docs: sync README + CHANGELOG with skills inject mode, output-chars knob, and litellm fix

- README/CHANGELOG: document `discovery="inject"` (read_skill-only +
  `render_skills_manifest` catalog injection) alongside the default list mode.
- README Configuration: add the `RLM_MAX_OUTPUT_CHARS` env knob to the list.
- CHANGELOG: add a Fixed entry for forcing litellm onto httpx
  (`disable_aiohttp_transport`) to silence the "Unclosed connector" warning.
- README: mark v0.2.0 as in-development (not yet tagged/published).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: qazbnm456

CHANGELOG.md

@@ -152,7 +152,11 @@ surfaced by dogfooding a real downstream consumer.
   validate → post-process pipeline; `model_as_tool` exposes extra models for LM-decided
   multi-model routing. *(Renamed from `make_middleware_lm` — see Changed.)*
 - **Skills-as-tools** (`skills.py`): `load_skills_as_tools` surfaces a Skills
-  directory to the RLM as `list_skills` / `read_skill` tools.
+  directory to the RLM. Default `discovery="list"` gives the LM `list_skills` /
+  `read_skill` (discover-then-read). `discovery="inject"` returns `read_skill`
+  only, and the caller injects the catalog into the prompt itself via
+  `render_skills_manifest(dir)` (or reads it structurally with `discover_skills`) —
+  skipping the `list_skills` round-trip when the skill set is small and fixed.
 - **Unified trajectory recording** (`trace.py`): `TraceRecorder` writes an
   append-only JSONL stream — main steps (`Prediction.trajectory`), every sub-LM
   call, every tool call — keyed by `run_id` + `step_id`. Optional Langfuse mirror.
@@ -193,6 +197,16 @@ surfaced by dogfooding a real downstream consumer.
 - **GEPA harness skeleton** (`optimize.py`): metric templates now; `compile_task`
   is a documented Phase-2 stub.
 
+### Fixed
+
+- **No more "Unclosed connector" warning from litellm** (`runtime.py`). litellm
+  (dspy's LM backend) defaults to an aiohttp transport whose pooled `ClientSession`
+  is bound to the per-run `asyncio.run` loop; when that loop closes, aiohttp logs a
+  noisy "Unclosed connector" through the loop's exception handler. `RLMTask` now sets
+  `litellm.disable_aiohttp_transport = True` before the first LM call, forcing litellm
+  onto httpx so no aiohttp session is created and nothing dangles. Best-effort and
+  idempotent — a litellm-free install just no-ops.
+
 ### Changed / Hardened (surfaced by dogfooding a consumer)
 
 - **`RLMConfig.max_retries` now defaults to `1` (was `3`) — no whole-RLM retry by default** (`config.py`;

README.md

@@ -187,6 +187,13 @@ body. The list→read split is two-level **progressive disclosure**; and because
 these as tools inside the REPL, "which knowledge did I load" lands in the trajectory (and the
 RL dataset).
 
+For a larger catalog you can skip the discovery round-trip: `load_skills_as_tools(dir,
+discovery="inject")` returns **only** `read_skill`, and you inject the catalog into the system
+prompt yourself with `render_skills_manifest(dir)`. The LM then sees every skill's
+`name` + `description` at startup (no `list_skills` call) and still pulls a full body
+just-in-time with `read_skill`. The default `discovery="list"` keeps the `list_skills` tool
+instead. See `examples/harness_run.py`.
+
 Scope & caveats:
 - **Knowledge only.** `read_skill` returns the markdown text — it does NOT execute bundled
   scripts or expand linked files. A "just instructions" skill works fully; a skill that ships
@@ -360,8 +367,8 @@ tends to retry verbatim. Isolation is unchanged; `RLMTask` owns the teardown.
 All via env (`RLMConfig.from_env()`): `RLM_MAIN_MODEL` (or `AI_MODEL_NAME`),
 `RLM_SUB_MODEL` (or `SUB_AI_MODEL_NAME`), `RLM_API_KEY` (or `AI_API_KEY`),
 `RLM_BASE_URL` (or `AI_BASE_URL`), `RLM_INTERPRETER`, `RLM_ADAPTER`,
-`RLM_MAX_TOKENS`, `RLM_ALLOW_INSECURE_SANDBOX`, `RLM_MAX_ITERATIONS`,
-`RLM_MAX_LLM_CALLS`, `RLM_MAX_RETRIES`, `RLM_OBSERVE`.
+`RLM_MAX_TOKENS`, `RLM_MAX_OUTPUT_CHARS`, `RLM_ALLOW_INSECURE_SANDBOX`,
+`RLM_MAX_ITERATIONS`, `RLM_MAX_LLM_CALLS`, `RLM_MAX_RETRIES`, `RLM_OBSERVE`.
 
 The `AI_*` fallbacks let the kit drop into projects already keyed on those vars
 without re-keying env; the `RLM_*` form wins when both are set.
@@ -420,7 +427,8 @@ invariants when modifying the kit.
 
 ## Status
 
-**v0.2.0** — scaffold + harness-engineering layer (sub-LM hook, skills-as-tools,
+**v0.2.0** (in development — not yet tagged or published to PyPI; the version is the
+target) — scaffold + harness-engineering layer (sub-LM hook, skills-as-tools,
 trajectory recording, replay, dataset export). Hardened by dogfooding against a
 real downstream consumer; the changes that surfaced are in [`CHANGELOG.md`](https://github.com/qazbnm456/rlm-kit/blob/main/CHANGELOG.md).