holycodex 0.16.4-dev.73.1 → 0.16.4-dev.78.1

package/dist/assets/plugin/skills/writing-for-agents/SKILL-MECHANICS.md

@@ -1,19 +1,9 @@
 # Skill mechanics
 
-This is the skill-specific branch of [writing-for-agents](SKILL.md): frontmatter, invocation choice, and router skills.
+Use a model-invoked skill when its description names a distinct model-facing
+branch. Keep the description trigger-first and short. Keep routing and
+invocation metadata in `agents/openai.yaml`; the body owns the behavior.
 
-## Invocation
-
-A **model-invoked** skill keeps a trigger-first `description` so an agent can select it and another skill can reach its shared reference. The always-loaded pointer spends context on every turn, so use one compact trigger per real branch and remove routing prose from the body.
-
-A **user-invoked** skill sets `disable-model-invocation: true`. Its description becomes a human-facing one-line summary with trigger lists removed. Use this when only a person should select the skill.
-
-Shared reference required by user-invoked skills belongs in a plain disclosed file that each can point to.
-
-## Splitting by invocation
-
-Split a model-invoked skill only when a distinct leading word should trigger it independently or another skill must reach it. The independent route must justify its permanent context pointer.
-
-## Router skills
-
-When many user-invoked skills exceed human cognitive load, add one user-invoked router that names them and their branches. A router can guide the person; it cannot autonomously invoke skills whose model invocation is disabled.
+Put uncommon mechanics in a linked reference only when a caller needs them.
+Keep each instruction in one source of truth and avoid copying body text into
+metadata or neighboring skills.

package/dist/assets/plugin/skills/writing-for-agents/SKILL.md

@@ -1,61 +1,35 @@
 ---
 name: writing-for-agents
-description: Use when Root prepares or reviews a subagent delegation, handoff, profile, skill, prompt, or agent-facing instruction; write a low-context independent contract. Load it before the first dispatch, then reuse it while the current context contains a complete usable load.
+description: Use when GPT-6 Astra Root writes or reviews a GPT-5.6 Luna specialist contract.
 ---
 
 # Writing for agents
 
-Treat each instruction as an interface with one contract:
-
-- **Owner:** the role permitted to act or decide.
-- **Boundary:** the exact scope, authority, constraints, and escalation point.
-- **Completion:** the observable outcome and proof that permit the agent to stop.
-- **Return:** the result Root needs to integrate: outcome or findings, changed
-  paths, verification, unresolved risk, and any requested decision.
-
-Write an objective, scope, retained constraints, evidence, exclusions, and the
-exact missing fact or material choice that returns to the owner. The receiver
-must be able to act, prove, and stop without reconstructing intent.
-
-Before any Root subagent dispatch, ensure this skill is fully loaded and applied
-in the current context. Reuse the loaded instructions for later dispatches
-while they remain complete and available; reload only when the current context
-no longer contains a complete usable load.
-
-## Dispatch contract
-
-Root owns dispatch. Every specialist invocation is a bounded Assignment whose
-objective, exact scope, exclusions, dependencies, acceptance criteria, proof
-route, and return format are explicit. Independent, non-overlapping
-Assignments may run concurrently; dependent work and writes to the same
-mutable seam remain ordered. Leaves execute only their Assignment: they do not
-delegate, message peers, mutate global Intent lifecycle, perform Git/VCS, or
-make material product or architecture decisions.
-
-## Information design
-
-A **context pointer** names out-of-context material and front-loads the branches that load it. Keep one trigger per genuine branch; synonyms spend permanent context without improving routing.
-
-Budget two loads:
-
-- **Context load:** always-loaded descriptions and repository rules.
-- **Cognitive load:** material a person must remember and select.
-
-Place information by immediacy: in-file steps, in-file reference, then disclosed reference. Inline what every invocation needs. Put branch-specific mechanics behind a clear pointer. Co-locate each concept with its rules and caveats; split real invocation or sequence branches when one file sprawls or later steps cause premature completion.
-
-Each step needs a clear, demanding completion criterion. Sharpen a vague bound before splitting the sequence. Use leading words with useful pretrained meaning to compress repeated concepts. Prefer positive steering; keep prohibitions for hard guardrails and pair them with the target behavior.
-
-## Pruning
-
-Keep each meaning in one source of truth. Treat code, configuration, directory layout, and `--help` as authoritative environment lookups; document only reasons, gotchas, or contracts the environment cannot reveal. Delete stale caches, irrelevant branches, sediment, and instructions that do not change model behavior.
-
-Use architectural role names: Root owns material decisions, Workers implement bounded seams, Reviewers repair their assigned findings to fixed point, and other specialists follow their literal authority. Continuations carry changed constraints, evidence, decisions, and semantic delta instead of replaying accepted context.
-
-Project the single repository surgical-mutation rule from `AGENTS.md` into
-every write-capable Assignment: minimize the edit/write surface and operation
-count while staying careful, complete, and evidence-driven. Do not create a
-weaker or competing mutation variant in a profile, skill, or dispatch prompt.
-
-When the document is a skill, read [SKILL-MECHANICS.md](SKILL-MECHANICS.md) for invocation and frontmatter mechanics.
-
-Completion means the instruction routes at the right branch, names authority and exclusions, prevents premature completion, supplies checkable proof, and contains no duplicated or no-op policy.
+This skill helps GPT-6 Astra Root author instructions for GPT-5.6 Luna. First
+identify Luna's effective context: the generated Luna baseline, repository
+`AGENTS.md`, the selected native Role.task, loaded skills, and relevant config
+or metadata. Do not restate a directive already visible to that receiver.
+
+Give each Assignment only the delta Luna needs:
+
+- objective and exact scope;
+- relevant context and authority;
+- unique hard constraints and exclusions;
+- acceptance criteria and required evidence;
+- the material choice or blocker that must escalate;
+- the return payload and stopping condition.
+
+Place new policy at the narrowest owner: Root/session behavior in Root
+instructions, repository conventions in `AGENTS.md`, Luna-global constraints in
+the generated baseline, task-specific behavior in Role.task, and conditional
+workflow in a skill or a disclosed reference. Keep invocation metadata as
+routing metadata.
+
+State completion before optional procedure. Use positive, literal boundaries,
+and disclose uncommon branches progressively. Let Astra infer routine safe
+defaults and use its judgment to decide what Luna actually needs; do not add
+competence scaffolding, generic recipes, or a broad questionnaire.
+
+When a contract is complete, it is independently actionable for Luna, contains
+no same-receiver semantic duplicate, and identifies the evidence that lets Root
+integrate the result.

package/dist/assets/plugin/skills/writing-for-agents/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Writing for agents"
   short_description: "Design compact, independently actionable agent instructions."
-  default_prompt: "When Root prepares a delegation or agent-facing instruction, ensure this skill is loaded and applied in the current context, then write or review a compact contract with behavior, authority, evidence, escalation, completion, and the repository surgical-mutation rule."
+  default_prompt: "Open the agent-instruction authoring workflow for this invocation."
 policy:
   allow_implicit_invocation: true