fix(harness): OpenClaw entry threshold 32,000 bytes so served skills also fit a 100K-200K-token tool-result budget (GST-110)

Co-authored-by: capy-ai[bot] <230910855+capy-ai[bot]@users.noreply.github.com>

CHANGELOG.md

@@ -5,7 +5,7 @@
 **OpenClaw opens gstack's large task skills again (GST-110).** #3555 made `hybrid` OpenClaw's default and registered review, investigate, qa, ship and plan-eng-review as skills of their own. OpenClaw 2026.10.6 serves a registered skill's `SKILL.md` whole or refuses it, with a 32,768-byte limit when its model catalog has no context window for the model. In HARN-5's panel the agent's first read of the 77,691-byte `gstack-review` file was refused in 4 of 5 trials, and opens fell from 5/5 to 0/5. Under the router layout the same file had been read as a plain file inside the router's folder, which OpenClaw pages.
 
 ### Fixed
-- **`gstack-harness register openclaw`** (`lib/harness-register.ts`): a skill whose rendered `SKILL.md` is over 32,768 bytes (`OPENCLAW_INSTRUCTION_MAX`, with the OpenClaw source lines it comes from) is registered as a short entry with the same frontmatter whose body says to read the skill's full `SKILL.md` in a plain copy of every skill and follow it. In hybrid that copy is the router's `gstack/skills/`, where the router layout read the same file, now including the five task skills; under `--skills all` it is `.gstack-skills/`, a folder OpenClaw's skill scan skips. Because the copy holds every skill, `sections/` paths and sibling paths such as `../gstack-qa/` and autoplan's `../gstack-plan-eng-review/SKILL.md` resolve to full texts. That covers four of hybrid's five direct skills and 26 of 63 skills in `--skills all` today. OpenClaw treats only a registered skill folder's `SKILL.md` as an instruction file. The ledger records the copies, so refresh, mode switches, upgrades from a #3555 registration and uninstall handle them. Skill bodies are unchanged, and Hermes, dots and `generic` registrations are byte-identical.
+- **`gstack-harness register openclaw`** (`lib/harness-register.ts`): a skill whose rendered `SKILL.md` is over 32,000 bytes (`OPENCLAW_INSTRUCTION_MAX`, under both OpenClaw's 32,768-byte limit and the 32,000-character budget it applies with a 100K-200K-token context window, with the OpenClaw source lines it comes from) is registered as a short entry with the same frontmatter whose body says to read the skill's full `SKILL.md` in a plain copy of every skill and follow it. In hybrid that copy is the router's `gstack/skills/`, where the router layout read the same file, now including the five task skills; under `--skills all` it is `.gstack-skills/`, a folder OpenClaw's skill scan skips. Because the copy holds every skill, `sections/` paths and sibling paths such as `../gstack-qa/` and autoplan's `../gstack-plan-eng-review/SKILL.md` resolve to full texts. That covers four of hybrid's five direct skills and 26 of 63 skills in `--skills all` today. OpenClaw treats only a registered skill folder's `SKILL.md` as an instruction file. The ledger records the copies, so refresh, mode switches, upgrades from a #3555 registration and uninstall handle them. Skill bodies are unchanged, and Hermes, dots and `generic` registrations are byte-identical.
 - docs/hosts.md and docs/OPENCLAW.md describe the entry and the limit, and docs/uninstall.md's manual OpenClaw removal command also removes `.gstack-skills/`.
 
 ### Tests

docs/OPENCLAW.md

@@ -189,8 +189,8 @@ workflows and helpers itself. Use ACP when the work should happen in a Claude
 Code session with the full install (hooks, Claude tool names); use native
 skills when the OpenClaw agent should do it. Both can be installed together.
 OpenClaw serves a registered skill's `SKILL.md` whole or refuses it, and its
-limit is 32,768 bytes when it does not know the model's context window, so a
+limit is 32,768 bytes when it does not know the model's context window
-larger skill (review, investigate, ship and plan-eng-review among the five) is
+(32,000 characters with a 100K-200K-token one), so a skill over 32,000 bytes (review, investigate, ship and plan-eng-review among the five) is
 registered as a short entry that tells the agent to read the skill's full
 `SKILL.md` in `gstack/skills/`, a plain copy of every skill that OpenClaw
 reads in pages (`.gstack-skills/` under `--skills all`).

docs/hosts.md

@@ -59,7 +59,7 @@ characters, the part of a description Hermes's skills index shows.
 
 | Harness | What gstack registers | Install | Verify | Remove |
 |---|---|---|---|---|
-| OpenClaw | default `--skills hybrid`: `gstack-review`, `gstack-investigate`, `gstack-qa`, `gstack-ship` and `gstack-plan-eng-review` as top-level skills plus one `gstack` router skill whose body lists every other skill and its path, in `${OPENCLAW_STATE_DIR:-~/.openclaw}/skills`. With OpenClaw 2026.10.6's 19 eligible bundled skills, hybrid's skills section is 10,290 characters, full format, under its 18,000-character budget; `--skills all` (17,971, compact format, descriptions cut) lists every skill as its own folder, `--skills router` (6,701) only the router. OpenClaw serves a registered skill's `SKILL.md` whole or refuses it ("Skill instructions cannot be partially served"), at most 32,768 bytes when its model catalog has no context window for the model, so any registered skill larger than that (four of the five in hybrid, 26 of 63 in `all` today) is a short entry `SKILL.md` with the same name and description that sends the agent to its full `SKILL.md` in a plain copy of every skill (the router's `gstack/skills/`, or `.gstack-skills/` under `--skills all`, a folder OpenClaw's skill scan skips), which OpenClaw reads in pages. With a known context window OpenClaw also caps a tool result (16,000 characters under 100K tokens), which this limit does not cover | `./setup --host openclaw`, or `./setup` inside OpenClaw | `openclaw skills list --json` shows `gstack` and the five task skills (source `openclaw-managed`, eligible) | `gstack-harness unregister openclaw` |
+| OpenClaw | default `--skills hybrid`: `gstack-review`, `gstack-investigate`, `gstack-qa`, `gstack-ship` and `gstack-plan-eng-review` as top-level skills plus one `gstack` router skill whose body lists every other skill and its path, in `${OPENCLAW_STATE_DIR:-~/.openclaw}/skills`. With OpenClaw 2026.10.6's 19 eligible bundled skills, hybrid's skills section is 10,290 characters, full format, under its 18,000-character budget; `--skills all` (17,971, compact format, descriptions cut) lists every skill as its own folder, `--skills router` (6,701) only the router. OpenClaw serves a registered skill's `SKILL.md` whole or refuses it ("Skill instructions cannot be partially served"), at most 32,768 bytes when its model catalog has no context window for the model and 32,000 characters with a 100K-200K-token window, so any registered skill over 32,000 bytes (four of the five in hybrid, 26 of 63 in `all` today) is a short entry `SKILL.md` with the same name and description that sends the agent to its full `SKILL.md` in a plain copy of every skill (the router's `gstack/skills/`, or `.gstack-skills/` under `--skills all`, a folder OpenClaw's skill scan skips), which OpenClaw reads in pages. A context window under 100K tokens lowers that budget to 16,000 characters, which this limit does not cover | `./setup --host openclaw`, or `./setup` inside OpenClaw | `openclaw skills list --json` shows `gstack` and the five task skills (source `openclaw-managed`, eligible) | `gstack-harness unregister openclaw` |
 | Hermes | every skill, as `gstack-<skill>` in one `gstack` category, in `${HERMES_HOME:-~/.hermes}/skills/gstack`. Hermes's skills index lists only skills registered as folders, so a router would leave the rest out of the listing (`skill_view` still loads an unlisted skill by name, but nothing tells the model it exists) | `./setup --host hermes`, or `./setup` inside Hermes | `hermes skills list` shows the `gstack-*` skills in category `gstack` | `gstack-harness unregister hermes` |
 | dots (feder-cr) | the `gstack` router in `/home/dot/skills/gstack`, which the Dot lists every turn | `./setup` on the Dot's computer, or `gstack-harness register dots` | the Dot's Skills page, or `GET /skills` | `gstack-harness unregister dots` |
 | QM | **not available yet.** QM imports a skill pack by cloning the whole repository and caps it at 32 MiB and 5,000 files; gstack's repository is about 73 MiB, so `gstack-harness register qm --yes` stops before sending anything ([why](troubleshooting.md#harness-qm-pack-too-large)) | meanwhile, paste `gstack-harness register generic --print` into the agent's instructions | ask it for a gstack skill | delete the pasted text |

lib/harness-register.ts

@@ -41,11 +41,13 @@ export const HYBRID_SKILLS = ['review', 'investigate', 'qa', 'ship', 'plan-eng-r
  * OpenClaw serves a registered skill's SKILL.md whole or refuses it ("Skill instructions cannot be partially
  * served"), with maxBytes from resolveAdaptiveReadMaxBytes: 32,768 (DEFAULT_READ_PAGE_MAX_BYTES) when the model's
  * context window is unknown (OpenClaw 2026.10.6 dist/agent-tools.read-BIxUGJy-.mjs:168, :189-194, :747, :798,
- * :819-821). A known window adds a live tool-result cap that can be lower (16,000 characters under 100K tokens,
+ * :819-821). A known window also applies the tool-result budget to that read (resolveToolResultBudget,
- * dist/tool-result-limits-DNBSu_jg.mjs:61-79); this limit does not cover it. A larger skill is registered as a short
+ * dist/tool-result-limits-DNBSu_jg.mjs:61-95): 16,000 characters under 100K tokens, 32,000 to 200K, 64,000 above.
+ * 32,000 bytes fits the unknown-window limit and a 100K-200K window; windows under 100K tokens are not covered. The
+ * post-delivery cap defaults to a 200K window (64,000 characters). A larger skill is registered as a short
  * entry that points at its full SKILL.md in a plain tree of every skill, which OpenClaw reads in pages.
  */
-export const OPENCLAW_INSTRUCTION_MAX = 32_768;
+export const OPENCLAW_INSTRUCTION_MAX = 32_000;
 /** The plain tree under --skills all: a dot folder, which OpenClaw's skill scan skips (dist/skill-root-discovery-C2QVZC77.mjs:44). Hybrid and router use the router's own gstack/skills, which the scan never enters because gstack/ holds a SKILL.md (:343). */
 export const OPENCLAW_ALL_TEXTS = '.gstack-skills';