memory.md tells agents to match tasks against skill descriptions but not how to learn which skills exist #73

New issue

Closed

opened 2026-09-30 22:05:22 +01:00 by allod-agent · 1 comment

allod-agent commented

2026-09-30 22:05:22 +01:00

Member

Agents are directed to discover skills from skills/ by matching the task against each skill's frontmatter description — but nothing tells them to enumerate the directory first, so an agent that has not already seen a skill's name can only match skills it already knows about. In practice this session read memory.md at start, was asked to "pick the next issue", and ran a manual tracker sweep before the owner pointed out the pick-next-issue skill existed.

Primary goals:

Enumerate before matching — the Skills paragraph instructs one command (grep -h '^description:' skills/*/SKILL.md) that prints every skill's trigger, run before task work, instead of only "when a task matches".
No duplicate index — memory.md keeps naming no individual skill (its byte budget and the hygiene rules both argue against mirroring README.md's list); the fix is a probe, not a copy.

Scope

The fix changes only the Skills paragraph in memory.md plus one ledger.md sighting recording the miss. README.md already indexes the skills for humans and is unchanged.

Agents are directed to discover skills from `skills/` by matching the task against each skill's frontmatter description — but nothing tells them to enumerate the directory first, so an agent that has not already seen a skill's name can only match skills it already knows about. In practice this session read `memory.md` at start, was asked to "pick the next issue", and ran a manual tracker sweep before the owner pointed out the `pick-next-issue` skill existed. Primary goals: - **Enumerate before matching** — the Skills paragraph instructs one command (`grep -h '^description:' skills/*/SKILL.md`) that prints every skill's trigger, run before task work, instead of only "when a task matches". - **No duplicate index** — `memory.md` keeps naming no individual skill (its byte budget and the hygiene rules both argue against mirroring `README.md`'s list); the fix is a probe, not a copy. ### Scope The fix changes only the Skills paragraph in `memory.md` plus one `ledger.md` sighting recording the miss. `README.md` already indexes the skills for humans and is unchanged.

allod-agent commented

2026-09-30 22:08:27 +01:00

Author

Member

Owner redirect mid-review: instead of teaching every agent the grep -h '^description:' incantation, add a read-only CLI command (allod skills list) in allod/tools that prints each skill's name and frontmatter description, defaulting to allod/memory's skills/ through the repository registry; memory.md then points at the command instead of at a shell pipeline. This issue now spans allod/tools (implementation) and allod/memory (pointer).

Owner redirect mid-review: instead of teaching every agent the `grep -h '^description:'` incantation, add a read-only CLI command (`allod skills list`) in `allod/tools` that prints each skill's name and frontmatter description, defaulting to `allod/memory`'s `skills/` through the repository registry; `memory.md` then points at the command instead of at a shell pipeline. This issue now spans `allod/tools` (implementation) and `allod/memory` (pointer).

allod-agent referenced this issue from allod/tools

2026-09-30 22:20:24 +01:00

allod skill: list agent skills, or print one in full #256

allod-agent referenced this issue from a pull request that will close it,

2026-09-30 22:20:49 +01:00

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