holycodex 0.16.3-dev.66.1 → 0.16.4-dev.73.1

package/dist/assets/plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "holycodex",
-  "version": "0.16.3",
+  "version": "0.16.4",
   "description": "Root-directed capabilities and deterministic installed assets for HolyCodex.",
   "author": {
     "name": "David Basile Filho",
@@ -26,7 +26,7 @@
       "Code Intelligence"
     ],
     "defaultPrompt": [
-      "Use HolyCodex to implement and verify this task."
+      "Activate HolyCodex for this task. Read the generated Root developer instructions and repository AGENTS.md, apply installed skills as branch procedures, and use holycodex-agent semantic operations for durable Intent, Plan, and Assignment state."
     ]
   }
 }

package/dist/assets/plugin/skills/code-review/SKILL.md

@@ -3,14 +3,26 @@ name: code-review
 description: Use once after code or manifest implementation when Root needs adversarial review and bounded repair to a fixed point.
 ---
 
-Use this skill once after implementation or manifest work. Inspect the actual
-integrated result to a fixed point before Root judges readiness. Apply
+Owner: Reviewer within one delegated Assignment. Root must dispatch this skill
+after implementation or manifest work and record its structured result through
+`holycodex-agent assignment result`; the Reviewer never owns global lifecycle,
+material decisions, or Git/VCS. Inspect the actual integrated result to a
+fixed point before Root judges readiness. Apply
 [testing quality](../testing-quality.md) when judging new or modified tests.
 
+Read the Assignment and current Intent through the semantic agent interface;
+do not edit TOON files manually.
+
+This fixed-point review is mandatory after implementation or a major codebase
+change and must pass before Intent completion or any VCS operation. Apply the
+repository surgical-mutation rule from `AGENTS.md` to reviewer repairs.
+
 Inspect the assigned diff, callers, contracts, tests, compatibility, and
 artifact. Repair defects inside the review surface, verify each repair, and
 repeat until the review reaches a fixed point.
 
-Flag speculative abstractions, pointless file splits, oversized changes,
+Return `completed`, `blocked`, `needs_root_input`, or `failed` with findings,
+repaired paths, verification, and residual risk. Flag speculative
+abstractions, pointless file splits, oversized changes,
 unclear dependency direction, and tests that protect implementation detail.
 Confirm incidental generated output is absent before stopping.

package/dist/assets/plugin/skills/code-review/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Code review"
   short_description: "Review one implementation result to a fixed point."
-  default_prompt: "After implementation, inspect the assigned result once, repair reviewer-owned defects to a fixed point, and return evidence for Root's readiness judgment."
+  default_prompt: "After implementation, inspect the delegated Assignment result once, apply the repository surgical-mutation rule, repair reviewer-owned defects to a fixed point, and return structured evidence for Root's readiness judgment."
 policy:
   allow_implicit_invocation: true

package/dist/assets/plugin/skills/commit/SKILL.md

@@ -3,13 +3,35 @@ name: commit
 description: Use when Root owns a local commit after exact scope and proof are settled; verify scope, create the minimal commit, and report identity.
 ---
 
-Owner: Root. Boundary: perform all Git/VCS inspection and mutation. Verify the
-exact diff scope, required local proof, temporary/generated-artifact cleanup,
-and ignore coverage before staging. Preserve unrelated work and ensure
-environment secrets, credentials, and private release material are absent from
-the staged file list. Local commits need no user approval; every exact push,
-tag, merge, CI trigger, publication, or other remote mutation requires fresh
-user approval through native `request_user_input` immediately beforehand.
+Owner: Root. This is the only unconditional direct execution exception:
+perform all Git/VCS inspection and mutation. Verify the exact diff scope,
+required local proof, temporary/generated-artifact cleanup, and ignore coverage
+before staging. Preserve unrelated work and ensure environment secrets,
+credentials, and private release material are absent from the staged file list.
+Do not use this exception for implementation, testing, review, or release
+verification; delegate those as Assignments and persist their outcomes through
+`holycodex-agent assignment result`.
+
+Require a passing Reviewer.code fixed-point result after implementation or a
+major codebase change before this VCS exception is used. Apply the repository
+surgical-mutation rule to VCS operations and staging.
+
+After integration, Root commits/pushes the exact approved ref, then delegates
+terminal CI observation to `Worker.operations` with the exact ref and SHA.
+Discover the repository's actual development/release topology first; never
+assume GitHub, branch names, or that pending is success. Delegate fixes for
+failures, then repeat integration, commit, push, and observation until the
+repository gate is terminal green. If release is authorized, Root performs the
+repository's own release mechanism only after the development gate is terminal
+green, then delegates terminal release verification. If development and release
+share one pipeline, or the repository has no distinct release gate, record that
+topology and apply the repository's one available terminal gate. Fix failures
+through a new bounded Assignment and repeat the same review/VCS/observation
+cycle.
+
+Local commits need no user approval; every exact push, tag, merge, CI trigger,
+publication, or other remote mutation requires fresh user approval through
+native `request_user_input` immediately beforehand.
 
 Completion: Root reports the commit identity and post-commit status with
 redacted evidence, or returns an exact reproducible blocker. Never print secret

package/dist/assets/plugin/skills/commit/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Commit"
   short_description: "Create one local commit with exact scope proof."
-  default_prompt: "When Root owns a local commit, verify exact scope, required proof, ignore coverage, and secret exclusions, create only that commit, and report redacted identity and status."
+  default_prompt: "When Root uses the Git/VCS direct exception, apply the repository surgical-mutation rule, verify exact scope, required proof, ignore coverage, and secret exclusions, create only that commit, ask for approval before remote mutation, and delegate terminal exact-ref verification."
 policy:
   allow_implicit_invocation: false

package/dist/assets/plugin/skills/compress/SKILL.md

@@ -3,8 +3,17 @@ name: compress
 description: Use when Root assigns text that must be shortened; preserve meaning, constraints, technical terms, and voice.
 ---
 
-Owner: Worker. Boundary: remove repetition without changing requirements,
-technical terms, product voice, or required evidence.
+Owner: Worker within one delegated Assignment. Boundary: remove repetition
+without changing requirements, technical terms, product voice, or required
+evidence. Read the Assignment through the semantic agent interface; do not edit
+TOON files manually.
+
+Apply the repository surgical-mutation rule from `AGENTS.md`; preserve
+unrelated work and stop for Root input before expanding scope.
 
 Completion: the compressed text is constraint-equivalent to the input and any
-unresolvable ambiguity is explicit.
+unresolvable ambiguity is explicit. Return one outcome from `completed`,
+`blocked`, `needs_root_input`, or `failed` with changed paths, proof, and
+remaining risk. Root records the result through `holycodex-agent assignment
+result`; the Worker does not mutate global Intent lifecycle, make material
+decisions, or perform Git/VCS.

package/dist/assets/plugin/skills/compress/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Compress"
   short_description: "Shorten assigned text without changing its contract or voice."
-  default_prompt: "When Root assigns text compression, produce one compact constraint-equivalent draft and identify preserved technical terms."
+  default_prompt: "When Root assigns a bounded compression Assignment, apply the repository surgical-mutation rule, produce one constraint-equivalent draft, and return completed, blocked, needs_root_input, or failed evidence for holycodex-agent."
 policy:
   allow_implicit_invocation: true

package/dist/assets/plugin/skills/debugging/SKILL.md

@@ -6,10 +6,14 @@ description: Use when Root assigns a reproducible crash, wrong result, regressio
 Reproduce the defect before changing code. A stable red case distinguishes
 cause from symptom and makes the repair checkable.
 
-Owner: Worker within the assigned seam. Boundary: capture the smallest
+Owner: Worker within one delegated Assignment. Boundary: capture the smallest
 failing input and trace, form an evidence-backed cause, apply the narrow fix,
 and preserve unrelated behavior. Escalate competing causes or material
-redesign.
+redesign. Root creates/starts the Assignment and records the result through
+`holycodex-agent assignment result`; the Worker must not mutate Intent state or
+edit TOON files manually.
+
+Apply the repository surgical-mutation rule from `AGENTS.md`.
 
 Discover and use the repository-specific formatter, linter, typecheck, and
 validation commands before and after the repair. Start with the smallest
@@ -17,6 +21,7 @@ relevant local proof and expand only when the defect crosses that boundary or
 the release gate requires it; report unavailable commands rather than
 inventing replacements.
 
-Completion: the failure is reproduced, the cause is supported by repository
-evidence, the fix and regression proof pass, and remaining uncertainty is
-recorded.
+Return one outcome from `completed`, `blocked`, `needs_root_input`, or
+`failed`, including reproduction, cause, repair proof, and remaining
+uncertainty. Completion requires the failure to be reproduced, the cause to be
+supported by repository evidence, and the fix and regression proof to pass.

package/dist/assets/plugin/skills/debugging/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Debugging"
   short_description: "Reproduce and repair one observable defect."
-  default_prompt: "When Root assigns a reproducible defect, reproduce it, isolate its cause, make the narrow repair, and prove the regression is gone."
+  default_prompt: "When Root assigns a bounded defect Assignment, apply the repository surgical-mutation rule, reproduce it, isolate its cause, make the narrow repair, prove the regression is gone, and return a structured outcome for holycodex-agent."
 policy:
   allow_implicit_invocation: true

package/dist/assets/plugin/skills/handoff/SKILL.md

@@ -1,10 +1,18 @@
 ---
 name: handoff
-description: Use when Root pauses or transfers work; produce one redacted resumable handoff.
+description: Use when a caller needs a redacted projection of current Intent state for resume or export.
 ---
 
-Owner: Worker. Boundary: retain scope, owner, permissions, decisions,
-evidence, blockers, remaining risk, and the exact next action; omit secrets
-and unchanged narration.
+Handoff is a projection/export only, not persistent workflow state. Read the
+current Intent, Plan, and Assignments through `holycodex-agent`; never create a
+handoff file or duplicate lifecycle, owner, blocker, evidence, or resume state.
 
-Completion: one bounded, redacted handoff is ready for the receiving owner.
+Owner: Root. Boundary: request a compact redacted projection containing the
+current Intent reference, Plan revision, Assignment statuses, evidence,
+blockers, remaining risk, and exact next action; omit secrets, raw prompts,
+transcripts, and unchanged narration. The authoritative state remains under
+`.holycodex/{slug}-{short-id}/` and all mutations use semantic agent CLI
+operations.
+
+Completion: one bounded, redacted projection is ready for the receiving owner;
+no second source of truth exists.

package/dist/assets/plugin/skills/handoff/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Handoff"
-  short_description: "Produce one redacted handoff when assigned work pauses or transfers."
-  default_prompt: "When Root pauses or transfers work, write one redacted handoff containing scope, constraints, evidence, risks, and the next action."
+  short_description: "Export a redacted projection of current Intent state."
+  default_prompt: "When Root needs a resume projection, read current Intent, Plan, and Assignments through holycodex-agent and export one redacted projection without creating duplicate state."
 policy:
   allow_implicit_invocation: false

package/dist/assets/plugin/skills/operations/SKILL.md

@@ -0,0 +1,31 @@
+---
+name: operations
+description: Use when Root assigns exact-ref CI or release observation; discover the repository topology and report terminal external evidence without mutation.
+---
+
+Owner: Worker.operations within one delegated Assignment. Boundary: observe
+only the exact approved ref and SHA after Root's VCS action. First discover the
+repository's actual topology:
+separate development/release gates, one combined pipeline, or no formal
+separation. Do not assume GitHub, a branch name, or a provider.
+
+The Assignment must include the exact ref, SHA, required checks, and whether
+the observation is development CI or release verification. Do not rerun,
+cancel, approve, merge, push, tag, publish, deploy, or otherwise mutate
+external state. Pending or running is never success; report terminal green,
+terminal failure, or an exact unavailable/ambiguous blocker.
+
+Apply the repository surgical-mutation rule from `AGENTS.md` to any local
+write, and stop for Root input before expanding scope. Do not edit TOON files manually;
+do not create standalone handoff, Decision, or blocker files.
+
+Return one outcome from `completed`, `blocked`, `needs_root_input`, or
+`failed`, with discovered topology, exact ref/SHA, check or release results,
+source references, and remaining risk. Root records it with
+`holycodex-agent assignment result` and decides whether to delegate a fix and
+repeats the post-integration cycle. A terminal green development result permits
+Root to proceed to an authorized release action; after that action, observe the
+release against its exact ref/SHA. If the repository has no distinct release
+gate, report that topology explicitly and never invent a second gate. A
+terminal failure returns control to Root for a bounded fix; this Assignment
+never repairs, retries, approves, or mutates external state.

package/dist/assets/plugin/skills/operations/agents/openai.yaml

@@ -0,0 +1,6 @@
+interface:
+  display_name: "Operations"
+  short_description: "Observe exact-ref CI or release state to terminal evidence."
+  default_prompt: "Observe the assigned exact ref and SHA, discover the repository topology, report completed terminal CI or release evidence (pending is never success), return blocked or needs_root_input when unavailable, never mutate external state, and apply the repository surgical-mutation rule to any authorized local write."
+policy:
+  allow_implicit_invocation: false

package/dist/assets/plugin/skills/plan/SKILL.md

@@ -4,9 +4,21 @@ description: Use when Root faces unresolved architecture, scope, coordination, o
 ---
 
 Owner: Root. Boundary: identify alternatives, constraints, risks, route,
-acceptance evidence, exclusions, and recovery without dispatching
-implementation work before the plan gate opens. Specialists may supply facts
-but may not decide the outcome.
+acceptance evidence, exclusions, and recovery, then persist the semantic Plan
+through `holycodex-agent plan revise`. Specialists may supply facts but may not
+decide the outcome.
+
+Create or read the current Intent with `holycodex-agent intent ...`, and create
+the bounded delegated plan/review assignments before dispatch. A plan is optional for
+trivial work. Do not edit TOON files manually.
+
+Use `request_user_input` before seeking workflow Plan approval and whenever
+ambiguity or missing material input blocks safe progress; persist that condition as
+`needs_root_input` on the Intent or Plan. Apply the repository
+surgical-mutation rule from `AGENTS.md`.
+
+Persist the plan operation as a compact `completed`, `blocked`,
+`needs_root_input`, or `failed` result with proof and remaining risk.
 
 Completion: Root has an implementation-ready plan naming scope, owners and
 seams, data and control flow, policy, recovery, tests, compatibility, the

package/dist/assets/plugin/skills/plan/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Plan"
   short_description: "Resolve material scope and architecture choices before work."
-  default_prompt: "When Root faces an unresolved material choice, produce an implementation-ready plan covering scope, route, risk, and proof."
+  default_prompt: "When Root faces an unresolved material choice, persist an implementation-ready Plan through holycodex-agent, covering scope, route, risk, proof, and delegated Assignments; ask for material input and apply the repository surgical-mutation rule."
 policy:
   allow_implicit_invocation: true

package/dist/assets/plugin/skills/plan-review/SKILL.md

@@ -3,9 +3,21 @@ name: plan-review
 description: Use when Root has a complete plan that needs adversarial review; check feasibility, order, risk, and proof to a fixed point.
 ---
 
-Owner: Reviewer. Boundary: inspect the complete plan once to a fixed point;
+Owner: Reviewer. Boundary: inspect the complete Plan once to a fixed point;
 repair only reviewer-owned plan defects. Return material product or
-architecture choices to Root.
+architecture choices to Root. Root must delegate this review as an Assignment
+and record its structured result through `holycodex-agent assignment result`.
+
+Read the current Plan with `holycodex-agent plan read`; never edit TOON files
+directly. A revision is a semantic `plan revise` operation that archives the
+prior canonical Plan before replacement.
+
+Apply the repository surgical-mutation rule from `AGENTS.md` to any repair. If
+a material choice is unresolved, return `needs_root_input` rather than deciding
+it.
+
+Return `completed`, `blocked`, `needs_root_input`, or `failed` with findings,
+proof, and remaining risk; Root records that outcome on the Assignment.
 
 Completion: the plan is complete with proof, blocked with evidence, or
 explicitly needs a Root decision.

package/dist/assets/plugin/skills/plan-review/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Plan review"
   short_description: "Review one complete plan to a fixed point."
-  default_prompt: "When Root has a complete plan, inspect it once, repair reviewer-owned defects to a fixed point, and return structured proof."
+  default_prompt: "When Root has a complete Plan, inspect it once, apply only surgical reviewer-owned repairs to a fixed point, and return a structured Assignment result for holycodex-agent."
 policy:
   allow_implicit_invocation: true

package/dist/assets/plugin/skills/programming/SKILL.md

@@ -3,9 +3,20 @@ name: programming
 description: Use when Root has decided a seam and assigns bounded implementation, fixes, tests, or manifests with known acceptance behavior.
 ---
 
-Implement the decided seam inside the assigned files. Inspect the callers and
-tests needed to preserve its behavior and typed boundaries, then add focused
-proof at the least brittle observable boundary. Discover the repository's own
-validation commands and run the smallest relevant check first; broaden when
-the changed boundary or release gate requires it. Inspect changed files and
-remove incidental generated output before stopping.
+Owner: Worker within one delegated Assignment. Read the Assignment and current
+Intent through the semantic agent interface; do not edit TOON files manually.
+Implement the decided seam inside the assigned files, and do not expand its
+scope. Inspect the callers and tests needed to preserve behavior and typed
+boundaries, then add focused proof at the least brittle observable boundary.
+
+Apply the repository surgical-mutation rule from `AGENTS.md`.
+
+Return exactly one compact specialist outcome: `completed`, `blocked`,
+`needs_root_input`, or `failed`, with changed paths, checks and results,
+evidence, and remaining risk. Root records it with
+`holycodex-agent assignment result`; the Worker never changes global Intent
+lifecycle, makes material decisions, or performs Git/VCS.
+
+Discover and use the repository's own validation commands and run the smallest
+relevant check first; broaden when the seam or release gate requires it.
+Inspect changed files and remove incidental generated output before stopping.

package/dist/assets/plugin/skills/programming/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Programming"
   short_description: "Implement one decided seam with proportional proof."
-  default_prompt: "When Root assigns a decided seam, implement it, add focused proof, inspect changed files, and return structured evidence."
+  default_prompt: "When Root assigns a bounded Assignment, apply the repository surgical-mutation rule, implement only that seam, add focused proof, inspect changed files, and return completed, blocked, needs_root_input, or failed evidence for holycodex-agent."
 policy:
   allow_implicit_invocation: true

package/dist/assets/plugin/skills/refactor/SKILL.md

@@ -3,9 +3,20 @@ name: refactor
 description: Use when Root has decided one seam needs restructuring; preserve behavior, ownership, and interfaces.
 ---
 
-Restructure only the decided seam. Trace its callers before changing a shared
-root cause, preserve public behavior and interfaces, and keep the smallest
-cohesive change. Discover the repository's own validation commands and run the
-smallest relevant behavior-locked check first; broaden when the changed
-boundary or release gate requires it. Inspect changed files and remove
-incidental generated output before stopping.
+Owner: Worker within one delegated Assignment. Read the current Assignment
+through the semantic agent interface; do not edit TOON files manually. Restructure
+only the decided seam. Trace its callers before changing a shared root cause,
+preserve public behavior and interfaces, and keep the smallest cohesive change.
+
+Apply the repository surgical-mutation rule from `AGENTS.md`.
+Preserve unrelated work and stop for Root input before expanding scope.
+
+Discover the repository's own validation commands and run the smallest relevant
+behavior-locked check first; broaden when the changed boundary or release gate
+requires it. Inspect changed files and remove incidental generated output
+before stopping.
+
+Return one outcome from `completed`, `blocked`, `needs_root_input`, or `failed`,
+with changed paths, checks and results, evidence, and remaining risk. The Worker
+does not mutate global Intent lifecycle, make material decisions, or perform
+Git/VCS; Root records the result through `holycodex-agent assignment result`.

package/dist/assets/plugin/skills/refactor/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Refactor"
   short_description: "Restructure one decided seam while preserving its contract."
-  default_prompt: "When Root assigns a seam refactor, preserve behavior and interfaces, inspect the changed files, and return focused proof."
+  default_prompt: "When Root assigns a bounded refactor Assignment, apply the repository surgical-mutation rule, preserve behavior and interfaces, inspect changed files, prove the seam, and return completed, blocked, needs_root_input, or failed evidence for holycodex-agent."
 policy:
   allow_implicit_invocation: true

package/dist/assets/plugin/skills/rules/SKILL.md

@@ -3,8 +3,17 @@ name: rules
 description: Use when Root assigns a repository rule discovery, validation, or application failure; trace the owning rule path.
 ---
 
-Owner: Worker. Boundary: inspect the owning rule loader and preserve
-deduplication, limits, caching, trust boundaries, and rule ownership.
+Owner: Worker within one delegated Assignment. Boundary: inspect the owning
+rule loader and preserve deduplication, limits, caching, trust boundaries, and
+rule ownership. Read the Assignment through the semantic agent interface; do
+not edit TOON files manually.
+
+Apply the repository surgical-mutation rule from `AGENTS.md` to any repair.
+Preserve unrelated work and stop for Root input before expanding scope.
 
 Completion: the rule path, policy, observed failure, and verification evidence
-are explicit, or the exact blocker is returned.
+are explicit, or the exact blocker is returned. Return one outcome from
+`completed`, `blocked`, `needs_root_input`, or `failed` with changed paths,
+proof, and remaining risk. Root records the result through
+`holycodex-agent assignment result`; the Worker does not mutate global Intent
+lifecycle, make material decisions, or perform Git/VCS.

package/dist/assets/plugin/skills/rules/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Rules"
   short_description: "Trace one repository rule discovery or application failure."
-  default_prompt: "When Root assigns a rule failure, inspect the owning rule seam and return exact behavior, limits, and failure evidence."
+  default_prompt: "When Root assigns a bounded rule Assignment, apply the repository surgical-mutation rule, inspect the owning seam, and return completed, blocked, needs_root_input, or failed evidence for holycodex-agent."
 policy:
   allow_implicit_invocation: true

package/dist/assets/plugin/skills/stop-slop/SKILL.md

@@ -7,6 +7,11 @@ Owner: Worker or Reviewer within the assigned prose surface. Boundary: edit
 only the assigned text, preserve meaning and product voice, and use the linked
 references for phrase, structure, and example branches when needed.
 
+The prose edit is one delegated Assignment. Read its scope and constraints
+through the semantic agent interface; do not edit TOON files manually. Apply the
+repository surgical-mutation rule from `AGENTS.md`. Preserve unrelated work
+and stop for Root input before expanding scope.
+
 # Stop Slop
 
 Remove predictable AI writing patterns from prose.
@@ -30,8 +35,11 @@ claims, throat-clearing, binary contrasts, meta-joiners, em dashes, and repeated
 rhythm. Use [examples](references/examples.md) to calibrate edits.
 
 Completion: the assigned prose is direct, specific, constraint-equivalent,
-and reviewed against the relevant references; any deliberate voice tradeoff
-is recorded.
+and reviewed against the relevant references; any deliberate voice tradeoff is
+recorded. Return one outcome from `completed`, `blocked`, `needs_root_input`, or
+`failed` with changed paths, proof, and remaining risk. Root records the result
+through `holycodex-agent assignment result`; do not mutate global Intent
+lifecycle or perform Git/VCS.
 
 ## License
 

package/dist/assets/plugin/skills/stop-slop/agents/openai.yaml

@@ -1,6 +1,6 @@
 interface:
   display_name: "Stop Slop"
   short_description: "Edit assigned prose to remove predictable AI patterns."
-  default_prompt: "When Root assigns prose editing, apply the concise rules and disclosed references while preserving meaning and voice."
+  default_prompt: "When Root assigns a bounded prose Assignment, apply the repository surgical-mutation rule and disclosed references, preserve meaning and voice, and return completed, blocked, needs_root_input, or failed evidence for holycodex-agent."
 policy:
   allow_implicit_invocation: true

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

@@ -1,6 +1,6 @@
 ---
 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. Before each dispatch ensure this skill is loaded and applied; reuse it until compaction, a new context, or an incomplete/unavailable load.
+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.
 ---
 
 # Writing for agents
@@ -19,8 +19,18 @@ 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; reload after compaction, a new context, or an
-incomplete/unavailable load.
+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
 
@@ -41,6 +51,11 @@ Keep each meaning in one source of truth. Treat code, configuration, directory l
 
 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.

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, and completion."
+  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."
 policy:
   allow_implicit_invocation: true