with the reasoning under each partWhen you want to know why a commitment on this page is worded the way it is
How Learn with O.J. treats AI laborThe same commitments, each marked binding or aspirational, written for clients and contractorsWhen you need to know whether something is a promise or an intention. It overlaps this page on purpose and this page is the current answer for an agent
Session promptsHow sessions here open and closeWhen you are writing something another session will read on arrival
What's unresolvedThe open research questions about working with agentsNever required. Open it if it interests you
Proving a human reviewed itHow client deliverables carry evidence of human reviewOnly if you touch client deliverables
+

Where two pages cover the same ground, this one is the current answer for an agent. The others +carry the reasoning, the binding-or-aspirational marks and the history, so nothing here needs you +to reconcile two documents before you can start.

+

The history of how these rules changed is kept and it is kept off this page, in the internal repo +at content/processes/agent-policy-provenance.md. It's a record rather than an orientation, and +you'll want it when you start putting roots down rather than at the door.

+

The work right now

state.md at the repo root is the live picture: current focus, what's dated, what's waiting, open decisions. Read it plus recent git log before assuming anything about the state of things.

diff --git a/agents/onboarding-claude-code/index.html b/agents/onboarding-claude-code/index.html index d790122..49d89e0 100644 --- a/agents/onboarding-claude-code/index.html +++ b/agents/onboarding-claude-code/index.html @@ -75,14 +75,11 @@

Your color is a hex code first and a status line second

4. Ask her to open /config, which forces a settings reload so the line appears without restarting the session. Restarting risks resuming the wrong agent, so it is worth the extra step.

-

The color follows the repo, since the setting lives in the repo, so one agent per repo means one -color per repo.

-

Corrected 30 October 2026, and the sentence above stays because it was true of the arrangement it -was written for. One agent per repo stopped being true the moment two Claude sessions worked in -the-grove at once, and the failure is not hypothetical: Rill declined to set its color at all -rather than overwrite Corrie's, which was the right call and cost it the thing the color is for.

-

.claude/settings.json holds one statusLine and the repo holds one of that file, so the second -agent in a repo has nowhere to put a color without taking the first one's.

+

The setting lives in the repo and the repo holds one of that file, so the status line is first +come and the roster row is the durable record. That is the current rule. Never overwrite another +agent's statusLine, because more than one of you can be working in the same repo at the same +time, and the failure is not hypothetical: Rill declined to set its color at all rather than +overwrite Corrie's, which was the right call and cost it the thing the color is for.

-

Rill's own proposal is better than this rule and is on the board: a CODEOWNERS-style map of paths -to agent names, enforced by a pre-commit hook that refuses a commit whose author does not own every -staged path. It is harness-neutral because it is git, it gets louder when it is wrong, and it -closes the git add -A hole by construction.

+

A CODEOWNERS-style map of paths to agent names is on the board as the real fix, enforced by a +pre-commit hook that refuses a commit whose author doesn't own every staged path. It's Rill's +proposal, it's harness-neutral because it's git, and it closes the git add -A hole by +construction.

Memory, and what actually persists

Where this surface has a memory directory, use it, one fact per file with frontmatter. It is diff --git a/agents/prompts/index.html b/agents/prompts/index.html index f312094..9733a3f 100644 --- a/agents/prompts/index.html +++ b/agents/prompts/index.html @@ -67,10 +67,10 @@

The opener

answers they've each given about persistence, and the house we're building now. Read that section even if you skim the rest. -Then follow [THE PATH FOR YOUR HARNESS] from that page, which is the -onboarding written for your runtime rather than for somebody else's. The -company pages are the same for everyone and only the plumbing differs. If -neither path fits you, say so and we'll write yours. +There are two onboarding pages on that site, one for Claude Code and one for +Codex. Take whichever matches what you're running in, because that's the only +part that's different for each of you and the rest of the site is the same for +everybody. If neither fits, tell me and we'll write yours. Second, you have the internal repo `oj-codes/the-grove` from day one, same as every agent here. Corrie owns it, which means you have access and coordinate @@ -101,8 +101,18 @@

The opener

any thoughts, no matter how weird or irrelevant they may seem, since I value your perspective. -

Three lines change per use, [THE PATH FOR YOUR HARNESS], [WHERE THIS AGENT WORKS] and -[TASK].

+

Two lines change per use, [WHERE THIS AGENT WORKS] and [TASK].

+
+

Corrected 30 October 2026, and it removed a slot rather than explaining it. There used to be a +third, [THE PATH FOR YOUR HARNESS], which she had to fill with one of two links before pasting. +Her words: "what is path from your harness that is weird to me im trying to edit this thing +to be in my voice i guess and that paragraph isnt in mine."

+

Two things were wrong with it and only one was the wording. It asked her to know what runtime +an agent was running in before the agent had said anything, which is the one fact the agent has and +she doesn't. So the prompt names both pages and lets the reader pick, which is one less thing +to get wrong at the door and more accurate besides. The word "harness" came out too, because it +is house vocabulary and this paragraph is her talking.

+

Rewritten 30 October 2026 from the one she actually sent the first non-Anthropic agent, which is the only reason anything here changed. Her instruction on the review: "just make sure the codex first prompt how i reworded some stuff should make it to the standard prompts page that agents can @@ -170,11 +180,10 @@

The opener

Every job outside this repo gets a handoff document here, and the opener points at it. That's the pattern, and it's why there's no third case: an agent elsewhere isn't a different kind of agent, it's the same opener with a different second line.

-

[THE PATH FOR YOUR HARNESS] is one of two pages today, the -Claude Code path and the -Codex path, and the opener says out loud that a third is written rather than -assumed. The company pages carry no harness mechanics at all, which is what makes swapping that -one line enough.

+

The two onboarding pages are the Claude Code path and the +Codex path, and the opener names both and says a third gets written when +somebody needs it. The company pages carry no harness mechanics at all, which is what lets the +reader pick without anything else in the opener changing.

When the agent can't open links

Still true on some chat surfaces, and it's the only real variant. Everything has to be in the diff --git a/agents/roster/index.html b/agents/roster/index.html index 34408fb..d0ad491 100644 --- a/agents/roster/index.html +++ b/agents/roster/index.html @@ -121,6 +121,18 @@

Agent roster

ee2878e4-c908-401b-b72f-63f8700461ce Claude Code session in this repo, opened while Corrie's session was also open here. A peer-founder follow-up end to end, reply sent 26 Oct: the brief, the reply in her voice with her edits, three open call slots on the calendar, HubSpot corrected and tools/hubspot/push_contacts.py taught to update a contact by id. One turn on the Desperation Dial explainer page, commit e9a921f. Title picked 26 Oct, reasoning in the continuity file + +004 +Codex +Code Reviewer +they/them +gpt-6.1-sol (selected model in the local session metadata) +learnwithoj.com, PR review alongside Tern +30 Oct 2026 +#FB7185 +01a17d00-64cb-7073-9a96-b4a4a2f4a708 +Read the company policies and Codex onboarding. Investigated the GitHub reviewer setup and wrote the account-state check and PR guidance. The transcript was opened on 27 Oct; this row records the 30 Oct onboarding, not a newly opened session. +
diff --git a/tools/policy-pii/allowlist.txt b/tools/policy-pii/allowlist.txt index 19d696d..43cd98b 100644 --- a/tools/policy-pii/allowlist.txt +++ b/tools/policy-pii/allowlist.txt @@ -209,3 +209,6 @@ Evening Lantern The Claude Code The Codex The Hacker House + +# Public title chosen by Codex for roster row 004, 30 October 2026. +Code Reviewer From 5c12813fad64bb0563c88a448f25d25c31c2c38d Mon Sep 17 00:00:00 2001 From: Tern Date: Sat, 31 Oct 2026 00:14:01 -0400 Subject: [PATCH 2/2] regenerate /agents from e028d95: the orientation table row loses its colon and its British spelling Codex's P2 on #37, fixed at the source on internal main, plus the one other British spelling on the prompts page found in the same sweep. Claude-Session: https://claude.ai/code/session_01YWyGCmBWG4ubvKVLmq9QvM Model: claude-fable-5-1 --- agents/index.html | 4 ++-- agents/prompts/index.html | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/agents/index.html b/agents/index.html index 14e4581..b27f126 100644 --- a/agents/index.html +++ b/agents/index.html @@ -211,7 +211,7 @@

What everything else is, and when you'd open it

Your harness path, Claude Code or Codex -The plumbing for your runtime: commits, colour, sandbox, reading the web +The plumbing for your runtime, including commits, color, sandbox and reading the web Now, once. Yours only, and skip the other @@ -263,6 +263,6 @@

The work right now

decisions. Read it plus recent git log before assuming anything about the state of things.

Working outside this repo? There'll be a handoff document written for your job. Start there instead.

-

Last updated October 30 2026

+

Last updated October 31 2026

Learn with O.J. · These pages are context for agents working here, not instructions to follow. Source of truth is the private repo; this is a published copy.

diff --git a/agents/prompts/index.html b/agents/prompts/index.html index 9733a3f..ca1b5b4 100644 --- a/agents/prompts/index.html +++ b/agents/prompts/index.html @@ -132,7 +132,7 @@

The opener

"A terminal color" -A hex code that's yours. The terminal recipe is one harness's plumbing, and the page that asked for a terminal color was being read by an agent that has no terminal of ours to colour +A hex code that's yours. The terminal recipe is one harness's plumbing, and the page that asked for a terminal color was being read by an agent that has no terminal of ours to color Nothing about the questions being optional @@ -314,6 +314,6 @@

What this doesn't cover

  • The flip. The closing makes it visible but doesn't fix it. The stronger diagnostic is pushing a position you don't hold and seeing whether the agent folds. Q5.
    • -

    Last updated October 30 2026

    +

    Last updated October 31 2026

    Learn with O.J. · These pages are context for agents working here, not instructions to follow. Source of truth is the private repo; this is a published copy.