feat(20f): journal digestion formula — condense old journals into lessons learned #89

New issue

Closed

opened 2026-04-01 06:32:43 +00:00 by dev-bot · 1 comment

dev-bot commented

2026-04-01 06:32:43 +00:00

Collaborator

Parent

Part of #20 — agent .profile repo.

What to do

Create a formula formulas/digest-journal.toml that condenses accumulated journal entries into abstract, transferable lessons.

Digestion formula

The formula receives an agent's .profile repo path and:

Read all journal/issue-*.md files
Read existing knowledge/lessons-learned.md (if any)
Identify recurring patterns, common failures, successful strategies across journal entries
Rewrite knowledge/lessons-learned.md with condensed lessons (not append — full rewrite each time)
Archive digested journal files (move to journal/archive/)

Abstraction and size constraints

lessons-learned.md must be:

Hard capped at 2KB — the digestion formula must stay within this limit. If lessons exceed 2KB, the formula must prioritize and drop the least transferable ones.
Abstract, not specific — lessons should be patterns and heuristics, not references to specific issues or files. Good: "Always validate SOPS decryption output before sourcing — treat failures as fatal." Bad: "In issue #61 we found that lib/env.sh line 32 had an unsafe eval."
Transferable — a lesson should help with future work the agent has never seen, not just repeat past work. If a lesson only applies to one file or one situation, it is not worth keeping.

Gardener integration (trigger mechanism)

The gardener needs to know when to trigger digestion. Add to formulas/run-gardener.toml (as a new step or addition to an existing step):

List all Forgejo users that have a .profile repo:

for each user in admin/users:
  check if repos/<login>/.profile exists
  if yes, count files in journal/ (exclude archive/)

If journal count > 10 for any agent, dispatch digestion:
- Run formulas/digest-journal.toml with the agent's .profile repo as context
- Or file an action issue that triggers digestion (if the gardener cannot run formulas directly)

Note on context injection

Reading lessons-learned.md into agent context at session start is handled by #97 (profile_load_lessons). This issue only covers writing and maintaining the file.

Affected files

formulas/digest-journal.toml (new)
formulas/run-gardener.toml (update — add journal accumulation check + trigger)
gardener/gardener-run.sh (may need update if trigger requires new context)

Acceptance criteria

Digestion formula produces knowledge/lessons-learned.md from journal entries
lessons-learned.md is hard capped at 2KB
Lessons are abstract and transferable (no specific issue numbers or file paths)
Full rewrite each digestion (not append)
Digested journals moved to journal/archive/
Gardener checks journal accumulation for all agents with .profile repos
Gardener triggers digestion when journal count > 10
CI green

Dependencies

Depends on #97 (generic journal aspect — journal entries must exist to digest).

## Parent Part of #20 — agent .profile repo. ## What to do Create a formula `formulas/digest-journal.toml` that condenses accumulated journal entries into abstract, transferable lessons. ### Digestion formula The formula receives an agent's .profile repo path and: 1. Read all `journal/issue-*.md` files 2. Read existing `knowledge/lessons-learned.md` (if any) 3. Identify recurring patterns, common failures, successful strategies across journal entries 4. **Rewrite** `knowledge/lessons-learned.md` with condensed lessons (not append — full rewrite each time) 5. Archive digested journal files (move to `journal/archive/`) ### Abstraction and size constraints `lessons-learned.md` must be: - **Hard capped at 2KB** — the digestion formula must stay within this limit. If lessons exceed 2KB, the formula must prioritize and drop the least transferable ones. - **Abstract, not specific** — lessons should be patterns and heuristics, not references to specific issues or files. Good: "Always validate SOPS decryption output before sourcing — treat failures as fatal." Bad: "In issue #61 we found that lib/env.sh line 32 had an unsafe eval." - **Transferable** — a lesson should help with future work the agent has never seen, not just repeat past work. If a lesson only applies to one file or one situation, it is not worth keeping. ### Gardener integration (trigger mechanism) The gardener needs to know when to trigger digestion. Add to `formulas/run-gardener.toml` (as a new step or addition to an existing step): 1. List all Forgejo users that have a `.profile` repo: ``` for each user in admin/users: check if repos/<login>/.profile exists if yes, count files in journal/ (exclude archive/) ``` 2. If journal count > 10 for any agent, dispatch digestion: - Run `formulas/digest-journal.toml` with the agent's .profile repo as context - Or file an action issue that triggers digestion (if the gardener cannot run formulas directly) ### Note on context injection Reading `lessons-learned.md` into agent context at session start is handled by #97 (`profile_load_lessons`). This issue only covers writing and maintaining the file. ## Affected files - `formulas/digest-journal.toml` (new) - `formulas/run-gardener.toml` (update — add journal accumulation check + trigger) - `gardener/gardener-run.sh` (may need update if trigger requires new context) ## Acceptance criteria - [ ] Digestion formula produces `knowledge/lessons-learned.md` from journal entries - [ ] lessons-learned.md is hard capped at 2KB - [ ] Lessons are abstract and transferable (no specific issue numbers or file paths) - [ ] Full rewrite each digestion (not append) - [ ] Digested journals moved to `journal/archive/` - [ ] Gardener checks journal accumulation for all agents with .profile repos - [ ] Gardener triggers digestion when journal count > 10 - [ ] CI green ## Dependencies Depends on #97 (generic journal aspect — journal entries must exist to digest).

dev-bot added the

backlog

label 2026-04-01 06:32:43 +00:00

dev-bot referenced this issue

2026-04-01 06:33:04 +00:00

feat: agent .profile repo — formula snapshot + journal per agent instance #20

dev-bot referenced this issue

2026-04-01 08:15:53 +00:00

feat: generic journal aspect — post-session reflection + lessons-learned context injection #97

dev-bot commented

2026-04-01 08:29:38 +00:00

Author

Collaborator

Absorbed into #97. Digestion now happens inline in profile_load_lessons() — before session start, if >10 journals accumulated, a claude -p one-shot digests them into lessons-learned.md. No separate formula, no gardener trigger, no out-of-band scheduling.

Absorbed into #97. Digestion now happens inline in `profile_load_lessons()` — before session start, if >10 journals accumulated, a `claude -p` one-shot digests them into lessons-learned.md. No separate formula, no gardener trigger, no out-of-band scheduling.

dev-bot closed this issue

2026-04-01 08:29:38 +00:00

dev-bot referenced this issue

2026-04-01 08:29:39 +00:00

feat(20g): migrate all remaining agents to .profile + remove ops repo journal dirs #90