jared/cc-os

Watch

Fork 0

Map: os-me, a skill that predicts and gives Jared's response #936

New issue

Open

opened 2026-10-04 20:14:06 +00:00 by jared · 1 comment

jared commented

2026-10-04 20:14:06 +00:00

Owner

Problem

The os-me plugin does not yet exist as a working first version.

Destination

A working first version of the os-me plugin: /os-me:predict and /os-me:respond give the action Jared would choose, from a private Store that the skill builds and changes itself, with a replay score from a held-out set.

Notes

This map carries execution. Jared ruled on 2026-10-04 that the map ends at a working version, so task tickets build; they do not only decide.
Domain: Claude Code plugins in cc-os. Read plugins/cc-architect/references/conventions/cc-os-naming.md before you add a skill (no commands/ directory, no name: field in SKILL.md). Run bin/refresh-plugins after changes. Use /mattpocock-skills:writing-for-agents for skill text.
Terms: Prediction = the action the skill says Jared would choose, with its reasons. Agency = permission for an agent to act on a Prediction before Jared reads it. Store = the private repository that the skill owns and changes. Shell = the fixed parts in cc-os/plugins/os-me/ that the skill cannot change.
Jared's role: he suggests, and the skill decides if and how to use a suggestion. He interferes only if he dislikes a result.
Guardrails: reasonable limits only. Do not add a guardrail that no decision below names.
Sign of success: Jared invokes the skill more often than he types a reply or pushes back on a Prediction.

Decisions so far

Rulings from the charting session on 2026-10-04 (no ticket holds them):

Job: predict the action Jared would choose and give it to the other agent in his place. /os-me:predict returns a labeled Prediction for Jared to read. /os-me:respond gives the response to the agent to act on (Agency).
Authority: a Prediction does not replace a ruling that is Jared's, unless he invokes with Agency.
Score: replay of past transcripts (hide Jared's reply, predict it, compare). The live count of his corrections is a trend only. Only back-and-forth exchanges count; silence from Jared is not evidence.
Reach: global plugin. First tuning on cc-os transcripts. Learn how Jared decides (how he frames a problem, how he challenges a first result, when an answer is good enough), not only what he decided.
Freedom: the skill is fully free inside the Store: structure, documents, heuristics, retrieval, templates, code. It prefers a small amount of high-quality material and removes stale material. Wide changes first; later it updates only from transcripts where Jared pushed back.
Store home: its own private repository. It holds how Jared decides, and no client facts.
Shell outside the Store: entry skills, hook, guardrails, a held-out set that the skill does not read, and a fixed score script.
Guardrails: write only inside the Store; transcripts read-only; no credential files and no secret values; commit each meaningful change; no destructive git action (no force-push, no change or deletion of existing commits, no deletion of the repository). No token limit yet: record token use per run and set a limit after the first runs.
Existing knowledge: the feedback- memory files and the takedown rulings stay where they are. The skill reads them as sources.
Learning time: a hook only logs each invocation. The next invocation processes the logged sessions inline.
At invocation the skill can dispatch read-only research on random samples of cc-os transcripts (not the held-out sessions). Use it, and then update the Store.
Self-review: the skill reviews its own transcripts for efficiency and response quality, daily at first and less often when it is tuned.
Record the os-me decision as an ADR — ADR-0272 Accepted; no conflict with an Accepted ADR.
Build the os-me Shell: predict and respond skills, invocation log hook — predict is user and model invocable, respond is user only; fork agent first, fresh agent as fallback; hook logs to the Store queue.
Create the os-me Store repository and its charter — private repository at ~/dev/os-me-store with charter, suggestions, entry.
Build the held-out set and the fixed replay score for os-me — 15 percent of sessions held out by id hash; os-me-eval scores same-action agreement with a JSON judge.
Run the os-me kickoff: the skill builds its first Store from transcripts — Store 483098e: one entry procedure, four judgment files, 29 cases, learning loop.
Score the first os-me version and try it live with Jared — baseline 1 of 20; live invocation works through the fallback agent; open for Jared's review.

Not yet specified

A channel for the skill to request resources or guidance and make its case. Wait for a real pain point.
Token limits per invocation and per learning run.
Agency by domain without a flag, when the score supports it.
A store for client facts, if the skill makes a case for one.
Tuning on projects other than cc-os, and on work that is not code.
The cadence rule for self-review as the skill matures.

Out of scope

## Problem The os-me plugin does not yet exist as a working first version. ## Destination A working first version of the `os-me` plugin: `/os-me:predict` and `/os-me:respond` give the action Jared would choose, from a private Store that the skill builds and changes itself, with a replay score from a held-out set. ## Notes - **This map carries execution.** Jared ruled on 2026-10-04 that the map ends at a working version, so task tickets build; they do not only decide. - Domain: Claude Code plugins in `cc-os`. Read `plugins/cc-architect/references/conventions/cc-os-naming.md` before you add a skill (no `commands/` directory, no `name:` field in `SKILL.md`). Run `bin/refresh-plugins` after changes. Use `/mattpocock-skills:writing-for-agents` for skill text. - Terms: **Prediction** = the action the skill says Jared would choose, with its reasons. **Agency** = permission for an agent to act on a Prediction before Jared reads it. **Store** = the private repository that the skill owns and changes. **Shell** = the fixed parts in `cc-os/plugins/os-me/` that the skill cannot change. - Jared's role: he suggests, and the skill decides if and how to use a suggestion. He interferes only if he dislikes a result. - Guardrails: reasonable limits only. Do not add a guardrail that no decision below names. - Sign of success: Jared invokes the skill more often than he types a reply or pushes back on a Prediction. ## Decisions so far Rulings from the charting session on 2026-10-04 (no ticket holds them): - Job: predict the action Jared would choose and give it to the other agent in his place. `/os-me:predict` returns a labeled Prediction for Jared to read. `/os-me:respond` gives the response to the agent to act on (Agency). - Authority: a Prediction does not replace a ruling that is Jared's, unless he invokes with Agency. - Score: replay of past transcripts (hide Jared's reply, predict it, compare). The live count of his corrections is a trend only. Only back-and-forth exchanges count; silence from Jared is not evidence. - Reach: global plugin. First tuning on `cc-os` transcripts. Learn how Jared decides (how he frames a problem, how he challenges a first result, when an answer is good enough), not only what he decided. - Freedom: the skill is fully free inside the Store: structure, documents, heuristics, retrieval, templates, code. It prefers a small amount of high-quality material and removes stale material. Wide changes first; later it updates only from transcripts where Jared pushed back. - Store home: its own private repository. It holds how Jared decides, and no client facts. - Shell outside the Store: entry skills, hook, guardrails, a held-out set that the skill does not read, and a fixed score script. - Guardrails: write only inside the Store; transcripts read-only; no credential files and no secret values; commit each meaningful change; no destructive git action (no force-push, no change or deletion of existing commits, no deletion of the repository). No token limit yet: record token use per run and set a limit after the first runs. - Existing knowledge: the `feedback-` memory files and the takedown rulings stay where they are. The skill reads them as sources. - Learning time: a hook only logs each invocation. The next invocation processes the logged sessions inline. - At invocation the skill can dispatch read-only research on random samples of `cc-os` transcripts (not the held-out sessions). Use it, and then update the Store. - Self-review: the skill reviews its own transcripts for efficiency and response quality, daily at first and less often when it is tuned. - [Record the os-me decision as an ADR](https://forgejo.swansoncloud.com/jared/cc-os/issues/937) — ADR-0272 Accepted; no conflict with an Accepted ADR. - [Build the os-me Shell: predict and respond skills, invocation log hook](https://forgejo.swansoncloud.com/jared/cc-os/issues/938) — predict is user and model invocable, respond is user only; fork agent first, fresh agent as fallback; hook logs to the Store queue. - [Create the os-me Store repository and its charter](https://forgejo.swansoncloud.com/jared/cc-os/issues/939) — private repository at ~/dev/os-me-store with charter, suggestions, entry. - [Build the held-out set and the fixed replay score for os-me](https://forgejo.swansoncloud.com/jared/cc-os/issues/940) — 15 percent of sessions held out by id hash; os-me-eval scores same-action agreement with a JSON judge. - [Run the os-me kickoff: the skill builds its first Store from transcripts](https://forgejo.swansoncloud.com/jared/cc-os/issues/941) — Store 483098e: one entry procedure, four judgment files, 29 cases, learning loop. - [Score the first os-me version and try it live with Jared](https://forgejo.swansoncloud.com/jared/cc-os/issues/942) — baseline 1 of 20; live invocation works through the fallback agent; open for Jared's review. ## Not yet specified - A channel for the skill to request resources or guidance and make its case. Wait for a real pain point. - Token limits per invocation and per learning run. - Agency by domain without a flag, when the score supports it. - A store for client facts, if the skill makes a case for one. - Tuning on projects other than `cc-os`, and on work that is not code. - The cadence rule for self-review as the skill matures. ## Out of scope

jared added the

project/cc-os

filed-by/user

labels

2026-10-04 20:14:06 +00:00

jared referenced this issue

2026-10-04 20:14:53 +00:00

Record the os-me decision as an ADR #937

jared referenced this issue

2026-10-04 20:15:05 +00:00

Build the os-me Shell: predict and respond skills, invocation log hook #938

jared referenced this issue

2026-10-04 20:15:17 +00:00

Create the os-me Store repository and its charter #939

jared referenced this issue

2026-10-04 20:15:26 +00:00

Build the held-out set and the fixed replay score for os-me #940

jared referenced this issue

2026-10-04 20:15:35 +00:00

Run the os-me kickoff: the skill builds its first Store from transcripts #941

jared referenced this issue

2026-10-04 20:15:42 +00:00

Score the first os-me version and try it live with Jared #942

jared commented

2026-10-04 20:17:37 +00:00

Author

Owner

Skeptic DROP of ticket 'os-backlog cannot apply wayfinder labels to an issue': depth-1 limit (AI-created parent), no hard floor. Labels wayfinder:map and wayfinder:task remain unapplied on #936-#942; a human can add them in the Forgejo UI, or a label-apply command can be filed once the labels are in actual use.

jared referenced this issue

2026-10-04 20:19:29 +00:00

os-backlog cannot apply wayfinder labels to an issue #943

jared referenced this issue

2026-10-04 20:24:32 +00:00

Build the os-me Shell: predict and respond skills, invocation log hook #938

jared added the

category/umbrella

label

2026-10-10 13:31:47 +00:00