nsaspy/a0-symbolics
1
1
Fork 0
Code Issues 140 Pull requests 7 Projects Releases Packages Wiki Activity Actions 1

feat(symbolics): route model turns through Prolog-RLM natively #91

Closed
nsaspy wants to merge 18 commits from fix/prolog-rlm-native-routing into main
pull from: fix/prolog-rlm-native-routing
merge into: nsaspy:main
Conversation 0 Commits 18 Files changed 55 +1448 -2020
nsaspy commented 2026-09-04 20:11:14 +00:00
Owner
Copy link

Summary

  • Fold _prolog_context_compiler into the single _prolog_rlm plugin: one fail-closed call_chat_model_turn extension compiles provider context in Prolog-RLM, invokes the model, and returns native Agent Zero tool calls through the normal host lifecycle.
  • Preserve Agent Zero as the visibility/tool-execution authority while Prolog-RLM owns symbolic selection, budgeting, SKILL.md normalization, skill-graph validation, and provider projection.
  • Keep response and skills_tool permanently visible so the model can always finish a turn or discover/load a skill; other tools remain relevance-selected by Prolog.
  • Project Agent Zero's already-filtered visible skill catalog as exact {name,path} packages plus the current-chat loaded-skill ledger. Hidden/shadowed skill names never cross the ABI.
  • Enable Prolog-RLM core skills by default. The canonical upstream adapter candidate is lost-rob0t/prolog-rlm#340 / issue #339: it parses those exact packages with rlm_skill, merges skill_default_catalog/1, validates the same normalized catalog with rlm_skill_graph, and returns active-skill/graph observability.
  • Do not reimplement metadata.prolog-rlm in Python and do not grant capability/authority from skill metadata.
  • Drop the old exec/git/patch plugin tool pack and standalone context-compiler plugin; shared transport remains under _prolog_rlm/helpers/transport.py.

Supersedes

  • Closed draft #51. Its toggle-based dual-plugin runtime-mode design is retired in favor of this single-router/single-plugin architecture.

Skill graph contract

  • Agent Zero owns project/profile/plugin precedence and hidden-skill policy.
  • Agent Zero sends exact admitted package directories, not broad roots.
  • selected_skills contains only currently loaded names that are still visible.
  • Prolog-RLM owns SKILL.md parsing, bounded resources/fingerprints, canonical metadata, selection, and graph validation.
  • Prolog-RLM core skills join the same graph when include_core_skills is enabled.
  • skills_tool remains available for runtime discovery/loading; a loaded skill is pinned through ordinary compiler semantics on later turns.

Verification

  • Added exact-head Symbolics RLM CI. It checks out the a0-symbolics PR head and the exact Prolog-RLM #340 candidate, temporarily resolves that candidate as the future flake pin, then runs the focused _prolog_rlm/prompt-budget pytest gate plus the full Prolog-RLM deterministic suite under a 4 GiB cap and checks flake evaluation.
  • Current upstream candidate under test: b12561cf888f62c5bfe692cc6669eae50eb1b634.
  • Do not merge this PR until that exact integration gate is green and the resulting Prolog-RLM revision is committed in both flake.nix and flake.lock.
## Summary - Fold `_prolog_context_compiler` into the single `_prolog_rlm` plugin: one fail-closed `call_chat_model_turn` extension compiles provider context in Prolog-RLM, invokes the model, and returns native Agent Zero tool calls through the normal host lifecycle. - Preserve Agent Zero as the visibility/tool-execution authority while Prolog-RLM owns symbolic selection, budgeting, SKILL.md normalization, skill-graph validation, and provider projection. - Keep `response` and `skills_tool` permanently visible so the model can always finish a turn or discover/load a skill; other tools remain relevance-selected by Prolog. - Project Agent Zero's already-filtered visible skill catalog as exact `{name,path}` packages plus the current-chat loaded-skill ledger. Hidden/shadowed skill names never cross the ABI. - Enable Prolog-RLM core skills by default. The canonical upstream adapter candidate is `lost-rob0t/prolog-rlm#340` / issue #339: it parses those exact packages with `rlm_skill`, merges `skill_default_catalog/1`, validates the same normalized catalog with `rlm_skill_graph`, and returns active-skill/graph observability. - Do not reimplement `metadata.prolog-rlm` in Python and do not grant capability/authority from skill metadata. - Drop the old exec/git/patch plugin tool pack and standalone context-compiler plugin; shared transport remains under `_prolog_rlm/helpers/transport.py`. ## Supersedes - Closed draft #51. Its toggle-based dual-plugin runtime-mode design is retired in favor of this single-router/single-plugin architecture. ## Skill graph contract - Agent Zero owns project/profile/plugin precedence and hidden-skill policy. - Agent Zero sends exact admitted package directories, not broad roots. - `selected_skills` contains only currently loaded names that are still visible. - Prolog-RLM owns SKILL.md parsing, bounded resources/fingerprints, canonical metadata, selection, and graph validation. - Prolog-RLM core skills join the same graph when `include_core_skills` is enabled. - `skills_tool` remains available for runtime discovery/loading; a loaded skill is pinned through ordinary compiler semantics on later turns. ## Verification - Added exact-head `Symbolics RLM` CI. It checks out the a0-symbolics PR head and the exact Prolog-RLM #340 candidate, temporarily resolves that candidate as the future flake pin, then runs the focused `_prolog_rlm`/prompt-budget pytest gate plus the full Prolog-RLM deterministic suite under a 4 GiB cap and checks flake evaluation. - Current upstream candidate under test: `b12561cf888f62c5bfe692cc6669eae50eb1b634`. - Do not merge this PR until that exact integration gate is green and the resulting Prolog-RLM revision is committed in both `flake.nix` and `flake.lock`.
Fold the context compiler into _prolog_rlm as a single fail-closed
call_chat_model_turn extension: Prolog compiles context and executes the
provider turn, returning native Agent Zero tool calls for normal
lifecycle execution. Drop the exec/git/patch tool pack and hide-eager
hooks, move transport under _prolog_rlm, pin prolog-rlm to b891c27 in
the flake, and update smoke checks plus plugin DOX accordingly.
Point PROLOG_RLM_TEST_ROOT at the pinned prolog-rlm pack checkout, add a
root conftest so repo-root modules import under bare pytest, and state
the 4 GB nix cap as an ulimit wrapper instead of systemd-run.
Make the sentence-transformers import in models.py lazy and gate the
torch-chain packages in the devshell Python behind A0_TEST_LITE so
impure evaluation can build a bounded test environment; document the
ulimit-wrapped test invocation in the root DOX.
docs(symbolics): document canonical skill graph bridge
Some checks failed
Symbolics RLM / skill-graph-integration (push) Failing after 10s
Details
Symbolics RLM / skill-graph-integration (pull_request) Failing after 10s
Details
56f7ffe974
ci(symbolics): fetch install-nix-action by full URL
Some checks failed
Symbolics RLM / skill-graph-integration (push) Has been cancelled
Details
Symbolics RLM / skill-graph-integration (pull_request) Failing after 1m0s
Details
4566258dc8 
cachix/install-nix-action is not mirrored on data.forgejo.org, so the
runner fails action resolution before any step runs. Reference the
action by its fully-qualified URL instead.
ci(symbolics): clone prolog-rlm from github and drop forgejo token for nix
Some checks failed
Symbolics RLM / skill-graph-integration (push) Has been cancelled
Details
Symbolics RLM / skill-graph-integration (pull_request) Has been cancelled
Details
cf53e46dde 
- actions/checkout with repository: lost-rob0t/prolog-rlm resolves
  against the local Forgejo instance where the repo does not exist;
  clone it from github.com with a fetch-by-sha step instead.
- secrets.GITHUB_TOKEN on Forgejo is an instance token, not a GitHub
  one; nix tarball fetches with it get HTTP 401 from api.github.com.
  Fetch the public flake inputs anonymously.
fix(nix): build nixpkgs mcp without its upstream test suite
Some checks failed
Symbolics RLM / skill-graph-integration (push) Has been cancelled
Details
Symbolics RLM / skill-graph-integration (pull_request) Has been cancelled
Details
95deae8362 
Skip mcp's upstream permission tests (they fail under root single-user
nix on the docker CI runner) and retrigger nix CI when nix/ changes.
nsaspy closed this pull request 2026-09-06 12:58:31 +00:00
Some checks failed
Symbolics RLM / skill-graph-integration (push) Has been cancelled
Details
Symbolics RLM / skill-graph-integration (pull_request) Has been cancelled
Details

Pull request closed

Please reopen this pull request to perform a merge.
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels   
accessibility
Barrier affecting people with disabilities

  
bug
Something isn't working

  
documentation
Improvements or additions to documentation

  
duplicate
This issue or pull request already exists

  
enhancement
New feature or request

  
good first issue
Good for newcomers

  
help wanted
Extra attention is needed

  
invalid
This doesn't seem right

  
question
Further information is requested

  
wontfix
This will not be worked on

No labels
accessibility
bug
documentation
duplicate
enhancement
good first issue
help wanted
invalid
question
wontfix
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
nsaspy/a0-symbolics!91
No description provided.