feat: per-project digests loaded into system prompt #43

Merged

joshtronic merged 1 commit from feat/project-digests into master

2026-05-20 02:14:03 +00:00

joshtronic commented

2026-05-20 02:13:10 +00:00

Owner

Companion to brain feat/project-digests. Wires the harness to load
the right digest file per tick based on the target repo.

brain_system_prompt now takes an optional second arg, the target
repo (e.g., "joshtronic/joshing.you"). When passed, it looks for
brain/memories/digest/_.md (slash replaced with
underscore for filesystem) and loads that file if present. The
digest goes LAST in the prompt so its churn doesn't invalidate
upstream cache.

All four invocation sites pass the relevant target:

tier-1 issue work: "$FORGEJO_REPO"
tier-3 discretionary website: "$W_REPO"
PR-review pickup: "$PR_REPO"
scheduled maintenance: "$target"

Isolation is the point: each tick only sees the digest for the
project it's working on. Other projects' digests stay on disk but
are invisible. Prevents the joshing.you context leaking into
certifiedtradejobs.com (or vice versa) -- the bleed concern Josh
raised when we designed this.

AGENTS.md adds a "Project digests" section explaining what goes
in vs out, the isolation rule (Claude must not write to another
project's digest from within a tick), when to update, the
first-tick behavior (create if useful, skip if not), and the size
target (~100-150 lines). Points at brain/memories/digest/README.md
for the full convention.

Companion brain PR scaffolds memories/digest/ with README.md and
adds the MEMORY.md index line. No seed digest files yet -- they
get created on first relevant tick.

Companion to brain feat/project-digests. Wires the harness to load the right digest file per tick based on the target repo. brain_system_prompt now takes an optional second arg, the target repo (e.g., "joshtronic/joshing.you"). When passed, it looks for brain/memories/digest/<owner>_<repo>.md (slash replaced with underscore for filesystem) and loads that file if present. The digest goes LAST in the prompt so its churn doesn't invalidate upstream cache. All four invocation sites pass the relevant target: - tier-1 issue work: "$FORGEJO_REPO" - tier-3 discretionary website: "$W_REPO" - PR-review pickup: "$PR_REPO" - scheduled maintenance: "$target" Isolation is the point: each tick only sees the digest for the project it's working on. Other projects' digests stay on disk but are invisible. Prevents the joshing.you context leaking into certifiedtradejobs.com (or vice versa) -- the bleed concern Josh raised when we designed this. AGENTS.md adds a "Project digests" section explaining what goes in vs out, the isolation rule (Claude must not write to another project's digest from within a tick), when to update, the first-tick behavior (create if useful, skip if not), and the size target (~100-150 lines). Points at brain/memories/digest/README.md for the full convention. Companion brain PR scaffolds memories/digest/ with README.md and adds the MEMORY.md index line. No seed digest files yet -- they get created on first relevant tick.

joshtronic added 1 commit

2026-05-20 02:13:10 +00:00

feat: per-project digests loaded into system prompt

Lint / check-sync (push) Successful in 4s

Details

Lint / check-sync (pull_request) Successful in 4s

Details

c98d65822b

Companion to brain feat/project-digests. Wires the harness to load
the right digest file per tick based on the target repo.

brain_system_prompt now takes an optional second arg, the target
repo (e.g., "joshtronic/joshing.you"). When passed, it looks for
brain/memories/digest/<owner>_<repo>.md (slash replaced with
underscore for filesystem) and loads that file if present. The
digest goes LAST in the prompt so its churn doesn't invalidate
upstream cache.

All four invocation sites pass the relevant target:
  - tier-1 issue work:           "$FORGEJO_REPO"
  - tier-3 discretionary website: "$W_REPO"
  - PR-review pickup:             "$PR_REPO"
  - scheduled maintenance:        "$target"

Isolation is the point: each tick only sees the digest for the
project it's working on. Other projects' digests stay on disk but
are invisible. Prevents the joshing.you context leaking into
certifiedtradejobs.com (or vice versa) -- the bleed concern Josh
raised when we designed this.

AGENTS.md adds a "Project digests" section explaining what goes
in vs out, the isolation rule (Claude must not write to another
project's digest from within a tick), when to update, the
first-tick behavior (create if useful, skip if not), and the size
target (~100-150 lines). Points at brain/memories/digest/README.md
for the full convention.

Companion brain PR scaffolds memories/digest/ with README.md and
adds the MEMORY.md index line. No seed digest files yet -- they
get created on first relevant tick.

joshtronic merged commit 5ff1837ed4 into master

2026-05-20 02:14:03 +00:00

joshtronic deleted branch feat/project-digests

2026-05-20 02:14:03 +00:00

joshtronic referenced this pull request from a commit

2026-05-20 02:14:04 +00:00

Merge pull request 'feat: per-project digests loaded into system prompt' (#43) from feat/project-digests into master