SUPPORTING KNOWLEDGE · 0019
Code archaeology: agents and skills¶
A name in the code is a lead, not an installed feature. The September 2026 audit followed agent and skill references through Berd, Buzz and Goose, then checked the relevant history. Some apparent gaps were real recoveries. Others were renamed files, examples, private dependencies or ideas that never became standalone prompts.
The Missing agents catalogue holds the descriptions and prompts. The skills evidence register records procedures, dependencies and unresolved names. Both distinguish recovered source from a proposed adaptation.
Why it matters to you¶
Copying a prompt is easy. Copying its environment is not. An instruction can name a reviewer, a search service or an organisational handbook without supplying any of them.
| Finding | What changed in our understanding |
|---|---|
| Planner, Reviewer, Automator, Explorer | Their original Berd prompts were two generated sentences, not lost long-form files. Longer versions are reconstructed expansions. |
| Solo, Scout, Ralph | Complete originals survive in Goose's removed goose2 tree. They are historical working styles, not newly discovered Berd built-ins. |
| Conductor | The role is referenced; no public original prompt was recovered. Similarity to Solo is not proof of a rename. |
| Sprout | An old filename for Agt. Builder, not another missing agent. |
block.md and feedback triage |
Deliberately removed from public distribution because their work depended on internal resources. |
Each conclusion has pinned source links in the catalogue or register. “Not recovered” describes the search scope, not every repository or branch that could exist.
How to apply it¶
Read a prompt in three passes: what job it defines, what authority it assumes, and what resources it needs. Preserve useful boundaries, then replace host-specific operations with ones your actual setup supports. Test a representative request before relying on the result.
For skills, check whether the procedure is absent at all. write-like-me is meant to be created from your writing. goose-help became Berd help; a standalone berdctl skill never existed in the searched history. Buzz's github-research is already a complete example. Reconstructing all four as “missing skills” would solve four different problems as if they were one.
What this suggests for a local Berd brain¶
The removed block.md agent points to authoritative knowledge: a maintained operating-model document and a broader organisational corpus. Feedback triage points to a procedure plus a live system: prioritisation rules, tracker configuration and issue records. Neither is evidence of a universal hidden memory store.
Design proposal, not an existing shared feature
Keep accepted knowledge separate from raw evidence, reusable skills and live tool state. Give records a source, date, audience and revision history. Let each app retrieve relevant material through an explicitly configured interface, with proposed additions reviewed before acceptance. This is a direction for a local brain, not an implementation recovered from the repositories. Checked: 2026-09-19.
Goose's shared .agents discovery offers a practical instruction-sharing seam. Memory files and chat recall are different mechanisms. Berd's inspected memory branch was unmerged, and using Goose from both Berd and Buzz does not automatically synchronise their conversations or permissions. The register explains those limits.
If you ignore this¶
A copied agent promises tools it cannot call, a source-only feature becomes a documented product capability, or a past conversation quietly becomes “what we know”. The text looks complete while the important dependency is still missing.
Examples¶
An old workshop manual can recover a method. It does not put the missing machine in your workshop.
Reuse Lev's read-only security-review boundary, replace its Buzz mention route with an authorised result destination, and test it on scoped material. Label the result a Berd adaptation.
Install Solo as “the original Conductor”, copy internal tracker identifiers into a new triage skill, and assume shared Goose usage gives every app the same memory.
Do it: Build an agent of your own · Write a skill and decide where it lives
Last checked: Berd v0.6.4, 2026-09-19 — documentation baseline only. The linked audit includes newer Berd/Buzz/Goose snapshots and historical or unmerged code; it is source-read, not an installed-feature or cross-app behaviour test.
Reference¶
- Missing agents: sources, limits and individual prompts
- Skills found in the code archaeology
- What survives the trip to Buzz — the existing handoff guide
- Verified
- Berd v0.6.4 · 2026-09-19