From 805606ed0ae022f21280c4a8eb3942f2c1b3d872 Mon Sep 17 00:00:00 2001 From: cafitac Date: Mon, 27 Apr 2026 23:52:31 +0900 Subject: [PATCH] docs: align canonical learning storage guidance --- .gitignore | 5 +++++ docs/adapter-convergence.md | 16 ++++++++++++++-- examples/consumer-repo-layout.md | 2 +- 3 files changed, 20 insertions(+), 3 deletions(-) diff --git a/.gitignore b/.gitignore index 238162e..1cfb9f6 100644 --- a/.gitignore +++ b/.gitignore @@ -13,3 +13,8 @@ node_modules/ frontend/node_modules/ .idea/ *.tgz +.agent-learner/learning/inbox/ +.agent-learner/history/ +.agent-learner/events/ +.agent-learner/candidates/ +.agent-learner/state/ diff --git a/docs/adapter-convergence.md b/docs/adapter-convergence.md index 08e7cac..b3cccb5 100644 --- a/docs/adapter-convergence.md +++ b/docs/adapter-convergence.md @@ -1,7 +1,7 @@ # Adapter Convergence Roadmap `agent-learner` should become the common control plane for learned feedback across -Codex and Claude instead of leaving the stronger runtime-specific logic trapped +Codex, Claude, and Hermes instead of leaving the stronger runtime-specific logic trapped inside adapter-local scripts. ## What the real local setups show @@ -22,6 +22,14 @@ After inspecting the current local environments: - prompt-time context injection is the natural place to apply approved learning - OMX adds routing/state/trace infrastructure, but not a shared learned-feedback control plane equivalent to Claude's `cc-learner` +### Hermes today (`.hermes` / `~/.hermes`) +- thin experimental adapter path built around project-local learned rules +- prompt-time context injection via the real Hermes `pre_llm_call` shell hook +- session-end capture via the real Hermes `on_session_end` shell hook and normalized `session_end` events +- shared retrieval and event-processing core reused instead of adding a Hermes-specific learner +- current recommended rollout is explicit opt-in and project scope first +- installer writes a project-local `config.yaml` only when missing and always emits a mergeable `config.agent-learner.yaml` snippet for fail-safe adoption + ## Principle Do not try to make Codex behave like Claude by copying ad hoc scripts forever. @@ -37,11 +45,14 @@ Instead: - transcript-aware candidate extraction into `.agent-learner/candidates/` - processed marker state under `.agent-learner/state/processed-events/` - Codex prompt-time learning context injection +- Hermes prompt-time learning context injection via `render-hermes-context` +- Hermes normalized `session_end` event capture and processing - adapter installers that stay thin while pointing back to shared CLI/core paths ## Must-have next capabilities - separation between raw event capture, extracted candidates, and promoted durable rules +- runtime-contract validation for Hermes hook payloads so the experimental adapter can harden without widening core branching ## Should-have soon @@ -50,10 +61,11 @@ Instead: - adapter-specific ranking hints layered on top of the shared retrieval core - processed marker metadata richer than a simple done file (processor version, timestamps) - promotion review UX that uses model validation/exclusion signals directly +- a common `render-context --adapter ` surface once Hermes/Codex output shapes can converge cleanly ## Later -- Claude and Codex parity for long-running refinement workers +- Claude, Codex, and Hermes parity for long-running refinement workers - cross-runtime promotion dashboards - autoresearch-assisted validation loops - promotion review UX that mirrors the stronger Claude portfolio-management workflow diff --git a/examples/consumer-repo-layout.md b/examples/consumer-repo-layout.md index 97a1fe2..37835bf 100644 --- a/examples/consumer-repo-layout.md +++ b/examples/consumer-repo-layout.md @@ -3,7 +3,7 @@ After running: ```bash -agent-learner bootstrap --target /path/to/consumer-repo +agent-learner bootstrap ``` you should expect a consumer repo layout similar to: