memory: discover skills by running allod skills list, not by already knowing them #74

Merged
vnprc merged 4 commits from agent/skills-discoverable into master 2026-09-30 23:12:41 +01:00
Member

An agent can now learn which skills exist by running allod skill, and read one with allod skill <name>; memory.md tells it to do that before task work instead of only matching tasks against descriptions it has never seen.

What changes for a running machine: the Skills paragraph in memory.md — the only skill surface read at session start — now instructs allod skill to list skills and allod skill <name> to read one (the allod/tools command in Refs below), with the old grep -h '^description:' line kept as a one-line fallback for machines that have not rebuilt yet. No skill is named or indexed in memory.md; README.md is untouched. One ledger.md sighting records the miss that motivated this: a session that read memory at start and still hand-rolled a tracker sweep until the owner pointed at pick-next-issue.

If it is wrong: revert restores the previous paragraph; the fallback grep keeps agents working even before the command reaches their machine.

Closes allod/memory#73.

Refs allod/tools#256 — the command this paragraph points at. Land that first: until a machine picks it up, the paragraph's fallback line is what agents run.

Risk

R0/R1. One paragraph and one ledger line in a docs repo; no runtime behavior of its own. The only integration uncertainty — the command not existing on an unrebuilt machine — is covered by the fallback line in the same paragraph.

Validation

  • .hooks/memory-budget --worktree passes (memory.md within its 16KB cap).
  • On this machine, the allod/tools branch's command lists all six skills in this repo's skills/ and prints pick-next-issue in full.
An agent can now learn which skills exist by running `allod skill`, and read one with `allod skill <name>`; `memory.md` tells it to do that before task work instead of only matching tasks against descriptions it has never seen. **What changes for a running machine:** the Skills paragraph in `memory.md` — the only skill surface read at session start — now instructs `allod skill` to list skills and `allod skill <name>` to read one (the `allod/tools` command in Refs below), with the old `grep -h '^description:'` line kept as a one-line fallback for machines that have not rebuilt yet. No skill is named or indexed in `memory.md`; `README.md` is untouched. One `ledger.md` sighting records the miss that motivated this: a session that read memory at start and still hand-rolled a tracker sweep until the owner pointed at `pick-next-issue`. **If it is wrong:** revert restores the previous paragraph; the fallback grep keeps agents working even before the command reaches their machine. Closes allod/memory#73. Refs allod/tools#256 — the command this paragraph points at. Land that first: until a machine picks it up, the paragraph's fallback line is what agents run. ## Risk R0/R1. One paragraph and one ledger line in a docs repo; no runtime behavior of its own. The only integration uncertainty — the command not existing on an unrebuilt machine — is covered by the fallback line in the same paragraph. ## Validation - `.hooks/memory-budget --worktree` passes (`memory.md` within its 16KB cap). - On this machine, the `allod/tools` branch's command lists all six skills in this repo's `skills/` and prints `pick-next-issue` in full.
Author
Member

Owner redirect: the paragraph will point at the singular form — bare allod skill lists, allod skill <name> prints the whole skill — so the grep fallback line stays only for machines without the command at all.

Owner redirect: the paragraph will point at the singular form — bare `allod skill` lists, `allod skill <name>` prints the whole skill — so the grep fallback line stays only for machines without the command at all.
vnprc merged commit c8e60db093 into master 2026-09-30 23:12:41 +01:00
vnprc deleted branch agent/skills-discoverable 2026-09-30 23:12:41 +01:00
Sign in to join this conversation.
No description provided.